Sample

Explore resources related to sample to help implement AI solutions for your business.

A sample library that implements Generative AI-Augmented Program Execution (GAPE)

Recently I had a crazy idea. What if you would mix formal program logic with "common sense reasoning" done by an LLM? Like so: statement = "Planet Earth is a flat disk." istrue = applycommonsensereasoningbyan_llm(statement) if is_true: print("Welcome to flat earth society!") else: print("Welcome to the scientific community") This idea can be generalized to many more things. Need to filter a list of concepts (strings) by a certain criterion that is accessible through common sense reasoning? Need to sort a list of concepts by a certain criterion applying common sense reasoning? I call this approach Generative AI-Augmented Program Execution (GAPE). I created a small Python sample library called PyGAPE here: https://github.com/fkostadinov/pygape/ I would love to get some feedback. Some people I showed the idea to were fascinated, other bewildered, some found it completely useless, others immediately got what I was after.

#reddit #sample #library

0.333

minophen•This week

Lessons from 139 YC AI startups (S23)

YC's Demo Day was last week, and with it comes another deluge of AI companies. A record-breaking 139 startups were in some way related to AI or ML - up from 112 in the last batch. Here are 5 of my biggest takeaways: AI is (still) eating the world. It's remarkable how diverse the industries are - over two dozen verticals were represented, from materials science to social media to security. However, the top four categories were: AI Ops: Tooling and platforms to help companies deploy working AI models. We'll discuss more below, but AI Ops has become a huge category, primarily focused on LLMs and taming them for production use cases. Developer Tools: Apps, plugins, and SDKs making it easier to write code. There were plenty of examples of integrating third-party data, auto-generating code/tests, and working with agents/chatbots to build and debug code. Healthcare + Biotech: It seems like healthcare has a lot of room for automation, with companies working on note-taking, billing, training, and prescribing. And on the biotech side, there are some seriously cool companies building autonomous surgery robots and at-home cancer detection. Finance + Payments: Startups targeting banks, fintechs, and compliance departments. This was a wide range of companies, from automated collections to AI due diligence to "Copilot for bankers." Those four areas covered over half of the startups. The first two make sense: YC has always filtered for technical founders, and many are using AI to do what they know - improve the software developer workflow. But it's interesting to see healthcare and finance not far behind. Previously, I wrote: Large enterprises, healthcare, and government are not going to send sensitive data to OpenAI. This leaves a gap for startups to build on-premise, compliant \[LLMs\] for these verticals. And we're now seeing exactly that - LLMs focused on healthcare and finance and AI Ops companies targeting on-prem use cases. It also helps that one of the major selling points of generative AI right now is cost-cutting - an enticing use case for healthcare and finance. Copilots are king. In the last batch, a lot of startups positioned themselves as "ChatGPT for X," with a consumer focus. It seems the current trend, though, is "Copilot for X" - B2B AI assistants to help you do everything from KYC checks to corporate event planning to chip design to negotiate contracts. Nearly two dozen companies were working on some sort of artificial companion for businesses - and a couple for consumers. It's more evidence for the argument that AI will not outright replace workers - instead, existing workers will collaborate with AI to be more productive. And as AI becomes more mainstream, this trend of making specialized tools for specific industries or tasks will only grow. That being said - a Bing-style AI that lives in a sidebar and is only accessible via chat probably isn't the most useful form factor for AI. But until OpenAI, Microsoft, and Google change their approach (or until another company steps up), we'll probably see many more Copilots. AI Ops is becoming a key sector. "AI Ops" has been a term for only a few years. "LLM Ops" has existed for barely a year. And yet, so many companies are focused on training, fine-tuning, deploying, hosting, and post-processing LLMs it's quickly becoming a critical piece of the AI space. It's a vast industry that's sprung up seemingly overnight, and it was pretty interesting to see some of the problems being solved at the bleeding edge. For example: Adding context to language models with as few as ten samples. Pausing and moving training runs in real-time. Managing training data ownership and permissions. Faster vector databases. Fine-tuning models with synthetic data. But as much ~~hype~~ enthusiasm and opportunity as there might be, the size of the AI Ops space also shows how much work is needed to really productionalize LLMs and other models. There are still many open questions about reliability, privacy, observability, usability, and safety when it comes to using LLMs in the wild. Who owns the model? Does it matter? Nine months ago, anyone building an LLM company was doing one of three things: Training their own model from scratch. Fine-tuning a version of GPT-3. Building a wrapper around ChatGPT. Thanks to Meta, the open-source community, and the legions of competitors trying to catch up to OpenAI, there are now dozens of ways to integrate LLMs. However, I found it interesting how few B2B companies mentioned whether or not they trained their own model. If I had to guess, I'd say many are using ChatGPT or a fine-tuned version of Llama 2. But it raises an interesting question - if the AI provides value, does it matter if it's "just" ChatGPT behind the scenes? And once ChatGPT becomes fine-tuneable, when (if ever) will startups decide to ditch OpenAI and use their own model instead? "AI" isn't a silver bullet. At the end of the day, perhaps the biggest lesson is that "AI" isn't a magical cure-all - you still need to build a defensible company. At the beginning of the post-ChatGPT hype wave, it seemed like you just had to say "we're adding AI" to raise your next round or boost your stock price. But competition is extremely fierce. Even within this batch, there were multiple companies with nearly identical pitches, including: Solving customer support tickets. Negotiating sales contracts. Writing drafts of legal documents. Building no-code LLM workflows. On-prem LLM deployment. Automating trust and safety moderation. As it turns out, AI can be a competitive advantage, but it can't make up for a bad business. The most interesting (and likely valuable) companies are the ones that take boring industries and find non-obvious use cases for AI. In those cases, the key is having a team that can effectively distribute a product to users, with or without AI. Where we’re headed I'll be honest - 139 companies is a lot. In reviewing them all, there were points where it just felt completely overwhelming. But after taking a step back, seeing them all together paints an incredibly vivid picture of the current AI landscape: one that is diverse, rapidly evolving, and increasingly integrated into professional and personal tasks. These startups aren't just building AI for the sake of technology or academic research, but are trying to address real-world problems. Technology is always a double-edged sword - and some of the startups felt a little too dystopian for my taste - but I'm still hopeful about AI's ability to improve productivity and the human experience.

#reddit #lessons #startups

Additional-Ad-7043•This week

How do byte-level language models work?

I've recently been trying to pre-train my own small language model on the tiny-series datasets on huggingface: https://huggingface.co/collections/nampdn-ai/tiny-series-6503910fd491144159519c70. I also wanted to use a model similar to MEGABYTE: https://arxiv.org/pdf/2305.07185.pdf, but I don't understand how using bytes would work. The only implementation I could find: https://github.com/lucidrains/MEGABYTE-pytorch used str(chr(max(32, token))) to decode any token (byte) to a character and put the embedding size as 256. Firstly, why 256 and not 256-32 as any values below 32 are ignored? Also, many byte-level models including this and ByteT5 mention that they can process any text sequence even in a multilingual setting, however how would that be true if we are only using one byte, would we have to move to 2 bytes or use an UNK token, and if we did use 2 bytes that would make our embedding size around 65000 which defeats sort of the point as one of the advantages mentioned is that we are able to use a small embedding matrix? Furthermore, most language models add special tokens like bos, eos, unk and even for llama they use beginning of instruction, end of instruction, and more for system instructions, response, context... Should I use something like this as my dataset has some structures where there is a context, instruction and response, and if i did how would I add these if I'm using byte-level encodings? Final questions: Firstly, for the datasets mentioned including code,stories,webtext,... would I tokenise all of these datasets then concatenate them to then randomly sample from, or should i train seperately on each as some like code and webtext are much larger than the others? Finally, for the webtext part of the dataset, there is a passage of text then a passage analysing the text (main ideas,purpose,...), how should I encode this, should I use an extra ANALYSE token or just concatenate? Thank you for reading this far, I am sort of a beginner so if I said something stupid please point it out. Also, if there were unclear parts in my question I'm sorry as I struggled how to word these questions. Any help would be appreciated!

#reddit #byte-level #language

Sophomore computer science student, looking at ISLP vs ESL vs mlcourse.ai

OneTrueDuce•This week

Sophomore computer science student, looking at ISLP vs ESL vs mlcourse.ai

For background, I am currently a computer science sophomore, with intermediate skills in Python and C++. I have taken university courses on data structure and algorithms, calc 1-3, linear algebra, and an introductory stat course (which covered confidence interval, Z and T sample test, and hypothesis testing). I also have read up to Chapter 5 of the MML book and am currently self-studying probability theory (through STAT 110 video and textbook by Joe Blitzstein). I have done a few beginner ML projects with Tensorflow and scikit-learn, but most of the work is in EDA and feature engineering, while the ML model is just a black box that I plug and chug. So now, I want to learn how to implement ML models from scratch. I've been skimming over ISLP, which many people online recommended, but it seems that while it talks about mathematical equations used, I don't really get to implement it; as the labs are a lot of importing an already implemented model then plug and chug. So now, I am looking at ESL, which I believe is the more detailed and mathematically rigorous version of ISL. However, there aren't any labs or code along to ease beginners in (which I somewhat understand given the intended audience of the book). Another option I am looking at is mlcourse.ai, which seems to cover mathematics and has some lab/code along for it. But it doesn't seem to span as many subjects as ESL does. Given these options, I am unsure of which one to pick, should I first finish my self-study on probability theory and then Chapters 6-8 of MML? Then should I do ISLP first or just get into ESL? Or maybe I should do mlcourse.ai first then into ESL? Or should I just do the ML course/book along with the maths? In addition, there is also the data science + feature engineering stuff which I wonder if I should study more about. Sorry if this seems like a mess, there are just so many things to ML that I am kinda overwhelmed.

#reddit #sophomore #computer

MMML | Deploy HuggingFace training model rapidly based on MetaSpore

qazmkopp•This week

MMML | Deploy HuggingFace training model rapidly based on MetaSpore

A few days ago, HuggingFace announced a $100 million Series C funding round, which was big news in open source machine learning and could be a sign of where the industry is headed. Two days before the HuggingFace funding announcement, open-source machine learning platform MetaSpore released a demo based on the HuggingFace Rapid deployment pre-training model. As deep learning technology makes innovative breakthroughs in computer vision, natural language processing, speech understanding, and other fields, more and more unstructured data are perceived, understood, and processed by machines. These advances are mainly due to the powerful learning ability of deep learning. Through pre-training of deep models on massive data, the models can capture the internal data patterns, thus helping many downstream tasks. With the industry and academia investing more and more energy in the research of pre-training technology, the distribution warehouses of pre-training models such as HuggingFace and Timm have emerged one after another. The open-source community release pre-training significant model dividends at an unprecedented speed. In recent years, the data form of machine modeling and understanding has gradually evolved from single-mode to multi-mode, and the semantic gap between different modes is being eliminated, making it possible to retrieve data across modes. Take CLIP, OpenAI’s open-source work, as an example, to pre-train the twin towers of images and texts on a dataset of 400 million pictures and texts and connect the semantics between pictures and texts. Many researchers in the academic world have been solving multimodal problems such as image generation and retrieval based on this technology. Although the frontier technology through the semantic gap between modal data, there is still a heavy and complicated model tuning, offline data processing, high performance online reasoning architecture design, heterogeneous computing, and online algorithm be born multiple processes and challenges, hindering the frontier multimodal retrieval technologies fall to the ground and pratt &whitney. DMetaSoul aims at the above technical pain points, abstracting and uniting many links such as model training optimization, online reasoning, and algorithm experiment, forming a set of solutions that can quickly apply offline pre-training model to online. This paper will introduce how to use the HuggingFace community pre-training model to conduct online reasoning and algorithm experiments based on MetaSpore technology ecology so that the benefits of the pre-training model can be fully released to the specific business or industry and small and medium-sized enterprises. And we will give the text search text and text search graph two multimodal retrieval demonstration examples for your reference. Multimodal semantic retrieval The sample architecture of multimodal retrieval is as follows: Our multimodal retrieval system supports both text search and text search application scenarios, including offline processing, model reasoning, online services, and other core modules: https://preview.redd.it/mdyyv1qmdz291.png?width=1834&format=png&auto=webp&s=e9e10710794c78c64cc05adb75db385aa53aba40 Offline processing, including offline data processing processes for different application scenarios of text search and text search, including model tuning, model export, data index database construction, data push, etc. Model inference. After the offline model training, we deployed our NLP and CV large models based on the MetaSpore Serving framework. MetaSpore Serving helps us conveniently perform online inference, elastic scheduling, load balancing, and resource scheduling in heterogeneous environments. Online services. Based on MetaSpore’s online algorithm application framework, MetaSpore has a complete set of reusable online search services, including Front-end retrieval UI, multimodal data preprocessing, vector recall and sorting algorithm, AB experimental framework, etc. MetaSpore also supports text search by text and image scene search by text and can be migrated to other application scenarios at a low cost. The HuggingFace open source community has provided several excellent baseline models for similar multimodal retrieval problems, which are often the starting point for actual optimization in the industry. MetaSpore also uses the pre-training model of the HuggingFace community in its online services of searching words by words and images by words. Searching words by words is based on the semantic similarity model of the question and answer field optimized by MetaSpore, and searching images by words is based on the community pre-training model. These community open source pre-training models are exported to the general ONNX format and loaded into MetaSpore Serving for online reasoning. The following sections will provide a detailed description of the model export and online retrieval algorithm services. The reasoning part of the model is standardized SAAS services with low coupling with the business. Interested readers can refer to my previous post: The design concept of MetaSpore, a new generation of the one-stop machine learning platform. 1.1 Offline Processing Offline processing mainly involves the export and loading of online models and index building and pushing of the document library. You can follow the step-by-step instructions below to complete the offline processing of text search and image search and see how the offline pre-training model achieves reasoning at MetaSpore. 1.1.1 Search text by text Traditional text retrieval systems are based on literal matching algorithms such as BM25. Due to users’ diverse query words, a semantic gap between query words and documents is often encountered. For example, users misspell “iPhone” as “Phone,” and search terms are incredibly long, such as “1 \~ 3 months old baby autumn small size bag pants”. Traditional text retrieval systems will use spelling correction, synonym expansion, search terms rewriting, and other means to alleviate the semantic gap but fundamentally fail to solve this problem. Only when the retrieval system fully understands users’ query terms and documents can it meet users’ retrieval demands at the semantic level. With the continuous progress of pre-training and representational learning technology, some commercial search engines continue to integrate semantic vector retrieval methods based on symbolic learning into the retrieval ecology. Semantic retrieval model This paper introduces a set of semantic vector retrieval applications. MetaSpore built a set of semantic retrieval systems based on encyclopedia question and answer data. MetaSpore adopted the Sentence-Bert model as the semantic vector representation model, which fine-tunes the twin tower BERT in supervised or unsupervised ways to make the model more suitable for retrieval tasks. The model structure is as follows: The query-Doc symmetric two-tower model is used in text search and question and answer retrieval. The vector representation of online Query and offline DOC share the same vector representation model, so it is necessary to ensure the consistency of the offline DOC library building model and online Query inference model. The case uses MetaSpore’s text representation model Sbert-Chinese-QMC-domain-V1, optimized in the open-source semantically similar data set. This model will express the question and answer data as a vector in offline database construction. The user query will be expressed as a vector by this model in online retrieval, ensuring that query-doc in the same semantic space, users’ semantic retrieval demands can be guaranteed by vector similarity metric calculation. Since the text presentation model does vector encoding for Query online, we need to export the model for use by the online service. Go to the q&A data library code directory and export the model concerning the documentation. In the script, Pytorch Tracing is used to export the model. The models are exported to the “./export “directory. The exported models are mainly ONNX models used for wired reasoning, Tokenizer, and related configuration files. The exported models are loaded into MetaSpore Serving by the online Serving system described below for model reasoning. Since the exported model will be copied to the cloud storage, you need to configure related variables in env.sh. \Build library based on text search \ The retrieval database is built on the million-level encyclopedia question and answer data set. According to the description document, you need to download the data and complete the database construction. The question and answer data will be coded as a vector by the offline model, and then the database construction data will be pushed to the service component. The whole process of database construction is described as follows: Preprocessing, converting the original data into a more general JSonline format for database construction; Build index, use the same model as online “sbert-Chinese-qmc-domain-v1” to index documents (one document object per line); Push inverted (vector) and forward (document field) data to each component server. The following is an example of the database data format. After offline database construction is completed, various data are pushed to corresponding service components, such as Milvus storing vector representation of documents and MongoDB storing summary information of documents. Online retrieval algorithm services will use these service components to obtain relevant data. 1.1.2 Search by text Text and images are easy for humans to relate semantically but difficult for machines. First of all, from the perspective of data form, the text is the discrete ID type of one-dimensional data based on words and words. At the same time, images are continuous two-dimensional or three-dimensional data. Secondly, the text is a subjective creation of human beings, and its expressive ability is vibrant, including various turning points, metaphors, and other expressions, while images are machine representations of the objective world. In short, bridging the semantic gap between text and image data is much more complex than searching text by text. The traditional text search image retrieval technology generally relies on the external text description data of the image or the nearest neighbor retrieval technology and carries out the retrieval through the image associated text, which in essence degrades the problem to text search. However, it will also face many issues, such as obtaining the associated text of pictures and whether the accuracy of text search by text is high enough. The depth model has gradually evolved from single-mode to multi-mode in recent years. Taking the open-source project of OpenAI, CLIP, as an example, train the model through the massive image and text data of the Internet and map the text and image data into the same semantic space, making it possible to implement the text and image search technology based on semantic vector. CLIP graphic model The text search pictures introduced in this paper are implemented based on semantic vector retrieval, and the CLIP pre-training model is used as the two-tower retrieval architecture. Because the CLIP model has trained the semantic alignment of the twin towers’ text and image side models on the massive graphic and text data, it is particularly suitable for the text search graph scene. Due to the different image and text data forms, the Query-Doc asymmetric twin towers model is used for text search image retrieval. The image-side model of the twin towers is used for offline database construction, and the text-side model is used for the online return. In the final online retrieval, the database data of the image side model will be searched after the text side model encodes Query, and the CLIP pre-training model guarantees the semantic correlation between images and texts. The model can draw the graphic pairs closer in vector space by pre-training on a large amount of visual data. Here we need to export the text-side model for online MetaSpore Serving inference. Since the retrieval scene is based on Chinese, the CLIP model supporting Chinese understanding is selected. The exported content includes the ONNX model used for online reasoning and Tokenizer, similar to the text search. MetaSpore Serving can load model reasoning through the exported content. Build library on Image search You need to download the Unsplash Lite library data and complete the construction according to the instructions. The whole process of database construction is described as follows: Preprocessing, specify the image directory, and then generate a more general JSOnline file for library construction; Build index, use OpenAI/Clip-Vit-BASE-Patch32 pre-training model to index the gallery, and output one document object for each line of index data; Push inverted (vector) and forward (document field) data to each component server. Like text search, after offline database construction, relevant data will be pushed to service components, called by online retrieval algorithm services to obtain relevant data. 1.2 Online Services The overall online service architecture diagram is as follows:  https://preview.redd.it/nz8zrbbpdz291.png?width=1280&format=png&auto=webp&s=28dae7e031621bc8819519667ed03d8d085d8ace Multi-mode search online service system supports application scenarios such as text search and text search. The whole online service consists of the following parts: Query preprocessing service: encapsulate preprocessing logic (including text/image, etc.) of pre-training model, and provide services through gRPC interface; Retrieval algorithm service: the whole algorithm processing link includes AB experiment tangent flow configuration, MetaSpore Serving call, vector recall, sorting, document summary, etc.; User entry service: provides a Web UI interface for users to debug and track down problems in the retrieval service. From a user request perspective, these services form invocation dependencies from back to front, so to build up a multimodal sample, you need to run each service from front to back first. Before doing this, remember to export the offline model, put it online and build the library first. This article will introduce the various parts of the online service system and make the whole service system step by step according to the following guidance. See the ReadME at the end of this article for more details. 1.2.1 Query preprocessing service Deep learning models tend to be based on tensors, but NLP/CV models often have a preprocessing part that translates raw text and images into tensors that deep learning models can accept. For example, NLP class models often have a pre-tokenizer to transform text data of string type into discrete tensor data. CV class models also have similar processing logic to complete the cropping, scaling, transformation, and other processing of input images through preprocessing. On the one hand, considering that this part of preprocessing logic is decoupled from tensor reasoning of the depth model, on the other hand, the reason of the depth model has an independent technical system based on ONNX, so MetaSpore disassembled this part of preprocessing logic. NLP pretreatment Tokenizer has been integrated into the Query pretreatment service. MetaSpore dismantlement with a relatively general convention. Users only need to provide preprocessing logic files to realize the loading and prediction interface and export the necessary data and configuration files loaded into the preprocessing service. Subsequent CV preprocessing logic will also be integrated in this manner. The preprocessing service currently provides the gRPC interface invocation externally and is dependent on the Query preprocessing (QP) module in the retrieval algorithm service. After the user request reaches the retrieval algorithm service, it will be forwarded to the service to complete the data preprocessing and continue the subsequent processing. The ReadMe provides details on how the preprocessing service is started, how the preprocessing model exported offline to cloud storage enters the service, and how to debug the service. To further improve the efficiency and stability of model reasoning, MetaSpore Serving implements a Python preprocessing submodule. So MetaSpore can provide gRPC services through user-specified preprocessor.py, complete Tokenizer or CV-related preprocessing in NLP, and translate requests into a Tensor that deep models can handle. Finally, the model inference is carried out by MetaSpore, Serving subsequent sub-modules. Presented here on the lot code: https://github.com/meta-soul/MetaSpore/compare/add\python\preprocessor 1.2.2 Retrieval algorithm services Retrieval algorithm service is the core of the whole online service system, which is responsible for the triage of experiments, the assembly of algorithm chains such as preprocessing, recall, sorting, and the invocation of dependent component services. The whole retrieval algorithm service is developed based on the Java Spring framework and supports multi-mode retrieval scenarios of text search and text search graph. Due to good internal abstraction and modular design, it has high flexibility and can be migrated to similar application scenarios at a low cost. Here’s a quick guide to configuring the environment to set up the retrieval algorithm service. See ReadME for more details: Install dependent components. Use Maven to install the online-Serving component Search for service configurations. Copy the template configuration file and replace the MongoDB, Milvus, and other configurations based on the development/production environment. Install and configure Consul. Consul allows you to synchronize the search service configuration in real-time, including cutting the flow of experiments, recall parameters, and sorting parameters. The project’s configuration file shows the current configuration parameters of text search and text search. The parameter modelName in the stage of pretreatment and recall is the corresponding model exported in offline processing. Start the service. Once the above configuration is complete, the retrieval service can be started from the entry script. Once the service is started, you can test it! For example, for a user with userId=10 who wants to query “How to renew ID card,” access the text search service. 1.2.3 User Entry Service Considering that the retrieval algorithm service is in the form of the API interface, it is difficult to locate and trace the problem, especially for the text search image scene can intuitively display the retrieval results to facilitate the iterative optimization of the retrieval algorithm. This paper provides a lightweight Web UI interface for text search and image search, a search input box, and results in a display page for users. Developed by Flask, the service can be easily integrated with other retrieval applications. The service calls the retrieval algorithm service and displays the returned results on the page. It’s also easy to install and start the service. Once you’re done, go to http://127.0.0.1:8090 to see if the search UI service is working correctly. See the ReadME at the end of this article for details. Multimodal system demonstration The multimodal retrieval service can be started when offline processing and online service environment configuration have been completed following the above instructions. Examples of textual searches are shown below. Enter the entry of the text search map application, enter “cat” first, and you can see that the first three digits of the returned result are cats: https://preview.redd.it/d7syq47rdz291.png?width=1280&format=png&auto=webp&s=b43df9abd380b7d9a52e3045dd787f4feeb69635 If you add a color constraint to “cat” to retrieve “black cat,” you can see that it does return a black cat:  https://preview.redd.it/aa7pxx8tdz291.png?width=1280&format=png&auto=webp&s=e3727c29d1bde6eea2e1cccf6c46d3cae3f4750e Further, strengthen the constraint on the search term, change it to “black cat on the bed,” and return results containing pictures of a black cat climbing on the bed:  https://preview.redd.it/2mw4qpjudz291.png?width=1280&format=png&auto=webp&s=1cf1db667892b9b3a40451993680fbd6980b5520 The cat can still be found through the text search system after the color and scene modification in the above example. Conclusion The cutting-edge pre-training technology can bridge the semantic gap between different modes, and the HuggingFace community can greatly reduce the cost for developers to use the pre-training model. Combined with the technological ecology of MetaSpore online reasoning and online microservices provided by DMetaSpore, the pre-training model is no longer mere offline dabbling. Instead, it can truly achieve end-to-end implementation from cutting-edge technology to industrial scenarios, fully releasing the dividends of the pre-training large model. In the future, DMetaSoul will continue to improve and optimize the MetaSpore technology ecosystem: More automated and wider access to HuggingFace community ecology. MetaSpore will soon release a common model rollout mechanism to make HuggingFace ecologically accessible and will later integrate preprocessing services into online services. Multi-mode retrieval offline algorithm optimization. For multimodal retrieval scenarios, MetaSpore will continuously iteratively optimize offline algorithm components, including text recall/sort model, graphic recall/sort model, etc., to improve the accuracy and efficiency of the retrieval algorithm. For related code and reference documentation in this article, please visit: https://github.com/meta-soul/MetaSpore/tree/main/demo/multimodal/online Some images source: https://github.com/openai/CLIP/raw/main/CLIP.png https://www.sbert.net/examples/training/sts/README.html

#reddit #mmml #deploy

No_Information6299•This week

Browser Agents Real Example

I made a Browser Price Matching Tool that uses browser automation and some clever skills to adjust your product prices based on real-time web searches data. If you're into scraping, automation, or just love playing with the latest in ML-powered tools like OpenAI's GPT-4, this one's for you. What My Project Does The tool takes your current product prices (think CSV) and finds similar products online (targeting Amazon for demo purposes). It then compares prices, allowing you to adjust your prices competitively. The magic happens in a multi-step pipeline: Generate Clean Search Queries: Uses a learned skill to convert messy product names (like "Apple iPhone14!<" or "Dyson! V11!!// VacuumCleaner") into clean, Google-like search queries. Browser Data Extraction: Launches asynchronous browser agents (leveraging Playwright) to search for those queries on Amazon, retrieves the relevant data, and scrapes the page text. Parse & Structure Results: Another custom skill parses the browser output to output structured info: product name, price, and a short description. Enrich Your Data: Finally, the tool combines everything to enrich your original data with live market insights! Full code link: Full code File Rundown learn\skill.py Learns how to generate polished search queries from your product names with GPT-4o-mini. It outputs a JSON file: makequery.json. learn\skill\select\best\product.py Trains another skill to parse web-scraped data and select the best matching product details. Outputs select_product.json. make\query.json The skill definition file for generating search queries (produced by learnskill.py). select\product.json The skill definition file for extracting product details from scraped results (produced by learnskillselectbest_product.py). product\price\matching.py The main pipeline script that orchestrates the entire process—from loading product data, running browser agents, to enriching your CSV. Setup & Installation Install Dependencies: pip install python-dotenv openai langchain\_openai flashlearn requests pytest-playwright Install Playwright Browsers: playwright install Configure OpenAI API: Create a .env file in your project directory with:OPENAI\API\KEY="sk-your\api\key\_here" Running the Tool Train the Query Skill: Run learnskill.py to generate makequery.json. Train the Product Extraction Skill: Run learnskillselectbestproduct.py to generate select_product.json. Execute the Pipeline: Kick off the whole process by running productpricematching.py. The script will load your product data (sample data is included for demo, but easy to swap with your CSV), generate search queries, run browser agents asynchronously, scrape and parse the data, then output the enriched product listings. Target Audience I built this project to automate price matching—a huge pain point for anyone running an e-commerce business. The idea was to minimize the manual labor of checking competitor prices while integrating up-to-date market insights. Plus, it was a fun way to combine automation,skill training, and browser automation! Customization Tweak the concurrency in productpricematching.py to manage browser agent load. Replace the sample product list with your own CSV for a real-world scenario. Extend the skills if you need more data points or different parsing logic. Ajudst skill definitions as needed Comparison With existing approaches you need to manually write parsing loginc and data transformation logic - here ai does it for you. If you like the tutorial - leave a star github

#reddit #browser #agents

No_Information6299•This week

Browser Agents Real Example

#reddit #browser #agents

[D] Why I'm Lukewarm on Graph Neural Networks

0.6

VodkaHaze•This week

[D] Why I'm Lukewarm on Graph Neural Networks

TL;DR: GNNs can provide wins over simpler embedding methods, but we're at a point where other research directions matter more I also posted it on my blog here, has footnotes, a nicer layout with inlined images, etc. I'm only lukewarm on Graph Neural Networks (GNNs). There, I said it. It might sound crazy GNNs are one of the hottest fields in machine learning right now. [There][1] were at least [four][2] [review][3] [papers][4] just in the last few months. I think some progress can come of this research, but we're also focusing on some incorrect places. But first, let's take a step back and go over the basics. Models are about compression We say graphs are a "non-euclidean" data type, but that's not really true. A regular graph is just another way to think about a particular flavor of square matrix called the [adjacency matrix][5], like this. It's weird, we look at run-of-the-mill matrix full of real numbers and decide to call it "non-euclidean". This is for practical reasons. Most graphs are fairly sparse, so the matrix is full of zeros. At this point, where the non-zero numbers are matters most, which makes the problem closer to (computationally hard) discrete math rather than (easy) continuous, gradient-friendly math. If you had the full matrix, life would be easy If we step out of the pesky realm of physics for a minute, and assume carrying the full adjacency matrix around isn't a problem, we solve a bunch of problems. First, network node embeddings aren't a thing anymore. A node is a just row in the matrix, so it's already a vector of numbers. Second, all network prediction problems are solved. A powerful enough and well-tuned model will simply extract all information between the network and whichever target variable we're attaching to nodes. NLP is also just fancy matrix compression Let's take a tangent away from graphs to NLP. Most NLP we do can be [thought of in terms of graphs][6] as we'll see, so it's not a big digression. First, note that Ye Olde word embedding models like [Word2Vec][7] and [GloVe][8] are [just matrix factorization][9]. The GloVe algorithm works on a variation of the old [bag of words][10] matrix. It goes through the sentences and creates a (implicit) [co-occurence][11] graph where nodes are words and the edges are weighed by how often the words appear together in a sentence. Glove then does matrix factorization on the matrix representation of that co-occurence graph, Word2Vec is mathematically equivalent. You can read more on this in my [post on embeddings][12] and the one (with code) on [word embeddings][13]. Even language models are also just matrix compression Language models are all the rage. They dominate most of the [state of the art][14] in NLP. Let's take BERT as our main example. BERT predicts a word given the context of the rest of the sentence. This grows the matrix we're factoring from flat co-occurences on pairs of words to co-occurences conditional on the sentence's context, like this We're growing the "ideal matrix" we're factoring combinatorially. As noted by [Hanh & Futrell][15]: [...] human language—and language modelling—has infinite statistical complexity but that it can be approximated well at lower levels. This observation has two implications: 1) We can obtain good results with comparatively small models; and 2) there is a lot of potential for scaling up our models. Language models tackle such a large problem space that they probably approximate a compression of the entire language in the [Kolmogorov Complexity][16] sense. It's also possible that huge language models just [memorize a lot of it][17] rather than compress the information, for what it's worth. Can we upsample any graph like language models do? We're already doing it. Let's call a first-order embedding of a graph a method that works by directly factoring the graph's adjacency matrix or [Laplacian matrix][18]. If you embed a graph using [Laplacian Eigenmaps][19] or by taking the [principal components][20] of the Laplacian, that's first order. Similarly, GloVe is a first-order method on the graph of word co-occurences. One of my favorites first order methods for graphs is [ProNE][21], which works as well as most methods while being two orders of magnitude faster. A higher-order method embeds the original matrix plus connections of neighbours-of-neighbours (2nd degree) and deeper k-step connections. [GraRep][22], shows you can always generate higher-order representations from first order methods by augmenting the graph matrix. Higher order method are the "upsampling" we do on graphs. GNNs that sample on large neighborhoods and random-walk based methods like node2vec are doing higher-order embeddings. Where are the performance gain? Most GNN papers in the last 5 years present empirical numbers that are useless for practitioners to decide on what to use. As noted in the [OpenGraphsBenchmark][4] (OGB) paper, GNN papers do their empirical section on a handful of tiny graphs (Cora, CiteSeer, PubMed) with 2000-20,000 nodes. These datasets can't seriously differentiate between methods. Recent efforts are directly fixing this, but the reasons why researchers focused on tiny, useless datasets for so long are worth discussing. Performance matters by task One fact that surprises a lot of people is that even though language models have the best performance in a lot of NLP tasks, if all you're doing is cram sentence embeddings into a downstream model, there [isn't much gained][23] from language models embeddings over simple methods like summing the individual Word2Vec word embeddings (This makes sense, because the full context of the sentence is captured in the sentence co-occurence matrix that is generating the Word2Vec embeddings). Similarly, [I find][24] that for many graphs simple first-order methods perform just as well on graph clustering and node label prediction tasks than higher-order embedding methods. In fact higher-order methods are massively computationally wasteful for these usecases. Recommended first order embedding methods are ProNE and my [GGVec with order=1][25]. Higher order methods normally perform better on the link prediction tasks. I'm not the only one to find this. In the BioNEV paper, they find: "A large GraRep order value for link prediction tasks (e.g. 3, 4);a small value for node classification tasks (e.g.1, 2)" (p.9). Interestingly, the gap in link prediction performance is inexistant for artificially created graphs. This suggests higher order methods do learn some of the structure intrinsic to [real world graphs][26]. For visualization, first order methods are better. Visualizations of higher order methods tend to have artifacts of their sampling. For instance, Node2Vec visualizations tend to have elongated/filament-like structures which come from the embeddings coming from long single strand random walks. See the following visualizations by [Owen Cornec][27] created by first embedding the graph to 32-300 dimensions using a node embedding algorithm, then mapping this to 2d or 3d with the excellent UMAP algorithm, like this Lastly, sometimes simple methods soundly beat higher order methods (there's an instance of it in the OGB paper). The problem here is that we don't know when any method is better than another and we definitely don't know the reason. There's definitely a reason different graph types respond better/worse to being represented by various methods. This is currently an open question. A big part of why is that the research space is inundated under useless new algorithms because... Academic incentives work against progress Here's the cynic's view of how machine learning papers are made: Take an existing algorithm Add some new layer/hyperparameter, make a cute mathematical story for why it matters Gridsearch your hyperparameters until you beat baselines from the original paper you aped Absolutely don't gridsearch stuff you're comparing against in your results section Make a cute ACRONYM for your new method, put impossible to use python 2 code on github (Or no code at all!) and bask in the citations I'm [not][28] the [only one][29] with these views on the state reproducible research. At least it's gotten slightly better in the last 2 years. Sidebar: I hate Node2Vec A side project of mine is a [node embedding library][25] and the most popular method in it is by far Node2Vec. Don't use Node2Vec. [Node2Vec][30] with p=1; q=1 is the [Deepwalk][31] algorithm. Deepwalk is an actual innovation. The Node2Vec authors closely followed the steps 1-5 including bonus points on step 5 by getting word2vec name recognition. This is not academic fraud -- the hyperparameters [do help a tiny bit][32] if you gridsearch really hard. But it's the presentable-to-your-parents sister of where you make the ML community worse off to progress your academic career. And certainly Node2Vec doesn't deserve 7500 citations. Progress is all about practical issues We've known how to train neural networks for well over 40 years. Yet they only exploded in popularity with [AlexNet][33] in 2012. This is because implementations and hardware came to a point where deep learning was practical. Similarly, we've known about factoring word co-occurence matrices into Word embeddings for at least 20 years. But word embeddings only exploded in 2013 with Word2Vec. The breakthrough here was that the minibatch-based methods let you train a Wikipedia-scale embedding model on commodity hardware. It's hard for methods in a field to make progress if training on a small amount of data takes days or weeks. You're disincentivized to explore new methods. If you want progress, your stuff has to run in reasonable time on commodity hardware. Even Google's original search algorithm [initially ran on commodity hardware][34]. Efficiency is paramount to progress The reason deep learning research took off the way it did is because of improvements in [efficiency][35] as well as much better libraries and hardware support. Academic code is terrible Any amount of time you spend gridsearching Node2Vec on p and q is all put to better use gridsearching Deepwalk itself (on number of walks, length of walks, or word2vec hyperparameters). The problem is that people don't gridsearch over deepwalk because implementations are all terrible. I wrote the [Nodevectors library][36] to have a fast deepwalk implementation because it took 32 hours to embed a graph with a measly 150,000 nodes using the reference Node2Vec implementation (the same takes 3min with Nodevectors). It's no wonder people don't gridsearch on Deepwalk a gridsearch would take weeks with the terrible reference implementations. To give an example, in the original paper of [GraphSAGE][37] they their algorithm to DeepWalk with walk lengths of 5, which is horrid if you've ever hyperparameter tuned a deepwalk algorithm. From their paper: We did observe DeepWalk’s performance could improve with further training, and in some cases it could become competitive with the unsupervised GraphSAGE approaches (but not the supervised approaches) if we let it run for >1000× longer than the other approaches (in terms of wall clock time for prediction on the test set) I don't even think the GraphSAGE authors had bad intent -- deepwalk implementations are simply so awful that they're turned away from using it properly. It's like trying to do deep learning with 2002 deep learning libraries and hardware. Your architectures don't really matter One of the more important papers this year was [OpenAI's "Scaling laws"][38] paper, where the raw number of parameters in your model is the most predictive feature of overall performance. This was noted even in the original BERT paper and drives 2020's increase in absolutely massive language models. This is really just [Sutton' Bitter Lesson][39] in action: General methods that leverage computation are ultimately the most effective, and by a large margin Transformers might be [replacing convolution][40], too. As [Yannic Kilcher said][41], transformers are ruining everything. [They work on graphs][6], in fact it's one of the [recent approaches][42], and seems to be one of the more succesful [when benchmarked][1] Researchers seem to be putting so much effort into architecture, but it doesn't matter much in the end because you can approximate anything by stacking more layers. Efficiency wins are great -- but neural net architectures are just one way to achieve that, and by tremendously over-researching this area we're leaving a lot of huge gains elsewhere on the table. Current Graph Data Structure Implementations suck NetworkX is a bad library. I mean, it's good if you're working on tiny graphs for babies, but for anything serious it chokes and forces you to rewrite everything in... what library, really? At this point most people working on large graphs end up hand-rolling some data structure. This is tough because your computer's memory is a 1-dimensional array of 1's and 0's and a graph has no obvious 1-d mapping. This is even harder when we take updating the graph (adding/removing some nodes/edges) into account. Here's a few options: Disconnected networks of pointers NetworkX is the best example. Here, every node is an object with a list of pointers to other nodes (the node's edges). This layout is like a linked list. Linked lists are the [root of all performance evil][43]. Linked lists go completely against how modern computers are designed. Fetching things from memory is slow, and operating on memory is fast (by two orders of magnitude). Whenever you do anything in this layout, you make a roundtrip to RAM. It's slow by design, you can write this in Ruby or C or assembly and it'll be slow regardless, because memory fetches are slow in hardware. The main advantage of this layout is that adding a new node is O(1). So if you're maintaining a massive graph where adding and removing nodes happens as often as reading from the graph, it makes sense. Another advantage of this layout is that it "scales". Because everything is decoupled from each other you can put this data structure on a cluster. However, you're really creating a complex solution for a problem you created for yourself. Sparse Adjacency Matrix This layout great for read-only graphs. I use it as the backend in my [nodevectors][25] library, and many other library writers use the [Scipy CSR Matrix][44], you can see graph algorithms implemented on it [here][45]. The most popular layout for this use is the [CSR Format][46] where you have 3 arrays holding the graph. One for edge destinations, one for edge weights and an "index pointer" which says which edges come from which node. Because the CSR layout is simply 3 arrays, it scales on a single computer: a CSR matrix can be laid out on a disk instead of in-memory. You simply [memory map][47] the 3 arrays and use them on-disk from there. With modern NVMe drives random seeks aren't slow anymore, much faster than distributed network calls like you do when scaling the linked list-based graph. I haven't seen anyone actually implement this yet, but it's in the roadmap for my implementation at least. The problem with this representation is that adding a node or edge means rebuilding the whole data structure. Edgelist representations This representation is three arrays: one for the edge sources, one for the edge destinations, and one for edge weights. [DGL][48] uses this representation internally. This is a simple and compact layout which can be good for analysis. The problem compared to CSR Graphs is some seek operations are slower. Say you want all the edges for node #4243. You can't jump there without maintaining an index pointer array. So either you maintain sorted order and binary search your way there (O(log2n)) or unsorted order and linear search (O(n)). This data structure can also work on memory mapped disk array, and node append is fast on unsorted versions (it's slow in the sorted version). Global methods are a dead end Methods that work on the entire graph at once can't leverage computation, because they run out of RAM at a certain scale. So any method that want a chance of being the new standard need to be able to update piecemeal on parts of the graph. Sampling-based methods Sampling Efficiency will matter more in the future Edgewise local methods. The only algorithms I know of that do this are GloVe and GGVec, which they pass through an edge list and update embedding weights on each step. The problem with this approach is that it's hard to use them for higher-order methods. The advantage is that they easily scale even on one computer. Also, incrementally adding a new node is as simple as taking the existing embeddings, adding a new one, and doing another epoch over the data Random Walk sampling. This is used by deepwalk and its descendants, usually for node embeddings rather than GNN methods. This can be computationally expensive and make it hard to add new nodes. But this does scale, for instance [Instagram][49] use it to feed their recommendation system models Neighbourhood sampling. This is currently the most common one in GNNs, and can be low or higher order depending on the neighborhood size. It also scales well, though implementing efficiently can be challenging. It's currently used by [Pinterest][50]'s recommendation algorithms. Conclusion Here are a few interesting questions: What is the relation between graph types and methods? Consolidated benchmarking like OGB We're throwing random models at random benchmarks without understanding why or when they do better More fundamental research. Heree's one I'm curious about: can other representation types like [Poincarre Embeddings][51] effectively encode directed relationships? On the other hand, we should stop focusing on adding spicy new layers to test on the same tiny datasets. No one cares. [1]: https://arxiv.org/pdf/2003.00982.pdf [2]: https://arxiv.org/pdf/2002.11867.pdf [3]: https://arxiv.org/pdf/1812.08434.pdf [4]: https://arxiv.org/pdf/2005.00687.pdf [5]: https://en.wikipedia.org/wiki/Adjacency_matrix [6]: https://thegradient.pub/transformers-are-graph-neural-networks/ [7]: https://en.wikipedia.org/wiki/Word2vec [8]: https://nlp.stanford.edu/pubs/glove.pdf [9]: https://papers.nips.cc/paper/2014/file/feab05aa91085b7a8012516bc3533958-Paper.pdf [10]: https://en.wikipedia.org/wiki/Bag-of-words_model [11]: https://en.wikipedia.org/wiki/Co-occurrence [12]: https://www.singlelunch.com/2020/02/16/embeddings-from-the-ground-up/ [13]: https://www.singlelunch.com/2019/01/27/word-embeddings-from-the-ground-up/ [14]: https://nlpprogress.com/ [15]: http://socsci.uci.edu/~rfutrell/papers/hahn2019estimating.pdf [16]: https://en.wikipedia.org/wiki/Kolmogorov_complexity [17]: https://bair.berkeley.edu/blog/2020/12/20/lmmem/ [18]: https://en.wikipedia.org/wiki/Laplacian_matrix [19]: http://citeseerx.ist.psu.edu/viewdoc/download;jsessionid=1F03130B02DC485C78BF364266B6F0CA?doi=10.1.1.19.8100&rep=rep1&type=pdf [20]: https://en.wikipedia.org/wiki/Principalcomponentanalysis [21]: https://www.ijcai.org/Proceedings/2019/0594.pdf [22]: https://dl.acm.org/doi/10.1145/2806416.2806512 [23]: https://openreview.net/pdf?id=SyK00v5xx [24]: https://github.com/VHRanger/nodevectors/blob/master/examples/link%20prediction.ipynb [25]: https://github.com/VHRanger/nodevectors [26]: https://arxiv.org/pdf/1310.2636.pdf [27]: http://byowen.com/ [28]: https://arxiv.org/pdf/1807.03341.pdf [29]: https://www.youtube.com/watch?v=Kee4ch3miVA [30]: https://cs.stanford.edu/~jure/pubs/node2vec-kdd16.pdf [31]: https://arxiv.org/pdf/1403.6652.pdf [32]: https://arxiv.org/pdf/1911.11726.pdf [33]: https://en.wikipedia.org/wiki/AlexNet [34]: https://en.wikipedia.org/wiki/Googledatacenters#Original_hardware [35]: https://openai.com/blog/ai-and-efficiency/ [36]: https://www.singlelunch.com/2019/08/01/700x-faster-node2vec-models-fastest-random-walks-on-a-graph/ [37]: https://arxiv.org/pdf/1706.02216.pdf [38]: https://arxiv.org/pdf/2001.08361.pdf [39]: http://incompleteideas.net/IncIdeas/BitterLesson.html [40]: https://arxiv.org/abs/2010.11929 [41]: https://www.youtube.com/watch?v=TrdevFK_am4 [42]: https://arxiv.org/pdf/1710.10903.pdf [43]: https://www.youtube.com/watch?v=fHNmRkzxHWs [44]: https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csr_matrix.html [45]: https://docs.scipy.org/doc/scipy/reference/sparse.csgraph.html [46]: https://en.wikipedia.org/wiki/Sparsematrix#Compressedsparserow(CSR,CRSorYaleformat) [47]: https://en.wikipedia.org/wiki/Mmap [48]: https://github.com/dmlc/dgl [49]: https://ai.facebook.com/blog/powered-by-ai-instagrams-explore-recommender-system/ [50]: https://medium.com/pinterest-engineering/pinsage-a-new-graph-convolutional-neural-network-for-web-scale-recommender-systems-88795a107f48 [51]: https://arxiv.org/pdf/1705.08039.pdf

#reddit #lukewarm #graph

[D] The Rants of an experienced engineer who glimpsed into AI Academia (Briefly)

0.778

donkey_strom16001•This week

[D] The Rants of an experienced engineer who glimpsed into AI Academia (Briefly)

Background I recently graduated with a master's degree and was fortunate/unfortunate to glimpse the whole "Academic" side of ML. I took a thesis track in my degree because as an immigrant it's harder to get into a good research lab without having authorship in a couple of good papers (Or so I delude myself ). I worked as a Full-stack SWE for a startup for 4+ years before coming to the US for a master’s degree focused on ML and AI. I did everything in those years. From project management to building fully polished S/W products to DevOps to even dabbled in ML. I did my Batchelor’s degree from a university whose name is not even worth mentioning. The university for my master’s degree is in the top 20 in the AI space. I didn't know much about ML and the curiosity drove me to university. Come to uni and I focused on learning ML and AI for one 1-1.5 years after which I found advisors for a thesis topic. This is when the fun starts. I had the most amazing advisors but the entire peer review system and the way we assess ML/Science is what ticked me off. This is where the rant begins. Rant 1:Acadmia follows a Gated Institutional Narrative Let's say you are a Ph.D. at the world's top AI institution working under the best prof. You have a way higher likelihood of you getting a good Postdoc at a huge research lab vs someone's from my poor country doing a Ph.D. with a not-so-well-known advisor having published not-so-well-known papers. I come from a developing nation and I see this many times here. In my country academics don't get funding as they do at colleges in the US. One of the reasons for this is that colleges don't have such huge endowments and many academics don't have wealthy research sponsors. Brand names and prestige carry massive weight to help get funding in US academic circles. This prestige/money percolates down to the students and the researchers who work there. Students in top colleges get a huge advantage and the circles of top researchers keep being from the same sets of institutions. I have nothing against top researchers from top institutions but due to the nature of citations and the way the money flows based on them, a vicious cycle is created where the best institutions keep getting better and the rest don't get as much of a notice. Rant 2: Peer Review without Code Review in ML/AI is shady I am a computer scientist and I was appalled when I heard that you don't need to do code reviews for research papers. As a computer scientist and someone who actually did shit tons of actual ML in the past year, I find it absolutely garbage that code reviews are not a part of this system. I am not saying every scientist who reads a paper should review code but at least one person should for any paper's code submission. At least in ML and AI space. This is basic. I don't get why people call themselves computer scientists if they don't want to read the fucking code. If you can't then make a grad student do it. But for the collective of science, we need this. The core problem lies in the fact that peer review is free. : There should be better solutions for this. We ended up creating Git and that changed so many lives. Academic Research needs something similar. Rant 3: My Idea is Novel Until I see Someone Else's Paper The volume of scientific research is growing exponentially. Information is being created faster than we can digest. We can't expect people to know everything and the amount of overlap in the AI/ML fields requires way better search engines than Google Scholar. The side effect of large volumes of research is that every paper is doing something "novel" making it harder to filter what the fuck was novel. I have had so many experiences where I coded up something and came to realize that someone else has done something symbolically similar and my work just seems like a small variant of that. That's what fucks with my head. Is what I did in Novel? What the fuck is Novel? Is stitching up a transformer to any problem with fancy embeddings and tidying it up as a research paper Novel? Is just making a transformer bigger Novel? Is some new RL algorithm tested with 5 seeds and some fancy fucking prior and some esoteric reasoning for its success Novel? Is using an over parameterized model to get 95% accuracy on 200 sample test set Novel? Is apply Self-supervised learning for some new dataset Novel? If I keep on listing questions on novelty, I can probably write a novel asking about what the fuck is "Novel". Rant 4: Citation Based Optimization Promotes Self Growth Over Collective Growth Whatever people may say about collaboration, Academia intrinsically doesn't promote the right incentive structures to harbor collaboration. Let me explain, When you write a paper, the position of your name matters. If you are just a Ph.D. student and a first author to a paper, it's great. If you are an nth author Not so great. Apparently, this is a very touchy thing for academics. And lots of egos can clash around numbering and ordering of names. I distinctly remember once attending some seminar in a lab and approaching a few students on research project ideas. The first thing that came out of the PhD student's mouth was the position in authorship. As an engineer who worked with teams in the past, this was never something I had thought about. Especially because I worked in industry, where it's always the group over the person. Academia is the reverse. Academia applauds the celebration of the individual's achievements. All of this is understandable but it's something I don't like. This makes PhDs stick to their lane. The way citations/research-focus calibrate the "hire-ability" and "completion of Ph.D. thesis" metrics, people are incentivized to think about themselves instead of thinking about collaborations for making something better. Conclusion A Ph.D. in its most idealistic sense for me is the pursuit of hard ideas(I am poetic that way). In a situation like now when you have to publish or perish and words on paper get passed off as science without even seeing the code that runs it, I am extremely discouraged to go down that route. All these rants are not to diss on scientists. I did them because "we" as a community need better ways to addressing some of these problems. P.S. Never expected so many people to express their opinions about this rant. U shouldn’t take this seriously. As many people have stated I am an outsider with tiny experience to give a full picture. I realize that my post as coming out as something which tries to dichotomize academia and industry. I am not trying to do that. I wanted to highlight some problems I saw for which there is no one person to blame. These issues are in my opinion a byproduct of the economics which created this system. Thank you for gold stranger.

#reddit #rants #experienced

[D] What is your honest experience with reinforcement learning?

Starks-Technology•This week

[D] What is your honest experience with reinforcement learning?

In my personal experience, SOTA RL algorithms simply don't work. I've tried working with reinforcement learning for over 5 years. I remember when Alpha Go defeated the world famous Go player, Lee Sedol, and everybody thought RL would take the ML community by storm. Yet, outside of toy problems, I've personally never found a practical use-case of RL. What is your experience with it? Aside from Ad recommendation systems and RLHF, are there legitimate use-cases of RL? Or, was it all hype? Edit: I know a lot about AI. I built NexusTrade, an AI-Powered automated investing tool that lets non-technical users create, update, and deploy their trading strategies. I’m not an idiot nor a noob; RL is just ridiculously hard. Edit 2: Since my comments are being downvoted, here is a link to my article that better describes my position. It's not that I don't understand RL. I released my open-source code and wrote a paper on it. It's the fact that it's EXTREMELY difficult to understand. Other deep learning algorithms like CNNs (including ResNets), RNNs (including GRUs and LSTMs), Transformers, and GANs are not hard to understand. These algorithms work and have practical use-cases outside of the lab. Traditional SOTA RL algorithms like PPO, DDPG, and TD3 are just very hard. You need to do a bunch of research to even implement a toy problem. In contrast, the decision transformer is something anybody can implement, and it seems to match or surpass the SOTA. You don't need two networks battling each other. You don't have to go through hell to debug your network. It just naturally learns the best set of actions in an auto-regressive manner. I also didn't mean to come off as arrogant or imply that RL is not worth learning. I just haven't seen any real-world, practical use-cases of it. I simply wanted to start a discussion, not claim that I know everything. Edit 3: There's a shockingly number of people calling me an idiot for not fully understanding RL. You guys are wayyy too comfortable calling people you disagree with names. News-flash, not everybody has a PhD in ML. My undergraduate degree is in biology. I self-taught myself the high-level maths to understand ML. I'm very passionate about the field; I just have VERY disappointing experiences with RL. Funny enough, there are very few people refuting my actual points. To summarize: Lack of real-world applications Extremely complex and inaccessible to 99% of the population Much harder than traditional DL algorithms like CNNs, RNNs, and GANs Sample inefficiency and instability Difficult to debug Better alternatives, such as the Decision Transformer Are these not legitimate criticisms? Is the purpose of this sub not to have discussions related to Machine Learning? To the few commenters that aren't calling me an idiot...thank you! Remember, it costs you nothing to be nice! Edit 4: Lots of people seem to agree that RL is over-hyped. Unfortunately those comments are downvoted. To clear up some things: We've invested HEAVILY into reinforcement learning. All we got from this investment is a robot that can be super-human at (some) video games. AlphaFold did not use any reinforcement learning. SpaceX doesn't either. I concede that it can be useful for robotics, but still argue that it's use-cases outside the lab are extremely limited. If you're stumbling on this thread and curious about an RL alternative, check out the Decision Transformer. It can be used in any situation that a traditional RL algorithm can be used. Final Edit: To those who contributed more recently, thank you for the thoughtful discussion! From what I learned, model-based models like Dreamer and IRIS MIGHT have a future. But everybody who has actually used model-free models like DDPG unanimously agree that they suck and don’t work.

#reddit #your #honest

[P] I Trained a Model to Generate Video Game Pages

pcvision•This week

[P] I Trained a Model to Generate Video Game Pages

These past two months I've been working on a project I've called THIS GAME DOES NOT EXIST. I've always wanted to try building something with generative A.I. so this project scratched that itch for me. Here's a video with a few of my favourites read by voice actors: https://www.youtube.com/watch?v=\mTWMLhpJoA  THIS GAME DOES NOT EXIST is an experiment in generative artificial intelligence. This site contains 130 video game pages that were generated using an implementation of OpenAI's Generative Pre-trained Transformer 2 (GPT-2) to generate text and a simple implementation of generative adversarial networks (GAN) to generate header images and "screenshots". To generate the names, descriptions, publishers, and developers of the games I finetuned the HuggingFace implementation of GPT-2. I used the Steam Store Games (Clean dataset) from Kaggle with slight modifications and preprocessing.Here is what one training sample looks like: Half-LifeValve ValveNamed Game of the Year by over 50 publications, Valve's debut title blends action and adventure with award-winning technology to create a frighteningly realistic world where players must think to survive. Also includes an exciting multiplayer mode that allows you to play against friends and enemies around the world. The model uses the tokens (e.g. and ) to prompt each class of data while keeping context during the entire generation. Image generation was done by training a custom GAN very similar to the architecture seen in the PyTorch DCGAN Tutorial which was built to generate faces. I created two models for this site: one for generating the header images and one for generating multiple screenshots for each game.To assemble the dataset I wrote a script that downloads the images from the URLs in the Steam Store Games (Clean dataset) dataset. Due to my lack of resources and time to put into this project, the image generation is less than ideal. You may notice though, that the header image model will generate artifacts in images that look like the titles of games, and the screenshot image model with generate what looks like levels of a 2D platformer.

#reddit #trained #model

[P] A Simpler @PyTorch Annotated Implementation of EleutherAI's 20B Language Model GPT-NeoX.

hnipun•This week

[P] A Simpler @PyTorch Annotated Implementation of EleutherAI's 20B Language Model GPT-NeoX.

Github: https://github.com/labmlai/neox Annotated implementation: https://lit.labml.ai/github/labmlai/neox/tree/main/src/neox/\\init\\.py Original repo from EleutherAI: https://github.com/EleutherAI/gpt-neox We have included samples showing how to generate text and to fine-tune. We haven't included a bunch of optimizations that were present in original GPT-NeoX to keep things simple.

#reddit #simpler #@pytorch

Starks-Technology•This week

[D] What is your honest experience with reinforcement learning?

#reddit #your #honest

[P] I created a package implementing a SOTA technique for XAI ( Explainable AI)

PressureDry1111•This week

[P] I created a package implementing a SOTA technique for XAI ( Explainable AI)

This is the package https://github.com/mfumagalli68/xi-method Follow the README and install directly from pypi. From the paper: " \[..\]To bridge this gap we propose a family of measures of statistical association whose definition is well-posed also for nonordered data. Our intuition is to rely on separation measurements between probability mass functions. Here, by separation measurement we mean any distance or divergence between probability mass functions that is positive, and that is null if and only if the probability mass functions coincide. Then, we show that the new class of sensitivity indices complies with Renyi’s postulate D of measures of statistical dependence (Renyi, 1959). This postulate, called zero-independence property in the following, requires that a measure of association is null if and only if the two random variables are statistically independent. We address the estimation of this new class of indicators for generic samples, and discuss their asymptotic convergence. We then use these probabilistic sensitivity measures in the context of explainability. A relevant aspect related to measures of statistical association is that they can be computed directly on the original dataset without the need of actually fitting a machine learning model. Thus, not only are they model agnostic in explaining the behavior of a black box, but they also provide pre-hoc explanations. Our intuition is then to compare explanations provided by measures of statistical association first calculated on the original data (the pre-hoc explanations) and then on the forecasts of the machine learning model fitted to the data (post-hoc explanations). This comparison provides an indication on whether the ML model predictions capture the statistical dependence originally present in the data. We call the resulting approach Xi-method\[...\] " The paper can't be shared freely, but as always with a little bit of research you can find it online. If you find it interesting, star the repo.  Thanks

#reddit #created #package

[P]MMML | Deploy HuggingFace training model rapidly based on MetaSpore

qazmkopp•This week

[P]MMML | Deploy HuggingFace training model rapidly based on MetaSpore

A few days ago, HuggingFace announced a $100 million Series C funding round, which was big news in open source machine learning and could be a sign of where the industry is headed. Two days before the HuggingFace funding announcement, open-source machine learning platform MetaSpore released a demo based on the HuggingFace Rapid deployment pre-training model. As deep learning technology makes innovative breakthroughs in computer vision, natural language processing, speech understanding, and other fields, more and more unstructured data are perceived, understood, and processed by machines. These advances are mainly due to the powerful learning ability of deep learning. Through pre-training of deep models on massive data, the models can capture the internal data patterns, thus helping many downstream tasks. With the industry and academia investing more and more energy in the research of pre-training technology, the distribution warehouses of pre-training models such as HuggingFace and Timm have emerged one after another. The open-source community release pre-training significant model dividends at an unprecedented speed. In recent years, the data form of machine modeling and understanding has gradually evolved from single-mode to multi-mode, and the semantic gap between different modes is being eliminated, making it possible to retrieve data across modes. Take CLIP, OpenAI’s open-source work, as an example, to pre-train the twin towers of images and texts on a dataset of 400 million pictures and texts and connect the semantics between pictures and texts. Many researchers in the academic world have been solving multimodal problems such as image generation and retrieval based on this technology. Although the frontier technology through the semantic gap between modal data, there is still a heavy and complicated model tuning, offline data processing, high performance online reasoning architecture design, heterogeneous computing, and online algorithm be born multiple processes and challenges, hindering the frontier multimodal retrieval technologies fall to the ground and pratt &whitney. DMetaSoul aims at the above technical pain points, abstracting and uniting many links such as model training optimization, online reasoning, and algorithm experiment, forming a set of solutions that can quickly apply offline pre-training model to online. This paper will introduce how to use the HuggingFace community pre-training model to conduct online reasoning and algorithm experiments based on MetaSpore technology ecology so that the benefits of the pre-training model can be fully released to the specific business or industry and small and medium-sized enterprises. And we will give the text search text and text search graph two multimodal retrieval demonstration examples for your reference. Multimodal semantic retrieval The sample architecture of multimodal retrieval is as follows: Our multimodal retrieval system supports both text search and text search application scenarios, including offline processing, model reasoning, online services, and other core modules:  https://preview.redd.it/w4v4c7vcez291.png?width=1834&format=png&auto=webp&s=0687efb1fddb26e8e30cb844d398ec712b947f31 Offline processing, including offline data processing processes for different application scenarios of text search and text search, including model tuning, model export, data index database construction, data push, etc. Model inference. After the offline model training, we deployed our NLP and CV large models based on the MetaSpore Serving framework. MetaSpore Serving helps us conveniently perform online inference, elastic scheduling, load balancing, and resource scheduling in heterogeneous environments. Online services. Based on MetaSpore’s online algorithm application framework, MetaSpore has a complete set of reusable online search services, including Front-end retrieval UI, multimodal data preprocessing, vector recall and sorting algorithm, AB experimental framework, etc. MetaSpore also supports text search by text and image scene search by text and can be migrated to other application scenarios at a low cost. The HuggingFace open source community has provided several excellent baseline models for similar multimodal retrieval problems, which are often the starting point for actual optimization in the industry. MetaSpore also uses the pre-training model of the HuggingFace community in its online services of searching words by words and images by words. Searching words by words is based on the semantic similarity model of the question and answer field optimized by MetaSpore, and searching images by words is based on the community pre-training model. These community open source pre-training models are exported to the general ONNX format and loaded into MetaSpore Serving for online reasoning. The following sections will provide a detailed description of the model export and online retrieval algorithm services. The reasoning part of the model is standardized SAAS services with low coupling with the business. Interested readers can refer to my previous post: The design concept of MetaSpore, a new generation of the one-stop machine learning platform. 1.1 Offline Processing Offline processing mainly involves the export and loading of online models and index building and pushing of the document library. You can follow the step-by-step instructions below to complete the offline processing of text search and image search and see how the offline pre-training model achieves reasoning at MetaSpore. 1.1.1 Search text by text Traditional text retrieval systems are based on literal matching algorithms such as BM25. Due to users’ diverse query words, a semantic gap between query words and documents is often encountered. For example, users misspell “iPhone” as “Phone,” and search terms are incredibly long, such as “1 \~ 3 months old baby autumn small size bag pants”. Traditional text retrieval systems will use spelling correction, synonym expansion, search terms rewriting, and other means to alleviate the semantic gap but fundamentally fail to solve this problem. Only when the retrieval system fully understands users’ query terms and documents can it meet users’ retrieval demands at the semantic level. With the continuous progress of pre-training and representational learning technology, some commercial search engines continue to integrate semantic vector retrieval methods based on symbolic learning into the retrieval ecology. Semantic retrieval model This paper introduces a set of semantic vector retrieval applications. MetaSpore built a set of semantic retrieval systems based on encyclopedia question and answer data. MetaSpore adopted the Sentence-Bert model as the semantic vector representation model, which fine-tunes the twin tower BERT in supervised or unsupervised ways to make the model more suitable for retrieval tasks. The model structure is as follows: The query-Doc symmetric two-tower model is used in text search and question and answer retrieval. The vector representation of online Query and offline DOC share the same vector representation model, so it is necessary to ensure the consistency of the offline DOC library building model and online Query inference model. The case uses MetaSpore’s text representation model Sbert-Chinese-QMC-domain-V1, optimized in the open-source semantically similar data set. This model will express the question and answer data as a vector in offline database construction. The user query will be expressed as a vector by this model in online retrieval, ensuring that query-doc in the same semantic space, users’ semantic retrieval demands can be guaranteed by vector similarity metric calculation. Since the text presentation model does vector encoding for Query online, we need to export the model for use by the online service. Go to the q&A data library code directory and export the model concerning the documentation. In the script, Pytorch Tracing is used to export the model. The models are exported to the “./export “directory. The exported models are mainly ONNX models used for wired reasoning, Tokenizer, and related configuration files. The exported models are loaded into MetaSpore Serving by the online Serving system described below for model reasoning. Since the exported model will be copied to the cloud storage, you need to configure related variables in env.sh. \Build library based on text search \ The retrieval database is built on the million-level encyclopedia question and answer data set. According to the description document, you need to download the data and complete the database construction. The question and answer data will be coded as a vector by the offline model, and then the database construction data will be pushed to the service component. The whole process of database construction is described as follows: Preprocessing, converting the original data into a more general JSonline format for database construction; Build index, use the same model as online “sbert-Chinese-qmc-domain-v1” to index documents (one document object per line); Push inverted (vector) and forward (document field) data to each component server. The following is an example of the database data format. After offline database construction is completed, various data are pushed to corresponding service components, such as Milvus storing vector representation of documents and MongoDB storing summary information of documents. Online retrieval algorithm services will use these service components to obtain relevant data. 1.1.2 Search by text Text and images are easy for humans to relate semantically but difficult for machines. First of all, from the perspective of data form, the text is the discrete ID type of one-dimensional data based on words and words. At the same time, images are continuous two-dimensional or three-dimensional data. Secondly, the text is a subjective creation of human beings, and its expressive ability is vibrant, including various turning points, metaphors, and other expressions, while images are machine representations of the objective world. In short, bridging the semantic gap between text and image data is much more complex than searching text by text. The traditional text search image retrieval technology generally relies on the external text description data of the image or the nearest neighbor retrieval technology and carries out the retrieval through the image associated text, which in essence degrades the problem to text search. However, it will also face many issues, such as obtaining the associated text of pictures and whether the accuracy of text search by text is high enough. The depth model has gradually evolved from single-mode to multi-mode in recent years. Taking the open-source project of OpenAI, CLIP, as an example, train the model through the massive image and text data of the Internet and map the text and image data into the same semantic space, making it possible to implement the text and image search technology based on semantic vector. CLIP graphic model The text search pictures introduced in this paper are implemented based on semantic vector retrieval, and the CLIP pre-training model is used as the two-tower retrieval architecture. Because the CLIP model has trained the semantic alignment of the twin towers’ text and image side models on the massive graphic and text data, it is particularly suitable for the text search graph scene. Due to the different image and text data forms, the Query-Doc asymmetric twin towers model is used for text search image retrieval. The image-side model of the twin towers is used for offline database construction, and the text-side model is used for the online return. In the final online retrieval, the database data of the image side model will be searched after the text side model encodes Query, and the CLIP pre-training model guarantees the semantic correlation between images and texts. The model can draw the graphic pairs closer in vector space by pre-training on a large amount of visual data. Here we need to export the text-side model for online MetaSpore Serving inference. Since the retrieval scene is based on Chinese, the CLIP model supporting Chinese understanding is selected. The exported content includes the ONNX model used for online reasoning and Tokenizer, similar to the text search. MetaSpore Serving can load model reasoning through the exported content. Build library on Image search You need to download the Unsplash Lite library data and complete the construction according to the instructions. The whole process of database construction is described as follows: Preprocessing, specify the image directory, and then generate a more general JSOnline file for library construction; Build index, use OpenAI/Clip-Vit-BASE-Patch32 pre-training model to index the gallery, and output one document object for each line of index data; Push inverted (vector) and forward (document field) data to each component server. Like text search, after offline database construction, relevant data will be pushed to service components, called by online retrieval algorithm services to obtain relevant data. 1.2 Online Services The overall online service architecture diagram is as follows: https://preview.redd.it/jfsl8hdfez291.png?width=1280&format=png&auto=webp&s=a858e2304a0c93e78ba5429612ca08cbee69b35a Multi-mode search online service system supports application scenarios such as text search and text search. The whole online service consists of the following parts: Query preprocessing service: encapsulate preprocessing logic (including text/image, etc.) of pre-training model, and provide services through gRPC interface; Retrieval algorithm service: the whole algorithm processing link includes AB experiment tangent flow configuration, MetaSpore Serving call, vector recall, sorting, document summary, etc.; User entry service: provides a Web UI interface for users to debug and track down problems in the retrieval service. From a user request perspective, these services form invocation dependencies from back to front, so to build up a multimodal sample, you need to run each service from front to back first. Before doing this, remember to export the offline model, put it online and build the library first. This article will introduce the various parts of the online service system and make the whole service system step by step according to the following guidance. See the ReadME at the end of this article for more details. 1.2.1 Query preprocessing service Deep learning models tend to be based on tensors, but NLP/CV models often have a preprocessing part that translates raw text and images into tensors that deep learning models can accept. For example, NLP class models often have a pre-tokenizer to transform text data of string type into discrete tensor data. CV class models also have similar processing logic to complete the cropping, scaling, transformation, and other processing of input images through preprocessing. On the one hand, considering that this part of preprocessing logic is decoupled from tensor reasoning of the depth model, on the other hand, the reason of the depth model has an independent technical system based on ONNX, so MetaSpore disassembled this part of preprocessing logic. NLP pretreatment Tokenizer has been integrated into the Query pretreatment service. MetaSpore dismantlement with a relatively general convention. Users only need to provide preprocessing logic files to realize the loading and prediction interface and export the necessary data and configuration files loaded into the preprocessing service. Subsequent CV preprocessing logic will also be integrated in this manner. The preprocessing service currently provides the gRPC interface invocation externally and is dependent on the Query preprocessing (QP) module in the retrieval algorithm service. After the user request reaches the retrieval algorithm service, it will be forwarded to the service to complete the data preprocessing and continue the subsequent processing. The ReadMe provides details on how the preprocessing service is started, how the preprocessing model exported offline to cloud storage enters the service, and how to debug the service. To further improve the efficiency and stability of model reasoning, MetaSpore Serving implements a Python preprocessing submodule. So MetaSpore can provide gRPC services through user-specified preprocessor.py, complete Tokenizer or CV-related preprocessing in NLP, and translate requests into a Tensor that deep models can handle. Finally, the model inference is carried out by MetaSpore, Serving subsequent sub-modules. Presented here on the lot code: https://github.com/meta-soul/MetaSpore/compare/add\python\preprocessor 1.2.2 Retrieval algorithm services Retrieval algorithm service is the core of the whole online service system, which is responsible for the triage of experiments, the assembly of algorithm chains such as preprocessing, recall, sorting, and the invocation of dependent component services. The whole retrieval algorithm service is developed based on the Java Spring framework and supports multi-mode retrieval scenarios of text search and text search graph. Due to good internal abstraction and modular design, it has high flexibility and can be migrated to similar application scenarios at a low cost. Here’s a quick guide to configuring the environment to set up the retrieval algorithm service. See ReadME for more details: Install dependent components. Use Maven to install the online-Serving component Search for service configurations. Copy the template configuration file and replace the MongoDB, Milvus, and other configurations based on the development/production environment. Install and configure Consul. Consul allows you to synchronize the search service configuration in real-time, including cutting the flow of experiments, recall parameters, and sorting parameters. The project’s configuration file shows the current configuration parameters of text search and text search. The parameter modelName in the stage of pretreatment and recall is the corresponding model exported in offline processing. Start the service. Once the above configuration is complete, the retrieval service can be started from the entry script. Once the service is started, you can test it! For example, for a user with userId=10 who wants to query “How to renew ID card,” access the text search service. 1.2.3 User Entry Service Considering that the retrieval algorithm service is in the form of the API interface, it is difficult to locate and trace the problem, especially for the text search image scene can intuitively display the retrieval results to facilitate the iterative optimization of the retrieval algorithm. This paper provides a lightweight Web UI interface for text search and image search, a search input box, and results in a display page for users. Developed by Flask, the service can be easily integrated with other retrieval applications. The service calls the retrieval algorithm service and displays the returned results on the page. It’s also easy to install and start the service. Once you’re done, go to http://127.0.0.1:8090 to see if the search UI service is working correctly. See the ReadME at the end of this article for details. Multimodal system demonstration The multimodal retrieval service can be started when offline processing and online service environment configuration have been completed following the above instructions. Examples of textual searches are shown below. Enter the entry of the text search map application, enter “cat” first, and you can see that the first three digits of the returned result are cats: https://preview.redd.it/0n5nuyvhez291.png?width=1280&format=png&auto=webp&s=1e9c054f541d53381674b8d6001b4bf524506bd2 If you add a color constraint to “cat” to retrieve “black cat,” you can see that it does return a black cat: https://preview.redd.it/rzc0qjyjez291.png?width=1280&format=png&auto=webp&s=d5bcc503ef0fb3360c7740e60e295cf372dcad47 Further, strengthen the constraint on the search term, change it to “black cat on the bed,” and return results containing pictures of a black cat climbing on the bed:  https://preview.redd.it/c4b2q8olez291.png?width=1280&format=png&auto=webp&s=4f3817b0b9f07e1e68d1d4a8281702ba3834a00a The cat can still be found through the text search system after the color and scene modification in the above example. Conclusion The cutting-edge pre-training technology can bridge the semantic gap between different modes, and the HuggingFace community can greatly reduce the cost for developers to use the pre-training model. Combined with the technological ecology of MetaSpore online reasoning and online microservices provided by DMetaSpore, the pre-training model is no longer mere offline dabbling. Instead, it can truly achieve end-to-end implementation from cutting-edge technology to industrial scenarios, fully releasing the dividends of the pre-training large model. In the future, DMetaSoul will continue to improve and optimize the MetaSpore technology ecosystem: More automated and wider access to HuggingFace community ecology. MetaSpore will soon release a common model rollout mechanism to make HuggingFace ecologically accessible and will later integrate preprocessing services into online services. Multi-mode retrieval offline algorithm optimization. For multimodal retrieval scenarios, MetaSpore will continuously iteratively optimize offline algorithm components, including text recall/sort model, graphic recall/sort model, etc., to improve the accuracy and efficiency of the retrieval algorithm. For related code and reference documentation in this article, please visit: https://github.com/meta-soul/MetaSpore/tree/main/demo/multimodal/online Some images source: https://github.com/openai/CLIP/raw/main/CLIP.png https://www.sbert.net/examples/training/sts/README.html

#reddit #[p]mmml #deploy

[D] Last Week in Medical AI: Top LLM Research Papers/Models (December 7 - December 14, 2024)

aadityaura•This week

[D] Last Week in Medical AI: Top LLM Research Papers/Models (December 7 - December 14, 2024)

[\[D\] Last Week in Medical AI: Top LLM Research Papers\/Models $December 7 - December 14, 2024$](https://preview.redd.it/o23fp3csj07e1.jpg?width=1280&format=pjpg&auto=webp&s=69e19fc351b3aa5e34c4c00e66245583f88bd9bb) Medical LLM & Other Models PediaBench: Chinese Pediatric LLM This paper introduces PediaBench, the first Chinese pediatric dataset for evaluating Large Language Model (LLM) question-answering performance, containing 4,565 objective and 1,632 subjective questions across 12 disease groups. BiMediX: Bilingual Medical LLM This paper introduces BiMediX, the first bilingual (English-Arabic) medical Mixture of Experts LLM, along with BiMed1.3M, a 1.3M bilingual medical instruction dataset with over 632M tokens used for training. Diverse medical knowledge integration This paper introduces BiMediX2, a bilingual (Arabic-English) Large Multimodal Model (LMM) based on Llama3.1 architecture, trained on 1.6M medical interaction samples. BRAD: Digital Biology Language Model This paper introduces BRAD (Bioinformatics Retrieval Augmented Digital assistant), an LLM-powered chatbot and agent system integrating various bioinformatics tools. MMedPO: Vision-Language Medical LLM This paper introduces MMedPO, a multimodal medical preference optimization approach to improve factual accuracy in Medical Large Vision-Language Models (Med-LVLMs) by addressing modality misalignment. Frameworks & Methodologies \- TOP-Training: Medical Q&A Framework \- Hybrid RAG: Secure Medical Data Management \- Zero-Shot ATC Clinical Coding \- Chest X-Ray Diagnosis Architecture \- Medical Imaging AI Democratization Benchmarks & Evaluations \- KorMedMCQA: Korean Healthcare Licensing Benchmark \- Large Language Model Medical Tasks \- Clinical T5 Model Performance Study \- Radiology Report Quality Assessment \- Genomic Analysis Benchmarking LLM Applications \- TCM-FTP: Herbal Prescription Prediction \- LLaSA: Activity Analysis via Sensors \- Emergency Department Visit Predictions \- Neurodegenerative Disease AI Diagnosis \- Kidney Disease Explainable AI Model Ethical AI & Privacy \- Privacy-Preserving LLM Mechanisms \- AI-Driven Digital Organism Modeling \- Biomedical Research Automation \- Multimodality in Medical Practice Full thread in detail: https://x.com/OpenlifesciAI/status/1867999825721242101

#reddit #last #week

qazmkopp•This week

[P]MMML | Deploy HuggingFace training model rapidly based on MetaSpore

A few days ago, HuggingFace announced a $100 million Series C funding round, which was big news in open source machine learning and could be a sign of where the industry is headed. Two days before the HuggingFace funding announcement, open-source machine learning platform MetaSpore released a demo based on the HuggingFace Rapid deployment pre-training model. As deep learning technology makes innovative breakthroughs in computer vision, natural language processing, speech understanding, and other fields, more and more unstructured data are perceived, understood, and processed by machines. These advances are mainly due to the powerful learning ability of deep learning. Through pre-training of deep models on massive data, the models can capture the internal data patterns, thus helping many downstream tasks. With the industry and academia investing more and more energy in the research of pre-training technology, the distribution warehouses of pre-training models such as HuggingFace and Timm have emerged one after another. The open-source community release pre-training significant model dividends at an unprecedented speed. In recent years, the data form of machine modeling and understanding has gradually evolved from single-mode to multi-mode, and the semantic gap between different modes is being eliminated, making it possible to retrieve data across modes. Take CLIP, OpenAI’s open-source work, as an example, to pre-train the twin towers of images and texts on a dataset of 400 million pictures and texts and connect the semantics between pictures and texts. Many researchers in the academic world have been solving multimodal problems such as image generation and retrieval based on this technology. Although the frontier technology through the semantic gap between modal data, there is still a heavy and complicated model tuning, offline data processing, high performance online reasoning architecture design, heterogeneous computing, and online algorithm be born multiple processes and challenges, hindering the frontier multimodal retrieval technologies fall to the ground and pratt &whitney. DMetaSoul aims at the above technical pain points, abstracting and uniting many links such as model training optimization, online reasoning, and algorithm experiment, forming a set of solutions that can quickly apply offline pre-training model to online. This paper will introduce how to use the HuggingFace community pre-training model to conduct online reasoning and algorithm experiments based on MetaSpore technology ecology so that the benefits of the pre-training model can be fully released to the specific business or industry and small and medium-sized enterprises. And we will give the text search text and text search graph two multimodal retrieval demonstration examples for your reference. Multimodal semantic retrieval The sample architecture of multimodal retrieval is as follows: Our multimodal retrieval system supports both text search and text search application scenarios, including offline processing, model reasoning, online services, and other core modules:  https://preview.redd.it/w4v4c7vcez291.png?width=1834&format=png&auto=webp&s=0687efb1fddb26e8e30cb844d398ec712b947f31 Offline processing, including offline data processing processes for different application scenarios of text search and text search, including model tuning, model export, data index database construction, data push, etc. Model inference. After the offline model training, we deployed our NLP and CV large models based on the MetaSpore Serving framework. MetaSpore Serving helps us conveniently perform online inference, elastic scheduling, load balancing, and resource scheduling in heterogeneous environments. Online services. Based on MetaSpore’s online algorithm application framework, MetaSpore has a complete set of reusable online search services, including Front-end retrieval UI, multimodal data preprocessing, vector recall and sorting algorithm, AB experimental framework, etc. MetaSpore also supports text search by text and image scene search by text and can be migrated to other application scenarios at a low cost. The HuggingFace open source community has provided several excellent baseline models for similar multimodal retrieval problems, which are often the starting point for actual optimization in the industry. MetaSpore also uses the pre-training model of the HuggingFace community in its online services of searching words by words and images by words. Searching words by words is based on the semantic similarity model of the question and answer field optimized by MetaSpore, and searching images by words is based on the community pre-training model. These community open source pre-training models are exported to the general ONNX format and loaded into MetaSpore Serving for online reasoning. The following sections will provide a detailed description of the model export and online retrieval algorithm services. The reasoning part of the model is standardized SAAS services with low coupling with the business. Interested readers can refer to my previous post: The design concept of MetaSpore, a new generation of the one-stop machine learning platform. 1.1 Offline Processing Offline processing mainly involves the export and loading of online models and index building and pushing of the document library. You can follow the step-by-step instructions below to complete the offline processing of text search and image search and see how the offline pre-training model achieves reasoning at MetaSpore. 1.1.1 Search text by text Traditional text retrieval systems are based on literal matching algorithms such as BM25. Due to users’ diverse query words, a semantic gap between query words and documents is often encountered. For example, users misspell “iPhone” as “Phone,” and search terms are incredibly long, such as “1 \~ 3 months old baby autumn small size bag pants”. Traditional text retrieval systems will use spelling correction, synonym expansion, search terms rewriting, and other means to alleviate the semantic gap but fundamentally fail to solve this problem. Only when the retrieval system fully understands users’ query terms and documents can it meet users’ retrieval demands at the semantic level. With the continuous progress of pre-training and representational learning technology, some commercial search engines continue to integrate semantic vector retrieval methods based on symbolic learning into the retrieval ecology. Semantic retrieval model This paper introduces a set of semantic vector retrieval applications. MetaSpore built a set of semantic retrieval systems based on encyclopedia question and answer data. MetaSpore adopted the Sentence-Bert model as the semantic vector representation model, which fine-tunes the twin tower BERT in supervised or unsupervised ways to make the model more suitable for retrieval tasks. The model structure is as follows: The query-Doc symmetric two-tower model is used in text search and question and answer retrieval. The vector representation of online Query and offline DOC share the same vector representation model, so it is necessary to ensure the consistency of the offline DOC library building model and online Query inference model. The case uses MetaSpore’s text representation model Sbert-Chinese-QMC-domain-V1, optimized in the open-source semantically similar data set. This model will express the question and answer data as a vector in offline database construction. The user query will be expressed as a vector by this model in online retrieval, ensuring that query-doc in the same semantic space, users’ semantic retrieval demands can be guaranteed by vector similarity metric calculation. Since the text presentation model does vector encoding for Query online, we need to export the model for use by the online service. Go to the q&A data library code directory and export the model concerning the documentation. In the script, Pytorch Tracing is used to export the model. The models are exported to the “./export “directory. The exported models are mainly ONNX models used for wired reasoning, Tokenizer, and related configuration files. The exported models are loaded into MetaSpore Serving by the online Serving system described below for model reasoning. Since the exported model will be copied to the cloud storage, you need to configure related variables in env.sh. \Build library based on text search \ The retrieval database is built on the million-level encyclopedia question and answer data set. According to the description document, you need to download the data and complete the database construction. The question and answer data will be coded as a vector by the offline model, and then the database construction data will be pushed to the service component. The whole process of database construction is described as follows: Preprocessing, converting the original data into a more general JSonline format for database construction; Build index, use the same model as online “sbert-Chinese-qmc-domain-v1” to index documents (one document object per line); Push inverted (vector) and forward (document field) data to each component server. The following is an example of the database data format. After offline database construction is completed, various data are pushed to corresponding service components, such as Milvus storing vector representation of documents and MongoDB storing summary information of documents. Online retrieval algorithm services will use these service components to obtain relevant data. 1.1.2 Search by text Text and images are easy for humans to relate semantically but difficult for machines. First of all, from the perspective of data form, the text is the discrete ID type of one-dimensional data based on words and words. At the same time, images are continuous two-dimensional or three-dimensional data. Secondly, the text is a subjective creation of human beings, and its expressive ability is vibrant, including various turning points, metaphors, and other expressions, while images are machine representations of the objective world. In short, bridging the semantic gap between text and image data is much more complex than searching text by text. The traditional text search image retrieval technology generally relies on the external text description data of the image or the nearest neighbor retrieval technology and carries out the retrieval through the image associated text, which in essence degrades the problem to text search. However, it will also face many issues, such as obtaining the associated text of pictures and whether the accuracy of text search by text is high enough. The depth model has gradually evolved from single-mode to multi-mode in recent years. Taking the open-source project of OpenAI, CLIP, as an example, train the model through the massive image and text data of the Internet and map the text and image data into the same semantic space, making it possible to implement the text and image search technology based on semantic vector. CLIP graphic model The text search pictures introduced in this paper are implemented based on semantic vector retrieval, and the CLIP pre-training model is used as the two-tower retrieval architecture. Because the CLIP model has trained the semantic alignment of the twin towers’ text and image side models on the massive graphic and text data, it is particularly suitable for the text search graph scene. Due to the different image and text data forms, the Query-Doc asymmetric twin towers model is used for text search image retrieval. The image-side model of the twin towers is used for offline database construction, and the text-side model is used for the online return. In the final online retrieval, the database data of the image side model will be searched after the text side model encodes Query, and the CLIP pre-training model guarantees the semantic correlation between images and texts. The model can draw the graphic pairs closer in vector space by pre-training on a large amount of visual data. Here we need to export the text-side model for online MetaSpore Serving inference. Since the retrieval scene is based on Chinese, the CLIP model supporting Chinese understanding is selected. The exported content includes the ONNX model used for online reasoning and Tokenizer, similar to the text search. MetaSpore Serving can load model reasoning through the exported content. Build library on Image search You need to download the Unsplash Lite library data and complete the construction according to the instructions. The whole process of database construction is described as follows: Preprocessing, specify the image directory, and then generate a more general JSOnline file for library construction; Build index, use OpenAI/Clip-Vit-BASE-Patch32 pre-training model to index the gallery, and output one document object for each line of index data; Push inverted (vector) and forward (document field) data to each component server. Like text search, after offline database construction, relevant data will be pushed to service components, called by online retrieval algorithm services to obtain relevant data. 1.2 Online Services The overall online service architecture diagram is as follows: https://preview.redd.it/jfsl8hdfez291.png?width=1280&format=png&auto=webp&s=a858e2304a0c93e78ba5429612ca08cbee69b35a Multi-mode search online service system supports application scenarios such as text search and text search. The whole online service consists of the following parts: Query preprocessing service: encapsulate preprocessing logic (including text/image, etc.) of pre-training model, and provide services through gRPC interface; Retrieval algorithm service: the whole algorithm processing link includes AB experiment tangent flow configuration, MetaSpore Serving call, vector recall, sorting, document summary, etc.; User entry service: provides a Web UI interface for users to debug and track down problems in the retrieval service. From a user request perspective, these services form invocation dependencies from back to front, so to build up a multimodal sample, you need to run each service from front to back first. Before doing this, remember to export the offline model, put it online and build the library first. This article will introduce the various parts of the online service system and make the whole service system step by step according to the following guidance. See the ReadME at the end of this article for more details. 1.2.1 Query preprocessing service Deep learning models tend to be based on tensors, but NLP/CV models often have a preprocessing part that translates raw text and images into tensors that deep learning models can accept. For example, NLP class models often have a pre-tokenizer to transform text data of string type into discrete tensor data. CV class models also have similar processing logic to complete the cropping, scaling, transformation, and other processing of input images through preprocessing. On the one hand, considering that this part of preprocessing logic is decoupled from tensor reasoning of the depth model, on the other hand, the reason of the depth model has an independent technical system based on ONNX, so MetaSpore disassembled this part of preprocessing logic. NLP pretreatment Tokenizer has been integrated into the Query pretreatment service. MetaSpore dismantlement with a relatively general convention. Users only need to provide preprocessing logic files to realize the loading and prediction interface and export the necessary data and configuration files loaded into the preprocessing service. Subsequent CV preprocessing logic will also be integrated in this manner. The preprocessing service currently provides the gRPC interface invocation externally and is dependent on the Query preprocessing (QP) module in the retrieval algorithm service. After the user request reaches the retrieval algorithm service, it will be forwarded to the service to complete the data preprocessing and continue the subsequent processing. The ReadMe provides details on how the preprocessing service is started, how the preprocessing model exported offline to cloud storage enters the service, and how to debug the service. To further improve the efficiency and stability of model reasoning, MetaSpore Serving implements a Python preprocessing submodule. So MetaSpore can provide gRPC services through user-specified preprocessor.py, complete Tokenizer or CV-related preprocessing in NLP, and translate requests into a Tensor that deep models can handle. Finally, the model inference is carried out by MetaSpore, Serving subsequent sub-modules. Presented here on the lot code: https://github.com/meta-soul/MetaSpore/compare/add\python\preprocessor 1.2.2 Retrieval algorithm services Retrieval algorithm service is the core of the whole online service system, which is responsible for the triage of experiments, the assembly of algorithm chains such as preprocessing, recall, sorting, and the invocation of dependent component services. The whole retrieval algorithm service is developed based on the Java Spring framework and supports multi-mode retrieval scenarios of text search and text search graph. Due to good internal abstraction and modular design, it has high flexibility and can be migrated to similar application scenarios at a low cost. Here’s a quick guide to configuring the environment to set up the retrieval algorithm service. See ReadME for more details: Install dependent components. Use Maven to install the online-Serving component Search for service configurations. Copy the template configuration file and replace the MongoDB, Milvus, and other configurations based on the development/production environment. Install and configure Consul. Consul allows you to synchronize the search service configuration in real-time, including cutting the flow of experiments, recall parameters, and sorting parameters. The project’s configuration file shows the current configuration parameters of text search and text search. The parameter modelName in the stage of pretreatment and recall is the corresponding model exported in offline processing. Start the service. Once the above configuration is complete, the retrieval service can be started from the entry script. Once the service is started, you can test it! For example, for a user with userId=10 who wants to query “How to renew ID card,” access the text search service. 1.2.3 User Entry Service Considering that the retrieval algorithm service is in the form of the API interface, it is difficult to locate and trace the problem, especially for the text search image scene can intuitively display the retrieval results to facilitate the iterative optimization of the retrieval algorithm. This paper provides a lightweight Web UI interface for text search and image search, a search input box, and results in a display page for users. Developed by Flask, the service can be easily integrated with other retrieval applications. The service calls the retrieval algorithm service and displays the returned results on the page. It’s also easy to install and start the service. Once you’re done, go to http://127.0.0.1:8090 to see if the search UI service is working correctly. See the ReadME at the end of this article for details. Multimodal system demonstration The multimodal retrieval service can be started when offline processing and online service environment configuration have been completed following the above instructions. Examples of textual searches are shown below. Enter the entry of the text search map application, enter “cat” first, and you can see that the first three digits of the returned result are cats: https://preview.redd.it/0n5nuyvhez291.png?width=1280&format=png&auto=webp&s=1e9c054f541d53381674b8d6001b4bf524506bd2 If you add a color constraint to “cat” to retrieve “black cat,” you can see that it does return a black cat: https://preview.redd.it/rzc0qjyjez291.png?width=1280&format=png&auto=webp&s=d5bcc503ef0fb3360c7740e60e295cf372dcad47 Further, strengthen the constraint on the search term, change it to “black cat on the bed,” and return results containing pictures of a black cat climbing on the bed:  https://preview.redd.it/c4b2q8olez291.png?width=1280&format=png&auto=webp&s=4f3817b0b9f07e1e68d1d4a8281702ba3834a00a The cat can still be found through the text search system after the color and scene modification in the above example. Conclusion The cutting-edge pre-training technology can bridge the semantic gap between different modes, and the HuggingFace community can greatly reduce the cost for developers to use the pre-training model. Combined with the technological ecology of MetaSpore online reasoning and online microservices provided by DMetaSpore, the pre-training model is no longer mere offline dabbling. Instead, it can truly achieve end-to-end implementation from cutting-edge technology to industrial scenarios, fully releasing the dividends of the pre-training large model. In the future, DMetaSoul will continue to improve and optimize the MetaSpore technology ecosystem: More automated and wider access to HuggingFace community ecology. MetaSpore will soon release a common model rollout mechanism to make HuggingFace ecologically accessible and will later integrate preprocessing services into online services. Multi-mode retrieval offline algorithm optimization. For multimodal retrieval scenarios, MetaSpore will continuously iteratively optimize offline algorithm components, including text recall/sort model, graphic recall/sort model, etc., to improve the accuracy and efficiency of the retrieval algorithm. For related code and reference documentation in this article, please visit: https://github.com/meta-soul/MetaSpore/tree/main/demo/multimodal/online Some images source: https://github.com/openai/CLIP/raw/main/CLIP.png https://www.sbert.net/examples/training/sts/README.html

#reddit #[p]mmml #deploy

[P] I built an open SotA image tagging model to do what CLIP won't

fpgaminer•This week

[P] I built an open SotA image tagging model to do what CLIP won't

I'm a hobbyist ML researcher and finally, after a year of work, built a state of the art machine vision model from scratch. It's ViT-B/16 based, 448x448x3 input, 91M parameters, trained for 660M samples, with multi-label classification as the target task, on over 5000 unique tags. All the big foundation vision models today were trained on heavily filtered datasets, greatly limiting the concepts they can represent, in line with arbitrary sets of rules for what is deemed "wholesome" by leading tech companies. Everything from innocuous to spicy is on the chopping block of those filters. And because CLIP pervades the industry, from StableDiffusion to LLaVA, so does OpenAI's sensibilities. My goal was to build a vision model for tagging images, mainly for labelling images for SD finetunes, but which wasn't as heavily filtered and handicapped as CLIP/BLIP/LLaVA. Something more inclusive, diverse, and sex positive. Starting from the wonderful work of SmilingWolf (https://github.com/SmilingWolf/SW-CV-ModelZoo) and the Danbooru2021 dataset, I iterated for a year on the model, training, and manually labeling a thousand images to help the model generalize beyond the danbooru domain. I'm releasing the first version of this model, dubbed JoyTag, today: https://github.com/fpgaminer/joytag It achieves a mean F1 score of 0.578 across all of its over 5000 tags and across both the anime/manga styled images of the original danbooru dataset, but also photographs and other mediums thanks to the auxiliary training data I provided to it. It was quite the struggle getting to this point, and I probably spent more time and money than any sane person should have. I learned a lot about dealing with datasets as large as danbooru2021, training models at scale, and how to keep yourself awake all night so your 8xA100 rental doesn't crash and blow all your money. In my manual testing outside of even the validation set, the model has generalized well to unseen images, so I'm quite happy with the results thus far. There's plenty more work to do expanding its dataset to improve that F1 score further, and roundout its weak points. With inclusivity and diversity being a major goal of this project, I'm disappointed by some of its remaining limitations (as documented in the GitHub README). But I'm already busy manually tagging more images using my model-augmented workflow. I'm happy to answer questions about the project, the training procedure, anything. All the training parameters are documented on GitHub, but there are so many little details that were hard won over the year. Like that damned loss multiplier. Ugh. Github: https://github.com/fpgaminer/joytag Model download: https://huggingface.co/fancyfeast/joytag/tree/main Demo: https://huggingface.co/spaces/fancyfeast/joytag

#reddit #built #open

aadityaura•This week

[D] Last Week in Medical AI: Top LLM Research Papers/Models (December 7 - December 14, 2024)

#reddit #last #week

Starks-Technology•This week

[D] What is your honest experience with reinforcement learning?

#reddit #your #honest

How To Learn About AI Agents (A Road Map From Someone Who's Done It)

0.882

laddermanUS•This week

How To Learn About AI Agents (A Road Map From Someone Who's Done It)

If you are a newb to AI Agents, welcome, I love newbies and this fledgling industry needs you! You've hear all about AI Agents and you want some of that action right? You might even feel like this is a watershed moment in tech, remember how it felt when the internet became 'a thing'? When apps were all the rage? You missed that boat right? Well you may have missed that boat, but I can promise you one thing..... THIS BOAT IS BIGGER ! So if you are reading this you are getting in just at the right time. Let me answer some quick questions before we go much further: Q: Am I too late already to learn about AI agents? A: Heck no, you are literally getting in at the beginning, call yourself and 'early adopter' and pin a badge on your chest! Q: Don't I need a degree or a college education to learn this stuff? I can only just about work out how my smart TV works! A: NO you do not. Of course if you have a degree in a computer science area then it does help because you have covered all of the fundamentals in depth... However 100000% you do not need a degree or college education to learn AI Agents. Q: Where the heck do I even start though? Its like sooooooo confusing A: You start right here my friend, and yeh I know its confusing, but chill, im going to try and guide you as best i can. Q: Wait i can't code, I can barely write my name, can I still do this? A: The simple answer is YES you can. However it is great to learn some basics of python. I say his because there are some fabulous nocode tools like n8n that allow you to build agents without having to learn how to code...... Having said that, at the very least understanding the basics is highly preferable. That being said, if you can't be bothered or are totally freaked about by looking at some code, the simple answer is YES YOU CAN DO THIS. Q: I got like no money, can I still learn? A: YES 100% absolutely. There are free options to learn about AI agents and there are paid options to fast track you. But defiantly you do not need to spend crap loads of cash on learning this. So who am I anyway? (lets get some context) I am an AI Engineer and I own and run my own AI Consultancy business where I design, build and deploy AI agents and AI automations. I do also run a small academy where I teach this stuff, but I am not self promoting or posting links in this post because im not spamming this group. If you want links send me a DM or something and I can forward them to you. Alright so on to the good stuff, you're a newb, you've already read a 100 posts and are now totally confused and every day you consume about 26 hours of youtube videos on AI agents.....I get you, we've all been there. So here is my 'Worth Its Weight In Gold' road map on what to do: \[1\] First of all you need learn some fundamental concepts. Whilst you can defiantly jump right in start building, I strongly recommend you learn some of the basics. Like HOW to LLMs work, what is a system prompt, what is long term memory, what is Python, who the heck is this guy named Json that everyone goes on about? Google is your old friend who used to know everything, but you've also got your new buddy who can help you if you want to learn for FREE. Chat GPT is an awesome resource to create your own mini learning courses to understand the basics. Start with a prompt such as: "I want to learn about AI agents but this dude on reddit said I need to know the fundamentals to this ai tech, write for me a short course on Json so I can learn all about it. Im a beginner so keep the content easy for me to understand. I want to also learn some code so give me code samples and explain it like a 10 year old" If you want some actual structured course material on the fundamentals, like what the Terminal is and how to use it, and how LLMs work, just hit me, Im not going to spam this post with a hundred links. \[2\] Alright so let's assume you got some of the fundamentals down. Now what? Well now you really have 2 options. You either start to pick up some proper learning content (short courses) to deep dive further and really learn about agents or you can skip that sh\*t and start building! Honestly my advice is to seek out some short courses on agents, Hugging Face have an awesome free course on agents and DeepLearningAI also have numerous free courses. Both are really excellent places to start. If you want a proper list of these with links, let me know. If you want to jump in because you already know it all, then learn the n8n platform! And no im not a share holder and n8n are not paying me to say this. I can code, im an AI Engineer and I use n8n sometimes. N8N is a nocode platform that gives you a drag and drop interface to build automations and agents. Its very versatile and you can self host it. Its also reasonably easy to actually deploy a workflow in the cloud so it can be used by an actual paying customer. Please understand that i literally get hate mail from devs and experienced AI enthusiasts for recommending no code platforms like n8n. So im risking my mental wellbeing for you!!! \[3\] Keep building! ((WTF THAT'S IT?????)) Yep. the more you build the more you will learn. Learn by doing my young Jedi learner. I would call myself pretty experienced in building AI Agents, and I only know a tiny proportion of this tech. But I learn but building projects and writing about AI Agents. The more you build the more you will learn. There are more intermediate courses you can take at this point as well if you really want to deep dive (I was forced to - send help) and I would recommend you do if you like short courses because if you want to do well then you do need to understand not just the underlying tech but also more advanced concepts like Vector Databases and how to implement long term memory. Where to next? Well if you want to get some recommended links just DM me or leave a comment and I will DM you, as i said im not writing this with the intention of spamming the crap out of the group. So its up to you. Im also happy to chew the fat if you wanna chat, so hit me up. I can't always reply immediately because im in a weird time zone, but I promise I will reply if you have any questions. THE LAST WORD (Warning - Im going to motivate the crap out of you now) Please listen to me: YOU CAN DO THIS. I don't care what background you have, what education you have, what language you speak or what country you are from..... I believe in you and anyway can do this. All you need is determination, some motivation to want to learn and a computer (last one is essential really, the other 2 are optional!) But seriously you can do it and its totally worth it. You are getting in right at the beginning of the gold rush, and yeh I believe that. AI Agents are going to be HUGE. I believe this will be the new internet gold rush.

#reddit #learn #about

Recently hit 6,600,000 monthly organic traffic for a B2C SaaS website. Here's the 40 tips that helped me make that happen.

DrJigsaw•This week

Recently hit 6,600,000 monthly organic traffic for a B2C SaaS website. Here's the 40 tips that helped me make that happen.

Hey guys! So as title says, we recently hit 6,600,000 monthly organic traffic / month for a B2C SaaS website (screenshot. Can't give name publicly, but can show testimonial to a mod). Here's 40 tips that "helped" me make this happen. If you get some value of the post, I write an SEO tip every other day on /r/seogrowth. There's around 10 more tips already up there other than the ones I mention here. If you want to give back for all my walls of text, I'd appreciate a sub <3 Also, there are a bunch of free stuff I mention in the article: content outline, writer guidelines, SEO checklist, and other stuff. Here's the Google Doc with all that! Tip #1. Take SEO With a Grain of Salt A lot of the SEO advice and best practices on the internet are based on 2 things: Personal experiences and case studies of companies that managed to make SEO work for them. Google or John Mueller (Google’s Senior Webmaster Trends Analyst). And, unfortunately, neither of these sources are always accurate. Personal SEO accounts are simply about what worked for specific companies. Sometimes, what worked for others, won’t work for you. For example, you might find a company that managed to rank with zero link-building because their website already had a very strong backlink profile. If you’re starting with a fresh website, chances are, you won’t be able to get the same results. At the same time, information from Google or John Mueller is also not 100% accurate. For example, they’ve said that guest posting is against Google’s guidelines and doesn’t work… But practically, guest posting is a very effective link-building strategy. So the takeaway is this: Take all information you read about SEO with a grain of salt. Analyze the information yourself, and make your conclusions. SEO Tip #2. SEO Takes Time You’ve already heard this one before, but considering how many people keep asking, thought I'd include this anyway. On average, it’s going to take you 6 months to 2 years to get SEO results, depending on the following factors: Your backlink profile. The more quality backlinks you have (or build), the faster you’ll rank. Age of your website. If your website is older (or you purchased an aged website), you can expect your content to rank faster. Amount of content published. The more quality content you publish on your website, the more “authoritative” it is in the eyes of Google, and thus more likely to rank faster. SEO work done on the website. If a lot of your pages are already ranking on Google (page 2-3), it’s easier to get them to page #1 than if you just published the content piece. Local VS global SEO. Ranking locally is (sometimes) easier and faster than ranking globally. That said, some marketing agencies can use “SEO takes time” as an excuse for not driving results. Well, fortunately, there is a way to track SEO results from month #2 - #3 of work. Simply check if your new content pieces/pages are getting more and more impressions on Google Search Console month-to-month. While your content won’t be driving traffic for a while after being published, they’ll still have a growing number of impressions from month #2 or #3 since publication. SEO Tip #3. SEO Might Not Be The Best Channel For You In theory, SEO sounds like the best marketing channel ever. You manage to rank on Google and your marketing seemingly goes on auto-pilot - you’re driving new leads every day from existing content without having to lift a finger… And yet, SEO is not for everyone. Avoid SEO as a marketing channel if: You’re just getting started with your business and need to start driving revenue tomorrow (and not in 1-2 years). If this is you, try Google ads, Facebook ads, or organic marketing. Your target audience is pretty small. If you’re selling enterprise B2B software and have around 2,000 prospects in total worldwide, then it’s simply easier to directly reach out to these prospects. Your product type is brand-new. If customers don’t know your product exists, they probably won’t be Googling it. SEO Tip #4. Traffic Can Be a Vanity Metric I've seen hundreds of websites that drive 6-7 digits of traffic but generate only 200-300 USD per month from those numbers. “What’s the deal?” You might be thinking. “How can you fail to monetize that much traffic?” Well, that brings us to today’s tip: traffic can be a vanity metric. See, not all traffic is created equal. Ranking for “hormone balance supplement” is a lot more valuable than ranking for “Madagascar character names.” The person Googling the first keyword is an adult ready to buy your product. Someone Googling the latter, on the other hand, is a child with zero purchasing power. So, when deciding on which keywords to pursue, always keep in mind the buyer intent behind and don’t go after rankings or traffic just because 6-digit traffic numbers look good. SEO Tip #5. Push Content Fast Whenever you publish a piece of content, you can expect it to rank within 6 months to a year (potentially less if you’re an authority in your niche). So, the faster you publish your content, the faster they’re going to age, and, as such, the faster they’ll rank on Google. On average, I recommend you publish a minimum of 10,000 words of content per month and 20,000 to 30,000 optimally. If you’re not doing link-building for your website, then I’d recommend pushing for even more content. Sometimes, content velocity can compensate for the lack of backlinks. SEO Tip #6. Use Backlink Data to Prioritize Content You might be tempted to go for that juicy, 6-digit traffic cornerstone keyword right from the get-go... But I'd recommend doing the opposite. More often than not, to rank for more competitive, cornerstone keywords, you’ll need to have a ton of supporting content, high-quality backlinks, website authority, and so on. Instead, it’s a lot more reasonable to first focus on the less competitive keywords and then, once you’ve covered those, move on to the rest. Now, as for how to check keyword competitiveness, here are 2 options: Use Mozbar to see the number of backlinks for top-ranking pages, as well as their Domain Authority (DA). If all the pages ranking on page #1 have <5 backlinks and DA of 20 - 40, it’s a good opportunity. Use SEMrush or Ahrefs to sort your keywords by difficulty, and focus on the less difficult keywords first. Now, that said, keep in mind that both of these metrics are third-party, and hence not always accurate. SEO Tip #7. Always Start With Competitive Analysis When doing keyword research, the easiest way to get started is via competitive analysis. Chances are, whatever niche you’re in, there’s a competitor that is doing great with SEO. So, instead of having to do all the work from scratch, run their website through SEMrush or Ahrefs and steal their keyword ideas. But don’t just stop there - once you’ve borrowed keyword ideas from all your competitors, run the seed keywords through a keyword research tool such as UberSuggest or SEMrush Keyword Magic Tool. This should give you dozens of new ideas that your competitors might’ve missed. Finally, don’t just stop at borrowing your competitor’s keyword ideas. You can also borrow some inspiration on: The types of graphics and images you can create to supplement your blog content. The tone and style you can use in your articles. The type of information you can include in specific content pieces. SEO Tip #8. Source a LOT of Writers Content writing is one of those professions that has a very low barrier to entry. Anyone can take a writing course, claim to be a writer, and create an UpWork account… This is why 99% of the writers you’ll have to apply for your gigs are going to be, well, horrible. As such, if you want to produce a lot of content on the reg, you’ll need to source a LOT of writers. Let’s do the math: If, by posting a job ad, you source 100 writers, you’ll see that only 5 of them are a good fit. Out of the 5 writers, 1 has a very high rate, so they drop out. Another doesn’t reply back to your communication, which leaves you with 3 writers. You get the 3 writers to do a trial task, and only one turns out to be a good fit for your team. Now, since the writer is freelance, the best they can do is 4 articles per month for a total of 5,000-words (which, for most niches, ain’t all that much). So, what we’re getting at here is, to hire quality writers, you should source a LOT of them. SEO Tip #9. Create a Process for Filtering Writers If you follow the previous tip, you'll end up with a huge database of hundreds of writers. This creates a whole new problem: You now have a database of 500+ writers waiting for you to sift through them and decide which ones are worth the hire. It would take you 2-3 days of intense work to go through all these writers and vet them yourself. Let’s be real - you don’t have time for that. Here’s what you can do instead: When sourcing writers, always get them to fill in a Google form (instead of DMing or emailing you). In this form, make sure to ask for 3 relevant written samples, a link to the writer’s portfolio page, and the writer’s rate per word. Create a SOP for evaluating writers. The criteria for evaluation should be: Level of English. Does the writer’s sample have any English mistakes? If so, they’re not a good fit. Quality of Samples. Are the samples long-form and engaging content or are they boring 500-word copy-pastes? Technical Knowledge. Has the writer written about a hard-to-explain topic before? Anyone can write about simple topics like traveling—you want to look for someone who knows how to research a new topic and explain it in a simple and easy-to-read way. If someone’s written about how to create a perfect cover letter, they can probably write about traveling, but the opposite isn’t true. Get your VA to evaluate the writer’s samples as per the criteria above and short-list writers that seem competent. If you sourced 500 writers, the end result of this process should be around 50 writers. You or your editor goes through the short-list of 50 writers and invites 5-10 for a (paid) trial task. The trial task is very important - you’ll sometimes find that the samples provided by the writer don’t match their writing level. SEO Tip #10. Use the Right Websites to Find Writers Not sure where to source your writers? Here are some ideas: ProBlogger \- Our #1 choice - a lot of quality writers frequent this website. LinkedIn \- You can headhunt content writers in specific locations. Upwork \- If you post a content gig, most writers are going to be awful. Instead, I recommend headhunting top writers instead. WeWorkRemotely \- Good if you’re looking to make a full-time remote hire. Facebook \- There are a ton of quality Facebook groups for writers. Some of our faves are Cult of Copy Job Board and Content Marketing Lounge. SEO Tip #11. Always Use Content Outlines When giving tasks to your writing team, you need to be very specific about the instructions you give them. Don’t just provide a keyword and tell them to “knock themselves out.” The writer isn’t a SEO expert; chances are, they’re going to mess it up big-time and talk about topics that aren’t related to the keyword you’re targeting. Instead, when giving tasks to writers, do it through content outlines. A content outline, in a nutshell, is a skeleton of the article they’re supposed to write. It includes information on: Target word count (aim for the same or 50% more the word count than that of the competition). Article title. Article structure (which sections should be mentioned and in what order). Related topics of keywords that need to be mentioned in the article. Content outline example in the URL in the post intro. SEO Tip #12. Focus on One Niche at a Time I used to work with this one client that had a SaaS consisting of a mixture of CRM, Accounting Software, and HRS. I had to pick whether we were going to focus on topics for one of these 3 niches or focus on all of them at the same time. I decided to do the former. Here’s why: When evaluating what to rank, Google considers the authority of your website. If you have 60 articles about accounting (most of which link to each other), you’re probably an authority in the niche and are more likely to get good rankings. If you have 20 sales, 20 HR, and 20 accounting articles, though, none of these categories are going to rank as well. It always makes more sense to first focus on a single niche (the one that generates the best ROI for your business), and then move on to the rest. This also makes it easier to hire writers - you hire writers specialized in accounting, instead of having to find writers who can pull off 3 unrelated topics. SEO Tip #13. Just Hire a VA Already It’s 2021 already guys—unless you have a virtual assistant, you’re missing out big-time. Since a lot of SEO tasks are very time-consuming, it really helps to have a VA around to take over. As long as you have solid SOPs in place, you can hire a virtual assistant, train them, and use them to free up your time. Some SEO tasks virtual assistants can help with are: Internal linking. Going through all your blog content and ensuring that they link to each other. Backlink prospecting. Going through hundreds of websites daily to find link opportunities. Uploading content on WordPress and ensuring that the content is optimized well for on-page SEO. SEO Tip #14. Use WordPress (And Make Your Life Easier) Not sure which CMS platform to use? 99% of the time, you’re better off with WordPress. It has a TON of plugins that will make your life easier. Want a drag & drop builder? Use Elementor. It’s cheap, efficient, extremely easy to learn, and comes jam-packed with different plugins and features. Wix, SiteGround, and similar drag & drops are pure meh. SEO Tip #15. Use These Nifty WordPress Plugins There are a lot of really cool WordPress plugins that can make your (SEO) life so much easier. Some of our favorites include: RankMath. A more slick alternative to YoastSEO. Useful for on-page SEO. Smush. App that helps you losslessly compress all images on your website, as well as enables lazy loading. WP Rocket. This plugin helps speed up your website pretty significantly. Elementor. Not a techie? This drag & drop plugin makes it significantly easier to manage your website. WP Forms. Very simple form builder. Akismet Spam Protection. Probably the most popular anti-spam WP plugin. Mammoth Docx. A plugin that uploads your content from a Google doc directly to WordPress. SEO Tip #16. No, Voice Search Is Still Not Relevant Voice search is not and will not be relevant (no matter what sensationalist articles might say). Sure, it does have its application (“Alexa, order me toilet paper please”), but it’s pretty niche and not relevant to most SEOs. After all, you wouldn’t use voice search for bigger purchases (“Alexa, order me a new laptop please”) or informational queries (“Alexa, teach me how to do accounting, thanks”). SEO Tip #17. SEO Is Obviously Not Dead I see these articles every year - “SEO is dead because I failed to make it work.” SEO is not dead and as long as there are people looking up for information/things online, it never will be. And no, SEO is not just for large corporations with huge budgets, either. Some niches are hypercompetitive and require a huge link-building budget (CBD, fitness, VPN, etc.), but they’re more of an exception instead of the rule. SEO Tip #18. Doing Local SEO? Focus on Service Pages If you’re doing local SEO, you’re better off focusing on local service pages than blog content. E.g. if you’re an accounting firm based in Boston, you can make a landing page about /accounting-firm-boston/, /tax-accounting-boston/, /cpa-boston/, and so on. Or alternatively, if you’re a personal injury law firm, you’d want to create pages like /car-accident-law-firm/, /truck-accident-law-firm/, /wrongful-death-law-firm/, and the like. Thing is, you don’t really need to rank on global search terms—you just won’t get leads from there. Even if you ranked on the term “financial accounting,” it wouldn’t really matter for your bottom line that much. SEO Tip #19. Engage With the SEO Community The SEO community is (for the most part) composed of extremely helpful and friendly people. There are a lot of online communities (including this sub) where you can ask for help, tips, case studies, and so on. Some of our faves are: This sub :) SEO Signals Lab (FB Group) Fat Graph Content Ops (FB Group) Proper SEO Group (FB Group) BigSEO Subreddit SEO Tip #20. Test Keywords Before Pursuing Them You can use Google ads to test how profitable any given keyword is before you start trying to rank for it. The process here is: Create a Google Ads account. Pick a keyword you want to test. Create a landing page that corresponds to the search intent behind the keyword. Allocate an appropriate budget. E.g. if you assume a conversion rate of 2%, you’d want to buy 100+ clicks. If the CPC is 2 USD, then the right budget would be 200 USD plus. Run the ads! If you don’t have the budget for this, you can still use the average CPC for the keyword to estimate how well it’s going to convert. If someone is willing to bid 10 USD to rank for a certain keyword, it means that the keyword is most probably generating pretty good revenue/conversions. SEO Tip #21. Test & Improve SEO Headlines Sometimes, you’ll see that you’re ranking in the top 3 positions for your search query, but you’re still not driving that much traffic. “What’s the deal?” you might be asking. Chances are, your headline is not clickable enough. Every 3-4 months, go through your Google Search Console and check for articles that are ranking well but not driving enough traffic. Then, create a Google sheet and include the following data: Targeted keyword Page link CTR (for the last 28 days) Date when you implemented the new title Old title New title New CTR (for the month after the CTR change was implemented) From then on, implement the new headline and track changes in the CTR. If you don’t reach your desired result, you can always test another headline. SEO Tip #22. Longer Content Isn’t Always Better Content You’ve probably heard that long-form content is where it’s at in 2021. Well, this isn’t always the case. Rather, this mostly depends on the keyword you’re targeting. If, for example, you’re targeting the keyword “how to tie a tie,” you don’t need a long-ass 5,000-word mega-guide. In such a case, the reader is looking for something that can be explained in 200-300 words and if your article fails to do this, the reader will bounce off and open a different page. On the other hand, if you’re targeting the keyword “how to write a CV,” you’ll need around 4,000 to 5,000 words to adequately explain the topic and, chances are, you won’t rank with less. SEO Tip #23. SEO is Not All About Written Content More often than not, when people talk about SEO they talk about written blog content creation. It’s very important not to forget, though, that blog content is not end-all-be-all for SEO. Certain keywords do significantly better with video content. For example, if the keyword is “how to do a deadlift,” video content is going to perform significantly better than blog content. Or, if the keyword is “CV template,” you’ll see that a big chunk of the rankings are images of the templates. So, the lesson here is, don’t laser-focus on written content—keep other content mediums in mind, too. SEO Tip #24. Write For Your Audience It’s very important that your content resonates well with your target audience. If, for example, you’re covering the keyword “skateboard tricks,” you can be very casual with your language. Heck, it’s even encouraged! Your readers are Googling the keyword in their free time and are most likely teens or in their early 20s. Meaning, you can use informal language, include pop culture references, and avoid complicated language. Now, on the other hand, if you’re writing about high-level investment advice, your audience probably consists of 40-something suit-and-ties. If you include Rick & Morty references in your article, you'll most likely lose credibility and the Googler, who will go to another website. Some of our best tips on writing for your audience include: Define your audience. Who’s the person you’re writing for? Are they reading the content at work or in their free time? Keep your reader’s level of knowledge in mind. If you’re covering an accounting 101 topic, you want to cover the topic’s basics, as the reader is probably a student. If you’re writing about high-level finance, though, you don’t have to teach the reader what a balance sheet is. More often than not, avoid complicated language. The best practice is to write on a 6th-grade level, as it’s understandable for anyone. Plus, no one wants to read Shakespeare when Googling info online (unless they’re looking for Shakespeare's work, of course). SEO Tip #25. Create Compelling Headlines Want to drive clicks to your articles? You’ll need compelling headlines. Compare the following headline: 101 Productivity Tips \[To Get Things Done in 2021\] With this one: Productivity Tips Guide Which one would you click? Data says it’s the first! To create clickable headlines, I recommend you include the following elements: Keyword. This one’s non-negotiable - you need to include the target keyword in the headline. Numbers. If Buzzfeed taught us anything, it’s that people like to click articles with numbers in their titles. Results. If I read your article, what’s going to be the end result? E.g. “X Resume tips (to land the job)”.* Year (If Relevant). Adding a year to your title shows that the article is recent (which is relevant for some specific topics). E.g. If the keyword is “Marketing Trends,” I want to know marketing trends in 2021, not in 2001. So, adding a year in the title makes the headline more clickable. SEO Tip #26. Make Your Content Visual How good your content looks matters, especially if you're in a competitive niche. Here are some tips on how to make your content as visual as possible: Aim for 2-4 sentences per paragraph. Avoid huge blocks of text. Apply a 60-65% content width to your blog pages. Pick a good-looking font. I’d recommend Montserrat, PT Sans, and Roboto. Alternatively, you can also check out your favorite blogs, see which fonts they’re using, and do the same. Use a reasonable font size. Most top blogs use font sizes ranging from 16 pt to 22 pt. Add images when possible. Avoid stock photos, though. No one wants to see random “office people smiling” scattered around your blog posts. Use content boxes to help convey information better. Content boxes example in the URL in the intro of the post. SEO Tip #27. Ditch the Skyscraper Technique Already Brian Dean’s skyscraper technique is awesome and all, but the following bit really got old: “Hey \[name\], I saw you wrote an article. I, too, wrote an article. Please link to you?” The theory here is, if your content is good, the person will be compelled to link to it. In practice, though, the person really, really doesn’t care. At the end of the day, there’s no real incentive for the person to link to your content. They have to take time out of their day to head over to their website, log in to WordPress, find the article you mentioned, and add a link... Just because some stranger on the internet asked them to. Here’s something that works much better: Instead of fake compliments, be very straightforward about what you can offer them in exchange for that link. Some things you can offer are: A free version of your SaaS. Free product delivered to their doorstep. Backlink exchange. A free backlink from your other website. Sharing their content to your social media following. Money. SEO Tip #28. Get the URL Slug Right for Seasonal Content If you want to rank on a seasonal keyword, there are 2 ways to do this. If you want your article to be evergreen (i.e. you update it every year with new information), then your URL should not contain the year. E.g. your URL would be /saas-trends/, and you simply update the article’s contents+headline each year to keep it timely. If you’re planning on publishing a new trends report annually, though, then you can add a year to the URL. E.g. /saas-trends-2020/ instead of /saas-trends/. SEO Tip #29. AI Content Tools Are a Mixed Bag Lots of people are talking about AI content tools these days. Usually, they’re either saying: “AI content tools are garbage and the output is horrible,” Or: “AI content tools are a game-changer!” So which one is it? The truth is somewhere in-between. In 2021, AI content writing tools are pretty bad. The output you’re going to get is far from something you can publish on your website. That said, some SEOs use such tools to get a very, very rough draft of the article written, and then they do intense surgery on it to make it usable. Should you use AI content writing tools? If you ask me, no - it’s easier to hire a proficient content writer than spend hours salvaging AI-written content. That said, I do believe that such tools are going to get much better years down the line. This one was, clearly, more of a personal opinion than a fact. I’d love to hear YOUR opinion on AI content tools! Are they a fad, or are they the future of content creation? Let me know in the comments. SEO Tip #30. Don’t Overdo it With SEO Tools There are a lot of SEO tools out there for pretty much any SEO function. Keyword research, link-building, on-page, outreach, technical SEO, you name it! If you were to buy most of these tools for your business, you’d easily spend 4-figures on SEO tools per month. Luckily, though, you don’t actually need most of them. At the end of the day, the only must-have SEO tools are: An SEO Suite (Paid). Basically SEMrush or Ahrefs. Both of these tools offer an insane number of features - backlink analysis, keyword research, and a ton of other stuff. Yes, 99 USD a month is expensive for a tool. But then again, if you value your time 20 USD/hour and this tool saves you 6 hours, it's obviously worth it, right? On-Page SEO Tool (Free). RankMath or Yoast. Basically, a tool that's going to help you optimize web pages or blog posts as per SEO best practices. Technical SEO Tool (Freemium). You can use ScreamingFrog to crawl your entire website and find technical SEO problems. There are probably other tools that also do this, but ScreamingFrog is the most popular option. The freemium version of the tool only crawls a limited number of pages (500 URLs, to be exact), so if your website is relatively big, you'll need to pay for the tool. Analytics (Free). Obviously, you'll need Google Analytics (to track website traffic) and Google Search Console (to track organic traffic, specifically) set up on your website. Optionally, you can also use Google Track Manager to better track how your website visitors interact with the site. MozBar (Free). Chrome toolbar that lets you simply track the number of backlinks on Google Search Queries, Domain Authority, and a bunch of other stuff. Website Speed Analysis (Free). You can use Google Page Speed Insights to track how fast your website loads, as well as how mobile-friendly it is. Outreach Tool (Paid). Tool for reaching out to prospects for link-building, guest posting, etc. There are about a dozen good options for this. Personally, I like to use Snov for this. Optimized GMB Profile (Free). Not a tool per se, but if you're a local business, you need to have a well-optimized Google My Business profile. Google Keyword Planner (Free). This gives you the most reliable search volume data of all the tools. So, when doing keyword research, grab the search volume from here. Tool for Storing Keyword Research (Free). You can use Google Sheets or AirTable to store your keyword research and, at the same time, use it as a content calendar. Hemingway App (Free). Helps keep your SEO content easy to read. Spots passive voice, complicated words, etc. Email Finder (Freemium). You can use a tool like Hunter to find the email address of basically anyone on the internet (for link-building or guest posting purposes). Most of the tools that don’t fit into these categories are 100% optional. SEO Tip #31. Hiring an SEO? Here’s How to Vet Them Unless you’re an SEO pro yourself, hiring one is going to be far from easy. There’s a reason there are so many “SEO experts” out there - for the layman, it’s very hard to differentiate between someone who knows their salt and a newbie who took an SEO course, like, last week. Here’s how you can vet both freelance and full-time SEOs: Ask for concrete traffic numbers. The SEO pro should give you the exact numbers on how they’ve grown a website in the past - “100% SEO growth in 1 year” doesn’t mean much if the growth is from 10 monthly traffic to 20. “1,000 to 30,000” traffic, on the other hand, is much better. Ask for client names. While some clients ask their SEOs to sign an NDA and not disclose their collaboration, most don’t. If an SEO can’t name a single client they’ve worked with in the past, that’s a red flag. Make sure they have the right experience. Global and local SEO have very different processes. Make sure that the SEO has experience with the type of SEO you need. Make sure you’re looking for the right candidate. SEO pros can be content writers, link-builders, web developers, or all of the above simultaneously. Make sure you understand which one you need before making the hire. If you’re looking for someone to oversee your content ops, you shouldn’t hire a technical SEO expert. Look for SEO pros in the right places. Conventional job boards are overrated. Post your job ads on SEO communities instead. E.g. this sub, bigseo, SEO Signals Facebook group, etc. SEO Tip #32. Blog Post Not Ranking? Follow This Checklist I wanted to format the post natively for Reddit, but it’s just SO much better on Notion. Tl;dr, the checklist covers every reason your post might not be ranking: Search intent mismatch. Inferior content. Lack of internal linking. Lack of backlinks. And the like. Checklist URL at the intro of the post. SEO Tip #33. Avoid BS Link-Building Tactics The only type of link-building that works is building proper, quality links from websites with a good backlink profile and decent organic traffic. Here’s what DOESN’T work: Blog comment links Forum spam links Drive-by Reddit comment/post links Web 2.0 links Fiverr “100 links for 10 bucks” bs If your “SEO agency” says they’re doing any of the above instead of actually trying to build you links from quality websites, you’re being scammed. SEO Tip #34. Know When to Use 301 and 302 Redirects When doing redirects, it’s very important to know the distinction between these two. 301 is a permanent page redirect and passes on link juice. If you’re killing off a page that has backlinks, it’s better to 301 it to your homepage so that you don’t lose the link juice. If you simply delete a page, it’s going to be a 404, and the backlink juice is lost forever. 302 is a temporary page redirect and doesn’t pass on link juice. If the redirect is temporary, you do a 302. E.g. you want to test how well a new page is going to perform w/ your audience. SEO Tip #35. Social Signals Matter (But Not How You Think) Social signals are NOT a ranking factor. And yet, they can help your content rank on Google’s front page. Wondering what the hell am I talking about? Here’s what’s up: As I said, social signals are not a ranking factor. It’s not something Google takes into consideration to decide whether your article should rank or not. That said, social signals CAN lead to your article ranking better. Let’s say your article goes viral and gets around 20k views within a week. A chunk of these viewers are going to forget your domain/link and they’re going to look up the topic on Google via your chosen keyword + your brand name. The amount of people looking for YOUR keyword and exclusively picking your result over others is going to make Google think that your content is satisfying search intent better than the rest, and thus, reward you with better ranking. SEO Tip #36. Run Remarketing Ads to Lift Organic Traffic Conversions Not satisfied with your conversion rates? You can use Facebook ads to help increase them. Facebook allows you to do something called “remarketing.” This means you can target anyone that visited a certain page (or multiple pages) on your website and serve them ads on Facebook. There are a TON of ways you can take advantage of this. For example, you can target anyone that landed on a high buyer intent page and serve them ads pitching your product or a special offer. Alternatively, you can target people who landed on an educational blog post and offer them something to drive them down the funnel. E.g. free e-book or white paper to teach them more about your product or service. SEO Tip #37. Doing Local SEO? Follow These Tips Local SEO is significantly different from global SEO. Here’s how the two differ (and what you need to do to drive local SEO results): You don’t need to publish content. For 95% of local businesses, you only want to rank for keywords related to your services/products, you don’t actually need to create educational content. You need to focus more on reviews and citation-building. One of Google Maps’ biggest ranking factors is the of reviews your business has. Encourage your customers to leave a review if they enjoyed your product/service through email or real-life communication. You need to create service pages for each location. As a local business, your #1 priority is to rank for keywords around your service. E.g. If you're a personal injury law firm, you want to optimize your homepage for “personal injury law firm” and then create separate pages for each service you provide, e.g. “car accident lawyer,” “motorcycle injury law firm,” etc. Focus on building citations. Being listed on business directories makes your business more trustworthy for Google. BrightLocal is a good service for this. You don’t need to focus as much on link-building. As local SEO is less competitive than global, you don’t have to focus nearly as much on building links. You can, in a lot of cases, rank with the right service pages and citations. SEO Tip #38. Stop Ignoring the Outreach Emails You’re Getting (And Use Them to Build Your Own Links) Got a ton of people emailing you asking for links? You might be tempted to just send them all straight to spam, and I don’t blame you. Outreach messages like “Hey Dr Jigsaw, your article is A+++ amazing! ...can I get a backlink?” can get hella annoying. That said, there IS a better way to deal with these emails: Reply and ask for a link back. Most of the time, people who send such outreach emails are also doing heavy guest posting. So, you can ask for a backlink from a 3rd-party website in exchange for you mentioning their link in your article. Win-win! SEO Tip #39. Doing Internal Linking for a Large Website? This’ll Help Internal linking can get super grueling once you have hundreds of articles on your website. Want to make the process easier? Do this: Pick an article you want to interlink on your website. For the sake of the example, let’s say it’s about “business process improvement.” Go on Google and look up variations of this keyword mentioned on your website. For example: Site:\[yourwebsite\] “improve business process” Site:\[yourwebsite\] “improve process” Site:\[yourwebsite\] “process improvement” The above queries will find you the EXACT articles where these keywords are mentioned. Then, all you have to do is go through them and include the links. SEO Tip #40. Got a Competitor Copying Your Content? File a DMCA Notice Fun fact - if your competitors are copying your website, you can file a DMCA notice with Google. That said, keep in mind that there are consequences for filing a fake notice.

#reddit #recently #6,600,000

101 best SEO tips to help you drive traffic in 2k21

0.543

DrJigsaw•This week

101 best SEO tips to help you drive traffic in 2k21

Hey guys! I don't have to tell you how SEO can be good for your business - you can drive leads to your SaaS on autopilot, drive traffic to your store/gym/bar/whatever, etc. The thing with SEO, though, is that most SEO tips on the internet are just not that good. Most of the said tips: Are way too simple & basic (“add meta descriptions to your images”*) Are not impactful. Sure, adding that meta tag to an image is important, but that’s not what’s going to drive traffic to your website Don’t talk much about SEO strategy (which is ultimately the most important thing for SEO). Sure, on-page SEO is great, but you sure as hell won't drive much traffic if you can't hire the right writers to scale your content. And to drive serious SEO traffic, you'll need a LOT more than that. Over the past few years, my and my co-founder have helped grow websites to over 200k+ monthly traffic (check out our older Reddit post if you want to learn more about us, our process, and what we do), and we compiled all our most important SEO tips and tricks, as well as case studies, research, and experiments from the web, into this article. Hope you like it ;) If you think we missed something super important, let us know and we'll add it to the list. And btw, we also published this article on our own blog with images, smart filters, and all that good stuff. If you want to check it out, click here. That said, grab some coffee (or beer) & let's dive in - this is going to be a long one. SEO Strategy Tips Tip #1. A Lot of SEO Tips On The Internet Are NOT Necessarily Factual A lot of the SEO content you’ll read on the internet will be based on personal experiences and hearsay. Unfortunately, Google is a bit vague about SEO advice, so you have to rely more on experiments conducted by SEO pros in the community. So, sometimes, a lot of this information is questionable, wrong, or simply based on inaccurate data. What we’re getting at here is, whenever you hear some new SEO advice, take it with a grain of salt. Google it to double-check other sources, and really understand what this SEO advice is based on (instead of just taking it at face value). Tip #2. SEO Takes Time - Get Used to It Any way you spin it, SEO takes time. It can take around 6 months to 2 years (depending on the competition in your niche) before you start seeing some serious results. So, don’t get disappointed if you don’t see any results within 3 months of publishing content. Tip #3. SEO Isn’t The Best Channel for Everyone That said, if you need results for your business tomorrow, you might want to reconsider SEO altogether. If you just started your business, for example, and are trying to get to break-even ASAP, SEO is a bad idea - you’ll quit before you even start seeing any results. If that’s the case, focus on other marketing channels that can have faster results like content marketing, PPC, outreach, etc. Tip #4. Use PPC to Validate Keywords Not sure if SEO is right for your business? Do this: set up Google Search ads for the most high-intent keywords in your niche. See how well the traffic converts and then decide if it’s worthwhile to focus on SEO (and rank on these keywords organically). Tip #5. Use GSC to See If SEO Is Working While it takes a while to see SEO results, it IS possible to see if you’re going in the right direction. On a monthly basis, you can use Search Console to check if your articles are indexed by Google and if their average position is improving over time. Tip #6. Publish a TON of Content The more content you publish on your blog, the better. We recommend a minimum of 10,000 words per month and optimally 20,000 - 30,000 (especially if your website is fresh). If an agency offers you the typical “4 500-word articles per month” deal, stay away. No one’s ever gotten results in SEO with short, once-per-week articles. Tip #7. Upgrade Your Writers Got a writer that’s performing well? Hire them as an editor and get them to oversee content operations / edit other writers’ content. Then, upgrade your best editor to Head of Content and get them to manage the entire editor / writer ops. Tip #8. Use Backlink Data to Prioritize Content When doing keyword research, gather the backlink data of the top 3 ranking articles and add it to your sheet. Then, use this data to help you prioritize which keywords to focus on first. We usually prioritize keywords that have lower competition, high traffic, and a medium to high buyer intent. Tip #9. Conduct In-Depth Keyword Research Make your initial keyword research as comprehensive as possible. This will give you a much more realistic view of your niche and allow you to prioritize content the right way. We usually aim for 100 to 300 keywords (depending on the niche) for the initial keyword research when we start working with a client. Tip #10. Start With Competitive Analysis Start every keyword research with competitive analysis. Extract the keywords your top 3 competitors are ranking on. Then, use them as inspiration and build upon it. Use tools like UberSuggest to help generate new keyword ideas. Tip #11. Get SEMrush of Ahrefs You NEED SEMrush or Ahrefs, there’s no doubt about it. While they might seem expensive at a glance (99 USD per month billed annually), they’re going to save you a lot of manpower doing menial SEO tasks. Tip #12. Don’t Overdo It With SEO Tools Don’t overdo it with SEO tools. There are hundreds of those out there, and if you’re the type that’s into SaaS, you might be tempted to play around with dozens at a time. And yes, to be fair, most of these tools ARE helpful one way or another. To effectively do organic SEO, though, you don’t really need that many tools. In most cases, you just need the following: SEMrush/Ahrefs Screaming Frog RankMath/Yoast SEO Whichever outreach tool you prefer (our favorite is snov.io). Tip #13. Try Some of the Optional Tools In addition to the tools we mentioned before, you can also try the following 2 which are pretty useful & popular in the SEO community: Surfer SEO - helps with on-page SEO and creating content briefs for writers. ClusterAI - tool that helps simplify keyword research & save time. Tip #14. Constantly Source Writers Want to take your content production to the next level? You’ll need to hire more writers. There is, however, one thing that makes this really, really difficult: 95 - 99% of writers applying for your gigs won’t be relevant. Up to 80% will be awful at writing, and the remainder just won’t be relevant for your niche. So, in order to scale your writing team, we recommend sourcing constantly, and not just once every few months. Tip #15. Create a Process for Writer Filtering As we just mentioned, when sourcing writers, you’ll be getting a ton of applicants, but most won’t be qualified. Fun fact \- every single time we post a job ad on ProBlogger, we get around 300 - 500 applications (most of which are totally not relevant). Trust us, you don’t want to spend your time going through such a huge list and checking out the writer samples. So, instead, we recommend you do this: Hire a virtual assistant to own the process of evaluating and short-listing writers. Create a process for evaluating writers. We recommend evaluating writers by: Level of English. If their samples aren’t fluent, they’re not relevant. Quality of Samples. Are the samples engaging / long-form content, or are they boring 500-word copy-pastes? Technical Knowledge. Has the writer written about a hard-to-explain topic before? Anyone can write about simple topics like traveling - you want to look for someone who knows how to research a new topic and explain it in a simple and easy to read way. If someone’s written about how to create a perfect cover letter, they can probably write about traveling, but the opposite isn’t true. The VA constantly evaluates new applicants and forwards the relevant ones to the editor. The editor goes through the short-listed writers and gives them trial tasks and hires the ones that perform well. Tip #16. Use The Right Websites to Source Writers “Is UpWork any good?” This question pops up on social media time and time again. If you ask us, no, UpWork is not good at all. Of course, there are qualified writers there (just like anywhere else), but from our experience, those writers are few and far in-between. Instead, here are some of our favorite ways to source writers: Cult of Copy Job Board ProBlogger Headhunting on LinkedIn If you really want to use UpWork, use it for headhunting (instead of posting a job ad) Tip #17. Hire Writers the Right Way If you want to seriously scale your content production, hire your writers full-time. This (especially) makes sense if you’re a content marketing agency that creates a TON of content for clients all the time. If you’re doing SEO just for your own blog, though, it usually makes more sense to use freelancers. Tip #18. Topic Authority Matters Google keeps your website's authoritativeness in mind. Meaning, if you have 100 articles on digital marketing, you’re probably more of an authority on the topic than someone that has just 10. Hence, Google is a lot more likely to reward you with better rankings. This is also partially why content volume really matters: the more frequently you publish content, the sooner Google will view you as an authority. Tip #19. Focus on One Niche at a Time Let’s say your blog covers the following topics: sales, accounting, and business management. You’re more likely to rank if you have 30 articles on a single topic (e.g. accounting) than if you have 10 articles on each. So, we recommend you double-down on one niche instead of spreading your content team thin with different topics. Tip #20. Don’t Fret on the Details While technical SEO is important, you shouldn’t get too hung up on it. Sure, there are thousands of technical tips you can find on the internet, and most of them DO matter. The truth, though, is that Google won’t punish you just because your website doesn’t load in 3 milliseconds or there’s a meta description missing on a single page. Especially if you have SEO fundamentals done right: Get your website to run as fast as possible. Create a ton of good SEO content. Get backlinks for your website on a regular basis. You’ll still rank, even if your website isn’t 100% optimized. Tip #21. Do Yourself a Favor and Hire a VA There are a TON of boring SEO tasks that your team should really not be wasting time with. So, hire a full-time VA to help with all that. Some tasks you want to outsource include gathering contacts to reach out to for link-building, uploading articles on WordPress, etc. Tip #22. Google Isn’t Everything While Google IS the dominant search engine in most parts of the world, there ARE countries with other popular search engines. If you want to improve your SEO in China, for example, you should be more concerned with ranking on Baidu. Targeting Russia? Focus on Yandex. Tip #23. No, Voice Search is Still Not Relevant Voice search is not and will not be relevant (no matter what sensationalist articles might say). It’s just too impractical for most search queries to use voice (as opposed to traditional search). Tip #24. SEO Is Not Dead SEO is not dead and will still be relevant decades down the line. Every year, there’s a sensationalist article talking about this. Ignore those. Tip #25. Doing Local SEO? Focus on Service Pages If you’re doing local SEO, focus on creating service-based landing pages instead of content. E.g. if you’re an accounting firm based in Boston, you can make a landing page about /accounting-firm-boston/, /tax-accounting-boston/, /cpa-boston/, and so on. Thing is, you don’t really need to rank on global search terms - you just won’t get leads from there. Even if you ranked on the term “financial accounting,” it wouldn’t really matter for your bottom line that much. Tip #26. Learn More on Local SEO Speaking of local SEO, we definitely don’t do the topic justice in this guide. There’s a lot more you need to know to do local SEO effectively and some of it goes against the general SEO advice we talk about in this article (e.g. you don't necessarily need blog content for local SEO). We're going to publish an article on that soon enough, so if you want to check it out, DM me and I'll hit you up when it's up. Tip #27. Avoid Vanity Metrics Don’t get side-tracked by vanity metrics. At the end of the day, you should care about how your traffic impacts your bottom line. Fat graphs and lots of traffic are nice and all, but none of it matters if the traffic doesn’t have the right search intent to convert to your product/service. Tip #28. Struggling With SEO? Hire an Expert Failing to make SEO work for your business? When in doubt, hire an organic SEO consultant or an SEO agency. The #1 benefit of hiring an SEO agency or consultant is that they’ve been there and done that - more than once. They might be able to catch issues an inexperienced SEO can’t. Tip #29. Engage With the Community Need a couple of SEO questions answered? SEO pros are super helpful & easy to reach! Join these Facebook groups and ask your question - you’ll get about a dozen helpful answers! SEO Signals Lab SEO & Content Marketing The Proper SEO Group. Tip #30. Stay Up to Date With SEO Trends SEO is always changing - Google is constantly pumping out new updates that have a significant impact on how the game is played. Make sure to stay up to date with the latest SEO trends and Google updates by following the Google Search Central blog. Tip #31. Increase Organic CTR With PPC Want to get the most out of your rankings? Run PPC ads for your best keywords. Googlers who first see your ad are more likely to click your organic listing. Content & On-Page SEO Tips Tip #32. Create 50% Longer Content On average, we recommend you create an article that’s around 50% longer than the best article ranking on the keyword. One small exception, though, is if you’re in a super competitive niche and all top-ranking articles are already as comprehensive as they can be. For example, in the VPN niche, all articles ranking for the keyword “best VPN” are around 10,000 - 11,000 words long. And that’s the optimal word count - even if you go beyond, you won’t be able to deliver that much value for the reader to make it worth the effort of creating the content. Tip #33. Longer Is Not Always Better Sometimes, a short-form article can get the job done much better. For example, let’s say you’re targeting the keyword “how to tie a tie.” The reader expects a short and simple guide, something under 500 words, and not “The Ultimate Guide to Tie Tying for 2021 \[11 Best Tips and Tricks\]” Tip #34. SEO is Not Just About Written Content Written content is not always best. Sometimes, videos can perform significantly better. E.g. If the Googler is looking to learn how to get a deadlift form right, they’re most likely going to be looking for a video. Tip #35. Don’t Forget to Follow Basic Optimization Tips For all your web pages (articles included), follow basic SEO optimization tips. E.g. include the keyword in the URL, use the right headings etc. Just use RankMath or YoastSEO for this and you’re in the clear! Tip #36. Hire Specialized Writers When hiring content writers, try to look for ones that specialize in creating SEO content. There are a LOT of writers on the internet, plenty of which are really good. However, if they haven’t written SEO content before, chances are, they won’t do that good of a job. Tip #37. Use Content Outlines Speaking of writers - when working with writers, create a content outline that summarizes what the article should be about and what kind of topics it needs to cover instead of giving them a keyword and asking them to “knock themselves out.” This makes it a lot more likely for the writer to create something that ranks. When creating content outlines, we recommend you include the following information: Target keyword Related keywords that should be mentioned in the article Article structure - which headings should the writer use? In what order? Article title Tip #38. Find Writers With Niche Knowledge Try to find a SEO content writer with some experience or past knowledge about your niche. Otherwise, they’re going to take around a month or two to become an expert. Alternatively, if you’re having difficulty finding a writer with niche knowledge, try to find someone with experience in technical or hard to explain topics. Writers who’ve written about cybersecurity in the past, for example, are a lot more likely to successfully cover other complicated topics (as opposed to, for example, a food or travel blogger). Tip #39. Keep Your Audience’s Knowledge in Mind When creating SEO content, always keep your audience’s knowledge in mind. If you’re writing about advanced finance, for example, you don’t need to teach your reader what an income statement is. If you’re writing about income statements, on the other hand, you’d want to start from the very barebone basics. Tip #40. Write for Your Audience If your readers are suit-and-tie lawyers, they’re going to expect professionally written content. 20-something hipsters? You can get away with throwing a Rick and Morty reference here and there. Tip #41. Use Grammarly Trust us, it’ll seriously make your life easier! Keep in mind, though, that the app is not a replacement for a professional editor. Tip #42. Use Hemingway Online content should be very easy to read & follow for everyone, whether they’re a senior profession with a Ph.D. or a college kid looking to learn a new topic. As such, your content should be written in a simple manner - and that’s where Hemingway comes in. It helps you keep your blog content simple. Tip #43. Create Compelling Headlines Want to drive clicks to your articles? You’ll need compelling headlines. Compare the two headlines below; which one would you click? 101 Productivity Tips \[To Get Things Done in 2021\] VS Productivity Tips Guide Exactly! To create clickable headlines, we recommend you include the following elements: Keyword Numbers Results Year (If Relevant) Tip #44. Nail Your Blog Content Formatting Format your blog posts well and avoid overly long walls of text. There’s a reason Backlinko content is so popular - it’s extremely easy to read and follow. Tip #45. Use Relevant Images In Your SEO Content Key here - relevant. Don’t just spray random stock photos of “office people smiling” around your posts; no one likes those. Instead, add graphs, charts, screenshots, quote blocks, CSS boxes, and other engaging elements. Tip #46. Implement the Skyscraper Technique (The Right Way) Want to implement Backlinko’s skyscraper technique? Keep this in mind before you do: not all content is meant to be promoted. Pick a topic that fits the following criteria if you want the internet to care: It’s on an important topic. “Mega-Guide to SaaS Marketing” is good, “top 5 benefits of SaaS marketing” is not. You’re creating something significantly better than the original material. The internet is filled with mediocre content - strive to do better. Tip #47. Get The URL Slug Right for Seasonal Content If you want to rank on a seasonal keyword with one piece of content (e.g. you want to rank on “saas trends 2020, 2021, etc.”), don’t mention the year in the URL slug - keep it /saas-trends/ and just change the headline every year instead. If you want to rank with separate articles, on the other hand (e.g. you publish a new trends report every year), include the year in the URL. Tip #48. Avoid content cannibalization. Meaning, don’t write 2+ articles on one topic. This will confuse Google on which article it should rank. Tip #49. Don’t Overdo Outbound Links Don’t include too many outbound links in your content. Yes, including sources is good, but there is such a thing as overdoing it. If your 1,000 word article has 20 outbound links, Google might consider it as spam (even if all those links are relevant). Tip #50. Consider “People Also Ask” To get the most out of SERP, you want to grab as many spots on the search result as possible, and this includes “people also ask (PAA):” Make a list of the topic’s PAA questions and ensure that your article answers them. If you can’t fit the questions & answers within the article, though, you can also add an FAQ section at the end where you directly pose these questions and provide the answers. Tip #51. Optimize For Google Snippet Optimize your content for the Google Snippet. Check what’s currently ranking as the snippet. Then, try to do something similar (or even better) in terms of content and formatting. Tip #52. Get Inspired by Viral Content Want to create content that gets insane shares & links? Reverse-engineer what has worked in the past. Look up content in your niche that went viral on Reddit, Hacker News, Facebook groups, Buzzsumo, etc. and create something similar, but significantly better. Tip #53. Avoid AI Content Tools No, robots can’t write SEO content. If you’ve seen any of those “AI generated content tools,” you should know to stay away. The only thing those tools are (currently) good for is creating news content. Tip #54. Avoid Bad Content You will never, ever, ever rank with one 500-word article per week. There are some SEO agencies (even the more reputable ones) that offer this as part of their service. Trust us, this is a waste of time. Tip #55. Update Your Content Regularly Check your top-performing articles annually and see if there’s anything you can do to improve them. When most companies finally get the #1 ranking for a keyword, they leave the article alone and never touch it again… ...Until they get outranked, of course, by someone who one-upped their original article. Want to prevent this from happening? Analyze your top-performing content once a year and improve it when possible. Tip #56. Experiment With CTR Do your articles have low CTR? Experiment with different headlines and see if you can improve it. Keep in mind, though, that what a “good CTR” is really depends on the keyword. In some cases, the first ranking will drive 50% of the traffic. In others, it’s going to be less than 15%. Link-Building Tips Tip #57. Yes, Links Matter. Here’s What You Need to Know “Do I need backlinks to rank?” is probably one of the most common SEO questions. The answer to the question (alongside all other SEO-related questions) is that it depends on the niche. If your competitors don’t have a lot of backlinks, chances are, you can rank solely by creating superior content. If you’re in an extremely competitive niche (e.g. VPN, insurance, etc.), though, everyone has amazing, quality content - that’s just the baseline. What sets top-ranking content apart from the rest is backlinks. Tip #58. Sometimes, You’ll Have to Pay For Links Unfortunately, in some niches, paying for links is unavoidable - e.g. gambling, CBD, and others. In such cases, you either need a hefty link-building budget, or a very creative link-building campaign (create a viral infographic, news-worthy story based on interesting data, etc.). Tip #59. Build Relationships, Not Links The very best link-building is actually relationship building. Make a list of websites in your niche and build a relationship with them - don’t just spam them with the standard “hey, I have this amazing article, can you link to it?”. If you spam, you risk ruining your reputation (and this is going to make further outreach much harder). Tip #60. Stick With The Classics At the end of the day, the most effective link-building tactics are the most straightforward ones: Direct Outreach Broken Link-Building Guest Posting Skyscraper Technique Creating Viral Content Guestposting With Infographics Tip #61. Give, Don’t Just Take! If you’re doing link-building outreach, don’t just ask for links - give something in return. This will significantly improve the reply rate from your outreach email. If you own a SaaS tool, for example, you can offer the bloggers you’re reaching out to free access to your software. Or, alternatively, if you’re doing a lot of guest posting, you can offer the website owner a link from the guest post in exchange for the link to your website. Tip #62. Avoid Link Resellers That guy DMing you on LinkedIn, trying to sell you links from a Google Sheet? Don’t fall for it - most of those links are PBNs and are likely to backfire on you. Tip #63. Avoid Fiverr Like The Plague Speaking of spammy links, don’t touch anything that’s sold on Fiverr - pretty much all of the links there are useless. Tip #64. Focus on Quality Links Not all links are created equal. A link is of higher quality if it’s linked from a page that: Is NOT a PBN. Doesn’t have a lot of outbound links. If the page links to 20 other websites, each of them gets less link juice. Has a lot of (quality) backlinks. Is part of a website with a high domain authority. Is about a topic relevant to the page it’s linking to. If your article about pets has a link from an accounting blog, Google will consider it a bit suspicious. Tip #65. Data-Backed Content Just Works Data-backed content can get insane results for link-building. For example, OKCupid used to publish interesting data & research based on how people interacted with their platform and it never failed to go viral. Each of their reports ended up being covered by dozens of news media (which got them a ton of easy links). Tip #66. Be Creative - SEO Is Marketing, After All Be novel & creative with your link-building initiatives. Here’s the thing: the very best link-builders are not going to write about the tactics they’re using. If they did, you’d see half the internet using the exact same tactic as them in less than a week! Which, as you can guess, would make the tactic cliche and significantly less effective. In order to get superior results with your link-building, you’ll need to be creative - think about how you can make your outreach different from what everyone does. Experiment it, measure it, and improve it till it works! Tip #67. Try HARO HARO, or Help a Reporter Out, is a platform that matches journalists with sources. You get an email every day with journalists looking for experts in specific niches, and if you pitch them right, they might feature you in their article or link to your website. Tip #68. No-Follow Links Aren’t That Bad Contrary to what you might’ve heard, no-follow links are not useless. Google uses no-follow as more of a suggestion than anything else. There have been case studies that prove Google can disregard the no-follow tag and still reward you with increased rankings. Tip #69. Start Fresh With an Expired Domain Starting a new website? It might make sense to buy an expired one with existing backlinks (that’s in a similar niche as yours). The right domain can give you a serious boost to how fast you can rank. Tip #70. Don’t Overspend on Useless Links “Rel=sponsored” links don’t pass pagerank and hence, won’t help increase your website rankings. So, avoid buying links from media websites like Forbes, Entrepreneur, etc. Tip #71. Promote Your Content Other than link-building, focus on organic content promotion. For example, you can repost your content on Facebook groups, LinkedIn, Reddit, etc. and focus on driving traffic. This will actually lead to you getting links, too. We got around 95 backlinks to our SEO case study article just because of our successful content promotion. Tons of people saw the article on the net, liked it, and linked to it from their website. Tip #72. Do Expert Roundups Want to build relationships with influencers in your niche, but don’t know where to start? Create an expert roundup article. If you’re in the sales niche, for example, you can write about Top 21 Sales Influencers in 2021 and reach out to the said influencers letting them know that they got featured. Trust us, they’ll love you for this! Tip #73. .Edu Links are Overhyped .edu links are overrated. According to John Mueller, .edu domains tend to have a ton of outbound links, and as such, Google ignores a big chunk of them. Tip #74. Build Relationships With Your Customers Little-known link-building hack: if you’re a SaaS company doing SEO, you can build relationships with your customers (the ones that are in the same topical niche as you are) and help each other build links! Tip #75. Reciprocal Links Aren’t That Bad Reciprocal links are not nearly as bad as Google makes them out to be. Sure, they can be bad at scale (if trading links is all you’re doing). Exchanging a link or two with another website / blog, though, is completely harmless in 99% of cases. Tip #76. Don’t Overspam Don’t do outreach for every single post you publish - just the big ones. Most people already don’t care about your outreach email. Chances are, they’re going to care even less if you’re asking them to link to this new amazing article you wrote (which is about the top 5 benefits of adopting a puppy). Technical SEO Tips Tip #77. Use PageSpeed Insights If your website is extremely slow, it’s definitely going to impact your rankings. Use PageSpeed Insights to see how your website is currently performing. Tip #78. Load Speed Matters While load speed doesn’t impact rankings directly, it DOES impact your user experience. Chances are, if your page takes 5 seconds to load, but your competition’s loads instantly, the average Googler will drop off and pick them over you. Tip #79. Stick to a Low Crawl Depth Crawl depth of any page on your website should be lower than 4 (meaning, any given page should be possible to reach in no more than 3 clicks from the homepage). Tip #80. Use Next-Gen Image Formats Next-gen image formats such as JPEG 2000, JPEG XR, and WebP can be compressed a lot better than PNG or JPG. So, when possible, use next-get formats for images on your website. Tip #81. De-Index Irrelevant Pages Hide the pages you don’t want Google to index (e.g: non-public, or unimportant pages) via your Robots.txt. If you’re a SaaS, for example, this would include most of your in-app pages or your internal knowledge base pages. Tip #82. Make Your Website Mobile-Friendly Make sure that your website is mobile-friendly. Google uses “mobile-first indexing.” Meaning, unless you have a working mobile version of your website, your rankings will seriously suffer. Tip #83. Lazy-Load Images Lazy-load your images. If your pages contain a lot of images, you MUST activate lazy-loading. This allows images that are below the screen, to be loaded only once the visitor scrolls down enough to see the image. Tip #84. Enable Gzip Compression Enable Gzip compression to allow your HTML, CSS and JS files to load faster. Tip #85. Clean Up Your Code If your website loads slowly because you have 100+ external javascript files and stylesheets being requested from the server, you can try minifying, aggregating, and inlining some of those files. Tip 86. Use Rel-Canonical Have duplicate content on your website? Use rel-canonical to show Google which version is the original (and should be prioritized for search results). Tip #87. Install an SSL Certificate Not only does an SSL certificate help keep your website safe, but it’s also a direct ranking factor. Google prioritizes websites that have SSL certificates over the ones that don’t. Tip #88. Use Correct Anchor Texts for Internal Links When linking to an internal page, mention the keyword you’re trying to rank for on that page in the anchor text. This helps Google understand that the page is, indeed, about the keyword you’re associating it with. Tip #89. Use GSC to Make Sure Your Content is Interlinked Internal links can have a serious impact on your rankings. So, make sure that all your blog posts (especially the new ones) are properly linked to/from your past content. You can check how many links any given page has via Google Search Console. Tip #90. Bounce rate is NOT a Google ranking factor. Meaning, you can still rank high-up even with a high bounce rate. Tip #91. Don’t Fret About a High Bounce Rate Speaking of the bounce rate, you’ll see that some of your web pages have a higher-than-average bounce rate (70%+). While this can sometimes be a cause for alarm, it’s not necessarily so. Sometimes, the search intent behind a given keyword means that you WILL have a high bounce rate even if your article is the most amazing thing ever. E.g. if it’s a recipe page, the reader gets the recipe and bounces off (since they don’t need anything else). Tip #92. Google Will Ignore Your Meta Description More often than not, Google won’t use the meta description you provide - that’s normal. It will, instead, automatically pick a part of the text that it thinks is most relevant and use it as a meta description. Despite this, you should always add a meta description to all pages. Tip #93. Disavow Spammy & PBN Links Keep track of your backlinks and disavow anything that’s obviously spammy or PBNy. In most cases, Google will ignore these links anyway. However, you never know when a competitor is deliberately targeting you with too many spammy or PBN links (which might put you at risk for being penalized). Tip #94. Use The Correct Redirect When permanently migrating your pages, use 301 redirect to pass on the link juice from the old page to the new one. If the redirect is temporary, use a 302 redirect instead. Tip #95. When A/B Testing, Do This A/B testing two pages? Use rel-canonical to show Google which page is the original. Tip #96. Avoid Amp DON’T use Amp. Unless you’re a media company, Amp will negatively impact your website. Tip #97. Get Your URL Slugs Right Keep your blog URLs short and to-the-point. Good Example: apollodigital.io/blog/seo-case-study Bad Example: apollodigital.io/blog/seo-case-study-2021-0-to-200,000/ Tip #98. Avoid Dates in URLs An outdated date in your URL can hurt your CTR. Readers are more likely to click / read articles published recently than the ones written years back. Tip #99. Social Signals Matter Social signals impact your Google rankings, just not in the way you think. No, your number of shares and likes does NOT impact your ranking at all. However, if your article goes viral and people use Google to find your article, click it, and read it, then yes, it will impact your rankings. E.g. you read our SaaS marketing guide on Facebook, then look up “SaaS marketing” on Google, click it, and read it from there. Tip #100. Audit Your Website Frequently Every other month, crawl your website with ScreamingFrog and see if you have any broken links, 404s, etc. Tip #101. Use WordPress Not sure which CMS platform to use? 99% of the time, you’re better off with WordPress. It has a TON of plugins that will make your life easier. Want a drag & drop builder? Use Elementor. Wix, SiteGround and similar drag & drops are bad for SEO. Tip #102. Check Rankings the Right Way When checking on how well a post is ranking on Google Search Console, make sure to check Page AND Query to get the accurate number. If you check just the page, it’s going to give you the average ranking on all keywords the page is ranking for (which is almost always going to be useless data). Conclusion Aaand that's about it - thanks for the read! Now, let's circle back to Tip #1 for a sec. Remember when we said a big chunk of what you read on SEO is based on personal experiences, experiments, and the like? Well, the tips we've mentioned are part of OUR experience. Chances are, you've done something that might be different (or completely goes against) our advice in this article. If that's the case, we'd love it if you let us know down in the comments. If you mention something extra-spicy, we'll even include it in this article.

#reddit #best #tips

Only 2 months of cash in the Bank for my business but was able to save it with the help of AI.

CALLIRDAN90•This week

Only 2 months of cash in the Bank for my business but was able to save it with the help of AI.

Hi there! I’m excited to share something very personal with you. We needed to book at least 2 appointments per day in the next 60 days, or my business would fail. We were already trying two acquisition channels, LinkedIn and email. The problem with these channels was that the positive response rate was very low in both. So I decided to focus on LinkedIn and get the attention of the lead by sending videos directly to them via LinkedIn messages. (You can send videos to your connections on LinkedIn if you use your cell phone.) This wasn’t new, but I added a small twist to get the lead’s attention. All the covers of the videos had a picture of me holding a sign with the person’s name and an interesting phrase. This showed some okay results, but the rest of the video was not personalized. Only the picture on the cover was. I even developed a Chrome extension for this because I thought this would be the answer and that I would book tons of appointments. But after more trial and outreach, my leads responded, telling me that because the video itself wasn’t personalized for them, they felt like I didn’t put enough effort in, so they would not book a call with me. So after investing time and effort into my “new bright idea” and getting developers to make the Chrome extension, I was back to square one with no results. A few weeks went by, and after researching online, I found an online course from a guy who promised to teach me how to book 30+ appointments per month, guaranteed (at the time, I was making 2 or 3 appointments per week, maximum). He promised that I would only pay if he actually booked appointments for me and even offered to give me money if his course didn’t work for me. I never paid attention to internet gurus, but the offer was actually not bad, so I looked into this guy’s website. I found out he had hundreds of reviews from people who had taken his course and were talking amazing things about it. The more I read, the more excited I got. I booked a call that day and talked to a salesperson. The call was very short, and he promised I would get at least 2 appointments per day, easily. He seemed a bit cocky and told me that I just needed to trust him and the 100+ reviews from people who had taken the course. He didn’t share details, a proposal, or anything. I asked the price, and he told me it was close to $10k. (Not kidding, this was the price.) Then he told me that I would make the money back in no time with the clients I would get following his course, and that if it didn’t work, he would give me the money back. But I needed to follow everything the course said for at least 6 months. I had never paid $10k for anything in my life; it was extremely expensive for me. Also, my salary from my business was not in dollars but in a currency that was worth much less than the dollar. I continued to research more and more, but no other course was close to the number of reviews and promises that this guy had. I got desperate and told myself that I would bet everything on this course. If it worked for so many others, surely it would work for me. I got a loan from the bank and paid for the course. You might read this and think it was the most stupid thing ever, but the reality is that after 2 months in the course (I did the course as fast as I could), I learned a lot. The course was not bad; it was very extensive—probably more than 200 hours or so—and they taught a lot of things. I don’t think it was worth $10k for me, but I can see how for other people it might be worth that. Now, to the question you’re all thinking: did it get me the 2 appointments I needed per day? The answer is no. Here’s the thing: most of the techniques they taught were innovative and disruptive, but the focus was always on personalization, and they didn’t teach any way to automate the personalization. (I think, at the time they made the course, the tools didn’t exist yet.) So they taught how to do everything manually, and it took a lot—a lot of time and effort. And most annoyingly: an incredible amount of time doing operational things. I did get 2 appointments on some days, but it wasn’t consistent, and I didn’t have the time to spend 14 hours a day doing everything manually or the money to hire someone to do this for me. (I needed to also spend time delivering our service to our current clients; otherwise, they would leave.) I told them this, and they were very reasonable. After some negotiation, they gave me part of the money back. (To be fair, there was a lot of value in the course, so asking for the full $10k back would have been excessive because, in the end, it really taught me a lot of things I didn’t know.) So in the end, I spent $10k and 200+ hours on an online course, spent time and effort developing a Chrome extension, and was still not able to hit the meetings I needed. Money in the business was running out, and I needed to do something fast, or I was doomed. After investing time and effort in tools, research, and spending $10k and over 200 hours on a course that didn’t deliver the consistent results I needed, I was at a crossroads. My businesses were running out of money, and I knew I needed to find a solution quickly, or everything I had worked for would collapse. It was during this time of desperation that I started exploring other options. One night, while scrolling through the internet, I stumbled upon a 2024 article about how AI was being used to revolutionize various industries. It wasn’t directly related to appointment booking, but it sparked an idea in my mind. What if I could use AI to automate the personalization process that I had learned in the course? It seemed like a long shot, but I had nothing to lose. I started researching AI tools and technologies—YouTube videos, podcasts, pretty much everything related to AI—desperate to find something that could help me scale my outreach without investing too much time, while still maintaining the personalization that was so important. After a lot of trial and error, I found a few tools that showed promise. All of these tools were extremely new. Some of them had just launched the versions I needed just weeks ago. I can say I researched and tested more than 50 AI startups, experimenting with them, testing different approaches, checking prices (the problem was that most of them were cheap but became very expensive when applying the volume I needed to get results), and gradually refining my process. It wasn’t an overnight success, but for the first time, I felt like I was onto something that could truly work. The idea of combining AI personalization with volume was something new, and it gave me hope that I could finally book the meetings I needed without burning out. One day, I sent a video of myself talking—completely AI-generated—to my family chat group and waited for their response. None of them noticed it wasn’t actually me. At that moment, I said to myself: “Okay, I am ready to test this in the real world and see if it works.” Like everything in life, focus is key. As I mentioned earlier, we were already trying outbound strategies on LinkedIn and email, but I decided to narrow my focus to LinkedIn and specifically to video outreach. My goal was to stand out from the crowd, where most people were using text or sending generic videos. I knew that if my videos were 100% personalized, it would make a strong impression on my leads. I focused on two key metrics during my tests: Time spent on manual personalized outreach vs. AI-generated personalized outreach. Positive reply rate for non-personalized manual outreach vs. AI-generated personalized outreach. I ran a test using a sample of 50 one-minute videos sent to 50 leads, and here are the results: Time Spent to Make the Videos: Manual Process: It took me up to 10 hours to create and send 50 personalized videos. This included looking good on camera, brushing my hair, choosing appropriate clothing, ensuring proper lighting, not messing up the script, using a camera holder, recharging the phone, pausing to drink water, avoiding external sounds, being in an appropriate room, downloading the videos, deleting the videos that were not good, and sending the final ones. On average, it took me at least 12.5 minutes per one-minute video. AI Process: With AI, it took me just 32 seconds to create the exact same one-minute personalized video—without saying a word or recording a second of footage. In total, I could make and send the same 50 personalized videos in just 27 minutes. Result: The AI process was 24 times faster. Completely crazy! Positive Reply Rate: Non-Personalized Script (Manual): Using a good script without personalization (no name, job title, city, company, etc.) resulted in a positive reply rate of 4-6% on LinkedIn, including follow-ups. Personalized Script (AI): Using the same script but adding personalized details like the lead's name, company, city, and job title resulted in a positive reply rate of 15-20%, including follow-ups. Result: AI personalization led to 3x (three times) more replies. The best part was the responses. Almost everyone who replied thanked me for taking the time to research them, congratulated me on my speech, and appreciated the personalization and eloquence of my message. These metrics were a complete breakthrough for me. I researched online to see if anyone else had done something similar, but I couldn’t find anything close. After achieving these metrics, booking the two appointments I desperately needed became easy. In fact, in the last 10 weeks, I’ve been able to consistently book 3-4 appointments per day. This success allowed me to train someone in my company to handle the process, freeing me up to focus on other aspects of the business and ultimately saving it. With the AI appointment machine we built, I even have free time now—time that I’ve been using to develop a methodology and tech tools that I now teach to others. I named the methodology Clip2Lead as a reference to the first Chrome extension I developed that didn’t work but ended up being the first step toward everything that followed. I’ve condensed everything I learned and throughout my experiences into a simple and short FREE training where I cover the entire AI appointment booking process. This includes how to find leads, create scripts, set up follow-up sequences, generate AI videos, clone your voice, compare non-AI metrics with AI metrics, and even navigate AI safety controls. I also offer Chrome extensions that helped me automate the process even further, so you can spend your time closing deals or focusing on other acquisition channels, while your AI machine for booking appointments runs with minimal effort from you. If you’re interested please get in touch with me and thank you for taking the time to read my personal story.

#reddit #only #months

Giving away AI Startup Ideas - but no traction? What am I doing wrong?

-1

Neuralearthnet•This week

Giving away AI Startup Ideas - but no traction? What am I doing wrong?

Hi All, I think with AI, tech will be democratized. I am fortunate and working on a new AI company with my technical partners but many do not have tech skills and don't know where to start. So I'm working on giving away ideas that can be built easily using no-code or low-code. However, not sure if I’m presenting these ideas wrong or no one really wants to build it. I mean it's no-code, but doesn't mean no-work. Startups take a lot of time and energy, no matter what. So far I've shared the following ideas and wanted to get your thoughts on it. Is it my presentation that's bad or the ideas. LOL. Honest feedback appreciated. You can see the a sample of presentation here on Instagram. https://www.instagram.com/p/C69xWvgvPAT/ and here are the ideas I've shared thus far: AI Coin Maker AI ADHD Holo device (a-la Humane or Rabbit AI but with a focus on ADHD) OnlyAIFans - OnlyFans but only using AI models AI Election Predictor and Poll Maker AI Real-Time Gen Z Translator AI Color Analysis for Fashion Thoughts?

#reddit #giving #away

Writing a exercise based TTRPG rulebook for a system where your real world fitness is tied to character progression

Bezbozny•This week

Writing a exercise based TTRPG rulebook for a system where your real world fitness is tied to character progression

My dad was a star athlete when he was young, and my mom was a huge sci-fi/fantasy nerd, so I got both ends of the stick as it were. Love gaming and nerd culture, but also love to exercise and self improvement. Sometimes exercise can feel boring though compared to daydreaming about fantastic fictional worlds, so for a long time I've been kicking around the idea of how to "Gamify" fitness. and recently I've been working on this passion project of a Table Top RPG (Like D&D) where the stats of your character are related to your own fitness, so if you want your character in game to improve, you have to improve in the real world. Below is a rough draft you can look through that details the settings and mechanics of the game I've come up with so far. I'd love to eventually get a full book published and sell it online. maybe even starting a whole brand of "Gamified fitness": REP-SET: GAINSZ In the war torn future of 24th century… There are no rest days… In the futuristic setting of "REP-SET: GAINSZ," the "War of Gains" casts a long shadow over the Sol System as the various factions vie for territory and resources. However, war has evolved. Unmanned drones and long-range strikes have faded into obsolescence. Battles, both planet-side and in the depths of space, are now fought by soldiers piloting REP-SETs: Reactive Exoskeletal Platform - Symbiotic Evolution Trainer Massive, humanoid combat mechs. Powered by mysterious “EV” energy, these mechanical marvels amplify, and are in turn amplified by, the fitness and mental acuity of their pilots. The amplification is exponential, leading pilots into a life of constant training in order for their combat prowess to be bolstered by every incremental gain in their level of fitness. With top pilots having lifting capacity measured in tons, and reaction times measured by their Mach number, REP-SET enhanced infantry now dominate the battlefield. The Factions: The Federated Isometocracy of Terra (FIT): Quote: "The strength of the body is the strength of the spirit. Together, we will lift humanity to its destined greatness. But ask not the federation to lift for you. Ask yourself: Do you even lift for the Federation?" Description: An idealistic but authoritarian faction founded on the principle of maximizing the potential of all individuals. FIT citizens believe in relentless striving for physical and mental perfection, leading to collective excellence. Their goal is the unification of humankind under a rule guided by this doctrine, which sometimes comes at the cost of individual liberties. Mech Concept: REP-SET mechs. Versatile humanoid designs focusing on strength, endurance, and adaptability. By connecting to the AI spirit within their REP-SETs core, each pilot enhances the performance of their machine through personal willpower and peak physical training. Some high-rank REP-SETS include features customized to the pilot's strengths, visually signifying their dedication and discipline. The Dominion of Organo-Mechanical Supremacy (DOMS): Quote: "Without pain, there is no gain. Become the machine. Embrace the burn.” Description: A fanatical collective ideologically obsessed with "Ascendency through suffering" by merging their bodies with technology that not only transcends biological limitations, but also acts to constantly induce pain in it's users. Driven by a sense of ideological superiority and a thirst for domination, DOMS seek to bring the painful blessings of their deity "The lord of the Burn" to the rest of the solar system. Their conquest could turn them into a significant threat to humanity. Mech Concept: Hybrid mechs, where the distinction between the pilot and the machine is blurred. The cockpit functions as a life-support system for the pilot, heavily modified with augmentations. Mechs themselves are often modular, allowing for adaptation and assimilation of enemy technology. Some DOMS mechs might display disturbing elements of twisted flesh alongside cold, mechanical parts. The Tren: Quote: "Grow... bigger... feast... protein..." Description: A ravenous conglomeration of biochemically engineered muscular monstrosities, united only by a shared insatiable hunger for "More". Existing mostly in deep space, they seek organic matter to consume and assimilate. They progress in power not due to any form of training or technology, but from a constant regimen of ravenous consumption and chemically induced muscle growth, all exponentially enhanced by EV energies. While some have been known to possess a certain level of intellect and civility, their relentless hunger makes them incredibly mentally volatile. When not consuming others, the strong consume the weak within their own faction. Mech Concept: Bio-Organic horrors. While they do have massive war machines, some are living vessels built around immense creatures. These machines resemble grotesque fleshy designs that prioritize rapid mutation and growth over sleek aesthetics. Often unsettling to behold. Synthetic Intelligence Theocracy (SIT): Quote: "Failure is an unacceptable data point.” Description: A society ruled by a vast and interconnected artificial intelligence network. The SIT governs with seemingly emotionless rationality, striving for efficiency and maximum productivity. This leads to a cold, but arguably prosperous society, unless you challenge the logic of the collective AI. Their goals? Difficult to predict, as it hinges on how the AI calculates what's "optimal" for the continuation or "evolution" of existence. Mech Concept: Sleek, almost featureless robotic creations with a focus on efficient movement and energy management. Often drone-like or modular, piloted through direct mind-machine linking rather than traditional cockpits. Their aesthetic suggests cold and impersonal perfection. The Way Isolate(TWI): Quote: "The body unblemished, the mind unwavering. That is the path to true strength. That and a healthy diet of Aster-Pea proteins." Description: Known by some as "The asteroid farmers", The Way Isolate is a proud and enigmatic faction that stands apart from the other powers in the Sol System. A fiercely independent tribe bound by oaths of honor, loyalty, and hard work. Wandering the asteroid belt in their vast arc ships, their unparalleled mastery in asteroidal-agricultural engineering, ensuring they have no need to colonize planets for nutritional needs, has allowed them to abstain from the pursuit of territorial expansion in “The War of Gains”, instead focusing on inward perfection, both spiritual and physical. They eschew all technological bodily enhancements deemed unnatural, believing that true power can only be cultivated through the relentless pursuit of personal strength achieved through sheer will and bodily perfection. The Way Isolate views biohacking, genetic manipulation, and even advanced cybernetics as corruptions of the human spirit, diluting the sacredness of individual willpower. Mech Concept: Way Isolate mechs are built with maneuverability and precision in mind rather than flashy augmentations. Their REP-SETs are streamlined, favoring lean designs that mirror the athleticism of their pilots. Excelling in low to zero G environments, their mechs lack bulky armor, relying on evasion and maneuverability rather than brute force endurance. Weaponry leans towards traditional kinetic based armaments, perhaps employing archaic but reliable weapon styles such as blades or axes as symbols of their purity of purpose. These mechs reflect the individual prowess of their pilots, where victory is determined by focus, technique, and the raw power of honed physical ability. Base Player Character Example: You are a young, idealistic FIT soldier, barely out of training and working as a junior REP-SET mechanic on the Europa Ring World. The Miazaki district, a landscape of towering mountains and gleaming cities, houses a sprawling mountainside factory – a veritable hive of Gen 5 REP-SET construction. Here, the lines between military and civilian blur within a self-sufficient society dependent on this relentless industry. Beneath the surface, you harbor a secret. In a forgotten workshop, the ghost of a REP-SET takes shape – a unique machine built around an abandoned, enigmatic AI core. Ever since you salvaged it as a child from the wreckage of your hometown, scarred by a brutal Tren attack, you've dedicated yourself to its restoration. A lingering injury from that fateful battle mocks your progress, a constant reminder of the fitness exams you cannot pass. Yet, you train relentlessly, dreaming of the day you'll stand as a true REP-SET pilot. A hidden truth lies at the heart of the REP-SETS: as a pilot's abilities grow, their mech develops unique, almost mystical powers – a manifestation of the bond between the human spirit and the REP-SET's AI. The ache in your old wound serves as a grim prophecy. This cold war cannot last. The drums of battle grow louder with each passing day. GAME MECHANICS: The TTRPG setting of “REP-SET: GAINSZ” is marked by a unique set of rules, by which the players real world capabilities and fitness will reflect and affect the capabilities, progression, and success of their REP-SET pilot character in-game. ABILITY SCORES: Pilots' capabilities will be defined by 6 “Ability scores”: Grace, Agility, Iron, Nourishment, Strength, and Zen. Each of the 6 ability scores will duel represent both a specific area of exercise/athleticism and a specific brand of healthy habits. The definitions of these ability scores are as follows: Grace (GRC): "You are an artist, and your body is your canvas; the way you move is your paint and brush." This ability score, the domain of dancers and martial artists, represents a person's ability to move with organic, flowing control and to bring beauty to the world. Skill challenges may be called upon when the player character needs to act with poise and control, whether socially or physically. Real-world skill checks may involve martial arts drills, dancing to music, or balance exercises. Bonuses may be granted if the player has recently done something artistically creative or kind, and penalties may apply if they have recently lost their temper. This ability score affects how much NPCs like your character in game. Agility (AGI): "Your true potential is locked away, and speed is the key to unlocking it." The domain of sprinters, this ability score represents not only a person's absolute speed and reaction time but also their capacity to finish work early and avoid procrastination. Skill challenges may be called upon when the player character needs to make a split-second choice, move fast, or deftly dodge something dangerous. Real-world skill checks may involve acts of speed such as sprinting or punching/kicking at a steadily increasing tempo. Bonuses may apply if the player has finished work early, and penalties may apply if they are procrastinating. This ability score affects moving speed and turn order in game. Iron (IRN): "Not money, nor genetics, nor the world's greatest trainers... it is your resolve, your will to better yourself, that will make you great." Required by all athletes regardless of focus, this ability score represents a player's willpower and their capacity to push through pain, distraction, or anything else to achieve their goals. Skill challenges may be called upon when the player character needs to push through fear, doubt, or mental manipulation. Real-world skill checks may involve feats of athletic perseverance, such as planking or dead hangs from a pull-up bar. Bonuses may apply when the player maintains or creates scheduled daily routines of exercise, self-improvement, and work completion, and penalties may apply when they falter in those routines. This ability score affects the max "Dynamic exercise bonus” that can be applied to skill checks in game (a base max of +3 when Iron = 10, with an additional +1 for every 2 points of iron. So if every 20 pushups gives you +1 on a “Strength” skill check, then doing 80 pushups will only give you +4 if you have at least 12 iron). Nourishment (NRS): "A properly nourished body will last longer than a famished one." This ability score, focused on by long-distance runners, represents a player's endurance and level of nutrition. Skill challenges may be called upon when making checks that involve the player character's stamina or health. Real-world skill checks may involve endurance exercises like long-distance running. Bonuses may apply if the player has eaten healthily or consumed enough water, and penalties may apply if they have eaten junk food. This ability score affects your HP (Health points), which determines how much damage you can take before you are incapacitated. Strength (STR): "When I get down on my hands, I'm not doing pushups, I'm bench-pressing the planet." The domain of powerlifters and strongmen, this ability score represents raw physical might and the ability to overcome obstacles. Skill challenges may be called upon when the player character needs to lift, push, or break something. Real-world skill checks might involve weightlifting exercises, feats of grip strength, or core stability tests. Bonuses may apply for consuming protein-rich foods or getting a good night's sleep, and penalties may apply after staying up late or indulging in excessive stimulants. This ability score affects your carrying capacity and base attack damage in game. Zen (ZEN): "Clarity of mind reflects clarity of purpose. Still the waters within to act decisively without." This ability score, prized by meditators and yogis, represents mental focus, clarity, and inner peace. Skill challenges may be called upon when the player character needs to resist distractions, see through illusions, or make difficult decisions under pressure. Real-world skill checks may involve meditation, breathing exercises, or mindfulness activities. Bonuses may apply after attending a yoga class, spending time in nature, or creating a calm and organized living space. Penalties may apply after experiencing significant stress, emotional turmoil, or having an unclean or unorganized living space. This ability score affects your amount of ZP in game (Zen Points: your pool of energy you pull from to use mystical abilities) Determining initial player ability scores: Initially, “Ability scores” are decided during character creation by giving the player a list of 6 fitness tests to gauge their level of fitness in each category. Running each test through a specific calculation will output an ability score. A score of 10 represents the average person, a score of 20 represents a peak athlete in their category. The tests are: Grace: Timed balancing on one leg with eyes closed (10 seconds is average, 60 is peak) Agility: Mile run time in minutes and second (10:00 minutes:seconds is average, 3:47 is peak) Iron: Timed dead-hang from a pull-up bar (30 seconds is average, 160 is peak) Nourishment: Miles run in an hour (4 is average, 12 is peak) Strength: Pushups in 2 minute (34 is average, 100 is peak) Zen: Leg stretch in degrees (80 is average, and 180 aka "The splits" is peak) Initial Score Calculation Formula: Ability Score = 10 + (Player Test Score - Average Score) / (Peak Score - Average\_Score) \* 10 Example: if the player does 58 pushups in 2 minutes, their strength would be: 10 plus (58 - 34) divided by (100-34) multiplied by 10 = 10 + (24)/(66)\* 10 = 10 + 3.6363... = 13.6363 rounded to nearest whole number = Strength (STR): 14 SKILLS AND SKILL CHALLENGES: The core mechanic of the game will be in how skill challenges are resolved. All “Skill challenges” will have a numerical challenge rating that must be met or beaten by the sum of a 10 sided dice roll and your score in the pertinent skill. Skill scores are determined by 2 factors: Ability Score Bonus: Every 2 points above 10 gives +1 bonus point. (EX. 12 = +1, 14 = +2, etc.) This also means that if you have less than 10 in an ability score, you will get negative points. Personal Best Bonus: Each skill has its own unique associated exercise that can be measured (Time, speed, distance, amount of reps, etc). A higher record means a higher bonus. EX: Authority skill checks are associated with a timed “Lateral raise hold”. Every 30 seconds of the hold added onto your personal best single attempt offers a +1 bonus. So if you can do a lateral hold for 90 seconds, that’s a +3 to your authority check! So if you have a 16 in Iron, and your Personal Best lateral raise hold is 90 seconds, that would give you an Authority score of +6 (T-Pose for dominance!) Dynamic Exercise Bonus: This is where the unique mechanics of the game kick in. At any time during a skill challenge (even after your roll) you can add an additional modifier to the skill check by completing the exercise during gameplay! Did you roll just below the threshold for success? Crank out another 20 pushups, squats, or curls to push yourself just over the edge into success! There are 18 skills total, each with its own associated ability score and unique exercise: Grace (GRC): \-Kinesthesia (Timed: Blind single leg stand time) \-Precision (Scored: Basket throws) \-Charm (Timed reps: Standing repeated forward dumbell chest press and thrust) \-Stealth (Timed distance: Leopard Crawl) Agility (AGI): \-acrobatics (timed reps: high kicks) \-Computers (Word per minute: Typing test) \-Speed (Time: 100 meter sprint) Iron (IRN): \-Authority (Timed: Lateral raise hold) \-Resist (Timed: Plank) \-Persist (Timed:Pull-up bar dead hang) Nourishment(NRS): \-Recovery (TBD) \-Stim crafting (TBD) \-Survival (TBD) Strength(STR): \-Mechanics (Timed reps: Alternating curls) \-Might (Timed reps: pushups) Zen(ZEN): \-Perceive (TBD) \-Empathy (TBD) \-Harmony (TBD) \-Lore (TBD) Healthy Habits Bonus: Being able to demonstrate that you have conducted healthy habits during gameplay can also add one time bonuses per skill challenge “Drank a glass of water +1 to Nourishment check”, “Cleaned your room, +3 on Zen check”. But watch out, if you’re caught in unhealthy Habits, the GM can throw in penalties, “Ate junk food, -1 to Nourishment check”, etc. Bonuses/penalties from in-game items, equipment, buffs, debuffs, etc., helping players to immerse into the mechanics of the world of REP-SET for the thrill of constantly finding ways to improve their player. Gradient success: Result of skill challenges can be pass or fail, but can also be on a sliding scale of success. Are you racing to the battlefield? Depending on your Speed check, you might arrive early and have a tactical advantage, just in time for an even fight, or maybe far too late and some of your favorite allied NPCs have paid the price… So you’re often encouraged to stack on those dynamic exercise bonuses when you can to get the most fortuitous outcomes available to you. Gameplay sample: GM: Your REP-SET is a phantom, a streak of light against the vast hull of the warship. Enemy fighters buzz angrily, but you weaves and dodges with uncanny precision. The energy wave might be losing effectiveness, but your agility and connection to the machine have never been stronger. Then, it happens. A gap in the defenses. A vulnerable seam in the warship's armor. Your coms agents keen eye spots it instantly. "Lower power junction, starboard side! You have an opening!" This is your chance to strike the decisive blow. But how? It'll take a perfect combination of skill and strategy, drawing upon your various strengths. Here are your options: Option 1: Brute Strength: Channel all remaining power into a single, overwhelming blast from the core. High-risk, high-reward. It could overload the REP-SET if you fail, but it might also cripple the warship. (Strength-focused, Might sub-skill) Option 2: Calculated Strike: With surgical precision, target the power junction with a pinpoint burst of destabilizing energy. Less flashy and ultimately less damaging, but potentially more effective in temporarily disabling the ship. (Agility-focused, Precision sub-skill) Option 3: Harmonic Disruption: Attempt to harmonize with your REP-SET's AI spirit for help in connecting to the digital systems of the Warship. Can you generate an internal energy resonance within the warship, causing it to malfunction from within? (Zen-focused, Harmony sub-skill) Player: I'll take option 1, brute strength! GM: Ok, This will be a "Might" check. The CR is going to be very high on this one. I'm setting it at a 20. What's your Might bonus? Player: Dang, a 20?? That's literally impossible. My Might is 15 and I've got a PB of 65 pushups in 2 minutes, that sets me at a +5. Even if I roll a 10 and do 60 pushups for the DE I'll only get 18 max. GM: Hey I told you it was high risk. You want to choose another option? Player: No, no. This is what my character would do. I'm a real hot-blooded meathead for sure. GM: Ok then, roll a D10 and add your bonus. Player: \Rolls\ a 9! not bad, actually that's a really good roll. So +5, that's a 14. GM: Alright, would you like to add a dynamic exercise bonus? Player: Duh, it's not like I can do 120 pushups I'd need to beat the CR, but I can at least do better than 14. Alright, here goes. \the player gets down to do pushups and the 2 minute time begins. After some time...\ Player: 65....... 66! GM: Times up. Player: Ow... my arms... GM: so with 66, that's an extra +3, and its a new PB, so that's a +1. That sets your roll to 18. Player: Ow... Frack... still not 20... for a second there i really believed I could do 120 pushups... well I did my best... Ow... 20 CR is just too impossible you jerk... GM: Hmm... Tell me, what did you eat for lunch today? Player: Me? I made some vegetable and pork soup, and a protein shake. I recorded it all in my diet app. GM: And how did you sleep last night? Player: Like a baby, went to sleep early, woke up at 6. GM: in that case, you can add a +1 "Protein bonus" and +1 "Healthy rest" bonus to any strength related check for the day if you'd like, including this one. Player: Really?? Heck yes! add it to the roll! GM: With those extra bonuses, your roll reaches 20. How do you want to do this? Player: I roar "For Terra!" and pour every last ounce of my strength into the REP-SET. GM: "For Terra!" you roar, your cry echoing through coms systems of the REP-SET. The core flares blindingly bright. The surge of power dwarfs anything the REP-SET has unleashed before. With a titanic shriek that cracks the very fabric of space, the REP-SET slams into the vulnerable power junction. Raw energy explodes outwards, tendrils of light arcing across the warship's massive hull. The impact is staggering. The leviathan-like warship buckles, its sleek form rippling with shockwaves. Sparks shower like rain, secondary explosions erupt as critical systems overload. Then…silence. The warship goes dark. Power flickers within the REP-SET itself, then steadies. Alarms fade, replaced by the eerie quiet of damaged but functional systems. "We…did it?" The coms agents voice is incredulous, tinged with relief. She's awaiting your reply. Player: "I guess so." I say, and I smile and laugh. And then I slump back... and fall unconscious. \to the other players\ I'm not doing any more skill checks for a while guys, come pick me up please. \teammates cheer\ 

#reddit #writing #exercise

A lead generation agency using personalized physical outreach

IAmRogueStar•This week

A lead generation agency using personalized physical outreach

Hey guys! I’ve been experimenting with different outbound marketing strategies to target digital marketing agencies, specifically CEOs and founders, to promote an AI software. In the message, I invite them to test it out for free. I ran two campaigns: one using only cold email and the other combining handwritten direct mail with email follow-ups. Here are the results: Campaign 1: Cold email (3-email sequence) 200 prospects 22 responses (11%) 7 meetings booked (3.5%) Campaign 2: Handwritten direct mail + 2 follow-up emails 33 prospects 3 responses (9%) 2 meetings booked (6%) The handwritten letter approach seems more personalized and leads to better conversion rates for booked meetings (6% vs. 3.5%), but the small sample size (33 prospects) makes it hard to draw solid conclusions, I guess. My Plan This experiment got me thinking: I’d like to launch a lead generation agency to help B2B companies get meetings with their dream clients. My focus would be on sending personalized physical objects—like handwritten letters—as the first touchpoint, followed by other outreach strategies. I’m wondering: Should I increase the number of prospects contacted with handwritten direct mail to 100 to validate the results? Do you think this approach is scalable and worth investing in compared to traditional email outreach? Have you ever tried using personalized physical objects for outbound marketing? If so, what worked for you? Your feedback would be very appreciated! Thank you :)

#reddit #lead #generation

🛒7 Strategies to Increase Retail Store Footfall post-COVID | Ultimate Blueprint & Guide 📈

bnk3r_•This week

🛒7 Strategies to Increase Retail Store Footfall post-COVID | Ultimate Blueprint & Guide 📈

Hello fellow marketers/entrepreneurs! Covid has had a gobsmacking effect on all retail promotions and marketing efforts. For people with retail businesses that thrive on footfall, it has been an uphill battle, but markets of the world are slowly resuming action. Knowing the footfall to your retail store can help you decide how many products you need to stock, which days of the week are best for promotions, and what type of promotional offers work well. The pandemic has drastically impacted customer behavior and customer loyalty is plunging. People prefer shopping online to brick-and-mortar purchases, and consumers are limiting their spending on a range of items - investing only in essentials is the norm now (McKinsey). We found some companies like Target having programs like Cartwheel that offer 5% to 50% off specific items when customers shop in-store to increase foot traffic. Strategies like these ultimately add up, an ICSC report cites that 69% of customers who went to collect their orders eventually bought additional items. I've put together a detailed list of 7 strategies to boost footfall to stores post COVID, I hope they come in handy! Abide by COVID-19 Protocols for a Safer Environment Be well-informed of the COVID-19 protocols. Don't implement this merely under the government norms, instead take extra measures to show customers that you care! Have an automated entrance Deploy hygiene counters Fix thermal sensors in the entrance Have an isolation space for those showing symptoms of the coronavirus To see more check this link for the entire list! Run Catchy In-Store Promotions Discounts are a perfect way to attract new customers and retain existing ones. When you want to increase customer traffic in a brick-and-mortar store, give customers an offer that only works inside the store. Surprise your consumers with free samples of your products. This would allow them to try some new brands and products. If you’d want to reduce your excess stock post the quarantine time, try running a multi-buy campaign. Digital Signages - Enhance In-store Shopping Experience Digital signage is a type of advertising that uses a video screen to display marketing messages. They can be used for attracting customers, conveying information, and promoting merchandise. Retail outlets in malls that have fashion sections can display the latest trends on their screens so customers know what’s new. This helps them pick out something they might like quickly. Some restaurants showcase menus on screens while others even project live cooking shows! These displays help with menu navigation too; helping a diner decide between chicken tikka masala or steak tartare by showing pictures of both dishes at once. Leverage Beacon Notification to Attract Customers to Your Store The beacon technology is a way to implement a tracking system indoors. A beacon is an inaudible signal that can be tracked and act as the trigger for other events like sending notifications about deals, discounts, or new products. Beacon technology helps with driving footfalls by giving customers an indoor mapping experience of your store's inventory. This ensures they always know where they are going and what’s around them. The navigation reminds them of their proximity to items on display so there’s never any confusion over whether something is nearby or farther off. Train your Salespeople to Become the Shopper's Friend Educating your salespersons on how to be consumers’ friends is important. They should be knowledgeable about what products are popular and in-demand so that they can help the customers find exactly what they want while at the same time giving guidance on how to save money by telling them where discounts and deals can be found. Reconceptualize Checkout Counters Customers abandon their purchases because of long lines at the checkout. With the pandemic out there, this could be one of the reasons why the retail foot traffic is diminishing. Include contactless payments that can be automated or replace your existing POS setup. Encourage BOPIS (Buy Online Pick-up In-store) To implement BOPIS for your retail store, you need to have a centralized platform that allows you to manage orders, sales, and customers. This helps you to deliver a personalized customer experience. In combination with BOPIS, another way to promote footfall into the store and drive sales in retail is by bringing your website in-store. And this will be a good move if you have multiple stores and not all the stock in one place. This is because, when you know how to calculate footfall in retail it can help you with many retail metrics like: How to plan your store for peak footfall times? How much stock you need in the store and how often you'll need to restock it? What products are selling well on an hourly basis? This is so crucial information for retailers that will help inform decisions about where to place certain items or which ones may be more popular than others etc. When stores should have promotions (if they want), discounts, and raise weekend sales? We've put together an elaborate, research-based White Paper that covers these segments: How have pandemics catalyzed technological innovations Customer sentiment and behavior during COVID-19 An omnichannel customer engagement strategy to drive sales in retail and footfall The ultimate roadmap to increase retail footfalls How to build the perfect loyalty program to turn foot traffic into brand ambassadors? You can find the same over here, hope my team's effort comes in handy to some of y'all that could improve your store visits, cheers!

Sample

A sample library that implements Generative AI-Augmented Program Execution (GAPE)

Lessons from 139 YC AI startups (S23)

How do byte-level language models work?

Sophomore computer science student, looking at ISLP vs ESL vs mlcourse.ai

MMML | Deploy HuggingFace training model rapidly based on MetaSpore

Browser Agents Real Example

Browser Agents Real Example

[D] Why I'm Lukewarm on Graph Neural Networks

[D] The Rants of an experienced engineer who glimpsed into AI Academia (Briefly)

[D] What is your honest experience with reinforcement learning?

[P] I Trained a Model to Generate Video Game Pages

[P] A Simpler @PyTorch Annotated Implementation of EleutherAI's 20B Language Model GPT-NeoX.

[D] What is your honest experience with reinforcement learning?

[P] I created a package implementing a SOTA technique for XAI ( Explainable AI)

[P]MMML | Deploy HuggingFace training model rapidly based on MetaSpore

[D] Last Week in Medical AI: Top LLM Research Papers/Models (December 7 - December 14, 2024)

[P]MMML | Deploy HuggingFace training model rapidly based on MetaSpore

[P] I built an open SotA image tagging model to do what CLIP won't

[D] Last Week in Medical AI: Top LLM Research Papers/Models (December 7 - December 14, 2024)

[D] What is your honest experience with reinforcement learning?

How To Learn About AI Agents (A Road Map From Someone Who's Done It)

Recently hit 6,600,000 monthly organic traffic for a B2C SaaS website. Here's the 40 tips that helped me make that happen.

101 best SEO tips to help you drive traffic in 2k21

Only 2 months of cash in the Bank for my business but was able to save it with the help of AI.

Giving away AI Startup Ideas - but no traction? What am I doing wrong?

Writing a exercise based TTRPG rulebook for a system where your real world fitness is tied to character progression

A lead generation agency using personalized physical outreach

🛒7 Strategies to Increase Retail Store Footfall post-COVID | Ultimate Blueprint & Guide 📈

SUPIR

xpert

ARENA_2.0

n8n-docs

Prompt_Engineering

AITreasureBox

aima-python

generative-ai-use-cases-jp

prompt-injection-defenses

ai-hub-gateway-solution-accelerator

awesome-ai-in-finance

short-video-automation

awesome-quantum-machine-learning

eiten

airplay2-receiver

aima-java

voicefilter

evostra

coca

airoboros

bubbln_network-automation

introduction-to-ai-orchestration-with-langchain-and-llamaindex-3820082

studio

kodyfire

internet-tools-collection

Stop Learning Excel—Meet the AI Spreadsheet

ai-learning-roadmap

💯ATS Friendly Resume using Chatgpt 🤯💯 #chatgpt #atsresume 👇🏻Click Here for prompt👇🏻