OpenAI’s decision to not do closed source OpenAI Chief Scientist Ilya S. explains that OpenAI is not doing closed source because of safety reasons, but because of competitive reasons. He believes models can be improved to the point in the future where it becomes a safety concern. Competitive reasons The competitive reasons mentioned by Ilya S. refer to Microsoft, which is a major investor in OpenAI. Microsoft has been known to use closed source models in the past, which gives them a competitive advantage.

Introduction Group chat transcript on Generative AI Chaotic discussion on various topics Learning Resources Link to a post on what transformers are Cohere has two of the best ML content creators List of top AI-themed newsletters shared 42papers.com recommended Ben’s Bites newsletter recommended Hackathon Invitation to join a team in Warpspeed GenAI hackathon Deterministic Output Discussion on making GPT2 or BLOOM model outputs consistent Two ways to make it deterministic: setting temp=0 or setting seeds Pro tip to pursue this line of reasoning and gain first-hand experience Visual Aesthetic Scoring Discussion on AVA dataset and image dataset curation methods Link to a space on huggingface for aesthetic predictor Use of clip interrogator to score generated captions Hallucination and determinism are not related Visual aesthetic scoring is largely a final layer problem Discussion on the limitations of the model Dataset Creation Request for good library references to create datasets for LLMs Discussion on creating a dataset based on few hand-written examples Suggestion to check the terms of service for commercial use Discussion on using a visual QA model to identify errors in a picture Workflow suggestion to detect errors in the image and fix them using models finetuned on those datasets Suggestion to use Andrew Ng’s landing.

OpenAI and Google models: Google models have no moat OpenAI may have a moat due to early mover advantage and alignment with Microsoft/Azure OpenAI’s moat may come from government and finance domains OpenAI may be able to maintain its lead for 2 to 5 years or more Open source models are far behind OpenAI currently, but may catch up eventually Network effects may not be a moat in this case Microsoft+OpenAI can make inference costs low and onboard other businesses on plugins Hugging Face:

Introduction and greetings Members introduce themselves and exchange pleasantries Generative AI for digital painting Discussion on the potential for creating artist-focused digital painting tools using Generative AI Current methods involve hacking with automatic tools and using Photoshop to achieve desired results Large scale dataset creation for LLM model training Member asks if anyone has experience creating large scale datasets Links The description and link can be mismatched because of extraction errors.

Topics Discussed in Chaotic Generative AI Group Chat Transcript AI Models and Tools GPT4 creates a vector DB Replit open source LLM just dropped Finetune on instruction dataset curated from geeksforgeeks Prompt Injection in less than 10 minutes (video, slides, transcripts) JSONFormer: guaranteed JSON output with Huggingface LLMs DeepFloyd or Multi-ControlNet? Is there a model (other than gpt4) that can extract info from an image in JSON format? Autogpt app Table-transformer might help if result cards are in tabular form AI Applications Abstract artwork made using stable diffusion Extracting values without OCR Parsing information from result cards Langflow AI Industry and Innovation India’s position in the generative AI race Chris Lattner’s new ML-focused language Ecosystem graphs for AI startups Networking and Collaboration Connecting with others in the group for contribution and collaboration Looking for journalists or media folks in the group Miscellaneous Group invite link request AI and art techniques OpenAI service on Azure Multi-modal GPT4 access Sanford C.

Article Bias and Twitter Discussion about biased articles and preference for tweets Shared link to a tweet: https://twitter.com/zoink/status/1653052807950536706 Learning with ChatGPT Discussion on how to use ChatGPT for learning Mention of yudbot.com as a resource API Level Caps Mention of developing an API level cap for a llm vault manager Question on whether API level caps are generally kept Shared link to yudbot.com as a potential resource Paywalls and Capitalism Discussion on paywalls and access to news articles Mention of capitalism in a political group context Conversational Memory with GPT-3.

Access to GPT Plugins: OpenAI demos it publicly Participants want access to the tool Suggestion to make a Twitter viral demo to get access One participant has access but cannot find the tool in the UI Web Development: Request for freelance Flask/Python web stack developers Link to GitHub repository for Eva: https://github.com/georgia-tech-db/eva AutoGPT: AutoGPT is emerging as a catch-all for automation Jason Calacanis has posted a bounty for outbound sales emails on Replit One participant was able to deploy something locally to generate personalized DM, email, and coffee conversation points for a LinkedIn profile as a non-coder Group Management:

Music and AI Collaboration between GPT and Bark for music creation “Bark is lit too” 🔥 AI Tools and Libraries pandas-ai library: https://github.com/gventuri/pandas-ai Issues with ChatOpenAI giving “Could not parse LLM output:” errors Code snippets for language chain helpers and embeddings utils: https://github.com/jerryjliu/llama_index/blob/590639a14dd7346b7f5cc00a21dd24ce0d35ae30/gpt_index/langchain_helpers/text_splitter.py#L240 https://github.com/openai/openai-python/blob/c556584eff3b36c92278e6af62cfe02ebb68fb65/openai/embeddings_utils.py#L21 Scale.com for content language: https://scale.com/content-language AI Models and Projects Cohere’s multilingual model Project idea for creating shortcuts to automate tasks Resources for creating charts/graphs from data, open-source “Chart-GPT” resources especially python based Implementing file search using Langchain’s drive connector Performance gains attributed to Rust vs Golang for HNSW algorithm AI Discussions and Queries Issues with limited RAM on Repl on Replit Measuring recall for file search Tuning recall/speed for HNSW algorithm Learners’ discussion at 4pm Feedback on a new product for generating design from prompts using a design system Generative AI impact on graphic design and marketing Interviewing product designers and graphic designers in marketing agencies AI Resources and Links LLM.

VectorDB and LlamaIndex Discussion on whether LlamaIndex or Langchain support VectorDB with Sources with GPT4 or GPT3.5-Turbo Sharing of GitHub link for LlamaIndex demo Suggestion to use response.source_nodes to get sources Discussion on using evaluation module or regex to get desired sources Mention of OpenAI models not citing references Sharing of Langchain link for QA with sources example Discussion on limitations of Langchain and need to extend chain functions Mention of LlamaIndex not having such limitations AI Models and Applications Discussion on models for aesthetic score of images Mention of ResNet and classification model for user-scored images Discussion on generating professional-level DSLR photos using AI Mention of LAMINI AI library for fine-tuning LLMs to custom domains Announcement of upcoming “Learning Transformers/NLP/ML” discussion Discussion on D-ID and SadTalker for generating content Mention of Whisper models for Hindi transcription Suggestion of Deepgram and Monster API for transcription Discussion on building generative model for legible pdf/image documents Mention of AI-generated content for social media and advertisements Request for transcription API that can handle Hindi and regional languages Miscellaneous Discussion on automating extraction of WhatsApp group data Mention of RunwayML Gen 2 and Text2Video-Zero Request for transcription API that can handle Hindi and regional languages Announcement of weekly/monthly newsletter Request for tech expert in Generative AI for Zoom meet-up Request for connections with those working on advertisements Discussion on building AI avatars and personalized content Mention of guardrails for cloned voices Discussion on building QnA over CSVs using Python code or SQL Links The description and link can be mismatched because of extraction errors.

Generative AI Meetup Nvidia-HF event on Saturday morning Online event, URL: https://sites.google.com/huggingface.co/generative-ai-meetup Need an experienced person to guide and answer queries, volunteer needed Twitter link for volunteer: https://twitter.com/thesephist/status/1651677221797371904?t=UAtNw7WFH00_AS5oGpirUw&s=19 AI Applications Using diffusion for detection/recognition/segmentation tasks Tesla using diffusion as part of their lane detection algorithm Transformers used for lane detection, URL: https://youtu.be/aVjDX5XshYo Text2motion Promptlayer Data exploration and feature engineering resources needed Sequential prompting for controlled conversation, e.g. roleplay AlignmentAI and PMGPT for guided chat with roleplay AI assistant/co-pilot for product managers/product teams GPT plugins AutoGPT AutoGPT for basic questions AutoGPT’s limitations AutoGPT’s marketing Large scale human feedback and nudges for AutoGPT Namefinder.

Weaviate Weaviate allows pre-filtered vector search based on metadata and builds separate indices for metadata. Their managed SaaS charges based on the number and dimensions of vectors, irrespective of the extra metadata stored. Learning LLMs/Transformers/ML Someone suggested doing a zoom session to discuss learning LLMs/Transformers/ML. People discussed how they are dealing with rate limits and 5xx errors with OpenAI. Some suggested using a leaky bucket to smoothen the API calling rate.

OpenAI: Doubtful of claim that they don’t need data, they still need more data for training Multinationals may start using OpenAI models OpenAI is looking for data partnerships, but currently only able to cater to proposals from big tech companies No list of data partners available OpenAI’s platform API data policy is public information OpenAI is small and has exhausted internet and academic datasets for training their models Replit Codegen model is better than OpenAI Codex in many human eval tasks and is smaller in size OpenAI’s models are a utility like EC2 and have greatly improved the UX of consuming models Many NLP models now work with OpenAI’s transformer library Stanford NLP is highly regarded Qualcomm has made progress in ML on edge Palantir has launched ChatGPT for war Anduril may come up with something similar OpenAI’s tech use cases range from disaster relief to drone warfare Music and Audio Generation:

ChatGPT Chrome Extensions Discussion on cool ChatGPT related Chrome extensions RL without Human Feedback Shared interesting read on potential to use RL without human feedback OSS Projects Discussion on interesting OSS projects in the space ShareGPT Mention of ShareGPT AI Dating Expert Discussion on a bot trained on Bumble chats to reply like the user Prompt Injection Attacks Warning about prompt injection attacks Vector Search Discussion on using vector search on past swipes and bios OpenAI Branding Discussion on OpenAI branding and trademark on GPT Qdrant Discussion on using Qdrant for vector search Fine Tuning OpenAI Models Discussion on fine tuning OpenAI models on code Tips and do’s/don’ts for fine tuning Use of embeddings and GPT-3.

OpenAI and Dense Embeddings: Currently using OpenAI and dense embeddings Planning to move to hybrid soon Using similarity search on a flat embedding space to retrieve context GitHub and LLM: Came across a GitHub where someone used PEFT to fine-tune an LLM based on their iMessage chats to impersonate and create a bot that talks like them Supabase: Discussion about Supabase and Postgres performance queries Suggestions to check RLS policies and use explain.

Anthropics pitch deck Does anyone have access to Anthropics pitch deck? RLHF for instruct models OpenAI’s John Schulman gave an interesting talk at Berkeley last week on why RLHF was needed to get the instruct models to behave nicely. Deepfakes and audio synthesis First instance of an Indian politician referring to deepfakes / audio synthesis? AI courses and resources course.fast.ai for being able to make sense of all of this — even as it changes in 2-3 months and we add VQA (Vision) to mainstream OpenAI APIs Lot of new work should come from STT and TTS side, including performance improvements like Whisper-JAX in the coming 4-6 month and more important, voice cloning, avatars and the like.

Cloud GPUs: Discussion on using GPUs for training models Google Colab and Kaggle Notebooks suggested as options Comparison table for cloud GPUs shared Cheapest option for model fine tuning mentioned as rtx5000 on runpod.io Pricing of cloud GPUs discussed, including economies of scale and interest rates on capital investment Vast.ai mentioned as a cloud GPU rental option Model training: Tips shared for reducing container image size when using Sentence Transformers for word embeddings Question asked about enabling multiple checkpoints in automatic1111 while fine tuning a model using dreambooth Privacy concerns:

Web and AI Tools: Langchain and WebGPT discussed Kubiya.ai resource shared and discussed Helm charts and K8s clusters can be managed via chat interface using natural language Discussion on the evolution of computer engineering from command lines to GUIs and back to command lines AI Techniques: Advanced stable diffusion techniques for inpainting and controlnet discussed Tutorial on creating consistent AI characters across images with SD shared Request for AI art/text to image group Proposal for a separate DeepMedia: Generative Art (Text to Images, Video, Music) group Illustrated Transformers and other resources shared Pinecone:

Meet-up: Generative AI meet-up in BLR on Saturday evening Link: https://hasgeek.com/generativeAI/april-meetup/ LLMs: Replit’s blog post on training their own LLMs Link: https://blog.replit.com/llm-training OpenAI Whisper had a team of 6 for their project Attention is All You Need paper had 8 authors Midjourney outsourced frontend to Discord Microsoft’s LayoutLMv3 does OCR Link: https://huggingface.co/microsoft/layoutlmv3-base MM-REACT uses reasoning capabilities of LLMs to extract information from visually rich documents Link: https://github.com/microsoft/MM-REACT Tools: Best practices for recording LLM experiments W&B is a familiar tool CohereAI tweet about model specificity and accessibility Link: https://twitter.

UI Design: Used UI elements from Koo app and dreamstudio Gradio space and playgroundai.com suggested as starting points Jadoosnap.com suggested for product design Data Analysis: Shapiro Wilk test explained as measure of normal distribution Normality not necessary for regression, other methods suggested Support vector regression and tree-based methods suggested for non-normal datasets Extreme Value Theory tools suggested for non-normal datasets Hill Estimator suggested for non-normal datasets Image Editing: Inpainting models suggested for image editing Qdrant and Chroma suggested for custom databases Vall-e and Descript suggested for voice cloning DeepSpeed evaluated and recommended Ada from OpenAI and sentence transformers from HuggingFace compared Language Models:

OpenAI Plugin Store: Adept.ai got copied as an in the OpenAI Plugin Store, called Multi-on Experience needed: People with experience in CLIP, BLIP-2, and VQA are needed for a project Someone with solid experience in multimodal vector similarity search is also needed Meeting invite: A meeting invite was shared in the WhatsApp group for the project Projects: Multimodal vector similarity search Personal search Generative AI for solving problems and building multi-modal systems Domain-specific feedback loop and fine-tuning Context-based search with cohere embeddings and OpenAI GPT3.

Langchain and Mendable.ai integration for building with Langchain; Kapa Langchain bot for asking doubts on Discord; Paradox.ai for recruiters; Skillate and Leoforce for automated resume screening; issues with bias and possible unethical use of AI solutions in recruitment; use of GPT during interviews; AI-generated music; discussion on GPT-4 and its performance on JEE questions; fine-tuning vs. prompt engineering davinci; use of LLMs for math problem solving and chain-of-reason prompting; LOL dataset for low light pictures; Huggingface’s PEFT wrapper for multiple fine-tuning methods; founder’s R&D work for Bewgle; LLM components for memory weighing in Langchain; publishing research papers on generative AI; controllability issues with LLMs and use of guardrails for structured outputs.

Sci-Fi and AI Discussion on sci-fi authors such as Asimov, Clarke, and Heinlein Comparison of Multivac from Asimov’s stories to today’s LLMs/Auto-GPT Sharing of a sci-fi story by Greg Egan Multimodality and its potential to improve LLMs’ understanding of reality Sharing of a paper on color clustering Appreciation for Asimov’s work and genius Recommendation of Cixin Liu as a favorite author Community Guidelines Drafting of community guidelines for the group Sharing of a website for sci-fi enthusiasts (Orion’s Arm) GPT-Related Topics Request for a service that allows using chatGPT interface with API key Sharing of a website for chatbot UI Discussion on context window handling of such services Recommendation to use GPT-4 with large context windows for cost efficiency Request for guidance on fine-tuning with custom data and sharing of a GitHub example Sharing of articles by Databricks on tuning Dolly LLM Discussion on deploying a Docker image with Sentence Transformers and OpenAI combined Recommendation to deploy the model separately behind a service endpoint to avoid distributing the model in application containers Discussion on the use of embeddings and OpenAI for cost reduction Sharing of a Twitter thread on fine-tuning and prompting Recommendation to use external vector databases like Pinecone or Weaviate instead of fine-tuning for cost efficiency Discussion on chunk size for converting text dump and documents into embeddings Sharing of Langchain’s conversation memory types for optimization Question on the possibility of specialized hardware speeding up embedding search Request for a summary of the GPT-related topics due to the increasing number of messages in the group Sharing of a post by Vespa founder on introducing embeddings and vector search Sharing of an IPython ChatGPT extension Discussion on Auto GPT’s intermediate summary phase for webpage parsing Experience with LLMs and plans to add new tasks in classifier and separate their workflow standalone Sharing of Langchain’s memory chains and conversation memory for summarization and retrieval Links The description and link can be mismatched because of extraction errors.

React Development: Someone is looking for React developers for a paid weekend project They ask for leads to be DM’d to them Controlnet Inpainting: Someone asks if anyone has gotten the latest controlnet v1-1 nightly release models working with inpainting They mention that they have gotten the models working with multi controlnet, but not for inpainting yet Azure Cognitive Search: Someone brings up MSFT’s recommendation to use Azure Cognitive Search with Azure OpenAI for enterprise document search in Azure They mention that it may be costly and ask for thoughts from the community on production level deployments Another person mentions the option of uploading document embeddings into a Vector DB and doing semantic search They share that they are currently using this method with Pinecone and have 10681 vectors in their DB GPT-5:

LangChain and Tenacity API LangChain uses Tenacity API +1 on using Tenacity API LangChain has an API for the idea of kids as superheroes Andrej Karpathy’s Neural Nets Video Finished watching Andrej Karpathy’s video on Neural Nets Highly recommended for anyone starting or in the field Link to the video playlist: https://youtube.com/playlist?list=PLAqhIrjkxbuWI23v9cThsA9GvCAUhRvKZ Image Inpainting and Diffusers Tried Img2img models for inpainting on an online image Asked for suggestions on using diffusers for inpainting Link to Stable Diffusion XL beta: https://stability.

Dolly 2.0 Databricks has released Dolly 2.0, a commercially viable LLM. The dataset was crowdsourced from Databricks employees. The training code, dataset, and model weights have been open-sourced and are suitable for commercial use. Link: https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm OpenAI’s Consistency Models OpenAI has proposed consistency models, a new family of generative models that achieve high sample quality without adversarial training. Link: https://www.marktechpost.com/2023/03/10/open-ai-proposes-consistency-models-a-new-family-of-generative-models-that-achieve-high-sample-quality-without-adversarial-training/ Vector Databases Participants discussed various vector databases, including Pinecone, Weaviate, Qdrant, Chroma, Redis Vector Cache, Vespa, pgvector on Supabase, and Milvus.

AI Embedding Models: Comparison between OpenAI’s “text-embedding-ada-002” and Huggingface models or Sentence Transformers Some prefer Huggingface or Sentence Transformers for their use case Recommended model: “sentence-transformers/all-mpnet-base-v2” or “multi-qa-mpnet-base-dot-v1” for semantic search Cross-encoder recommended for better scoring mechanisms AI in News Media: Traditional News Media is frightened with AI Buzz is hot and controversial Interviews conducted on AI opinions AI Chatroom Plugins: ChatGPT Plugins SF hackathon winners announced Winners have higher usability than expected Indians are no longer playing catch up AI Conversational Memory:

Generative AI Discussion about a paper on generative agents and a demo at https://reverie.herokuapp.com/arXiv_Demo/ Mention of Andrej Karpathy’s tweet Sharing of a link to a website that uses flat GPT output without crawling their pages Mention of a bot that doesn’t have memory and a classic NER ambiguity mistake Discussion about ChatGPT plugins and OpenAI account access Speculation on how OpenAI determines a user’s country and suggestions for contacting OpenAI Mention of Ojasvi leading AI at mydukaan.

Hackathon Ideas Discussion about a thread of hackathon ideas and similarities to their own hackathon Political action project was the first thing they sold to a client back in 2022 Creators of the thread haven’t been mentioned, but can tweet to Joseph or @swyx to ask Mention of knowing @swyx from the react world because of a talk Variety of chatbots in their hackathon GPT 3.5 Discussion about problems with limiting output length Question about changing length units to token instead of characters Recommended way to do this is to use Guardrails and turn on re-ask Mention of Guardrails creator doing a demo at their hackathon venue Link to Guardrails GitHub page Mention of Indians no longer playing catch up in the tech world Kor and Guardrails Discussion about Kor being a general purpose parser for any text/schema Confusion about why the schema format in Guardrails is not the same as the output format Balancing language/lib design POV Resizing prompts work well when input and output language is English BPE tokeniser is notoriously unstable for CJK and Indian languages Prompt engineering is like casting a spell Suggestion to tweet and tag OpenAI folks for more information on GPT4 GPT4 and GPT3 Discussion about customers raising support tickets for being fooled about GPT4 access Suggestion to improve with a system prompt Chat models are instruction finetuned, sanitised and scaled forks of large base models Suggestion to have GPT3.

SAAS Company Sale Discussion about a tweet by Dan Martell regarding selling a SAAS company. Generative AI Tools Discussion about a tool for generating desi lofi girls. Mention of a tool built by [PHONE REMOVED] and friends for YouTube videos. Mention of sitegpt.ai, which does something similar. Mention of Microsoft Edge’s copilot option for generating text. Mention of jsonlines format for solving text generation challenges. Vector DBs Discussion about starting a vector DB company.

Integration of chatGPT/LLMs into smart home devices Query about integration of chatGPT/LLMs into Alexa/Nest type of smart home devices Mention of tools that can do it, but openai key is a barrier to adoption Paddlespeech Query about anyone dabbling with paddlespeech ChatGPT talk by Anil Ananthaswamy Video of the talk by Anil Ananthaswamy on ChatGPT held at BIC is published Link: https://youtu.be/WF28ZwhUCc4 Automatic1111 and DrawThings App Suggestion to use Automatic1111 on colab Suggestion to download DrawThings App on MacBook GPU memory issue, try smaller models Generative AI on local machine Discussion about running generative AI workflows locally on a Mac Recommendation to either go all the way or get the Macbook air Suggestion to get as big of a machine as your wallet permits Mention of text-to-video taking off in the future Optimization of models for running on Apple RAM Query about how optimized the models are for running on Apple RAM as opposed to Nvidia GPUs Serverless Vector DB and pricing Discussion about serverless Vector DB and pricing around it Link: https://twitter.