Generative AI Meet-up and LLM Developments

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.com/CohereAI/status/1649097293201547264?t=UsFrQQNyNhdkoqPz8AgcrA&s=19

Models:

  • Controlnet inpaint model
  • Diffusers library
  • Minigpt-4 is better than stability’s latest stable diffusion model
  • Llava-vl is a good model for GPT-4 samples
  • Link: https://llava-vl.github.io/

Other:

  • EyeQuant founder talks about text2film
  • Link: https://twitter.com/fabianstelzer/status/1648700767992180737?s=48
  • Original song by Ariana Grande
  • Link: https://www.youtube.com/watch?v=DOJremEQw88
  • Martin Shkreli’s AI launch
  • Link: https://twitter.com/marty_catboy/status/1649032460573745152
  • Weights & Biases LLMOps London event
  • Link: https://www.youtube.com/watch?v=YfBtytGNEKE

The description and link can be mismatched because of extraction errors.

  • https://hasgeek.com/generativeAI/april-meetup/ - This link provides information about a meet-up happening in BLR on Saturday evening, which the person is asking about in the message.
  • https://hasgeek.com/generativeAI/april-meetup/ BLR, Saturday evening: A meetup event for generative AI in Bangalore on a Saturday evening. No direct relation to the second URL provided.
  • https://blog.replit.com/llm-training A good blog post by Replit where they give us high-level description of how they train their own LLMs. The blog post mentions that the Replit team is <5 people, which is amazing.
  • Fabian Stelzer’s tweet about talent density and head count, with a heart and fire emoji, and a link to a video where the EyeQuant founder talks about text2film. (https://twitter.com/fabianstelzer/status/1648700767992180737?s=48)
  • https://simonwillison.net/2022/Sep/17/prompt-injection-more-ai/ - The message “midjourney outsourced frontend to discord 😂” and “Cracked crazy adoption” are related to the link.
  • Original song by Ariana Grande: https://www.youtube.com/watch?v=DOJremEQw88 (related to a message about cracked crazy adoption and a question about the technology used)
  • https://twitter.com/vboykis/status/1648756882679427072?t=JDfSjZx03Rlj9JHp1IbBpA&s=19 - A tweet discussing the idea of setting up different teams for different keys to track usage per organization basis by OpenAI, with a related analogy.
  • The user mentions a GitHub repository for OCR called “donut” (https://github.com/clovaai/donut) and questions the performance of Tesseract in production environments.
  • Microsoft’s LayoutLMv3 already does OCR: https://huggingface.co/microsoft/layoutlmv3-base, as discussed in a conversation about using a multimodal LLM for OCR instead of Tesseract or AWS/GCP API.
  • https://github.com/microsoft/MM-REACT and https://github.com/Layout-Parser/layout-parser are mentioned in relation to the author’s experience using MM-REACT, which utilizes the reasoning capabilities of LLMs to extract information from visually rich documents and is found to work better than Donut.
  • https://github.com/Layout-Parser/layout-parser: Last time I checked, Langchain had a layout-parser integration for PDFs.
  • https://blog.eleuther.ai/transformer-math/?s=08: Langchain had a layout-parser integration for PDFs and an optimized implementation for Whisper, which is faster than real-time.
  • https://twitter.com/CohereAI/status/1649097293201547264?t=UsFrQQNyNhdkoqPz8AgcrA&s=19: A tweet asking for suggestions on tools or best practices for recording LLM experiments and tracking metrics like accuracy/precision improvements. The message also emphasizes the importance of being open and accessible for more acceptance and traction.
  • https://llava-vl.github.io/ - A website that performs well on gpt-4 samples, mentioned in a tweet comparing it to minigpt-4. The tweet can be found at https://twitter.com/marty_catboy/status/1649032460573745152.
  • https://llava-vl.github.io/ is a very good website that performs well on GPT-4 samples, as mentioned in a tweet by Marty Catboy about Martin Shkreli’s AI launch.
  • https://www.youtube.com/watch?v=YfBtytGNEKE - Link to live talks at Weights & Biases LLMOps London event mentioned in a message about Martin Shkreli’s AI launch.
← Exploring AI Tools, Techniques, and Pinecone
AI and Data Analysis Techniques in UI Design, Image Editing, and Language Models →