Generative AI Group Chat: Hackathon and Alpaca AI

Introduction

  • Group chat discussing Generative AI
  • Non-techie member seeking help to keep up with the discussion

Participants

  • [PHONE REMOVED] [PHONE REMOVED] [PHONE REMOVED] are PMs
  • [PHONE REMOVED] [PHONE REMOVED] are founder and VC respectively
  • [PHONE REMOVED] is Mumbai based creative agency person
  • Manjot [PHONE REMOVED] is with Lightspeed VC currently

Hackathon

  • Prize money split is 1L minimum per track, split between all members of the winning team equally
  • Evaluation criteria is left to the jury’s discretion
  • Some members prefer leaving the criteria to the judges while others prefer having a clear set of criteria
  • Tracks are announced for the hackathon

Alpaca AI

  • https://alpaca-ai-custom4.ngrok.io/ is slow to try prompts
  • It looks more like davinci-002 than the 003 series in terms of ability to follow logic over multiple hops
  • No weights, so it’s iffy on what is going inside training
  • The instruct dataset is key
  • The PR they have raised with Huggingface gives you the training code
  • Built for enterprise
  • 3x the price of other AI models
  • 12 cents per 1K token completions is quite expensive
  • GPT4 capabilities have a FOSS counterpart from Amazon: github.com/amazon-science/mm-cot

Miscellaneous

  • Too much happening in this space in a day, difficult to keep up with things
  • Be My Eyes demo is hilarious
  • IST timing is odd for some members
  • Thank you for adding me to the community

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

  • https://alpaca-ai-custom4.ngrok.io/ - The message discusses the slow performance of the website for trying prompts and compares its ability to follow logic to the davinci-002 series. It also mentions uncertainty about the training process due to the lack of weights.
  • https://twitter.com/OpenAI/status/1635687373060317185?t=nhm2BaNejunBE9-syFgEeA&s=19 - The tweet mentions that the instruct dataset is key and provides a link to the PR raised with Huggingface for the training code. It also suggests that the project is built for enterprise.
  • https://www.youtube.com/live/outcGtbnMuQ?feature=share - The message mentions the “be my eyes demo” being hilarious and thanks someone, but also notes odd timing and waiting for a summary from others.
← Exploring OpenAI's Impact, Learning Resources, and Research Topics
On-Premises Deployment for OpenAI LLMs and Google's FLan-T5 →