2024-10-29

Owned by Tom Qin
Last updated: Oct 30, 2024 by Wilson Wang
3 min read

Agenda

  1. AI Hachathon, @Kandan Kathirvel, @Tina Tsou , 15min

  2. https://lf-edge.atlassian.net/wiki/x/OQLP , Integrating quantized model https://ai.meta.com/blog/meta-llama-quantized-lightweight-models/ , @Kris He , @Chen Lai, 15min

  3. Update on AI Agent Platform, @Wilson Wang , @Borui Li (李博睿) , 10min

  4. Brainstorming AI Agent Marketplace, Noe Otero, 5min

Participants

@Tom Qin @Tina Tsou @Wilson Wang @Qi Tang @Moshe Shadmon @Kris He @Borui Li (李博睿) Chun Liu, @saiyan68 , @AKRAM SHERIFF

 

Recording

https://bytedance.us.larkoffice.com/minutes/obusme696v152p39p2398k15?from=from_copylink

Summary

The meeting discussed the progress and plans of various work streams within the Infinite Edge AI Technical Steering Committee, including the organization and hosting of a hackathon, updates on the AI agent platform, and discussions on the scalable and performance edge agent runtime. The main contents included:

  • Hackathon: Tina Tsou, Kenneth, and others discussed the organization of a hackathon with Google, TikTok, and other parties. They are considering November or December for the event and will send an email to include all relevant parties.

  • AI agent platform: Leia shared the latest quantized model, Llama 3.2, and they are evaluating its accuracy. They also discussed the possibility of running a large language model on a cell phone and converting natural language into device APIs.

  • Spear architecture: Bo Rui presented the preliminary architecture of the scalable and performance edge agent runtime, which includes a user space and a kernel space, with an API to bridge the two. They are looking forward to comments and advice.

  • AI agent marketplace: Tina Tsou and Noah discussed the idea of an AI agent marketplace, where different types of agents could be categorized and provided. They encouraged interested parties to join the discussion.

  • Other work streams: Qi updated on the progress of work stream 2, focusing on improving the accuracy of the current large language model using retrieval and fine-tuning methods. They also discussed the differences and overlaps between work streams 2 and 3.

Chapters

10:29 Meeting Kickoff

Meeting Kickoff

12:02 Infinite Edge AI Technical Steering Committee Meeting: Hacker Song and Work Stream Updates

This section is about an Infinite Edge AI Technical Steering Committee meeting. They are planning a hackathon (hacker song) with multiple parties including Google and TikTok. It was initially planned for October but is likely to be in November or December. Also, there are updates on running Llama models on cell phones and work stream updates on the AI Agent platform, including vector DB API progress.

24:33 Introduction to the Sphere Architecture in Work Stream 3 and Related Discussions

This section is about Li Borui presenting the preliminary architecture of work Stream 3 named Sphere. He explains the legend of the architecture diagram, splitting Sphere into user and kernel spaces. In the user space, there are layers including applications developed with various SDKs. The kernel space has system services. Tina and Wilson then give some comments and questions regarding components like vector DB and edge ingress.

36:08 Discussion on Figure Coordination and Moving to Next Topic in a Meeting

This section focuses on a figure that needs to be reshared. Tina Tsou has difficulty connecting the current figure with the one on the wiki page. Li Borui acknowledges the issue and will discuss with Wilson Wang to update both figures to make them more coordinated. Wilson mentions that the core parts are the same. After this discussion, they are ready to move to the next topic if there are no more questions from the audience.

40:12 Discussion on AI Agent Marketplace, Work Streams, and Upcoming Events

This section mainly focuses on the AI agent marketplace. Tina Tsou mentions Noah's ideas for it and encourages interested parties to discuss or contact relevant people. There are also discussions about code links on wiki pages, work stream updates (including short - term goals for work stream 2), potential overlaps between work - streams, and upcoming events.