Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

AttendeesAttendees

Recording

https://bytedance.us.larkoffice.com/minutes/obusq8pb1jx4ylqd9tip3q64?from=from_copylink

Summary

The meeting discussed the progress of the edge computing project and the related work streams, the main contents included:

  • Federated learning demo: Presented by Lark, will show how to enable federated learning at the edge on the model training and inference side.

  • TikTok presentation: Shared by Qi, introduced the idea of a data adapter for user-centric, lightweight, and self-describing features.

  • Work stream updates: Tom Qin updated on the work stream, including the integration of lightweight machine to machine in Shifu and the progress of Lom.

  • Retail edge apps: Tina Tsou shared two edge apps for the work stream, one for real-time translation and the other for AI fashion.

  • Next steps: Wilson Wang mentioned that he will set up his meetings to attend and give more updates next week.

Chapters

00:26 Meeting Updates with Team Members including Tina's Late Joining

This section is about a meeting. Wilson Wang sends out the meeting invitation. Tina Tsou joins the meeting though she is in Dublin at 2 a.m. An unknown speaker gives an update on federated learning and the plan to show a demo at the end of the month. Tom Qin shares updates on lightweight machine - to - machine integration in Shifu, and mentions work on Lom. TikTok is also present and is about to present.

09:14 Accuracy Improvement in AI Competitions and the Need for Edge Database in Specific User Scenarios

This section is about the speaker's experiences in two competitions. They mention that the accuracy of foundation models in these challenging LLM competitions is low, but can be improved with proper context, retrieval systems, algorithms, and human hints. They also discuss the need for an edge database interoperable with other LLMs, and how in user - held data scenarios, accuracy should be improved.

13:26 Proposing a user - centric data adapter for agent services in data processing

This section discusses a missing user tool. For cloud platform agent services, users currently search for agents and upload data to the platform. A proposed data adapter, not officially named yet, is user - centric, lightweight and self - describing. It would be a thin layer with data context, description, cached index, and callable operators. There are also ideas about building agents by the LLM, and the adapter would be like description files with code for data processing. The demo is in progress.

17:31 Discussion on Rack Systems and Self - descriptive Files for LLM Services

This section discusses aspects related to rack systems. It's agnostic currently. If users have the OpenAI API, the rack or index part may be similar to the Lama index. The normal index is complex with many dependencies. It mentions algorithms like Facebook's FAISS and using Hugging Face embedding models. The idea of a self - description file is proposed for users to manage pre - processing data and use it across different LLM services regardless of backend changes.

21:36 Discussion on data labeling in a project phase and use of cached data, plus reliance on open - source community for implementation.

This section first discusses the data labeling in Phase 2 of a project. Phase 1 doesn't need labels mainly for data description for LMS retrieval etc. Creating a pipeline for user feedback on results can be useful for later model labeling/fine - tuning. Then, in response to a question, the speaker says they will do quick Python - based prototyping at a high level rather than integrating lower - layer implementations like Titan, relying on the open - source community instead.

25:39 Discussion on Training and Inference Framework Phases for TikTok's AI and Edge Computing

This section mainly focuses on the training or inference framework. For phase 1, it only involves prompts and doesn't deal with tuning parameters or models, just using rack. Phase 2 may involve training and fine - tuning based on certain data conditions. Phase 3 might include RL. Tina Tsou also suggests listing the phases so that interested parties can prepare to contribute, and mentions potential contributors from Hong Kong universities.

29:37 Lisa's Introduction and the Start of Work Stream 1 Discussion

This section starts with ensuring all attendees are present. A person named Lisa is introduced. Lisa gives a self - introduction, stating it's her first time logging into Lark. She has worked in AI R&D for different companies, was previously more focused on configuration and now on Lark long model. She joined Tina's WeChat group recently. She also mentions her past connection with Tina. Then Tina asks Tom to share the screen for work stream one which has two edge apps.

32:39 Discussion on Edge Apps and Other Work Updates

This section mainly focuses on some edge apps. There are two edge apps mentioned, one for live translation and the other for AI fashion. Tina Tsou introduced how the live translation app works on the geo - distributed edge influence cloud and its cost - effectiveness. The AI fashion app was also briefly described. Then there were some discussions about LFIDs and other topics like using quantized Lama on Snapdragon chips, and updates being postponed due to meeting conflicts.