Attendees

• Alaska
• Pratik
• Philipp

Topics

• Announcements

  • LLM for triaging
    • please review this doc so we can justify the project to HQ planning

• Pending TODOs/followups (wont get an active topic if no news have to be announced

  • Open PR list growing, can Triagers help in that?
  • Usage of LLMs
    • is the setup at the HQ up and running as expected?
  • talk about https://devtalk.blender.org/t/proposal-label-changes/38984
  • Alaska has made a tool to automatically notify people of old Needs information from user reports: https://projects.blender.org/Alaska/needs-information-reports/issues
  • Feedback regarding our triaging infrastructure
    • We use a script to help write our weekly reports. There have been a few issues with it's output recently (E.g. Duplicate data, missing data from PRs merged by others)
      • Yimming was planning to look into this
  • Wizard
    • Gather a list of two or three examples from "out there" that work similar to what we want to achieve (add that to https://projects.blender.org/infrastructure/blender-projects-platform/issues/78) is still pending, some more discussion in https://projects.blender.org/blender/blender/issues/121260 though regarding Richards experiments
  • Could we add a feature that lets users know the location of crash logs after Blender crashes so they can more easily provide it in a bug report?
    • There has been progress on the crash window pull request
  • Discussion about exposing dependency cycles to the end user (an idea we have discussed in previous meetings)
    • Design task: https://projects.blender.org/blender/blender/issues/134659
    • PR: https://projects.blender.org/blender/blender/pulls/134672
  • Discussion about Tests:
    • reply in devtalk

Meeting Notes

• TODOs/followups
  • LLM for triaging
    • We quickly went over the LLM document
      • Probably best to put a "plan" section (at the end?), shifting the database to HQ, also generally stating we are in tryout phase… and also put a date there for re-evaluation of goals/status etc.
    • Open question: how often do we (as Triaging) need to download new models and how is the current process of doing it (requires Arnd/Bart each time?)
      • Are there other things we can do to reduce the impact on the people at the head quarters (Time needed to maintain the machine, and the network impact for people at the head quarters while downloading large models)?
    • At the moment there are some issues with with the LLMs (Germano mentioned they can get confused about who they're responding too).
      • Their proposed solution was to fine tune the model. But could we explore some other ideas (Adjusting system prompts, adding more databases and more tools to the LLM, etc, to create "agents") to see if we can resolve these issues without having to resort to finetuning?
  • We didn't talk about other topics since the people incharge of them were away.