Skip to main content

discussion 2024-06-27

Summary​

In the Discord chat, Samuel expressed interest in forming a team while Andy Brock joined the server. Yikesawjeez shared an article on achieving high scores with GPT-4 on arc-agi games and showed willingness to contribute until their contract ends. Kabir Kumar announced a mech interp speedrun scheduled for July, and loveofdoing discussed transformer memorization capabilities in response to various input/output combinations of Hilbert wave curves. Shaw shared an arXiv paper related to the topic and confirmed George Hotz's suggestion that Transformers with reinforcement learning could be a potential solution. Virgile Blais joined the server, Corona_Chan proposed applying facial video-to-ECG/PPG prediction models for arc tasks using Hilbert wave curves as inputs/outputs, and Arch2230 offered to collaborate on live hack sessions over weekends with goo14_.

FAQ​

  • How can we form a team?

  • SSamuel: The user expressed interest in forming a team but did not receive any direct answers or suggestions on how to proceed with the formation of the team.

  • What is icymi, and where can I find more information about it?

    • yikesawjeez: icymi refers to an article on LessWrong titled "Getting 50 Sota on Arc AGI With GPT-4o." The user shared a link for further reading.

  • What is the purpose of holding a mech interp speedrun in July?

    • Kabir Kumar: The user mentioned organizing an event called "mech interp speedrun" but did not provide details on its purpose or how it relates to the chat's context.

  • How does the transformer memorize input data, and is this a concern for our project?

    • loveofdoing: The user asked about the potential of a transformer model memorizing input data but did not receive any direct answers in the chat transcript.

  • What are some relevant papers or resources to learn more about the topic discussed?

  • How would George Hotz approach solving the problem at hand?

    • shaw: Shaw mentioned that they asked George Hotz for his opinion, and he suggested using transformers with some kind of reinforcement learning (RL) reward function. However, this is not a confirmed solution but rather an interesting idea shared by George Hotz.

Who Helped Who​

  • yikesawjeez helped Andy Brock with joining a team by expressing interest in participating and offering to keep up to date until their contract ends.
  • Kabir Kumar helped community members interested in mech interpretation speedruns by announcing an event scheduled for July, providing them with information on how they can get involved.
  • Virgile Blais joined the server, contributing to the sense of community and collaboration among its members.
  • Corona_Chan proposed a novel approach using facial video data to predict ECG/PPG wave models, potentially helping others in their research or projects related to this topic by sharing relevant resources and ideas.

Action Items​

Technical Tasks:

  • Form a team and join the server (mentioned by SSamuel, Andy Brock)
  • Keep up to date till contract ends and see where can help out after (yikesawjeez)
  • Holding a mech interp speedrun in July (Kabir Kumar)
  • Apply framework for arc task with hilbert wave curves as inputs/outputs (Corona_Chan, Shaw)
  • Explore Transformers and RL reward function for solving tasks (Shaw)
  • Documentation Needs: None explicitly requested.
  • Feature Requests:
  • Community Tasks:
    • Live hack sessions on weekends (Arch2230, goo14_)