- ARC-AGI Material Link List
- Brainstorming/Ideas (Dec 2024)
- tgi-eval - using ARC-AGI-1 to evaluate LLMs
- Annotated RE-ARC
2024-12-10 andreaskoepf
Welcome fellow AI enthusiast!
You have probably heard about François Chollet's ARC-Price competition. The 2024 ARC-Price is now over and it is time to prepare for next year's ARC-AGI-2. I therefore started the OpenThough ARC-AGI team.
Our mission is to develop a cognitive core to solve ARC-AGI-2 🙂 - and to do this as much as possible open-source.
My preliminary plan is to follow a transformer based transduction approach, which involves:
- generating high quality synthetic data (like my riddle-generator+verifier pairs: https://x.com/neurosp1ke/status/1854796093801775215 )
- fine-tuning best open-weight LLM/VLMs
- training multi-turn reasoning (meta-cognition, planning, process-feedback-models, reflection, correction/back-tracking, etc.)
- applying test-time fine-tuning
Any help building the team and project is of course welcome. I will try to break down the project into many small work chunks which can be developed open-source (also to help other teams). The idea is to allow collaboration even without long-term commitment.
If you have feedback or want to chat about reasoning models, self-improvement, test-time fine-tuning or ARC-AGI in general, don't hesitate to contact me in the #arc-agi-2 channel on GPU mode.