Three Approaches to Solving ARC AGI
ARC AGI is a set of challenges that are straightforward for humans but difficult for LLMs - particularly the latest v2, for which there is a substantial prize outstanding.
I cover the three main techniques from past years of competitions, including:
Domain Specific Language (icecuber approach)
LLM generated programs (Ryan Greenblatt approach)
Test time training (MindsAI / Tufa Labs / Jack Cole / Mohamed Osman and 2024 winners Jan Disselhoff and Daniel Franzen - as covered on MLStreetTalk )
I explain things at a high level and review detailed scripts, which I have made open source.
Scripts are open source in the TrelisResearch/minimal-arc Github repo!
And Trelis Research is submitting a team to the 2025 competition, apply here:
Cheers, Ronan
Video Links:
Other Trelis Links:
🤝 Are you a talented developer? Work for Trelis
💡 Need Technical or Market Assistance? Book a Consult Here
💸 Starting a New Project/Venture? Apply for a Trelis Grant