Problem brief:

ARC AGI (Abstract reasoning corpus- Artificial general intelligence) is a benchmark designed to test fluid intelligence of AI models. It consists of puzzles easy for humans to solve but hard for AI models. Puzzles which even state-of-the-art models fail to solve, and if they do so — they burn millions of tokens on one puzzle.

In this project:

We will mechanistically interpret and advance small models built on elegant approaches for ARC AGI 1 and 2.

What to expect:

• A learning curve as steep as you want it to be- brainstorming ideas, experiments, and theory with an extremely passionate group of people

• Thinking foundationally: The project is yours, and the trajectories and ideas you work on are to your freedom

• Reading a lot of research papers from AI frontiers, connecting ideas cross domains and obviously, having a lot of fun along the way :)