GlassBox

Inspiration

ARC-AGI benchmark puts agents in unknown situations and measures intelligence by how well an agent is able to adapt. Additionally, in many AI safety scenarios, an agents knows little about the environment and has to figure out rules in this world. Often these rules learnt by an agent are implicit in the neural network weights. The idea behind GlassBox is for the agent to externalize these rules as code. This is possible now because models are very good at generating code.

What it does

Given any (tile-based) game, an agent plays the game and writes code to reverse engineers the game. Essentially, one can clone a game by just playing it. The demo can do this for any tile game based game the user has created at run time

How we built it

The idea is that an agent forms a hypothesis of the rules of the system and then generates code for this game. It then iteratively refines this hypothesis based on code and tests it generates while playing the game. It converges once most tests agree for a user set amount of time.

Challenges we ran into

The initial goal was to to do this for Pokemon Red but taking in actions and state via vision is very hard and so had to narrow down to games which have well defined actions and state.

Accomplishments that we're proud of

Read some papers in this area and found a cool application for this counterexample based agent improvement idea. Implemented start to finish in a couple of hours.

What we learned

hard to extend to perception out of the box. Need better ideas.

What's next for GlassBox

Extend to more complex environments - the plan is to do this for APIs

Built With

• python