joe-esquibel 2 hours ago

What type of problems is this well suited for? Ive found that I can just linearly improve on one idea and get further to my goal. Like in 10 directed back and forths. I feel like this would create x options for every step that need to be independently evaluated with some Oracle and then again for the next steps, I feel like tokens are too expensive to random walk an llm through a solution space.

Id like to see a validation of taking a small AI and then have to chug through a tough problem like a chess engine to a specific standard and compare that to just one shot it with fable.

Maybe we could use cheap poor thinking llm models with this harness to solve problems we don't think they can currently solve.

Bataino 3 hours ago

Interesting...