// experiments

experiments

Weekly experiments at the intersection of dual-process LLMs, causal RL, and game-playing agents. Code open-sourced with each ship.

apr 29
I gave a 0.5B model the option to phone a 3B friend
0.5B Qwen drafts. 3B Qwen critiques and revises. Four gate variants, 340 items across GSM8K, MMLU, MT-Bench. Three findings, one paper, code public on ship day.
experiment-001dual-process