Athon Millane

What the Future Teaches

Athon Millane, 2026

Reinforcement learning on language models has done best where an answer can be checked straight away. Consequential decisions tend to resolve slowly, and when one does, a model trained on the outcome gets one bit: what happened, luck included. Hindsight knows much more, from what could have been known at the time to which evidence turned out to matter. This essay asks how much of that a model can learn. The follow-up, Step by Step, looks at whether hindsight can grade the reasoning as well as the answer. Both include the parts that didn't work.

Read it

To read both essays, email me for the PIN.