Ankit Singh

Curious about how language models learn and behave. Building, experimenting, and trying to understand what happens inside.

[Ankit Singh @ankitsingh435]

I built Lichess-GPT from scratch to study whether next-move prediction can produce legal chess.

Scaling improved prediction, but not autonomy.

Final model:
• Median first illegal: 4
• Survival@10: 0%

I’ve frozen scaling and started investigating why.

https://t.co/oFujszeFOP