Back to home

Blog

Field notes on AI, LLMs and software architecture. End-to-end projects, with the numbers, the trade-offs and the mistakes.

·50 min read·In FR

I trained an AI model from scratch in a weekend, and it learned chess on its own

51 million parameters, two hours on a gaming GPU, and a single objective: predict the next move. The full write-up — decisions, reasoning, mistakes, and every command in the order I ran it.

LLMPyTorchTransformerGPUÉchecs
Read the article