How ChatGPT works

ChatGPT(Product), built by OpenAI(Company), is powered by a large language model in the GPT family — currently GPT-5(Model).

Pre-training

First the model is trained to predict the next token on a large text corpus. This is where it learns grammar, facts, and code.

Then it is fine-tuned with human feedback — a form of Reinforcement Learning(Concept) — so its answers become more helpful and less harmful.

At inference time, requests are routed to fleets of GPU(Hardware)s running CUDA(Platform) kernels.

Entities mentioned