OpenAI

Contribution to AI

Scaling was a hunch before it was arithmetic. The 2020 paper on scaling laws made it the latter, showing loss falling as a smooth power law in parameters, data and compute, so the return on a training run could be estimated before anyone spent the money on it. GPT-3 was the demonstration: tasks handled from a few examples placed in the prompt rather than from a fine-tuning set, which weakened the assumption that every new task needed its own labelled corpus and its own model. Shipping it through a paid API instead of a weights download established model-as-a-service as the industry's default commercial shape, and made prompt engineering a job. InstructGPT mattered more than its size suggested — raters preferred a 1.3-billion-parameter model tuned on human feedback to the 175-billion base model, which split capability from usability and turned alignment research into product engineering. ChatGPT, released on 30 November 2022, was a thin interface over that result and found an audience estimated in the hundreds of millions within months, fixing the chat box as the way ordinary people meet a language model and redirecting capital and legislative attention worldwide. Earlier work is quieter but still in use: Gym gave reinforcement learning a shared environment interface in 2016, PPO became the default policy-gradient method, and the staged release of GPT-2 opened the still-unsettled argument over publication norms.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).