ChatGPTChatGPT

On November 30, 2022, OpenAI released ChatGPT with modest expectations, describing it internally as a low-key research preview. The underlying technology was not new: it was a conversational fine-tune of GPT-3.5, a model family whose ancestor GPT-3 had been available to developers for over two years.

What was new was the packaging. A free chat box on a web page removed every barrier between an ordinary person and a large language model. Within five days ChatGPT had a million users, and analysts at UBS estimated it reached about 100 million monthly users within two months, making it the fastest-growing consumer application in history at the time.

The technical ingredient that made conversation feel natural was reinforcement learning from human feedback, or RLHF. Human labelers ranked candidate responses, a reward model learned those preferences, and the language model was tuned against it, the recipe OpenAI had published earlier in 2022 with InstructGPT. The result followed instructions, declined harmful requests, and admitted mistakes far more gracefully than raw GPT-3.

The reaction rippled outward fast. Google reportedly declared an internal 'code red', Microsoft deepened its multibillion-dollar OpenAI partnership and pushed AI into Bing and Office, and virtually every major technology company reorganized around generative AI within a year.

Schools scrambled over essays, programmers adopted it as a daily tool, and 'prompt' entered everyday vocabulary. Hallucination, the model's tendency to state falsehoods fluently, went from research jargon to dinner-table topic.

ChatGPT did not mark a leap in machine intelligence so much as a leap in access. It compressed years of gradual AI progress into a single moment of public awareness, and the industry has been sprinting to keep up with the expectations it set ever since.