The GPT-3 moment

In 2020, OpenAI released GPT-3 — at the time, the largest language model ever trained. It had 175 billion parameters.

The reactions were immediate and polarizing. Researchers and writers who got early access reported that it could write convincing prose, answer obscure questions, translate languages, write code, and mimic the style of specific authors — all from a short prompt, with no fine-tuning.

It could also be confidently wrong in bizarre ways. It would fabricate citations to papers that didn't exist. It would contradict itself across paragraphs. It would answer questions about current events with complete authority based on training data that was already out of date.

GPT-3 was the first time the general public encountered a model that felt genuinely surprising — not a narrow task-specific tool, but something with a broad, unpredictable range. It forced a reckoning with what "understanding language" might actually mean for a machine.

ChatGPT, released in 2022 on a similar architecture, brought this to hundreds of millions of people. The AI moment had arrived.