GPT-3

OpenAI's 2020 language model — the first AI system to demonstrate convincingly that scaling transformer models produces qualitative leaps in capability, triggering the modern AI race.

The scaling revelation: GPT-3 (175B parameters) demonstrated that language models at sufficient scale exhibit few-shot and zero-shot learning — performing tasks they weren't explicitly trained for, using only a few examples in the prompt. This was not predicted by the architecture alone; it appeared to emerge from scale. The paper 'Language Models are Few-Shot Learners' reframed AI development: the path to general capability might be through scale and data rather than explicit programming of reasoning. Triggered the Anthropic founding (safety concerns about GPT-3 scale), the foundation model industry, and the race to GPT-4, Claude, Gemini.