GPT-2 and the Too Dangerous to Release Moment

In February 2019, OpenAI announced GPT-2, a 1.5 billion parameter Transformer trained on about 8 million web pages. Its coherent, occasionally eerie prose was a step change from GPT-1 a few months earlier.

OpenAI did something unusual: it declined to release the full model, citing concerns about misuse for automated disinformation, spam, and impersonation. It released a smaller 124 million parameter version, then a 355 million, and eventually the full weights by November 2019.

What GPT-2 could do

• Finish a prompt in plausible prose across many styles
• Perform unseen tasks zero-shot if phrased as text continuation
• Summarize, translate, and answer questions without task-specific training
• Produce confident nonsense and hallucinated facts in equal measure

GPT-2 also sketched what became OpenAI's scaling thesis: larger models, trained on more data, acquire unexpected abilities. The company doubled down, and GPT-3 followed eighteen months later.

“Due to concerns about malicious applications, we are not releasing the trained model.”

The big idea: GPT-2 showed that language models were becoming powerful enough to raise real deployment questions. The tradeoff between openness and caution remains an open problem.