In 2019, OpenAI said GPT-2 was too dangerous to release. The thread argues both sides were right

The artifact being discussed

The retrospective, written in December 2022 just as ChatGPT launched, recounts OpenAI's February 2019 decision to withhold the full 1.5-billion-parameter GPT-2 model — "Due to our concerns about malicious applications of the technology, we are not releasing the trained model" — and to do a staged "responsible disclosure" release instead. It summarizes OpenAI's nine-month follow-up: humans found the output convincing, the model could be fine-tuned for harm, generated text was detectable at roughly 95%, but there was no strong evidence of real-world misuse. The full model, 48 decoder blocks, was finally released with code and weights on November 5, 2019.

Vindicated, or the original hype move?

The thread split almost down the middle. One side reads 2019's caution as prescient given what synthetic content did to the information ecosystem:

"In hindsight, they were entirely correct. The social damage caused by low cost content generation that's hard to distinguish from human authorship is astronomical." — a commenter arguing the caution was justified

The other side reads the same history as the template for years of self-flattering danger marketing:

"Seven years of this insufferable brand of 'Oh it's so dangerous, I sure hope no one gives us a ton of money and takes us seriously' marketing and people are still falling for it at scale." — a commenter on the danger-as-marketing pattern

The synthesis from someone who built on GPT-2

The most cited middle position separated the 2019 context from the 2026 reflex.

"Back in 2019, it was more fair to have caution around the larger GPT-2 models since robust text generation … was a complete unknown. For something like Mythos in 2026 … it's more fair to call it … a marketing gimmick." — Max Woolf, a senior data scientist and creator of gpt-2-simple

The historical-pattern crowd had the most fun, dredging up a precedent that makes the whole "too dangerous" genre feel less novel:

"Same vibes: In 2000 Sony 'declared that the company's PlayStation2 has been hit with export restriction because it could be used for military purposes.'" — the author of a graphics library, citing Variety

One detail aged strangely: in 2019, nobody in the original GPT-2 discussion mentioned code generation at all — the feared harms were spam and misinformation, not the security-vulnerability worries that dominate the Mythos discourse today (on what the 2019 fears actually were). Both readings turn out to be true at once: the caution was reasonable when the capability was unknown, and the same script became a gimmick once the unknown became a product.