OpenAI published GPT-2 in February. They also published a blog post explaining why they weren't releasing the full model. The stated reason was concern about misuse: the model generates sufficiently coherent text that releasing it could enable disinformation at scale.
This is the first time a major AI lab has declined to release a model on safety grounds. It's worth sitting with that for a moment.
What they released
The 117M parameter model. A small version of the full thing, which has 1.5 billion parameters. The small version generates plausible text. It's clearly not GPT-2's full capability. You can tell when you run it that you're seeing a constrained demonstration.
I've been running it on pitch transcript data, partly out of curiosity and partly because I wanted to understand what it can actually do. Short answer: it completes sentences with impressive coherence and falls apart on longer continuations. The text stays grammatical. The meaning drifts. By the third paragraph of a generated continuation, the model has wandered away from the original topic while maintaining the superficial appearance of staying on it.
What it means for the field
GPT-2 is a language model trained on a large corpus of internet text with no fine-tuning for any specific task. The generation quality is better than anything I've seen from a general-purpose model. BERT is a better fit for classification and extraction tasks, but GPT-2 is showing something different: that scale alone, applied to next-token prediction, produces capabilities nobody specifically trained for.
The practical question is what this means for applied NLP work. The answer in the short term is: not much changes. The 117M model isn't good enough to replace the specific pipelines I've built. The full model might be, for some tasks. We won't know until it's released.
The more interesting question is whether this is a step change or a point on a curve. If you extrapolate the GPT series, the next model will be larger and better. If the pattern holds, at some point the generation quality crosses a threshold where it's genuinely useful for production tasks, not just impressive demonstrations.
I don't know where that threshold is. I don't think anyone does.
The safety argument
I'm genuinely uncertain about OpenAI's decision. On one hand, the concern about misuse is legitimate. Good enough text generation lowers the cost of producing convincing disinformation. That's a real risk and acknowledging it publicly is reasonable.
On the other hand, the decision to withhold is unusual in a field that has operated with relatively open publication norms. The 117M model generates text that other groups could have produced. It's not obvious that withholding the larger model actually prevents capable actors from building something comparable.
What the decision does do is signal that the lab takes the capability trajectory seriously. That signal matters independent of whether the specific safety action is correct.
What I'm watching
Whether the staged release plan holds. OpenAI said they'd release larger versions gradually while monitoring for misuse. If that happens, we'll have a clearer picture of what the full model can do by year end.
Whether other labs respond. Google has large language models. Researchers at universities have been scaling similar architectures. The release decision is OpenAI's to make. The underlying capability isn't theirs to contain.
Whatever happens next, something shifted in February. The question of what these systems can do when scaled up became a lot less theoretical.
With gusto, Fatih.