My cynical impression is that OpenAI figured out a way to get tons of free PR, with the ultimate “humblebrag” of all time. I don’t think they honestly believe that what they’re doing has a material impact, beyond media coverage and people talking about it as we are here. This situation cries out for Occam’s Razor, and when you cut away the fluff it looks and smells like clever marketing, not in the least because it’s worked.
Hi! OpenAI employee and paper co-author here, speaking entirely on my own (nobody else in the company knows I'm posting this). Occam's razor is a great principle, and I completely see how it might point towards your viewpoint. But I'd like to make a few points which will hopefully correct some misconceptions:
1. When we made the decision to do partial release, the talking points were about dangers and benefits of release - reputational benefits was one of the four main benefits we listed and not talked about much. Everyone in the room, including mostly technical non-comms folk, agreed that caution was a reasonable decision, given our uncertainty about malicious use.
2. I was at every meeting we had with the press, prior to release. We had written and rehearsed our major talking points beforehand (and stuck to them). Here they are: what we did technically, how it's different/new, how it's still limited; generality and zero-shot behavior of the model; how the trends of scaling are steady; the realism of the synthetic text, and how we're potentially on a trajectory similar to the synthetic images (we specifically compared GPT-2 to Alec's 2015 work on DCGAN, not the photorealistic stuff we see now); malicious use cases and policy implications; framing our partial release as an experiment and a discussion-starter. Ultimately, we didn't have fine-grained control over the actual articles written.
3. My impression is our samples are a qualitative jump from anything publicly known prior to our release. Most people (whether academics or journalists) who have seen our demo come away impressed; a small number had more muted responses. I would be pretty surprised if there weren't use cases this immediately enabled that previous language models didn't; whether they are more economically-valuable than malicious, I'm unsure (although I have a guess).
4. I personally don't have any evidence that the people I worked with on the release aren't well-intentioned, aren't thoughtful, or aren't smart.
Stepping back a bit, there are definitely aspects of what we did that I already regret, and perhaps more aspects I will regret in the future. I wish we'd had more time to think through every aspect of release, but opening up discussion to the community doesn't seem like a terrible outcome, regardless of how well thought out our decisions were.
