Open letter to OpenAI
TL;DR: I think it is safer to release advanced AI technology early, so that humanity will have time to explore it and adapt to it.
Dear OpenAI,
Thank you for all the amazing work you’re doing in fundamental AI research and in safety. The quality of GPT-2 texts is astonishing and significantly changed my estimates of AGI timeline. I’m writing this letter to provide my opinion on your current approach to AI safety.
It seems to me that we are currently following the scenario of the slow takeoff. That is, multiple actors posess roughly the same ability to implement SOTA AI technology. Since you used only publicly available data, it is very likely that your results (any results, in fact) could be replicated by a sufficiently well-funded corporate or state actor. To keep this from happening you would have to not publish any results at all to not direct the attention of potentially malicious actors (and researchers/engineers close to them) would be kept elsewhere. Even in this scenario, I would guess the information will leak one way or the other.
So, the real question is this: is it beneficial to have the latest AI/ML technology in an easy-to-use form available to the wide public? In my opinion yes, it is. True, this technology could and will be used for malicious purposes. But before it will cause too much harm, they will also be used in a variety of apps, pranks, and interactive demos, and people will get used to it and will have time to adapt. In this sense, Fake News and DeepFake phenomena were net-positive for humanity: most people are aware of it now and will not be fooled by these kinds of things.
Same goes for practically any kind of dangerous AI/ML research and tech: not releasing it does not really change the probability that malicious actors will be able to replicate and use it. On the other hand, it is less likely that their victims will be able to adapt by building technological, political or psychological defenses.
Hope you will consider my opinion,
Best regards,
Matvey Ezhov,
Co-founder and CTO at Diagnocat
