I’m really curious what people’s theories are on why openai released this and not o3?
My old main theory was that they would have to charge so much for o3 that it would create bad PR, but this is now much less likely.
My first remaining guess is that they don’t want competitors extracting full o3 reasoning traces to train on. I guess its also possible that o3 is just dangerous. On the other side of capabilities, its technically possible that o3 is benchmark gamed so hard that its outputs are not usable.
My model was just that o3 was undergoing safety evals still, and quite plausibly running into some issues with the preparedness framework. My model of OpenAI Preparedness (epistemic status: anecdata+vibes) is that they are not Prepared for the hard things as we scale to ASI, but they are relatively competent at implementing the preparedness framework and slowing down releases if there are issues. It seems intuitively plausible that it’s possible to badly jailbreak o3 into doing dangerous things in the “high” risk category.
AFAIK the only info about the (lack of) release of o3 comes from this tweet—and it does seem like o3 will be available explicitly via the API. I think it’s what you say—pricing causing bad PR. It seems that even for pro users, if you use o1 pro too much they will silently switch you to using o3 mini without telling you even if you select o1 pro. So I think they want to continue doing this while not looking so bad doing it.
I’m really curious what people’s theories are on why openai released this and not o3?
My old main theory was that they would have to charge so much for o3 that it would create bad PR, but this is now much less likely.
My first remaining guess is that they don’t want competitors extracting full o3 reasoning traces to train on. I guess its also possible that o3 is just dangerous. On the other side of capabilities, its technically possible that o3 is benchmark gamed so hard that its outputs are not usable.
My model was just that o3 was undergoing safety evals still, and quite plausibly running into some issues with the preparedness framework. My model of OpenAI Preparedness (epistemic status: anecdata+vibes) is that they are not Prepared for the hard things as we scale to ASI, but they are relatively competent at implementing the preparedness framework and slowing down releases if there are issues. It seems intuitively plausible that it’s possible to badly jailbreak o3 into doing dangerous things in the “high” risk category.
AFAIK the only info about the (lack of) release of o3 comes from this tweet—and it does seem like o3 will be available explicitly via the API.
I think it’s what you say—pricing causing bad PR. It seems that even for pro users, if you use o1 pro too much they will silently switch you to using o3 mini without telling you even if you select o1 pro.
So I think they want to continue doing this while not looking so bad doing it.