OpenAI plans staged releases for cyber-capable models

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 1:04:08

   Yeah, there's a million questions I want to ask with the time we have left. But one of them, we touched on this, like making sure a bunch of models are in the hands of a bunch of people, and they have that powerful technology. But we've reached a point where now some of these models are too powerful, we're being told, and they're gated for only certain companies. Claude and Mythos have really made a lot of headlines and I think a lot more fear.

2. 1:04:32

   And I'm wondering what you guys think of this new moment. Are we getting to a point where we need to start keeping these powerful models behind the scenes rather than in the hands of everybody? There are people in the world who for a long time have wanted to keep AI in the hands of a smaller group of people. You can justify that in a lot of different ways. And some of it's real. There are going to be legitimate safety concerns.

3. 1:04:56

   um but i expect but if what you want is like we need control of ai just us because we're the trustworthy people i think the the fear-based marketing is probably the most effective way to justify that um that doesn't mean it's not legitimate in some cases uh but it is you know clearly incredible marketing to say we had built a bomb we're about to drop it on your head we will sell you a bomb shelter for a hundred million dollars you need it to like run across all your stuff, but only if we pick you as a customer. The way that we view balancing these new capabilities that are going to come with still our belief that the world needs to get and use and understand and come up with new ideas for this technology is not always easy. We have had cybersecurity in our preparedness framework for a long time, and we have been building mitigations to figure out

4. 1:05:48

   how we release this, how we put these models in the hands of a trusted access program, how we then put more capable models in the hands of everybody. But there will be a lot more rhetoric about models that are too dangerous to release. There will also be very dangerous models that will have to be released in different ways. But to the point Greg was making about the goal here is to benefit everybody and also to i don't want to say market this in a way but but like get the world to come along on this journey with us where where it's like we are going to give you more powerful technology there's going to be responsibility that goes along with that we are going to help set up the world for success as much as we can um but we will try to avoid the fear-based marketing as much as we can. So I'll just ask you directly, do you think Mythos is

5. 1:06:40

   just a lot of marketing? No, I'm sure it's a great cybersecurity. We've been talking about this for a long time. This has been in our model, but there's a way of saying, like our version of this is to say, these models are going to get much better at cyber. We have this preparedness framework category. Here is our plan for how we deploy this into the world. Here's how this trusted access program looks like. Here's the mitigations we put on the models. I think anthropic has not had it as one of the categories in their preparedness framework uh so i'm sure mythos is a great model uh for cyber security but i think we have a plan we feel good about for how we put this kind of capability out into the world when the anthropic stuff was going down with the