Uncapped with Jack Altman
00:44:34 · Jack Altman, Ev Randle
[00:44:34] S0: Mhmm. [00:44:35] S1: So an obliterated model is an open weight model that has its internal sort of safety guards removed and... To help help with its work. Because, you know, ultimately, like, when the Hugging Face incident happened, they had some complaints where they're like, we had to use an open source model because we get the Frontier Labs to turn off the safety guards in time to help us with the hack. And so obliterated models are literally open weight models that have no safety guards. So it's very scary, and I think one of these is actually what some groups in China have been using I think it's called RTX have been using to commit some of these hacks globally. It's also an incredible defense tool for enterprises. Because you take an obliterated model, you do what that CEO told us, and you just say, hack us. Like, hack us to death. But obviously, you're the one using it, and so nothing actually happens from it. And that is, I think, going to be an incredible tool. You need these models that will because guess what? The adversaries have obliterated models. The adversaries are not working with safety controls. You need to see the tools that they're using and use it on your own infrastructure. And I feel like we're going to hear a lot more about this idea of obliterated models.
00:45:46 · Jack Altman, Ev Randle
[00:45:46] S0: By the way, I think the tech world, which I think is mostly a good thing, at first, I would say reacted negatively to pacing the frontier type stuff. And there's much more of just a let it rip mindset, which is very libertarian, tech forward good. I think that's basically a good thing. But this security dynamic is an instance where it not going too fast would be very helpful because you actually do want the defenses to penetrate before we have four generations forward of models. That's right. Because you do need things to get a little time to harden against this stuff. [00:46:21] S1: It is, but this is the crazy thing is how long do we think it's going to take to harden all this stuff? [00:46:27] S0: It's going take a [00:46:27] S1: long time. Think about I don't know, think about a hospital system in West Virginia. How long is it going to take for them to because I think if you talk to some of these cybersecurity professionals, they're like, we basically have to re architect. This is the biggest code modernization project in all of history. [00:46:43] S0: That's what we need cognition for. [00:46:44] S1: That's right. That's what we need. That's what we need cognition for. [00:46:48] S0: But I I think like I was going [00:46:50] S1: to try to do an ad read there. Yeah. Cognition. Update your Cognition. Update your code. And security.
00:46:56 · Jack Altman
[00:46:56] S0: That's good. But it's going to take a long time. But the hope I have is that, like we were talking about in the cloud era, it's like people bought Workday and Salesforce on these really elongated timelines. And they were slow. And if you were talking to a government or a hospital or a law firm, it was like thirty six month cycles on some of this stuff. That's not really how it's going right now on the purchasing of software side. I think people are pretty alert. So the hope would be that people see the security stuff for real, people are making it a top priority, And everybody reacts a lot faster than you think. I still don't think that that means it happens in three months. But maybe in twelve months, people could do a lot. That'd be amazing. I think people see the urgency. Yeah, Hopefully. And maybe as a as a last topic, the... Think what's cool is the bio thing is interesting to me because I've kind of been of the view... I still am of the view that all the software improvements that we're making are awesome, but there is a ceiling to how much that can improve the world. And if you touch some grass and you get out of San Francisco and you're like, what will actually make the world a better place? It's not that we're shipping software way faster and way more easily. It's things like people have better housing, or people live longer, or people live healthier, or people have food, or people have transportation. It's physical things. I just think there's a limit to what you can get when it's all in the screen. And bio, to me, would be one of the very best possible things if we're discovering drug targets and finding amazing new drugs quickly or at all. That would be... What are laughing about? So anyway So you
00:48:48 · Ev Randle, Jack Altman
[00:48:48] S1: were on a good tirade. [00:48:48] S0: I was on a good tirade. So anyway... But I think that it is both really promising, and it seems like it's on the horizon. Same as robotics, by the way. [00:48:55] S1: Yeah. Similar. Those are [00:48:57] S0: the two to me where you see the labs both talking about it, building teams around it, purchasing huge amount of data around it. They're writing about it publicly. And so I think that that to me is likely an indicator that it's there. I mean, I haven't yet seen the labs be wrong when they say something's coming. [00:49:16] S1: That's [00:49:17] S0: right. Getting the timeline right is really difficult. With self driving, took Now it's fully here, but it took a while. And self driving, Tesla's obviously dominant. But a lot of that was because of just infinite data. And once you get enough data at these things, maybe it just works. [00:49:32] S1: I think that's... And that's the... There's this really weird sort of, like, inverse relationship between... I completely agree. I feel like you you you... Do you want, like, the... Do you want the rest of America who doesn't have, like, an amazing, I think, view on AI? Or at least a lot of people don't throughout the world and in America don't have an amazing view on AI. Do you want to change perception of AI, solve disease, save people's lives, help people live longer, build more housing so housing prices are lower, create physical abundance? [00:50:01] S0: That's what makes a better world.
00:50:02 · Ev Randle
[00:50:02] S1: That's what makes a better And the weird thing, though, is that there's almost this inverse relationship with those things, the things that when you actually touch the physical world, could actually make the physical world a better place, and the things that have been able to get better very quickly from AI. AI just tends to improve fastest when it can generate large amounts of trustworthy feedback cheaply. I think it's a good way to abstract it. So if you look at code, it's this incredible area for this because it's extremely verifiable. Can do a unit test. You know if it works or not. You can scale up experiments extremely cheaply. They're infinitely scalable. It's all in the digital space. And so when you think about Claude Code or how Cursor improved its models very quickly or even ChatGPT when they started getting feedback from users, all of these were in a digital realm where there was an API and customers could come on at the click of a button and start basically contributing into the leviathan of training data for these labs. And in the real world, robotics and now bio, the tough thing is that you have real bottlenecks to that feedback loop. In robotics, obviously, robots are expensive. They're expensive to deploy, expensive to maintain. There isn't a big customer base that can press a button. You could order a robot, and you could start doing these things, and Py has an API now, and you could do the Sunday beta program or any of these things. But it isn't like ChatGPT, where you can get 100,000,000 people in a few months via a website. And so that's just very different. Bio, you also have natural limiters. The equivalent of running a unit test in biology is a physical experiment. And and even that experiment might only answer a small part of the question you actually care
00:51:43 · Jack Altman, Ev Randle
[00:51:43] S0: And a lot things do just take time. [00:51:44] S1: They just take time. It's like, I don't know, a physical experiment. Even if you speed it up making your own virtual [00:51:50] S0: You have to parallelize. And I think so in bio, it's like you can imagine a virtual or like a real lab, but where you're kind of run somewhat autonomously or at massive scale, where you're just doing a ridiculous number of thoughtful trials at once. [00:52:06] S1: 100%. [00:52:06] S0: I think that could be time will hold you back, but the paralyzing will still speed things up quite a lot, I think. [00:52:13] S1: Yeah. But I think this and this is I know you have an investment that you recently did in this general space for robotics. We've been very interested in things happening in bio as well. The infrastructure to we've obviously seen it with the AI labs as well. Mercor as a business has grown to an unbelievable amount of scale, [00:52:31] S0: because there's there's just no limit Very unintuitively, by the way. Everybody thought that the data thing would run out, they would buy the data, and then they bought the data, and now they can go And the movie's gone. Just turned out to be wrong.