WEBVTT 00:00:00.000 --> 00:00:03.357 And there's been this pattern where these massive models come out uh they 00:00:03.440 --> 00:00:06.357 get distilled uh by open source model 00:00:06.440 --> 00:00:10.077 makers and then open source is just a 00:00:10.160 --> 00:00:14.077 couple months behind the leading foundational models. Um 00:00:14.160 --> 00:00:17.517 and you know, I guess like when the investment was smaller, being a couple 00:00:17.600 --> 00:00:21.437 months ahead, uh you know, mattered a lot. But I'm curious now that the 00:00:21.520 --> 00:00:24.477 investment is so big um and the models 00:00:24.560 --> 00:00:26.757 are capabilities are increasing you 00:00:26.840 --> 00:00:28.957 know, fairly dramatically you know, uh 00:00:29.040 --> 00:00:32.357 as you go. Um how is this defensible in 00:00:32.440 --> 00:00:36.277 the long term if you're just going to have that pattern repeat over and over? 00:00:36.360 --> 00:00:37.837 Well, I look at it a little differently. 00:00:37.920 --> 00:00:42.357 Like I think that the real investment that we are making is into 00:00:42.440 --> 00:00:46.717 that end-to-end co-design, right? Of having a system a system of people, 00:00:46.800 --> 00:00:49.957 right? Who are producing this technology, right? A way of working 00:00:50.040 --> 00:00:53.277 together and some of this is about how you leverage these massive 00:00:53.360 --> 00:00:55.837 supercomputers to produce these models. 00:00:55.920 --> 00:01:00.797 Now, it is also the case that it's not as simple as you can take the output of 00:01:00.880 --> 00:01:04.717 these models and distill and you have exactly the model of the same capability. 00:01:04.800 --> 00:01:07.797 It's just smaller and can run fast. If that were the case, we would just do 00:01:07.880 --> 00:01:11.837 that and then we would also have a model that would be, you know, much more uh 00:01:11.920 --> 00:01:15.517 easy to serve in in many ways. And of course, there's a lot of art behind 00:01:15.600 --> 00:01:17.837 distillation. There's a lot of great things there. But the point that I'm 00:01:17.920 --> 00:01:21.757 getting at is that the real thing that we are investing in is the machine that 00:01:21.840 --> 00:01:26.197 makes the machine. Now, the at the deployment side, we think a lot about 00:01:26.280 --> 00:01:30.077 safeguards. We think a lot about mitigations and we do that for many many 00:01:30.160 --> 00:01:32.157 different aspects of how these models 00:01:32.240 --> 00:01:37.317 could be misused um in real situations. And that's something 00:01:37.400 --> 00:01:40.637 that we have been investing in for many years and we think about that across 00:01:40.720 --> 00:01:42.477 areas like cyber or thinking about that 00:01:42.560 --> 00:01:48.397 areas like bio that we have a I a long-standing effort that you can see in 00:01:48.480 --> 00:01:52.157 our preparedness framework, which is public, about how we approach these 00:01:52.240 --> 00:01:55.637 kinds of uses of the model and how we try to make maximize the benefits, 00:01:55.720 --> 00:01:59.877 mitigate the risks. And so I think it's a real motion that every piece of what 00:01:59.960 --> 00:02:03.117 we do needs to connect to the question of how do we 00:02:03.200 --> 00:02:06.357 continue to make progress, but also how do we make these models broadly 00:02:06.440 --> 00:02:08.917 available? Because that's something that we really believe in, that we believe 00:02:09.000 --> 00:02:13.037 this technology empowers people, and that we want it to benefit people and 00:02:13.120 --> 00:02:17.237 lift everyone up. Yeah, but just to go back on that, um 00:02:17.320 --> 00:02:19.037 the pricing on this model is, I think, 00:02:19.120 --> 00:02:23.277 double the last model, GPT-5.4. 00:02:23.360 --> 00:02:27.277 Um and so, from an economics or a business standpoint, the question would 00:02:27.360 --> 00:02:30.957 be, you know, let's say you keep on progressing, but because there's been 00:02:31.040 --> 00:02:33.757 all this infra- infrastructure that's been put towards 00:02:33.840 --> 00:02:36.597 training the models, if open source uh 00:02:36.680 --> 00:02:38.637 can deliver not as good performance, but 00:02:38.720 --> 00:02:42.317 almost as good, uh and do it cheaper, um 00:02:42.400 --> 00:02:44.357 how do you handle that threat? 00:02:44.440 --> 00:02:47.637 Well, again, I look at it a little differently. So, first of all, if you 00:02:47.720 --> 00:02:51.397 look at our history, which really is not driven by anything 00:02:51.480 --> 00:02:54.797 in competition, it's just like our our own sort of 00:02:54.880 --> 00:02:57.557 progress and and desire, we have dropped 00:02:57.640 --> 00:03:02.277 prices on the same level of intelligence year over year, sometimes by literally a 00:03:02.360 --> 00:03:05.237 factor of a hundred. Right? It's like, at least an order of magnitude year over 00:03:05.320 --> 00:03:06.997 year, sometimes literally a hundred. 00:03:07.080 --> 00:03:09.517 But, the thing that keeps happening, it's real Jevons paradox, where it's 00:03:09.600 --> 00:03:12.077 like you lower the cost of something, 00:03:12.160 --> 00:03:14.877 way more activity happens, right? And I 00:03:14.960 --> 00:03:19.517 think that what we keep seeing is that there are returns to intelligence, 00:03:19.600 --> 00:03:23.477 right? That for the kinds of tasks that these models are now capable of doing, 00:03:23.560 --> 00:03:26.997 that a little bit more intelligence goes a long way. And I think that is the 00:03:27.080 --> 00:03:31.677 story of 5.5, that in some ways, you can almost look at it as like, "Oh, there's 00:03:31.760 --> 00:03:34.197 just an incremental improvement in intelligence." But, I think there's 00:03:34.280 --> 00:03:37.917 going to be a massive improvement in terms of what people use it for. And by 00:03:38.000 --> 00:03:40.997 the way, I actually think that incremental is actually very much an 00:03:41.080 --> 00:03:44.757 understatement for this model relative to 5.5. You know, it's a point one 00:03:44.840 --> 00:03:48.837 improvement in some ways, um but I think that I that that actually really 00:03:48.920 --> 00:03:52.717 undersells the magic that we see within this model, and that the that our early 00:03:52.800 --> 00:03:57.320 testers have have really [music] seen in their practical work.