WEBVTT - AI’s Biggest Players Debate a Frontier Slowdown 0:00:02.529 --> 0:00:04.210 Bloomberg Audio Studios. 0:00:04.670 --> 0:00:05.370 Podcasts. 0:00:05.690 --> 0:00:06.210 Radio. 0:00:06.650 --> 0:00:07.050 News. 0:00:13.630 --> 0:00:17.150 Bloomberg Tech is live from the heart of Silicon Valley 0:00:17.690 --> 0:00:19.950 with Ed Ludlow in San Francisco. 0:00:22.570 --> 0:00:25.310 This is Bloomberg Tech. AI leaders are calling for a 0:00:25.430 --> 0:00:29.310 slower pace at the frontier. Anthropic CEO Dario Amadei saying 0:00:29.370 --> 0:00:31.950 over the weekend, quote, we must slow the pace at 0:00:31.990 --> 0:00:35.530 which we improve the capabilities of AI models. His argument, 0:00:35.909 --> 0:00:38.989 safety needs time to keep up. OpenAI's Sam Altman and 0:00:39.050 --> 0:00:42.720 SpaceX AI's Elon Musk quickly backing the idea. 0:00:42.820 --> 0:00:44.040 Anthropic's now opening. 0:00:43.820 --> 0:00:48.340 Its doors to independent evaluators, while Altman says OpenAI will 0:00:48.400 --> 0:00:53.070 not go public this year as it prioritizes safety. This 0:00:53.090 --> 0:00:55.610 has prompted a response from President Trump this morning, saying 0:00:55.630 --> 0:00:58.170 in a True Social post that his administration has, quote, 0:00:58.550 --> 0:01:01.710 stopped AI people from doing bad things and they will 0:01:01.730 --> 0:01:05.470 continue to do so. And the debate's hitting markets, NVIDIA, Micron, 0:01:05.550 --> 0:01:08.750 other AI-linked stocks all under pressure as investors ask what 0:01:08.790 --> 0:01:12.140 a slower frontier could mean for the AI spending boom. 0:01:12.230 --> 0:01:16.300 Throughout the hour, what pacing actually means, the industry response, 0:01:16.640 --> 0:01:20.459 and what comes next. From Washington, Bloomberg's AI reporter Rachel 0:01:20.500 --> 0:01:23.080 Metz And tech editor in D.C., Mike Shepard, join us 0:01:23.120 --> 0:01:25.500 for the details. Rachel, let's start with the essay from 0:01:25.560 --> 0:01:28.540 Anthropic CEO Dario Amadei, which was published this Saturday. 0:01:28.640 --> 0:01:29.320 What do we need to know? 0:01:29.920 --> 0:01:32.540 I mean, he's talking a lot about some of the 0:01:32.580 --> 0:01:35.760 wonderful things that AI can lead to, but also how 0:01:35.880 --> 0:01:39.300 recent events such as an inadvertent hack by open AI 0:01:39.360 --> 0:01:43.340 models of hugging face are leading to this moment where 0:01:43.500 --> 0:01:46.820 he believes that some sort of pacing is necessary in 0:01:46.880 --> 0:01:48.380 order to ensure safety of AI. 0:01:49.240 --> 0:01:51.690 Shep, you were on hand all weekend on duty to 0:01:51.730 --> 0:01:55.590 basically bring together all of the responses that came in. First, 0:01:55.650 --> 0:01:57.830 it was Sam Altman, then Elon Musk, and then lots 0:01:57.870 --> 0:02:00.110 of investors and politicians weighed in, too. 0:02:00.390 --> 0:02:01.270 Try and summarize it. 0:02:02.670 --> 0:02:07.610 Well, you know, quickly, we saw the arch rivals of Anthropic, 0:02:08.000 --> 0:02:12.260 Sam Altman from OpenAI, Elon Musk from SpaceX AI, quickly 0:02:12.320 --> 0:02:15.880 embrace in broad terms the Dario Amadei's call for a slowdown. 0:02:16.380 --> 0:02:18.720 When you get into the specifics, it's a little bit 0:02:19.160 --> 0:02:22.280 unclear how closely aligned they all are. 0:02:22.610 --> 0:02:24.320 But on the broad strokes, they back it. 0:02:24.389 --> 0:02:27.550 Investors are a little bit skeptical about how far the 0:02:27.590 --> 0:02:30.989 companies will actually go in slowing this down. And for 0:02:31.030 --> 0:02:34.410 their part, Dario Amadei and Sam Altman are insisting that 0:02:34.970 --> 0:02:38.160 slowing the pace at the frontier does not mean stopping 0:02:38.220 --> 0:02:41.600 spending or slowing down work or growth in their businesses. 0:02:42.840 --> 0:02:46.430 There is no commitment to end training runs or reduce spending. 0:02:46.450 --> 0:02:47.310 I think that's really important. 0:02:47.350 --> 0:02:50.830 Rachel, I think it's worth going back over the timeline 0:02:50.870 --> 0:02:54.109 of the last seven days. When we started this story, 0:02:54.510 --> 0:02:57.790 it was because an anthropic researcher had resigned. And we 0:02:57.830 --> 0:03:00.790 also had reporting that Sam Altman, who's the CEO of OpenAI, 0:03:01.110 --> 0:03:04.130 had already gone company-wide to say, we think we should 0:03:04.169 --> 0:03:06.060 slow the pace, fill the gaps. 0:03:06.800 --> 0:03:07.100 Okay. 0:03:07.680 --> 0:03:09.260 So we had this resignation. 0:03:09.540 --> 0:03:11.300 It seemed to gain ground. 0:03:11.340 --> 0:03:15.040 It gained a lot of attention, not just within AI, 0:03:15.060 --> 0:03:17.660 but also outside of AI. And then Sam Altman, as 0:03:17.700 --> 0:03:21.470 you mentioned, our reporting indicated that he had said that 0:03:21.510 --> 0:03:26.049 the company was interested. um, in pacing, um, and was 0:03:26.070 --> 0:03:29.810 hoping other companies would join as well. And then since then, 0:03:29.930 --> 0:03:33.410 we're just seeing, um, this sort of snowball effect where 0:03:33.470 --> 0:03:38.790 we're getting more and more talking toward this idea of pacing. 0:03:38.910 --> 0:03:41.560 But as you guys both mentioned, it's not. 0:03:41.440 --> 0:03:42.500 Clear what that would mean. 0:03:42.560 --> 0:03:48.510 And it to actually slow down innovation internally, perhaps more externally. 0:03:48.740 --> 0:03:52.580 And also they're talking about using some outside companies as 0:03:52.620 --> 0:03:55.200 well to look over their models and that kind of thing. 0:03:55.440 --> 0:03:58.280 So a lot of unanswered questions still, I would say. 0:03:58.620 --> 0:04:00.340 We've had the reaction from the president this morning. 0:04:00.410 --> 0:04:03.450 Over the weekend, the reaction from Congress, Democratic Senator Chris 0:04:03.490 --> 0:04:05.790 Coons of Delaware speaking to Bloomberg this weekend. 0:04:06.980 --> 0:04:09.640 Now that we've got agreement from Elon Musk and Sam 0:04:09.700 --> 0:04:13.480 Altman and Dario Amadei, we need agreement from Republicans and 0:04:13.540 --> 0:04:18.270 Democrats in Congress to put aside our partisan pre-election bickering 0:04:18.730 --> 0:04:22.130 and put in place some real guardrails before AI goes 0:04:22.230 --> 0:04:22.809 off the rails. 0:04:24.370 --> 0:04:27.600 Shep, what's happening on Capitol Hill right now? Well, there's 0:04:27.620 --> 0:04:28.130 a lot of talk. 0:04:28.180 --> 0:04:31.260 And we heard Chris Coons really lay out his vision 0:04:31.279 --> 0:04:34.980 for trying to get both parties together quickly before the 0:04:35.040 --> 0:04:37.740 election to find some consensus on this. But in a 0:04:37.860 --> 0:04:41.070 bitterly divided House and Senate, it's hard to imagine that 0:04:41.089 --> 0:04:43.250 they would be able to agree on something that's so 0:04:43.650 --> 0:04:47.450 technically difficult to wrap their heads around. In that interview, 0:04:47.529 --> 0:04:51.640 Coons also pointed to a relative lack of tech expertise 0:04:51.839 --> 0:04:55.680 on Capitol Hill. And then meanwhile, going down Pennsylvania Avenue 0:04:55.779 --> 0:04:58.740 inside the Trump administration, we heard The president just now 0:04:58.800 --> 0:05:02.560 in social media really going after Dario Amadei directly over 0:05:02.620 --> 0:05:05.919 this call for guardrails, insisting that he as president is 0:05:05.960 --> 0:05:09.620 the best guardrail of all. And, you know, you look 0:05:09.680 --> 0:05:13.690 at this, stepping back for a moment, AI development is 0:05:13.760 --> 0:05:17.190 central to the president's economic agenda. And it's also critical 0:05:17.450 --> 0:05:21.409 to this upcoming meeting with Chinese leader Xi Jinping. The president, 0:05:21.970 --> 0:05:24.830 Donald Trump, has put AI at the center of the 0:05:25.089 --> 0:05:28.400 economic competition with China. And he doesn't want anything holding 0:05:28.420 --> 0:05:29.670 the U.S. back in this moment. 0:05:31.190 --> 0:05:34.210 Bloomberg's Rachel Metz here in San Francisco, Mike Shepard in Washington, D.C. 0:05:34.270 --> 0:05:37.109 Thank you very much. The AI trade's under pressure as 0:05:37.150 --> 0:05:40.600 investors are asking whether slowing AI development could also slow 0:05:40.620 --> 0:05:44.300 the massive spending boom behind it. Uday Thiruvu, portfolio manager 0:05:44.580 --> 0:05:47.279 at Harding Lova, joins us now, and a portfolio manager 0:05:47.339 --> 0:05:49.620 who's focused on some of the names being hit hardest 0:05:49.920 --> 0:05:52.400 this morning, NVIDIA on the compute side, the memory names, 0:05:52.720 --> 0:05:55.120 some of the Neo clouds. Uday, just start with your 0:05:55.140 --> 0:05:58.299 reaction to the weekend's events as somebody that is directly 0:05:58.339 --> 0:06:00.900 invested in this corner of the technology market. 0:06:01.520 --> 0:06:03.440 So I think the weekend's events sets in play that 0:06:03.820 --> 0:06:07.380 it's not just about demand and demand growth, but realistic 0:06:07.820 --> 0:06:11.520 timeline for this and how implementation of AI is really 0:06:11.600 --> 0:06:14.730 going to happen. We're going from the face of super 0:06:14.770 --> 0:06:17.770 optimistic blue sky to, hey, hang on, there's a few 0:06:17.810 --> 0:06:19.350 more hurdles here that we need to get through. And 0:06:19.790 --> 0:06:21.630 I think that's what investors are grappling with at this 0:06:21.690 --> 0:06:24.609 point in time is how much of this current pacing 0:06:25.050 --> 0:06:27.570 is going to impact the demand environment over the next two, 0:06:27.610 --> 0:06:29.570 three years, but also over. 0:06:29.490 --> 0:06:30.090 The longer term. 0:06:32.089 --> 0:06:35.450 There is a debate about what pacing actually means, right? 0:06:35.490 --> 0:06:37.490 There have been many voices. We're going to go to 0:06:37.570 --> 0:06:40.440 one of them later in the program, Greg Brockman, OpenAI president, 0:06:40.800 --> 0:06:43.659 saying that all the open source work continues. It is 0:06:43.700 --> 0:06:48.060 the absolute frontier cutting edge. As an investor in the space, 0:06:48.260 --> 0:06:50.580 is it linear that if they were to slow down 0:06:50.620 --> 0:06:54.300 at the frontier, they would reduce spending in aggregate on 0:06:54.320 --> 0:06:55.140 the compute needed? 0:06:55.730 --> 0:06:58.570 No, absolutely not. Because if you think about it, most 0:06:58.610 --> 0:07:01.110 of the frontier models are not being used on a 0:07:01.150 --> 0:07:04.660 wide spaces now anyway. If you talk to most enterprises, 0:07:04.680 --> 0:07:07.800 they're looking at how to basically optimize their costs by 0:07:07.900 --> 0:07:10.290 using some open source models, some of the older models. 0:07:10.790 --> 0:07:13.670 And maybe for extremes and tasks that require a lot 0:07:13.690 --> 0:07:16.190 of reasoning, the latest model. So what this means is 0:07:16.210 --> 0:07:19.980 that slowing down the frontier doesn't actually lead to a 0:07:20.040 --> 0:07:22.960 slowdown of the total demand curve, but just basically spreads 0:07:23.000 --> 0:07:25.660 the curve out over a longer period of time. To us, 0:07:25.680 --> 0:07:28.180 that's what investors would need to focus their minds on 0:07:28.260 --> 0:07:31.060 is this is a timing issue as opposed to a 0:07:31.080 --> 0:07:32.570 total scale of demand issue. 0:07:34.140 --> 0:07:37.360 There was one response from David Sachs who said, okay, 0:07:38.020 --> 0:07:40.520 if you think you, the Frontier Labs, should slow the 0:07:40.540 --> 0:07:43.400 pace at the frontier, just do it. You're in control 0:07:43.440 --> 0:07:45.980 of your own destiny. And the sort of other side 0:07:46.000 --> 0:07:48.580 of that coin is the need for government to come 0:07:48.610 --> 0:07:50.850 in and regulate. You've got a very interesting point of 0:07:50.900 --> 0:07:54.010 view on the Frontier Labs and the need for regulation. 0:07:54.510 --> 0:07:57.310 Yeah, look, from a Frontier Labs perspective, there's two things, right? Firstly, 0:07:57.670 --> 0:08:00.990 they're talking about how The technology can be used for 0:08:01.000 --> 0:08:04.210 dangerous purposes and therefore regulation needs to come in because 0:08:04.750 --> 0:08:06.890 if one frontier lab allows them to do it and 0:08:06.910 --> 0:08:08.650 the other doesn't, the one who doesn't is at a 0:08:08.670 --> 0:08:11.270 competitive disadvantage. So that's why what they want is a 0:08:11.450 --> 0:08:14.420 set of rules that applies to everyone. So therefore no 0:08:14.480 --> 0:08:17.320 one has an incentive or a disincentive to slow down 0:08:17.400 --> 0:08:17.990 or speed up. 0:08:18.300 --> 0:08:19.100 So that's the first one. 0:08:19.340 --> 0:08:21.640 But then also the second aspect of this is if 0:08:21.660 --> 0:08:24.680 there are regulations in place, then the companies who are 0:08:24.800 --> 0:08:29.690 already ahead, the Frontier Labs and OpenAI, as well as Anthropic, Gemini, 0:08:29.790 --> 0:08:33.010 all of them, they have a strategic advantage because anyone 0:08:33.050 --> 0:08:35.650 else behind them has to go at a much slower 0:08:35.690 --> 0:08:38.250 pace than they do. So from both these perspectives, what 0:08:38.270 --> 0:08:41.500 they want is rules across to everyone, which stops everyone 0:08:41.660 --> 0:08:45.020 and also benefits them. So from my perspective, there's a 0:08:45.140 --> 0:08:47.620 bit of self-serving from the Frontier Labs that they do 0:08:47.660 --> 0:08:49.620 want rules because that is going to help them. 0:08:51.360 --> 0:08:55.620 You worried that an anthropic IPO or an open AI 0:08:55.679 --> 0:08:58.949 IPO gets delayed, kicked into the long grass because of 0:08:58.990 --> 0:08:59.850 the environment we're in? 0:09:00.730 --> 0:09:02.390 If you're one of these two companies, you'd want the 0:09:02.429 --> 0:09:05.030 revenue environment set and the cost structure set before you 0:09:05.070 --> 0:09:07.610 come to an IPO. What you don't want investors is 0:09:07.670 --> 0:09:10.970 talking about what the demand curve slowdown might look like, 0:09:11.030 --> 0:09:12.929 might not look like. So from that perspective, maybe it 0:09:12.970 --> 0:09:15.890 is better for them maybe into next year. But also 0:09:15.929 --> 0:09:18.140 the second part of this is What are their returns 0:09:18.179 --> 0:09:20.980 looking like right now? We still don't know the profitability 0:09:21.000 --> 0:09:24.140 of these companies. They do provide that they're going EBITDA positive, 0:09:24.320 --> 0:09:27.730 that they're increasing their profitability. But the pace of that 0:09:27.850 --> 0:09:30.770 might not be enough to be able to attract investors 0:09:30.809 --> 0:09:33.670 to the level at which that they want. So the 0:09:33.710 --> 0:09:37.510 pacing argument actually helps them get a clearer IPO story. 0:09:39.460 --> 0:09:41.840 I'm just going to jump in with some other reporting 0:09:41.860 --> 0:09:43.920 from the weekend, which came from the Financial Times. The 0:09:43.960 --> 0:09:46.559 Anthropic told some investors it was on track for a 0:09:46.580 --> 0:09:52.130 second consecutive quarter of profit, where margins are at 80%, 8-0%. 0:09:52.450 --> 0:09:55.510 But that doesn't take into account revenue share, which they're 0:09:55.530 --> 0:09:57.770 on the hook for with a number of parties, and 0:09:57.830 --> 0:10:00.480 also cost of compute, which would seem. to be something 0:10:00.520 --> 0:10:04.199 quite big to factor in. How do you interpret that? 0:10:04.730 --> 0:10:06.170 So I look at it from the cost of compute 0:10:06.230 --> 0:10:08.410 is going up for these companies, because what you're coming 0:10:08.429 --> 0:10:10.990 to an environment where if you go back to a 0:10:10.990 --> 0:10:14.030 year and a half ago, buying the same servers... is 0:10:14.130 --> 0:10:18.189 so much more expensive now. So the next infrastructure that's 0:10:18.230 --> 0:10:21.290 coming online comes at a much higher cost. So the 0:10:21.410 --> 0:10:24.320 revenue from that must be much higher too. But then 0:10:24.340 --> 0:10:26.059 when you look at the top line, that's being affected 0:10:26.080 --> 0:10:29.280 by token prices coming down. So we're looking at that 0:10:29.320 --> 0:10:31.980 sort of environment where revenue per token is coming down, 0:10:32.000 --> 0:10:35.870 but potentially the cost per token And the total depreciation 0:10:35.910 --> 0:10:39.250 that comes through is increasing. So maybe what you saw 0:10:39.290 --> 0:10:42.070 from 80% is not likely to be the same number 0:10:42.090 --> 0:10:44.210 going forward. It could be better or it could be worse, 0:10:44.250 --> 0:10:45.890 but we just don't know. And that's where I think 0:10:45.929 --> 0:10:48.750 in my perspective, the clarity that comes through over the 0:10:48.790 --> 0:10:51.429 next few months as we see this new infrastructure come 0:10:51.490 --> 0:10:55.150 online and the earnings and profitability come through, that might 0:10:55.170 --> 0:10:57.820 be a better story for Anthropic to be IPO-ing into. 0:10:59.290 --> 0:11:01.220 Uday, let's go back to this through the lens of NVIDIA. 0:11:01.320 --> 0:11:03.860 We know what's going to happen next year. NVIDIA told 0:11:03.940 --> 0:11:08.540 us 70% growth based on supply. Does that change at 0:11:08.640 --> 0:11:10.040 all or is this a 2028 story? 0:11:10.400 --> 0:11:13.510 No, I think this is a 2028 story because 2027, 0:11:13.510 --> 0:11:16.030 most of the companies have bought their demand in. In fact, 0:11:16.290 --> 0:11:20.630 there's a significant amount of overordering because companies know that 0:11:20.670 --> 0:11:24.179 they need these chips to do the agentic AI that 0:11:24.220 --> 0:11:28.520 they want to achieve. Now, The performance of models is improving. 0:11:28.559 --> 0:11:32.620 So 2028, maybe there's a different type of infrastructure that's required. 0:11:32.940 --> 0:11:35.400 But for 2027, if you hadn't locked in by now, 0:11:36.080 --> 0:11:37.780 you are not going to get them. And if you 0:11:37.800 --> 0:11:39.699 have them, you're not going to let them out of 0:11:39.740 --> 0:11:42.460 your hand. So that's why I think 2027 is most 0:11:42.540 --> 0:11:44.740 likely locked in for NVIDIA and put across the whole 0:11:44.780 --> 0:11:48.360 supply chain, including the NEO clouds, including everyone. But then 0:11:48.400 --> 0:11:51.910 as we get into this time next year, that's when 0:11:52.290 --> 0:11:57.750 the pacing, AI regulations... IPOs of these companies, all these 0:11:57.850 --> 0:11:58.569 come into play. 0:11:58.990 --> 0:11:59.690 But over the next. 0:11:59.530 --> 0:12:01.490 12 months, I don't think we should worry too much 0:12:01.550 --> 0:12:03.880 about revenue growth and profitability. 0:12:04.820 --> 0:12:07.520 Uday Charuva of Harding Lovner, back on the show. Thank 0:12:07.540 --> 0:12:10.080 you very much. Tune in later today. We're going to 0:12:10.100 --> 0:12:12.360 speak to David Sachs, Chair of the President's Council of 0:12:12.380 --> 0:12:16.200 Advisors on Science and Technology, as he has and continues 0:12:16.240 --> 0:12:20.820 to weigh in on what happened over the weekend. One 0:12:20.840 --> 0:12:24.570 of Dario Amadei's most concrete proposals for AI safety is 0:12:24.610 --> 0:12:27.850 independent oversight, and Fropik says it will bring in third-party 0:12:27.870 --> 0:12:33.079 evaluators with near-employee-level access, allowing outsiders to scrutinize not just 0:12:33.160 --> 0:12:37.020 finished models, but how they're being developed. Rohan Krishnan is 0:12:37.040 --> 0:12:40.160 the founder and CEO of Vows AI. Vows independently tests 0:12:40.240 --> 0:12:45.290 and benchmarks AI models. including unreleased models from OpenAI and Anthropic. 0:12:45.330 --> 0:12:48.790 He says, embedded evaluators will promote an even playing field 0:12:49.070 --> 0:12:52.060 and public trust. And he joins us now. You had 0:12:52.080 --> 0:12:55.400 a response to Dario Amadei's essay. Let's start with that response. 0:12:55.780 --> 0:12:58.850 What was your conclusion? Yeah, well, I think. 0:12:59.350 --> 0:13:00.770 I mean, to be honest, I think there's been a 0:13:00.809 --> 0:13:02.930 lot of extreme statements being made over the last week. 0:13:03.490 --> 0:13:05.589 And I can understand where they're coming from, but I 0:13:06.010 --> 0:13:08.770 think they can also cause undue anxiety for the general public. 0:13:09.130 --> 0:13:11.700 And so, really, I think in response to this, I 0:13:11.740 --> 0:13:14.000 found reason to be optimistic. From what I see at 0:13:14.040 --> 0:13:16.959 Val's AI, boots on the ground, we've actually been able 0:13:17.020 --> 0:13:19.600 to see the labs be very forthcoming. And we engage 0:13:19.640 --> 0:13:22.059 across all the major foundation model labs, as well as 0:13:22.080 --> 0:13:24.069 with members of Congress and government agencies and some of 0:13:24.090 --> 0:13:27.550 the largest enterprises. And so I actually think there's a 0:13:27.870 --> 0:13:30.390 lot of conflict happening on the public stage, but on 0:13:30.410 --> 0:13:32.710 the whole, we have reason to be optimistic that coordinated 0:13:32.730 --> 0:13:33.390 effort is possible. 0:13:33.450 --> 0:13:35.089 So let's talk about what Valve sees, right? 0:13:35.150 --> 0:13:38.230 If we were trying to really simplify what Dario Amodei 0:13:38.350 --> 0:13:42.260 and others are calling for, it is to slow down 0:13:42.640 --> 0:13:48.380 capability development, the pace of capability development, allowing safety alignment 0:13:48.400 --> 0:13:51.580 to catch up. You see these models in the training 0:13:51.620 --> 0:13:53.100 phase and before they're released, right? 0:13:53.780 --> 0:13:56.500 Yeah, well, I think what we see is that there's 0:13:56.540 --> 0:13:59.250 really been an outsized investment towards the generation or the 0:13:59.270 --> 0:14:02.110 capabilities of the models without that same parallel investment in 0:14:02.130 --> 0:14:05.190 the testing and evaluation. And so our ability to create 0:14:05.230 --> 0:14:08.210 better models now outpaces our ability to actually understand them. 0:14:08.250 --> 0:14:10.640 That's the issue at hand. A lot of people are asking, 0:14:10.720 --> 0:14:11.240 why now? 0:14:11.600 --> 0:14:15.670 And what Dario Amodei wrote is about the ability of 0:14:15.850 --> 0:14:19.910 AI to aid in the research and development of future 0:14:19.950 --> 0:14:23.730 generations of AI, so-called recursive improvement. That was a very 0:14:23.810 --> 0:14:26.840 clear piece of the essay. It is something that you 0:14:26.890 --> 0:14:29.800 look at at VOWS through the RSI index. So I 0:14:29.820 --> 0:14:32.880 think we're going to bring up that index. Explain what 0:14:32.900 --> 0:14:36.980 we're looking at, the methodology, and what that means. tells 0:14:37.060 --> 0:14:40.770 you about an AI's ability to contribute to the development 0:14:40.790 --> 0:14:41.229 of AI. 0:14:41.450 --> 0:14:41.850 There it is. 0:14:42.110 --> 0:14:42.270 Right. 0:14:43.010 --> 0:14:44.750 Well, so this is, yeah, this is the mainline figure 0:14:44.830 --> 0:14:47.600 tracking the progress of the major frontier models towards RSI, 0:14:47.830 --> 0:14:51.840 recursive self-improvement. And that describes the model's ability to make 0:14:52.100 --> 0:14:54.380 the successor version of itself autonomously. 0:14:54.860 --> 0:14:56.360 So wholly without human input. 0:14:56.820 --> 0:14:59.200 And that's, I think, to contrast today, where models are 0:14:59.300 --> 0:15:01.770 used as an active part, but humans are still involved. 0:15:02.230 --> 0:15:05.030 And so this would be akin to GPT-6 being responsible 0:15:05.050 --> 0:15:07.910 for making GPT-7, or Claude making the next version of 0:15:07.950 --> 0:15:10.660 itself without any human researchers. And so what you see 0:15:10.700 --> 0:15:13.680 on the plot is that we have... some early signs 0:15:13.740 --> 0:15:16.340 that models are exhibiting RSI, but they're still far away 0:15:16.400 --> 0:15:19.480 from the human frontier. And in extrapolating out these trends, 0:15:19.580 --> 0:15:22.460 we're actually making the prediction that we'll see models eclipse 0:15:22.480 --> 0:15:24.970 human researchers probably around August of 2027. 0:15:24.970 --> 0:15:26.280 August of 2027. 0:15:26.270 --> 0:15:28.330 So the question right now, like loads of people pose 0:15:28.350 --> 0:15:30.950 this is, Can an AI model act like an AI 0:15:31.010 --> 0:15:35.410 researcher right now? Can it run experiments? Can it improve 0:15:35.470 --> 0:15:39.950 AI systems with little or no human help? And right 0:15:40.030 --> 0:15:42.330 now the answer is no, but you have line of 0:15:42.390 --> 0:15:43.770 sight to when the answer would be yes. 0:15:43.910 --> 0:15:46.070 To be clear, we're testing a set of proxies and 0:15:46.090 --> 0:15:49.170 the publicly available models. And so the indication that we 0:15:49.230 --> 0:15:50.690 see is that the models are actually very good at 0:15:50.770 --> 0:15:54.050 executing experiments and operating as engineers. But they lack kind 0:15:54.070 --> 0:15:57.710 of the intuition to develop new experiments and have fundamental breakthroughs. 0:15:58.230 --> 0:15:59.990 But I think it's also important to add that we're 0:16:00.030 --> 0:16:03.410 not testing any unreleased model systems. We're not testing specialized 0:16:03.510 --> 0:16:06.700 agents or even multi-agent systems. And so I think the 0:16:06.720 --> 0:16:09.120 role of event evaluators over time could be to get 0:16:09.200 --> 0:16:10.680 closer to these internal systems. 0:16:11.320 --> 0:16:16.820 Again, the kind of concrete proposal from Anthropic is to have... 0:16:17.540 --> 0:16:21.280 independent evaluation, but inside the company. It sounds very close 0:16:21.330 --> 0:16:24.610 to what you're already doing with that company. This would 0:16:24.650 --> 0:16:27.550 create a demand for your services. So how would that work? 0:16:27.610 --> 0:16:30.750 And Fropic would pay you. How would that impact your 0:16:30.770 --> 0:16:34.130 independence to make an assessment of a yet unreleased model? 0:16:34.410 --> 0:16:36.210 Yeah, I think on the whole, it's going to take 0:16:36.250 --> 0:16:38.530 a diversity of approaches. So it'd be a failure to 0:16:38.570 --> 0:16:41.330 say that this will be the end-all be-all. But I 0:16:41.350 --> 0:16:43.350 think there's a historical precedent when you look at things 0:16:43.390 --> 0:16:46.530 like ratings agencies or auditing firms who have been for-profit 0:16:46.550 --> 0:16:49.490 companies and and have been responsible for testing and evaluation 0:16:49.530 --> 0:16:52.650 in other industries, I think there's natural conflicts that arise. 0:16:52.820 --> 0:16:54.760 And so the main one that we want to avoid 0:16:54.780 --> 0:16:57.220 is a situation where the same group auditing is also 0:16:57.260 --> 0:17:00.040 responsible for remediating or fixing. And then you end up 0:17:00.100 --> 0:17:02.800 in situations like Enron. And so we've set the line 0:17:02.840 --> 0:17:05.520 very clearly where although we make our methodologies public, we 0:17:05.560 --> 0:17:08.160 have test sets that are kept private. And we also 0:17:08.200 --> 0:17:10.240 ensure that we don't sell training data or the solution 0:17:10.260 --> 0:17:12.199 to labs while we do the testing process. 0:17:12.859 --> 0:17:15.880 I guess to end our conversation, the obvious next question 0:17:15.920 --> 0:17:20.230 is what happens next? You know, people have taken different approaches, 0:17:20.330 --> 0:17:22.590 but it seems that this is only going to work 0:17:23.010 --> 0:17:25.480 if all of the frontier labs are able to cooperate 0:17:25.500 --> 0:17:27.639 with one another to set a standard. 0:17:28.160 --> 0:17:31.460 You work with them. Do you see that happening? Honestly, 0:17:32.200 --> 0:17:33.899 I have a lot of reason to be optimistic. I 0:17:33.940 --> 0:17:37.040 think that people at these labs are rational actors, and 0:17:37.080 --> 0:17:39.000 so they see that the way things are headed, it 0:17:39.040 --> 0:17:42.330 will take some coordinated effort. And so I think with 0:17:42.350 --> 0:17:45.470 or without regulation, I think market-based solutions will have a 0:17:45.510 --> 0:17:49.020 part to play. Ryan Krishnan, co-founder and CEO of Vow's AI, 0:17:49.580 --> 0:17:53.460 key player in benchmarking testing unreleased models on Bloomberg Tech. 0:17:53.510 --> 0:17:56.830 Now coming up, executives may be calling for an AI slowdown, 0:17:56.850 --> 0:17:59.369 but the global race to dominate the technology is still 0:17:59.410 --> 0:18:02.230 top of mind. For US and Chinese leaders, we have 0:18:02.250 --> 0:18:04.410 the latest next. This is Bloomberg Tech. 0:18:12.130 --> 0:18:15.470 We're leading China in AI. We're the most sophisticated country 0:18:15.510 --> 0:18:17.850 in the world. And frankly, I want to keep it 0:18:17.869 --> 0:18:21.600 that way because whoever wins, AI wins. And we can 0:18:21.960 --> 0:18:23.859 put guardrails, we can do this and that, but I 0:18:23.880 --> 0:18:25.639 think you have a lot of negative forces that are 0:18:25.680 --> 0:18:28.580 bringing it up that shouldn't be. 0:18:28.500 --> 0:18:29.040 Bringing it up. 0:18:30.310 --> 0:18:33.370 President Trump there responding to calls from anthropic Zaria Amodei 0:18:33.410 --> 0:18:36.490 and other AI executives to slow the pace of the 0:18:36.560 --> 0:18:41.639 technology's development. Chinese leaders also criticized those warnings, with one 0:18:41.660 --> 0:18:46.100 official calling them fear-mongering and saying confrontation and vicious competition 0:18:46.380 --> 0:18:48.980 would disrupt efforts for global AI governance. Let's get the 0:18:49.000 --> 0:18:52.780 latest with Bloomberg Tech Editor. Peter Elstrom. There's a sort 0:18:52.800 --> 0:18:55.629 of out-of-order chronology of that. The president was posed the 0:18:55.670 --> 0:19:00.490 question after the essay from CEO Dariya Medea of Anthropic, 0:19:00.530 --> 0:19:04.100 but then we heard from Chinese officials. Summarize China's position 0:19:04.140 --> 0:19:04.479 right now. 0:19:06.530 --> 0:19:10.369 So China's position here is quite interesting. The spokesman for 0:19:10.390 --> 0:19:12.689 the foreign ministry came out and said the U.S. is 0:19:12.730 --> 0:19:16.060 starting this fear-mongering around AI. A couple things to remember. 0:19:16.200 --> 0:19:19.179 AI is much more popular with the citizens in China 0:19:19.200 --> 0:19:21.240 than it is in the U.S. That's for a number 0:19:21.260 --> 0:19:25.510 of different reasons, not coincidentally because... the press is quite 0:19:25.550 --> 0:19:28.690 controlled there. But also, the AI companies that are much smaller, 0:19:28.750 --> 0:19:30.930 they haven't had the kind of big data center build-outs 0:19:30.970 --> 0:19:32.620 that you've seen in the U.S., and they don't have 0:19:32.650 --> 0:19:36.240 the kind of power that the American AI companies have. 0:19:36.520 --> 0:19:37.260 They're much smaller. 0:19:37.280 --> 0:19:39.360 They're not really at the frontier the way we see 0:19:39.380 --> 0:19:43.160 with OpenAI and Anthropic. They've been fast followers in this 0:19:43.220 --> 0:19:45.410 market so far, and they don't have nearly the kind 0:19:45.430 --> 0:19:48.770 of compute that you do have in the US right now. 0:19:48.869 --> 0:19:51.450 So the Communist Party has held a pretty tight rein 0:19:51.510 --> 0:19:52.430 on the AI companies. 0:19:52.470 --> 0:19:53.710 It's also been supporting them. 0:19:53.990 --> 0:19:56.310 It's been pushing them towards this open source model that 0:19:56.330 --> 0:19:58.170 we've talked about a few times in the past that's 0:19:58.590 --> 0:20:01.930 less expensive. It allows companies to adopt the technology a 0:20:01.990 --> 0:20:04.710 bit more quickly too. So they're not nearly the big 0:20:05.210 --> 0:20:07.580 players in the market that you see. 0:20:07.500 --> 0:20:08.220 In the United States. 0:20:09.300 --> 0:20:11.800 We've covered in the program so far the idea that 0:20:11.880 --> 0:20:16.020