00:00Our equity markets are just trying to wrap their heads around currently as we understand the broader implications of the
00:05story.
00:06Perhaps to help us fill in some of those gaps, our next guest heads artificial analysis,
00:10which helps developers and users compare these large language models.
00:13His co-founder and CEO, Micah Hill-Smith, joins us right now from San Francisco.
00:18Micah, happy Sunday afternoon where you are. Good morning from the Asia-Pacific.
00:22Simple question for you. Should we believe what they said, the three?
00:30Yes, but we should read what they've said carefully.
00:33They haven't said that they are committing now to a broad AI pause.
00:37We're talking about mechanisms to pace the frontier, to do more safety work in order to make sure that they
00:43can safely release these models.
00:44They're not saying they're not going to release them.
00:46This comes as a capstone call-out on several months of increasing temperature on the conversation around all of these
00:54models.
00:54We've been seeing capability thresholds, especially around cyber capabilities being crossed and being of interest to a broad range of
01:02players.
01:03And we've seen several incidents now in the last couple of months that have got a lot of attention,
01:09including the OpenAI hugging face incident, that have made it very clear that the building and deployment of these models
01:15is creating real risks.
01:19And how do you see this?
01:21I mean, obviously, they're the biggest three and they're, you know, competitors.
01:25I mean, obviously, there's one question how far they will go imposing these sort of limits.
01:29But will others follow, you think, in the industry?
01:34There is a particular significance to the top players who are building the models that push the frontier, making this
01:41commitment.
01:41So I do expect other players in the industry to join this, but it is significant that all of the
01:48leading American players just this weekend have made this commitment to,
01:52in Dario's words, pace the frontier.
01:56We very much expect to see continued progress.
01:59This is not a commitment to stop progress.
02:02It's not a commitment to stop releasing models.
02:04It hopefully will just be this mechanism that allows enough time to be taken to make sure that the right
02:15safety work is done alongside releasing the models.
02:19And when you say others would and does that include the Chinese labs, how might this affect their behavior?
02:28And arguably, they are obviously motivated by a different set of incentives, you know, suffice to say.
02:34Just help us understand the context now.
02:38Yeah.
02:39So first, just some background is that the race for the AI model part of the AI stack is more
02:47competitive now than it has been at any point in history.
02:50There are not – it's not just the top three companies.
02:53We talk about about 10 to 50 companies that we benchmark who are creating models that, in various ways, push
03:01different Pareto frontiers between metrics.
03:04At the very top end for the most intelligent models, those are the ones that are the most important for
03:08these risks.
03:09And I'd say that as of now, it is unclear how it's going to play out globally with any efforts
03:15to pace the frontier.
03:16But I would say that it's not clear that there's going to be enough slowing with the top U.S.
03:23players that the U.S. will be giving up its lead.
03:25I think there have been statements from the lab leaders and statements from U.S. politicians this weekend making clear
03:32that the U.S. is intending to maintain the current lead in model intelligence that the U.S. has had
03:37over China.
03:40Especially, President Trump said, too, he doesn't want to see the slowing of AI progress because there is that risk
03:45that they could lose out to China.
03:48What about in China – I mean, the fact that China has its own sort of ecosystem, right, cheap models,
03:53and the fact that there is a bit of a decoupling in this whole AI race.
03:59Net-net, then, do you think that it's not going to slow the pace of development among some of these
04:04Chinese startups?
04:04Because that speed culture within these startups is very much ingrained in the DNA of some of these companies.
04:14Yeah.
04:15So, there has been an aspect to date that none of the Chinese companies have got into a point of
04:22having the smartest model in any given time.
04:24And there is an untested question here of whether it would be possible for them to achieve that.
04:31Putting that aside, though, I very much expect them to continue competing.
04:36One of the aspects that is most interesting in the setup today are that many of the leading Chinese AI
04:41labs are being much more aggressive about giving away open-weights copies of their models than the leading American labs,
04:48leading to the strongest open-weights models coming from the Chinese labs.
04:53That is of particular interest in a lot of these conversations because while there is a story around being able
05:00to own your own destiny and run those models on your own hardware instead of needing to trust a third
05:04party with all of your data, there is another side to it that once a model has been released open
05:11-weights, there's nowhere to take that back.
05:13Which means the risk profile is different of releasing an open-weights model.
05:17The safeguards that can be run at deployment time alongside products and APIs offered by labs offering proprietary models don't
05:25apply as much to the open-weights models.
05:27That means that the models that we see being released open-weights,
Comments