00:00Let's start with this executive order, as we just laid out. It's interesting because to a lot of
00:06folks, this looks like a 180 from the Trump administration's previous approach to regulating
00:12AI, which was very hands-off. I mean, it is a 180, but it's a 180 from something that could
00:19never
00:19really last, never really made sense to something that does make more sense. So it was untenable.
00:25It was always going to be untenable to say no regulation on AI. And in fact, I told Senator
00:31Kennedy when I testified in the U.S. Senate three years ago that the number one thing we needed was
00:35like a pre-flight testing to make sure that things are safe before they were released. And for a while,
00:41the Trump administration seemed to have no interest in doing that at all. And then I think it's because
00:45mythos came from Anthropoc that they woke up to the reality that, hey, we might let something out there
00:50that we shouldn't have let out, and we are going to be blamed for it. And I think that that
00:54realization
00:55turned them around, and I think turned them around in the right direction.
00:58Well, I'm curious to hear whether you think this should have gone farther, because again,
01:03this is based on voluntary access here, basically, that the U.S. government access would be granted
01:11on a voluntary basis. And I wonder, again, if that goes far enough for you.
01:16I don't think it does. And in any case, it's only an executive order. Congress still needs to do
01:21something. I hope that Congress will eventually pass a real law that says you do need a mandatory,
01:27maybe it's 30 days, maybe it's 90, whatever, but you need some kind of FDA-like process where you
01:32look at new systems and say, do they pose new risks that we need to worry about? Do the costs
01:37of these things outweigh the benefits?
01:40Gary, I do want to get your thoughts just on this idea that we seem, I don't know if I
01:45would say we,
01:45but the idea that there seems to be a large cohort of people that have accepted the idea that AI
01:50is
01:50our future. And we've talked a lot in this program, not only about the risks, but obviously a lot about
01:55the pushback that we've seen from certain folks out there, certain generations that look at what
02:00we're doing right now, not necessarily saying it shouldn't be done, but maybe it shouldn't be done
02:05at this pace. Does this executive order maybe address that or maybe slow the process at least
02:11just a step so we can have a better assessment, not just of what maybe the national security
02:16implications might be, but what the impact on society might be?
02:20I think there are a lot of different risks of AI, and I don't know that slowing things is quite
02:25the
02:26right answer, but we do need to address all of the downside risks. This executive order will help a
02:33little bit. I think we need to do more than what is just here. There's actually now two different
02:38backlashes. One's around cost, and maybe we'll get to that. And the other is about young people worried
02:43about jobs, worried about the ways in which AI affects society. We really need to move towards a
02:49positive AI outcome. I think the default here has been just unfettered capitalism without any regard
02:57to what's happening to the citizens. There's an old saying about privatizing the gains and socializing
03:02the costs. That's kind of what's been happening, and that's what we need to fight here so that
03:07everybody gains in some way. I'm not endorsing what Sanders said yesterday, but we do need to
03:12find a set of AI policies where the consumers and prospective employees and so forth have some
03:19protection from the risks. But when you use the phrase unfettered capitalism, and we've certainly
03:23seen that with regards to the amount of money that's being spent by some of the big corporations
03:27out there, small corporations too. And I am curious if it's a little bit too late. Given the amount of
03:32money invested, you now have this sunk cost and this idea that there are certain companies and
03:37certain cohorts of our society that aren't going to let this go for one reason or another. The main
03:41reason is they don't want to lose what they invested. Yeah, I think there's a kind of insanity right
03:47now where people are putting in literally trillions of dollars thinking that this is going to be a
03:51winner-take-all market like we saw with Google and Search where they got like 90% of the market.
03:56And I don't think it's going to turn out that way, which means I think a lot of people are
03:59going to
04:00lose a lot of money. If you put in trillions of dollars, you need to make a return where you're
04:06making a profit of $100 billion, $200 billion every year, or else it's just an insane investment.
04:14And we're not seeing that, right? Now, the only party that's really making money is NVIDIA
04:20on the chips. And all the LLM providers, the large language model providers, like OpenAI and
04:26Anthropoc are actually losing money. We could talk about one quarter that might be a little different.
04:30And so you're putting all this money in, and if these bets don't pay off, the results may be epic
04:37for all of society, really. Like a lot of retirement funds are going to be pressed to participate in
04:42these IPOs. And if the IPOs don't sustain themselves, then everybody's retirement fund takes a hit.
04:49And we may eventually get in a position where people say in these companies, hey, we're not
04:53really making money. We need some kind of bailout. And then the taxpayers pay for it. I think it could
04:57be really quite bad that would lead to a recession or worse.
05:00Well, certainly some frightening ripple effects there. But I want to talk again about sort of the
05:06starting point for that. You said that this has been treated as if it's going to be a winner-take
05:12-all
05:12market. And I want to talk about why it won't be necessarily. You mentioned Google
05:18with search. And I can imagine having a conversation decades ago that why would Google necessarily
05:25become the incumbent in search? Why wouldn't you see many different search companies come out and
05:30be able to have a more evenly split pie there? Obviously, that's not how it ended up being. But
05:36why does AI necessarily have to fall into that same path?
05:41It doesn't absolutely have to. But what's happening right now is that everybody's building
05:46the same technology. And so it's very hard to get a lead. It's not like Facebook, where they got a
05:52moat from their social network. So once enough people were there, that was reason for them to
05:56stay and not go elsewhere. But you don't really have that kind of lock in here. So if people don't
06:01like Anthropic, they switch over to OpenAI. If they don't like that, they switch over to Google
06:04and so forth. Because everybody's basically offering the same product. They're all using
06:09a technology known as a large language model. They're all training them on essentially the
06:13same data, which is essentially the entire internet. And so there's very little product
06:17differentiation. And that means that these things are commodities. It's hard for anybody
06:22to get a sustained moat. And what we see is like somebody's in the lead for a month and then
06:26somebody else comes in. Somebody else isn't. We haven't seen anybody take a sustained lead.
06:31And so it's kind of, I think, wishful thinking for anybody to think that they're going to dominate
06:35this market. And don't forget, China is also involved in this. And for many purposes, people
06:40will just go over to China if the U.S. doesn't have a competitive product. So I expect the prices
06:45to
06:45consumers will go down. But profits are going to be really hard for these companies to find.
06:50Is that does that sort of disrupt the build out or I guess the adoption of a lot of these
06:56tools
06:56longer term? I mean, if some of these companies not just fail, but at least they start to be hobbled
07:03by a lack of profitability, by squeeze margins, what incentive would they have to continue down that
07:08road? What I think we're going to see is that large language models stick around. People continue
07:13to use that technology. But A, there will be better technologies. B, the current vendors, a lot of them
07:18are going to lose a lot of money. And C, we may find other approaches that, for example, people can
07:22run
07:22on their laptops. And so they don't need all of this cloud. And so we may find that all of
07:27these data
07:27structures are kind of built for nothing, even if the technology itself, some of it survives.
07:32And probably it will be displaced. I mean, one way to think about it is these systems take like
07:37megawatts to answer simple queries that a human brain can answer with 20 watts. It's got to be possible
07:42to do it more efficient with better technology. And once it is, these data centers may look a bit foolish.
Comments