Steve Hsu: I don't think fast takeoff is realistic. However, I do feel confident that over a five to 10 year timescale, we are definitely gonna produce things that the average person, not a highfalutin AI scientist, but average people are gonna view this thing that we're able to build in the next day, five years as really like an AGI.

All right. Welcome to the very first Manifold Live Q&A. We're having this on our Discord channel, which Alf was nice enough to set up for us. The story behind this channel is that over the past few years, I've been having very informal meetups. Whenever I'm traveling to a big city where there are potentially lots of listeners to the podcast, I usually sort of tweet out that I'm around and that I can meet up.

And those gatherings have been really a lot of fun. And I think the most remarkable thing to me is that, A, the, the people who attend tend to be really interesting people, and B, they all live in the same city and often usually none of them know each other until the meeting. And after the meeting a lot of them tend to remain friends or, or, you know, interact more.

And so I thought it would be great to create that functionality on the internet. And so at the last New York City meetup, there was one guy, who actually is a traitor at DeShaw. He suggested that we set this up in discord, which is what we've done. And so the main purpose of this channel is for you all to just meet each other and get to know each other and hopefully have the same kind of serendipity that we've had at the live meetups.

To kick things off, I thought I would just do a Q&A. and so. In the Q&A July 12th channel on the Discord that people have been posting questions about, I've picked a few questions that I'm gonna answer. And if we have time, we'll do some more. I think the format we're gonna go for is

I will ask and answer the question. So I'll give a paraphrase of one of the questions that's in the thread and then I'll answer it. And then, we'll leave time for one sort of clarifying question and further response because I think there's high value to that for clarifying the issue. And then we'll move on to the next question.

So hopefully we'll do a fair number of interesting, Q&A, on, on different topics in the next hour or so.

So the first question that I picked out of the thread is someone asked me, Hey, if there were a billionaire out there who was gonna contribute to genomics, research and fund some kind of study or data effort, what would you go for?

Or what would you ask for? And the answer to that is pretty simple. I would try to create a data set which had of order millions of people, at least a million people, with their genomes and with a well-characterized psychometric score, so well-characterized IQ score and possibly personality scores if those were also available.

And I've had a longstanding theoretical prediction that if we had that data, we'd be able to build a pretty good, predictor, of psychometric ability IQ, if you like, or G. I think the closest we've come to that is a recent paper, which I'll try to stick a link to that. I think I tweeted about this paper not so long ago.

I'll stick that into the thread so people can look at it. It's a recent paper by Abdel Abdellaoui and his collaborators, and what they did is they went to the UK Biobank, which only has a very weak. Psychometric score called Fluid int or Fluid Intelligence, and that is like a two minute test and it only has something like a dozen questions.

So basically about 200,000 participants in UK Biobank have taken this test. It only takes like two minutes to take this test, but still only about a couple hundred thousand UK Biobank participants have taken it. What Abdel Abdellaoui and collaborators did is they tried to estimate what the score of the other 300,000 people would be on that test, using other phenotypic information about the other people like their socioeconomic status or their educational status or profession, et cetera.

So, they have this very noisy set of roughly 200,000 scores and then they have another roughly 300,000 crudely estimated additional scores, but, and so that's in total only half a million people. But using that training data, they were able to create a predictor, which captures about 15% of variants, 16% of variants of, of the original FIS score, which again, is itself only a noisy measure.

The way they do this is they hold back some data set of people that they have scored for, and then they do their training, and then they test on the holdback set how well the predictor performs, on that holdback set. And, it correlates about 0.4 with the actual FIS score, which is non-trivial. I mean, it's good enough that if someone were an extreme negative or positive outlier on that genetic score, there's a, you know, significantly elevated chance that that person is well above average or well below average for g.

And you know if they had a stronger measurement. So the test retest correlation of the fluid in score for yuca valve bank is pretty low. I think it's only like 0.6 or 0.7. Most good IQ tests have test retest correlations of over 0.9. And so even the indicator itself is quite noisy, but the fact that you can predict the indicator with some success. Having trained on, you know, not even half a million people. More like 200,000 people, plus, you know, this, this estimated additional 300,000, it suggests you could do much, much better if you wanted to.

And I don't think there's any sign that the quality of the predictor is reaching some asymptotic limit. An additional thing these guys did is, in building the predictor, they used family data to make sure that the resulting predictor is predicting what are called direct genetic effects. So it's not

predicting effects which make you a better parent and therefore elevate the eventual G score of your kids. It's actually trying to estimate what are called direct genetic effects on, you know, more or less kind of the brain function of the individual person. So I would predict that if you could get high quality G scores, say a half hour inventory, which a half hour test, which COR has test retest correlation of 0.9 or something like that, you, you would be able to do substantially better than  Abdel Abdellaoui did.

If you could do that for, you know, I would say one or 2 million people, I think you could get to a very high score, you know, above 0.5 8.6 or higher, correlation between the predicted score and the actual score. And that would enhance quite a bit the ability of people to do embryo selection.

So I think the cost to do this is not that high because there are existing biobanks with a million people in them of at least an order of a million people. And if you could just psychometrically evaluate, most of the people or all the people in that sample, that would be enough to, do qualitatively better than what  Abdel Abdellaoui and were able to do in their paper.

If someone were just to hand me a big chunk of money, this would probably cost, I would guess, something like tens of millions of dollars to really be able to get a few million people scored and their genotypes, you know, using some variety of, approaches like where you either already have their DNA or you already have their genotype and you're just trying to get the psychometric score

Maybe you have their psychometric score and you're trying to get their DNA, but assembling a data set on the order of a million people, I think that would dramatically change the perspectives of the public and scientists regarding this particular area because, a predictor, which is, you know, correlating 0.5 or 0.6, with actual well characterized g you know, which shift the game, shift the discussion a lot.

So let me stop there and take a follow up or clarifying question on this, and then, I'll respond to that.

Alf: You described that level of genomic technology and that level of prediction of cognitive ability, how far off would you say that kind of technology is?

Steve Hsu: I think it's very close. I think if we had the data set I described, so of order a million people with each of whom we've, we would have taken a stronger psychometric assessment.

Assessment, like a half hour assessment instead of a two hour assessment. I think we would have, you know, a predictor that correlates above 0.5, maybe 0.6, or even higher with the actual assessment score. And, you know, the analysis for that, the pipelines and all that are already well established. You know, we could follow something along the lines of what Abdel Abdellaoui did or my lab could do it very easily.

I think the consequence would be quite significant in the way that people think about these things.

Alf: All right, let me go on to the next question that I picked out of the pile. I think this was from David Kane, who's quite an amazing guy, former hedge fund manager and lecturer at Harvard. And well-known, I guess you'd call him a conservative thinker. He asked me about my efforts to help the Trump administration recruit some kind of high ability weirdos and misfits, that kind of thing.

Steve Hsu: I was working on that earlier this year. I guess I would turn that question also into some, you know, somewhat broader commentary on how I think the Trump Administration is performing so far. Let me just say that there was an appetite for sort of weirdos and misfits to come in and do things in the Trump administration.

Kind of similar to what Dom tried to do, Dominic Cummings tried to do when he was the chief advisor to Boris. This is quite similar to what DOGE was doing because DOGE also definitely recruited a bunch of people of that type. My contacts in the administration were not specifically for DOGE, so it was for other things as well.

And all that I was doing was, you know, you may have seen my tweets on X, just asking people to, send me their, you know, resume information or whatever and I would forward it along. I wasn't doing much more than helping them just filter these resumes and send them along. I really don't know how many people.

Made it all the way through that process and started working in the administration. One of the barriers was that, the people I was in touch with were specifically looking for people who were willing to go full-time live in DC and go in and actually fight battles within the bureaucracy, which, you know, for example, the DOGE people are willing to do, like, you know, spending your days, you know, in some office, you know, in Washington, you know, dealing with existing civil servants and bureaucrats and things like this.

And so that limited quite a bit. A lot of people contacted me saying, Hey, I'm a leading professor of X, Y, Z and I'm pro-Trump and I wanna help out. Oh, but I can't move because my family's in Ohio or something, and et cetera. So, so many people wanted to participate but they didn't really fit the requirements of what the administration was looking for.

So I think ultimately only a limited number of people went in. I hope those people have a big impact. I do know people who are the kinds of people who will actually literally listen to Manifold or have been Manifold listeners for a long time and who have taken very high level positions in this administration.

So I, I don't want to put them, but there are people at very high level positions with really serious impact who are part of what I would consider the generalized community, for this Discord channel. So that's sort of the good news. As a general commentary of how Trump is doing, I think I was surprised at how hard his team was willing to go on certain things, like, for example, the battle they're having with Harvard and with universities in general. That sort of exceeded my expectations. It turns out there's a specific reason why they're going so hard in this particular area.

Some of the things they've chosen to do, I don't agree with like, cutting science funding for academic research I think is not a great idea. I think that's actually gonna hurt us in our ability to compete with China. Other things like one of the reasons I supported Trump was I thought he didn't want the US to get into more useless wars.

I sort of feel that his performance on this is mixed. I think he has tried to keep us out o f Ukraine, of the Israel Iran conflict, et cetera. But, you know, sort of with mixed success, I mean, not as, not as successfully as I would've hoped. So I, I think the jury's still out on how this administration is gonna perform.

I think another interesting thing that's happening, which is very topical, is the Epstein situation because I know for a fact some of the people who went in there really wanted to release the Epstein information and what is going on right now is extremely interesting. Like, it'll be interesting to see whether someone like Dan, Dan Bungino or Cash Patel resigns if they are not actually more forthcoming with information about Epstein. And if you're curious about Epstein, look at my recent timeline on X because I posted a bunch of old, old stuff because I think the story behind Epstein in a way has been known for a long time and I just sort of reposted a bunch of old stuff that I had blogged about, you know, even five, 10 years ago.

Let me stop there and take any clarifying question that anyone has.

Alf: Very interesting one here from Triple Helix who asked, what are the limits of embryo selection for IQ enhancement where we just select from natural variants?

Does it max out around the smartest people to ever live or can we go way beyond that?

Steve Hsu: You can potentially go way beyond. I mean, if you just mean one shot, like one round of embryo selection, it's really limited by two things. The number of embryos you have and the quality of her predictor. Even in an extreme case, let's suppose the mother and father had a hundred embryos to choose from and you were able to literally just nail down the top embryo out of a hundred.

You're not gonna get necessarily anywhere near, someone like, a Van Neumann or Terry Tau. You're gonna get a big enhancement for that family, but you're not gonna necessarily push it beyond what has been seen, you know, in terms of historical geniuses. To reliably produce something like that you'd wanna have mastered the capability to actually do edits. Lots of edits.

Alf: Do you see things moving in a direction closer to what you want the Trump administration to go in? Or do you see it moving in a less desirable direction based on the current stage of play?

Steve Hsu: Yeah, I think most of what they want to do is okay. I think the execution is extremely poor. So I forgot to mention the tariff thing.

The tariff thing really shocked me because it was, I think it's been done so incompetently. You know, they might try to do something like make the US more competitive to China, like Reshore manufacturing, but they might not execute well. That's always the problem with the government doing it. I think it is more common for the government to screw it up than to actually get it right.

I think the dark part of this Epstein thing is the question is are we gonna do what's in, what I consider U.S. interest, which is more of a pivot to Asia versus are we gonna do what Israel wants us to do, which is to get involved in solving problems for them in the Middle East or what the Ukrainians want us to do and get involved in solving problems for them in Europe.

So, it's totally unclear how this is gonna go.

Alf: And I guess we've seen recently somewhat of a pivot from Trump on the Ukraine issue. I saw some reports that he's fast tracking to Ukraine, when obviously a key page has been. I mean, I think, I think he had a meeting with Victor Orban last year where Orban said that Trump had said that he wasn't sending over a penny.

More so I don't know what you make of that.

Steve Hsu: I think the initial mood, among Trump and all of, almost all of his key appointees was we're gonna get out of Ukraine as soon as possible. And you saw a lot of momentum in that direction. But I think what they didn't understand is they thought they would be able to mollify Putin very easily.

But I think Putin has good reason not to trust the US administration. So, I think Putin is not in a mood to settle in a way that's palatable to Washington. And I think there is this thing as, as some people say online, this is gonna be settled on the battlefield.

Alf: A follow up question here regarding genomics, if you want me to read that out.

Steve Hsu: Editing at scale, is that the one?

Alf: Yeah. Yeah.

Editing is, yeah.

Steve Hsu: I can't really go fully into that because it's so complicated. But there are startups that are trying to do this and generally the favorite approach now involves induced pluripotency in stem cells. So if you could have cells that live in vitro, like in a Petri dish.

And you're able to make whatever edits you want on them. So then if you make mistakes in the edits you can throw them out, right? No harm. It's not an embryo. You edit those cells to the DNA target where you want them to be, but then you use this induced pluripotency to make them into egg or embryo cells.

And then use that to produce babies. I was just at a conference in Berkeley on this called the Reproductive Frontiers meeting. And that's what the most aggressive teams are trying to do, is, the whole thing is intertwined now with the ability to control induced pluripotency, which there have been big advances in that subfield in the last few years.

And then I think about two other problems that really need to be solved before you could do this, like really aggressive editing is number one, you need better gene editing tools in CRISPR. CRISPR still has a significant off target rate and it's not really that precise, in terms of doing what you want to do.

Again, this is contrary to what the public generally perceives, but CRISPR's actually not good enough. And then secondly, a lot of the actual causal variance for things which influence a particular trait or not known, even though we might have really good predictors for that trait, the algorithms might be leveraging snips that tag the causal variant without actually telling us which of several possibilities is the actual causal variant.

And so that you'd wanna overcome that at least to some degree before you actually go ahead and edit things. If you edit the tag snip, it doesn't, it may not do anything for you, right? So you really need to know which one is causal. I don't really have time to do full justice to this particular question. We'd probably almost have to do a full episode of Manifold to discuss it.

Someone asked about the Needham question. So this is relevant to the US China technology competition.

So Joseph Needham was a Cambridge academic who spent a long time, I think originally his field was biochemistry, but he spent a long time studying the history of science and technology in China. Mm-hmm. And the Needham question is, given that they were ahead of the West in a lot of areas of engineering and technology and science for a long time.

Why did they not first have the Industrial Revolution there? Why did it happen first in England? And this is a very deep question that all kinds of historians, economic historians, historians of science work on. There's, there's various, proposals for what is the answer to the Needham question or Needham paradox?

I wouldn't say it's clear at all to me what the answer is. It's probably multifactorial. The comment that I wanted to make is that what we're seeing right now, so people who follow me or listen to my, then listen to Manifold know that I, I see the US kind of slipping, slipping in. Its lead over China in terms of technology to the point where there are a lot of areas now where the Chinese are actually ahead.

Or they're comparable at a comparable level, but their first derivative is much higher than the United States. And you can imagine a situation where, like a hundred years from now, if the Chinese end up way ahead of the Americans, the historians of that era, a hundred years from now we'll be looking back and saying, how, how did this happen?

Like, what a mystery. Like, instead of the Needham question, it would be like the, you know, it'll be like the, the, the song question. Some Chinese historians are trying to ask like, what, what, like, the Americans are way ahead. How did they fall behind? And because we're living in this era, it's easy to understand what some of these multifactorial influences could be. Like one society gets very rich and they decide they don't need to manufacture things anymore, and they lose a whole bunch of tacit knowledge associated with the production at scale of really complicated technologies. Or it's sort of the, they're heading into the decadent phase of empire and most people are concerned with becoming rich or fighting over the remaining spoils and they're not that interested in doing really hard stuff that's nevertheless in the best interest of the society as a whole.

And so there are a lot of subtle things which I personally being, you know, pretty old now, having lived through the Cold War and, and, you know, 20, 30 years of US hegemony. I can see how these factors have shifted from generation to generation.

When I was a 20-year-old kid, how I looked at the world versus how current 20 year olds look at the world. I think all these factors are at play and, but they're the kind of factors that are not easy for a future historian to look back and analyze correctly, at a remove of a hundred years or more.

I guess I would just say I don't really have an answer for the Needham question for why the Chinese who were probably at parity with the West in science and technology up until about 1800 then fell way behind for about 200 years and now are finally catching back up. I don't have a simple, sharp answer for what exactly caused that.

But I think if you think about our current situation, if you accept my characterization of the current situation, you can understand why this is a very complex phenomenon and how a society that is in the lead and producing lots of innovation and stuff can, can like, slow down dramatically just in a couple of generations, almost stagnate in a few generations.

And some rapidly rising competitors can surpass them. Because I think that's what we're actually seeing now.

Alf: Okay. Okay. And as a little follow up question, kind of a follow up question, it's a, it's an adjacent topic, but, it's from true seeking is fun and it's related to China. I guess it's related to one aspect. where China and America are diverging right now, and it's on, it's on the cost of living actually.

So saying currently on a small trip through China and its eyeopening, what is your internal model for why China is able to have such dramatically lower prices for almost everything, for equal, sometimes better quality? Is there more to it than currency manipulation, if that still exists?

Steve Hsu: Yeah, that's a great question.

I was thinking about answering that one as well. I would reference the two podcasts I recorded when I was in, on my China trip late 2024. One is called Letter from Beijing and the other one is called Letter from Shanghai. And I actually talked to this, one of the people, the guy that I recorded the Beijing episode with, Han Fata, who's his pseudonym.

He writes. A column for Asia Times. He's a person of American background, but he lives in Beijing and has been an investment banker in Beijing for a long time. So, we talked about this question. So, number one there, like if you just do like normie economics, so you look at the World Bank PPP adjustment for China relative to the United States, that adjustment could be 1.5 x plausibly 2 x.

So in other words, if you just, if you just wander around and you're just buying stuff at the grocery store or at the pharmacy, or you buy a jacket or something in China, there's easily like a kind of 2 x price difference that you see in, in, in real dollars. So what you can buy for the same dollar in China could be 2 x what you can buy for that same dollar in terms of physical goods and even some services, many services actually.

So there's already a factor like that that's widely acknowledged. Like I think any good economist who understands what PPP purchasing power parity is, would immediately say the exchange rate comparison between r and b and dollar.

So about seven r and b per dollar is wildly off in terms of purchasing power parity terms. And so that's one thing that's just noticeable to anybody who travels in China. Now the difference could be even more extreme if you're talking about buying a jet fighter or buying a hypersonic missile or buying, you know, an AI engineer's time.

It could be more like 3 x or 5 x or something like that. And so there, there are even more distortions than that, but just at a purely like normy level and studied by big institutions like the World Bank or, you know, major, investment banks and things like that, things like that.

There's already a well-known PPP factor, which could be as large as two.

And, and by the way, if you, if you make that PPP adjustment, the Chinese economy is substantially larger than the US economy.

Alf: Okay. And I mean, if you'd like to move on to, I think you said you had a fourth. pre-select the question if you'd like to move on.

Steve Hsu: Yeah. So my fourth one, a lot of people are asking about AI and fast takeoff of AI and what I should do for my career to not be replaced by a robot or an AGI. Obviously that is the kind of question that would take an entire episode of Manifold to try to grapple with, but I just wanna make one comment, which is that I do not agree with the fast takeoff scenarios that some people have pushed.

Alf: Is that the 2027 scenario that summer suggested?

Steve Hsu: Right. So probably the one that's best known. In fact, I'm actually good friends withScott Alexander. That's not his real name, but Scott Alexander is, I think, one of the co-authors of that report.

And, you know, the people who wrote it are all. Berkeley people. Yeah, I, I, I disagree with that projection. I don't think 2027 is a plausible year by which we'll actually reach AGI. I think actually if you talk, if you talk to actual researchers at the labs, and I'm not talking about Dario Amodei, who has to like to talk about his book all the time 'cause he's always raising money for Anthropic.

I'm talking about the actual guys who do the research, the kind of guys who get a hundred million dollars offers from Zuckerberg to move to Meta. If you talk to those guys, there's not a single guy that, and I've probably talked to dozens of guys like that. I don't think I've talked to a single one who believes in 2027.

So I just don't think it's plausible. Now, I think a better question is more the Yann LeCun question, which is, is the existing transformer architecture enough given additional hyper scaling to get us to AGI or ASI. And of course, you have to properly define what you mean by AGI or ASI. But Yann LeCun, who obviously is one of the pioneers in this field, a Turing award winner and early pioneer in neural nets, deep learning and stuff like this, pointed out that there are intrinsic shortcomings to the transformer architecture.

And this is something I've talked about as well. For example, it's not good at what's called test time learning. It can't really learn without adjusting its weights and adjusting its weights involves, like, using trillions of tokens of data, or at least, you know, very expensive, cumbersome activity.

It can't suddenly have a realization about something the way your grad student can in your office. Like your grad, my grad student could come into the office and he or she's confused about something and I sort of confuse them and then suddenly they can do useful work on a particular topic. You don't see that really with the models.

So, I think a more interesting question is, you know, hyper scaling alone without a significant change in existing transformer architecture, how far can that get you? A sub-question, even within that framework is, are we currently bottlenecked by training data limitations that are limiting how far we can get in the pre-training phase for these models?

So I, I think if you, if you get into the nitty gritty, I think you can see lots of reasons not to think that fast take off is realistic. I'm in that camp. I don't think a fast takeoff is realistic. However, I do feel confident that over a five to 10 year timescale, we are definitely gonna produce things that the average person, not a highfalutin AI scientist, but average people are gonna view this thing that we're able to build in the next day, five years as really like an AGI. Like it will be able to do all kinds of useful stuff.

Alf: It'll be a useful, daily companion, daily confidant, daily advisor. It'll be able to replace a lot of work that is currently being done by humans. So I think those things are in the cards. Yeah. And that's, a question here that I think is quite interesting, from Connor who asked, is the gold golden dome. Obviously that's Trump's suggestion for a missile defense system clearly modeled on the Israeli iron dome. He's asking, is the Golden Dome doomed to fail? Given the long known since Cold War limitations of missile defense, why are hacks in the Pentagon, going such lengths to conceal success rates, or like thereof of US Israeli defense systems?

How do we make component strategy for air and missile defense in particular, with reference to a potentially US China conflict in the Western Pacific?

Steve Hsu: Well, that's a great question. Let me try to answer that and then if while I'm answering that some follow up on the AI stuff shows up, I can, I can sort of loop back.

At the cost of pedagogy. We can, we can loop back to AI if we need to.

Alf: Sorry if we're darting all across the place. I'm, I'm just going through and seeing what, seeing what questions on the channel seem most interesting.

Steve Hsu: No, I think that's a great question. It's also something that I've been posting a lot about, online, I mean, on X and other places.

So, there's this very old question of how well missile defense can work, and this goes all the way back to the heart of the Cold War. And so if I could give like a very crude history of it, some very smart people who were working in, for DOD say, you know, going back to the time when, you know, lots and lots of physicists were working there, including people like Hugh Everett, if you're a physicist, you know who, who Hugh Everett is.

You know, they concluded very early on that it was gonna be very difficult to do missile defense. There are of course incredible technical challenges for having a. Supersonic, bullet hitting another supersonic bullet, which is effectively what you're asking the interceptor to do. And, in addition to that, even if you could solve those technical challenges, I think we've solved a lot of those technical challenges.

It still could be economically unviable because the opponent can use lots of countermeasures like decoys that have a similar radar signature to the warhead that you actually wanna intercept. All kinds of techniques that are cheaper for the person attacking with missiles to implement, than for the defender to defeat.

And this culminated into a, in a huge kind, actually kind of beneath the surface battle between academic physicists and people who wanted in government, who wanted to implement Ronald Reagan's Star Wars program in the 1980s. And so I was like, I had a front row seat for all this stuff 'cause I was a physics student at the time.

And so you, to me, the current situation is very like the da, it's like deja vu for me because what happened then was a bunch of people who didn't have cogent scientific arguments for why this was all gonna work or why the economics made sense, got into power in government and started pouring a lot of money into it.

And of course, the guy in charge, Ronald Reagan, or in this case Donald Trump, they just like the vision. They like the vision that yes, this is something positive. Instead of building, you know, more and more things whose only purpose is to kill millions of people, we're building things that are protecting you from atomic weapons or hydrogen bombs.

So the vision is nice, but the technical issues still remain from going all the way back to the original Star Wars program in the 1980s. And, they've never really been overcome. Like, the, the, the capabilities that the Star Wars people wanted to build, they were saying they could build in the eighties, still have not been built today.

So, you can see how, like, unrealistic, their, you know, their planning or their extrapolation was even back then. So I don't think the US can build a golden dome. I think, whether the goal is to defend yourself against conventional warheads on anti-ship ballistic missiles or hypersonic missiles, that is extremely hard to do in an economically viable way.

And even harder is to shoot down enemy ICBMs that are armed with hydrogen bombs. I don't think either of those is actually viable right now, but having seen what happened in the 1980s, I can easily imagine the US government just wasting tons of money on this again. And, unfortunately that's a, it's a pretty cynical view, but the physics, underlying physics of all this stuff hasn't changed.

And so, in fact, it's gotten worse for the defender now because we have very super maneuverable, hypersonic weapons, in the hands of some countries now. And so I just think this is very hard and if you don't believe me, like you don't have to trust any kind of theoretical argument here. You can just actually, if you, if you're willing to do the work of trying to figure out how well did Patriot Missile systems work in Ukraine? How well did missile defense work in the Iran Israel conflict? I think you'll, you can convince yourself that it's nowhere near foolproof. And the Iran Israel conflict is very interesting because it pits the best anti-missile technology that the Israelis and the Americans could field.

Okay. Because we had Arley Burke, destroyers, with Ageist systems. We had fad and then we had all the Israeli stuff, like the Aero systems and David Slingshot, et cetera, Iron Dome. We had all of that stuff, trying to defeat very primitively compared to what the Russians and the Chinese can field.

Very primitive ballistic missiles from Iran. And so, and if you look carefully, you'll, I think you'll convince yourself very easily that the initial claims, like during the conflict that they were getting 90% kill rates and stuff like this are totally fake. And there's a whole history of fake reporting on the success of anti-missile systems.

I recommend people listen to the episode I did with a guy called Ted Postel, who is an MIT, or he's an emeritus, MIT professor who specialized in this particular topic. And he got in big trouble because during the Iraq war, Raytheon claimed the original Patriot system, which has been upgraded several times now, but the original Patriot system was claimed to be able to shoot down scud missiles from Iraq.

Now, these are the most primitive, you know, like, like, Soviet era scud missiles that Saddam had. It was claimed that the Patriot system could shoot those down, and Postel did, as a professor, did a very detailed analysis of this and showed the Pentagon and Raytheon were just lying. And, this culminated in pressure on, I think Charles Vest was the president of MIT at the time.

There was pressure on MIT to shut Postel up. And in fact, I think they threatened to cut off the funding to Lincoln Labs. Most people don't realize, but the majority of MIT funding comes from the Defense Department. And, they were gonna, Lincoln Labs is a defense lab run by MIT and they were threatening to shut down Lincoln Labs.

And so, Postel basically fought this huge battle. This was a kind of battle that for people who grew up in the seventies and eighties, it's very common. Like, we still had a free press, like, and you know, we still had investigative journalists and we still had professors with backbone. And so a lone professor could go up against the DOD and point out that like something that they're saying on the front page of every newspaper is a lie is just wrong and prevail.

And this actually happened. And so I think we're kind of in a similar situation now where people, you know, who have the wrong incentives are claiming that missile defense works much better than it actually does.

Alf: That's interesting. You mentioned there are many examples from the nineties up until today of both America and Israel going to real, really great extents to cover up the failures of their Ms whole defense system systems. Why do you think they go so far as to hide the failures of their system? I mean, the recent example is I think the, with Saudi Arabia and the Houthis in the Yemeni in the, in their intervention in the Yemeni Civil War, where the Patriot Missile Defense system completely failed.

And it's got, and the reason the Saudis spent so much money buying, buying so many of them up, was to defend the Aramco oil fields and they completely failed in that. So I don't, so why do you think that the Pentagon as well as Israel go to such lengths to hide the failures of these systems.

Steve Hsu: Well, there's a short term incentive and a long-term incentive.

So the short term incentive is that, and this is again like something that, like the think young people don't understand because you, you haven't lived through enough wars, you haven't seen a war all the way through, and so you're still stuck in the short term dynamics. In the short term dynamics, there's a completely rational perspective, which is that, look, we're in a war and we have to win and we're gonna bend the truth.

We have soldiers whose lives are on the line here. We're, we're, of course we're gonna bend the truth. Like if we have to lie about the outcome of this battle, or we have to lie about the outcome of this weapon, or the efficacy of this weapon system. It's all part of war, right? Propaganda is part of war and both sides use it.

So in Vietnam, the stuff that the generals and the Pentagon were saying about what was actually happening in Vietnam or in Laos or in Cambodia, was just complete lies. Everything was complete lies. But it takes a generation, it takes a group of people who are paying attention to it and then pass through it, and then later this stuff is not so urgent.

It's not the short term dynamics, the war's over, but you can start digging in and saying, wait, did that really happen? What really happened there was, was there really a Tonkin golf incident or was there not? Was that all made up? You know? So only over time do a few careful thinkers get to the true reality of what they live through.

But in the short run, both sides have every reason. And you could argue, if, if I have my armies, my soldiers are on the line, of course I'm gonna use propaganda. Of course I'm gonna lie to my own people because I have to. 'cause we have to win this war, right? That's the short term dynamic. Longer term dynamic is stuff like Raytheon is still selling patriot missiles, right?

So they don't want you to know that it doesn't work. And there are Pentagon officers and officials who are heavily invested in these programs who have been testifying in Congress for, you know, many decades saying, this works, this is awesome. This is American ingenuity at work. And they don't want to be caught, finally by some expose saying, yeah, this doesn't really work.

Just to elaborate Alf on what you were saying about Saudi Arabia and Yemen, I think most people, listeners are not familiar with this. So, prior to all this fun and games we're having now with the Houthis, there was an attempt by the Saudis and the Americans to dislodge the Houthis from Yemen.

So there was a war in Yemen and a lot of people were killed,

Alf: I believe in 2015, basically sign of the capital,

Steve Hsu: Right. So when the Ukraine war started, because it was happening in Europe and closer to places like, quote, people actually care about, everyone went ape shit over Ukraine. But at that time, the war in Yemen was still going on, and like many, many more people were being killed in Yemen or starving because of famine in Yemen because of this war.

But nobody paid any attention to it because it wasn't on the doorstep of quote Europe, right?

Alf: There was a large famine as well as a massive cholera outbreak. Yeah,

Steve Hsu: no, it was a waste. Yeah. It was horrible. It just showed that, I mean, if you take a leftist perspective or an anti-colonial perspective, it's just like, you know, people in the west don't care about brown lives, basically, because you had a very good example of this, because like relatively small number of people were dying in Ukraine and many, many, many more people were dying in Yemen. But nobody cared at all about what was happening in Yemen.

But the way that war got resolved is the way Alf explained it, which is that the Yemenis, the Houthis demonstrated that using cheap drones and cheap missiles that, you know, presumably they were getting from Iran, they could easily destroy Saudi oil infrastructure.

And there was nothing, these crappy patriot missile systems and all the stuff that the Americans gave the Saudis and all the, all the billions of dollars of shit that the Saudis bought from us. None of that could stop the Yemenis, the Houthis, from destroying the oil infrastructure in KSA in the Kingdom of Saudi Arabia.

So they quickly negotiated it into the war and that's it, that's why the Houthis are still there. But it just shows you how shitty missile defense is. Some guys in sandals with 30, 40-year-old technology can cause you big problems. 'cause it is fundamentally hard to shoot down a ballistic missile. And so,

Alf: if I remember rightly as well, they were even landing hits on the, on the uae.

Steve Hsu: Yeah, exactly. So I mean, anyway, yeah, the, if, if you hear me sounding really exasperated or like shrill on X about this, it's because like you go on TV and you, you, you go on YouTube and you listen to, there's some meeting at the. You know, the AEI think tank or the Brookings, you know, think tank or whatever, or Rand and these people supposedly know what they're talking about, but almost none of them in the modern era have any technical background.

So you could have these pointy headed guys with political science degrees pretending to talk about like, missile defense or, or this balance of power, you know, in the Middle East, or, and they just don't have they don't know what they're talking about. It's just insane. It wasn't like that when I was growing up because anybody, if you went into a meeting like that in the Cold War, there were bound to be a couple of physicists sitting around and they were not gonna get their facts wrong.

Right. And so, and they would hold the other, they would terrify the other guys, right? Who was in the meeting. So you just didn't have as many talking idiots, purporting to discuss grand strategy. Back then. Now it's almost impossible to find someone who isn't technically incompetent talking about these things.

Alf: Very, very interesting. I think it is very illustrative of the decline that we've seen in the caliber of, I suppose you could say the political class, but especially the diplomatic class. I mean, there was an example I saw someone give on X, which was, well obviously when Ted Cruz did his infamous interview with Tucker Carlson where he, he didn't know the population of Iran or its ethnic makeup and someone pointed out that no matter your thoughts on the British Empire, it would've been unthinkable for Lord Kon, that one of the governor General, one of the vice lawyers of India, during the British RA to not know that kind of information about India.

Right. And it just shows the fact that, I think this partially stems from the fact that America is an empire that denies it's an empire, whereas obviously the British Empire never denied that. inhibits the ability of the political class of the elite in America to acquire the necessary skillset that is needed to run an empire because there's still this massive denial that it is an empire.

Steve Hsu: Yeah, I think you're right about that. Is one of the factors contributing to what we see today. Another factor is that if you really are so, the US has not really been in a technological competition for 30 years. We've been number one for 30 years since the Soviet Union fell apart. Yeah. And the Chinese only recently became a peer competitor like 10, 20 years ago.

We were still way ahead of them, right? So we've had this period of being fat and happy and we just didn't need, we didn't need like, you know, some very dour von Neumann like guy or you know, Hugh, some nerdy Hugh Everett guy doing a bunch of calculations to figure out like, are we, is there a missile gap with the Soviet Union or is there not a missile gap?

Just recently I tweeted out some, I found this old report. So, Hugh Everett, who was a legitimate genius in in physics, left academic physics and worked at the Pentagon in something called the Weapons Systems Evaluation Group, WSEG, and I dug up some, there was a very famous WSEG 30 report, report number 30, which really laid the whole basis for mutually assured destruction and all kinds of things, which later became like the way that the Cold War was actually fought.

And one of the amusing things in that report, which I dug out, was a statement from them also talking about how missile defense was just not gonna work.

And this was like, published in the late fifties. So it's been a long time.

Alf: Okay.

we've got a question here. I don't know if you wanna round off fairly soon. So I thought there are these two questions here that might be nice to round off on. Firstly we've got from Georgia who asked, how do you personally develop good intellectual taste?

So clearly you have it, or else the pod would be boring. It seems that having good taste and judgment is far more important. Now, given as you ask, do you read a lot? Do you talk with people? Do you have specific, deeply technical or domain experts that you turn to for various questions, perhaps build things yourself or tinker, what did you use to do?

What did you used to do? And what are you doing these days? Or maybe you don't think about it that hard, have you seen others develop good intellectual taste, maybe PhDs or postdocs? Or maybe you are just mostly born with it.

Steve Hsu: That is a great question, Alf. Thanks for pulling that one to the top. I saw that one in there and I should have made a mental note to try to answer that one.

That's a tough one to answer actually. So first of all, some of this is hardwired. You know, even among top scientists, some people have good taste and others don't. And it's, it's, of course it's very subjective and sometimes you don't know who has good taste until the dust has settled, you know, long afterwards.

But all I can do is describe my own thinking framework, which is what I always do, and this is pretty much just like, in a way, describing Bayesian learning. It's like you're trying to develop a causal model of how the world works and any particular node in that structure that you're building, you realize, I could be wrong about this.

Like I have some confidence level in this assumption. This is how missiles work, right? It could be like, I don't know anything about how missiles work. Let me just make that completely low confidence. Right? Or, oh, no, I have spent a long time studying this and I've actually talked to real experts like TED Postel and stuff like this.

So you need to be aware of that, that sort of lattice of nodes in this picture of how the world works. You need to be aware of your confidence or conviction level in each of those nodes. And what are the open questions. Like if you happen to be at an event and you happen to sit down with someone who works on quantum computing or something, and there's something, you always wanted to ask that person to sort things out.

You just have to be disciplined and say like, oh, this is my one chance, this is my opportunity to try to get, try to try to increase my conviction level or confidence level in one of the, you know, nodes or hypotheses, which are a part of this big causal structure that I'm building in my head. And serious people, you know, that would include like a lot of, for example, physics professors who are, you know, my age and have been at this a long time.

They've been trying to build a, you know, map of the universe, not just in physics, but including other areas like biology or AI or technology or politics. They've been building this for a long time. It's always a pleasure to like meet somebody like that and compare notes and, you know, maybe we'd find, oh, here's something we disagree on, like, let's talk about why we, why I think A and you think not A or confidence levels are very far apart on this hy hypothesis A. So I think developing good taste is sort of that. It's all in service of building this framework that you're using to evaluate the world. And of course you could specialize and say, look, look, I'm only right now thinking about semiconductor physics.

I'm not trying to understand genomics. I just weigh everything in genomics with low confidence because I haven't spent time trying to do it. I'm focused on just making sure my map for semiconductor physics is really good, right? So you just have to be intellectually honest with yourself.

One little anecdote. When I first met Robin Hansen, we were both at a conference at HP Labs and Robin was really interested in something called the alman agreement theorem, which if you're familiar with like some kind of abstract economics or decision theory. Aumann agreement or game theory, Aumann agreement theorem is like a big deal, right?

And Aumann's agreement theorem is sort of like two rational bayesians can't disagree because if we start disagreeing, you know, I should be able to show you my evidence and you, and, and you show me your evidence, and that should cause convergence in our confidence levels and our hypothesis and stuff.

And I was sitting there with, with, Robin, I just met Robin and we were just, we were, and so we immediately got into the meat of this. And there was another guy with us who was a PhD in math. and I just said to Robin, look, the problem with this Aumann agreement theory, I'm sure it's all very nice.

I'm sure economists can get really, like, big erections over this or something. But, the problem is I cannot open up my priors and my evidence for you. There's no way you can actually, you, you didn't sit next to the Israeli, physicist who worked on their system X who explained to you on a napkin, you know, 10 years ago how it works and what its weaknesses are.

You, you weren't there. I could summarize it a little bit for you, but that isn't causing the same update in your weights in your brain as it causes in mine. Because I was actually there talking to this guy and I know this guy and I've spent, you know, time talking about other problems in physics with this guy.

So, the whole idea that we have to all agree is kind of a fantasy. It's based on a very unrealistic set of assumptions about, you know, the evidence or information that we have to build our sort of world model. And so you're always gonna have people walking around with different world models.

If I find someone who I really admire and trust and their world model differs drastically from mine, I'm really interested in spending time just trying to resolve that, trying to figure out why, why do we disagree about this?

Alf: And I think to round off, that's a question here from Edward Anderson who asked another question for Steve. How do you spend your time? How do you spend your time, you productivity wizard with a system that makes you work more than others? And has this changed from your college days to now? And he said thank you.

Steve Hsu: I think that there have been times in my life when I was younger, when I had to be a productivity wizard. Like if you're in college and you're taking very challenging courses and you're also trying to maybe play sports or at least keep in shape and you're also trying to be girls and stuff, like to get through that, you have to be a productivity wizard.

You have to really have lots of tricks for managing things and doing things efficiently. Also I think for sure, when I was a first time tech startup CEO, I had to really manage time because there was so, so, so many demands on your time. If you're the CEO, you're the nexus of a bunch of information flows in the company and you have to be really, really,clever about how you manage your time and how you manage your interactions for efficiency with other people in the company, and investors and outside people.

At this stage in my life, it's a little different because I've accumulated a lot of knowledge and I have pretty deep interests in, across a number of areas. So my day is kind of like this. I get up in the morning, I work out for, you know, 45 minutes or something. I drink my protein smoothie.

I've, I've given the recipe out for that before I can give it out again. And then the rest of the day I'm kind of just working because my kids are grown up, my kids are in college. I. And so my wife doesn't bother me too much, and so I can basically be working all day. I can be reading a physics paper or I can be watching a seminar on YouTube, or I can, someone on X points me at some AI benchmarks that I wanna look at.

You know, I can basically be working all day and even at night, like in the evening, I can, you know, continue working on stuff. I'm interested, there's enough interesting stuff for me that I can kind of just be productive all day, You know, with breaks to do other stuff. But, so I currently don't really use any special productivity hacks.

Maybe the one I would mention is if you're trying to sort something out or you have to write something, often the best way to do it, or at least I find this effective, is to, you think really hard about it.You know, you do your calc, you write equations, or you read, or you do, you do all the things that you need to do, but then, you know, at the end you're like trying to solve something, some problem, or you've gotta write like an essay or an article about this at the end.

But then you just put it out of your head and you do something completely different. Like walk, go for a walk, right? Or ride your bike or something. And somehow in the background, the brain or my brain is processing this.

And then some, then at some later point, maybe not that day, maybe the next day I'm able to sit down and then the brain has done the like, kind of subterranean processing and I'm able to just pound out the article or, or maybe like write out the solution to the problem. And I think other scientists, particularly mathematicians, have actually commented on this as a productive way to do things.

Alf: I think we're coming to the end of what has been just over an hour long and brilliant. Live Q&A. I dunno. How, how have you found it, Steve?

Steve Hsu: Well it's just been me talking. I hope people found what I said. Interesting. But yeah, no, it's been great. It's been fun. I just, I wanna say again to everybody who's in this Discord that I hope it becomes a lively community. And, you know, you may feel isolated. Maybe you live in some small town in the US or the UK, but there are other people with brains like yours and they actually want to meet you. So, you know, hopefully people will use the channel productively.

Alf: That's great. And thanks to everyone for tuning in and goodbye.