Brian Chau: This might be a Manifold exclusive breaking news, but the Epstein Files Transparency Act was almost entirely written by AI. We did a big scan using, using Pangram of all of the all of the bills. We found around 6.5% of bills were written by AI. But of all of them, the only one that is now law, the only law that rules us that is written almost entirely by AI is is, is the Epstein Files Transparency Act

Steve Hsu: Welcome to Manifold. My guest today is Brian Chau, a former math and coding prodigy who's all grown up now and trying to change the world. Brian, welcome to the show.

Brian Chau: Great to here, Steve.

Steve Hsu: Now, Brian, I've known you for a long time. I think, like, maybe when we first met, were you still in college or you had just finished college?

Brian Chau: Yeah, I think so. It was right on the tail end.

Steve Hsu: Yeah. You had a podcast and I think we recorded some long episodes on your podcast. And then since then we kept in touch and I think did we only meet IRL in oh we probably met at some conferences like Less Online or Manifest, but, but then like I hung out with you at Network School. Was it, how long ago was it? It was like maybe earlier this year, was

Brian Chau: four months ago, five months ago.

Steve Hsu: Yeah.

Brian Chau: Oh, huh. Maybe it was more than I thought. Yeah

Steve Hsu: Yeah. So welcome to the show. I think of you as one of these super insightful younger guys who's very agentic. And the reason I invited you for this episode is I heard what you were doing with Effort News, which is your latest venture.

Brian Chau: Yeah. It's been an amazing time since the launch. I was basically fooling around with a bunch of AI, actually like financial auditing tools. I was developing this and I was like, "Okay. Well, what am I gonna audit?" 'Cause at this point I had left Network School. I'm no longer working for them. I didn't have any revenue back then, obviously.

So I thought, "Okay, let's throw it at some government databases." And then I found some really, like, mind-breaking and ultimately, like, incredibly successful stories. I started asking around. I thought, "Surely someone must know about these stories," and it was just crickets on all of them. No one had reported on this.

It was, like, brand-new stuff. And at that point I thought, "Okay, I have to do this full time." And I was just sprinting 24/7 up until our launch, and then it's been an amazing time since. Just like every, every week we get, like, a ton more traction.

Steve Hsu: And, and you launched like a month ago, early August, is that right?

Brian Chau: Yeah. We launched almost exactly a month ago as of recording. So it was on August 10th was our launch.

Steve Hsu: Got it.

Before we get into Effort News, 'cause I, I think maybe we've whetted the audience's appetite a little bit let's just go back and talk about Network School, because the last time I saw you, which was earlier in 2026, we were both on this little island off the coast of Singapore where Balaji Srinivasan had established a thing called Network School.

It was on this artificial island which was built by, you know, byproduct of the Chinese property bubble of five or 10 years ago and was almost like a ghost city, like 10% occupancy in this huge apartment complex right on the beach facing Singapore. There's a kind of four or five-star hotel there which Balaji had converted into Network School.

And for people who don't know about it, Network School is a place for entrepreneurs and digital nomads to kind of lock in, be very healthy, super large gym, healthy food available all the time, and you can just lock in and work on your startup or just learn, attend events and stuff like that. And so Brian, you were an NS employee, right?

Brian Chau: Yeah, so I worked for Network School from May 2025 to July 2026.

Steve Hsu: Okay. So when I was there, there was all this talk about getting ready to spin up another location, another instance of it, and trying to get the sort of permanent or long-term population up to, I forgot the number, was it 300 or 1,000 or something? And so things were pretty optimistic. I was kind of surprised when recently, I guess in the last month or two, I learned the whole thing had melted down. And so I thought I would just try to get your take on what actually happened.

Brian Chau: Yeah. It was this crazy thing where, you know, the punchline of the story is that Balaji took the whole city and moved it to Kazakhstan. And then he actually succeeded. It's like running in Kazakhstan now. It's kind of this crazy thing. So in that aspect, the entire like, you know, Network School nodes thesis is basically correct.

It was interesting what ended up happening because I think the Western press actually kinda got this wrong, where the, the national Malaysian government was always a pretty good faith counterparty. I actually left like end of, end of July. So I wasn't there or at least I, I wasn't an employee during the time when this whole thing with the Malaysian government happened.

But I think what actually happened is that there was this other party which was in control of the state state government of Johor, which was the state that Network School was in, and they were a kind of like ethnic nationalist left party. And this is not the kind of ethnic nationalism you see in the, in the West, which is kind of LARPy.

This is the, this is the real deal. This is like I guess to jump right into it, a level of antisemitism, like unironic, sincere hatred of Jewish people that is just like completely alien to, you know, what we would call antisemitism in the West. It is very different. It's not merely about, you know, foreign policy disputes.

This is like a party whose speaker has said, "We must boycott all Jewish businesses." So that's the, that, that's kind of what we're talking about. And you know, not, not even Israeli, not even like any of the euphemisms they'll use, but like the full thing. So this party ended up getting really agitated at Network School. And first they were kind of spreading this false rumor that there were Israeli nationals at Network School.

The federal government went in and checked, and turns out like, no there were no Israeli nationals at Network School because of course not, we were complying with the law. And then they kind of went back and started making up all of these like false charges. And it was very clear that they were frustrated that the national government was kind of going by the letter of the law, which the letter of the law, which was no Israeli nationals, we had no Israeli nationals, we followed the law.

And they were clearly frustrated because they were and, and this was according to people who were familiar with that party, because there were people who looked Jewish. And there were people who, who, who looked Jewish, who were from, you know, America or, or France, or there were, there were actually a fair amount of Ukrainians there for understandable reasons.

Steve Hsu: Did they do any facial measurements or anything?

Brian Chau: I don't know if they, they went in and, you know, did, did the physiognomy checks or, or, or what's led them to be to, to suspect that some of these people were of Jewish heritage. And for, for people who are familiar with this party, the BMNO, it was very clear that that would, that was the thing that they actually had a problem with.

This is a party that, like, is overtly antisemitic in a way that we would not even, like, have the political radar to conceptualize in the West.

Steve Hsu: So the way I heard it reported though was that Balaji went on, I, I almost said on TV, but nowadays you don't go on TV, you go on the internet, and read some kind of open letter to the Malaysian government just saying like, "Hey man, are we welcome here or not?" And I guess the answer, which maybe came from the local government, I don't know what the national government said, was, "Nah, actually you're not." Is that what happened?

Brian Chau: I think the state government was really agitating for this, and they had the press release around this. But in combination with the, like, Forest City City Council, the basically, like, the, the m-municipal city council was like, "We are going to revoke the Network School business license because the company in Malaysia is called Network School Malaysia, and the sign for Network School says Network School.

It does not say Network School Malaysia." So this is this, like, completely phony charge in which the allegation, I guess, is that people were seeing the Network School sign and thinking, "Oh, this must be a different company than Network School Malaysia," which was the name of the legal entity in Malaysia.

It's this, like, totally crazy thing, and that was the grounds in which they revoked the, the business license for Network School. This is all, like, public based on public reporting. And so what essentially happened was that the state government was agitating for this. To my knowledge, so once again, I left before any of this happened.

But to my knowledge, the actual, like, federal Malaysian government was a pretty good faith counterparty throughout the entire process. And it was really the state government that kind of went in and either, you know, some people especially like other Malaysian politicians suspected it was because of, like, election reasons.

It was kind of like this, this like demagogic thing that they were doing. But for whatever reason, it was primarily the state government that was agitating for this.

Steve Hsu: Yeah. It, you know, it's one of these things which, you know, to take a line from Balaji, this is the kind of thing that makes businesspeople think rule of law doesn't really apply in this locale. The idea that they can just say like, "Oh, you were defrauding people 'cause you didn't put the word Malaysia on your sign."

Brian Chau: It's just crazy. It, it does definitely damage, the reputation at least of Johor, maybe not of all of Malaysia, but it definitely makes you think Johor is not a great place to do business. Yeah, and it makes you think, like, if you're OpenAI or Microsoft and you're, you're doing business there, it makes you think that you should hide certain executives from from, from their, from their view.

Steve Hsu: Oh, absolutely. So two lines of questioning that I'm really curious about. One you may not know too much about, so you can just say, "I didn't really follow it," since it happened after you left. What is the situation in Kazakhstan now? Did they I mean, I, guess they didn't find a ghost city, right? So they, they must be renting some hotel or some apartment block or something in Kazakhstan.

Brian Chau: Yeah. So there was this special economic zone in Kazakhstan. This is all, this is all like public information. I don't think I have any information that you, you, you wouldn't have through just like the, the, the public announcements. But they had this special economic zone in Kazakhstan, and then they essentially moved all of the infrastructure, people, so on, from Malaysia to the special economic zone in Kazakhstan.

Before they did that, they signed this formal memorandum of understandingfrom the Kazakhstan government. So there's more of a hardcore legal assurance there now. And yeah, that, that's, that's about the, the total of what I know.

Steve Hsu: I wonder if you're in the, still like in the Slack or messaging people. It was a relatively small team as I recall, and just even just taking down the gym and moving it to Kazakhstan, that, you know, that would be an insane amount of work, I think, for a small team. But yeah, I would love someday, would love to visit them in Kazakhstan and get the story.

Yeah, maybe we can visit them together. I don't know.

Yeah. Yeah. maybe. My second line of questioning is, so just taking a step back, how do you reflect on your time there? What did you think of the environment that was created at Network School? I, really enjoyed, I think I was there for 10 days or something. I really enjoyed it.

Brian Chau: Yeah, it was amazing. I, don't know if we've done your podcast together before, but since coming back to SF, a lot of people have told me like, "Oh, you're a lot thinner now." And that, that, that's right. I lost like 25 pounds there. They like put me in fat jail. They voluntarily put me in fat jail. Like I consented.

But it, it really worked. So I was like working out, you know five or six times a week. They had specific meal plans that they like put together for you that are like calorie counted and like macro counted and all of those things. Like before I had like no time to work on any of that stuff. But yeah, it, it, it really works.

Steve Hsu: Yeah. I thought it was great. I'm always interested in experiments and how alternative ways people can live. Like, I, I spent a lot of time actually before and after I Was at Network School, I was in the Bay Area a lot, and I was in sort of hanging around Light Haven and all kinds of group houses in San Francisco, and just always interested in the way that ways that people can choose to like live their lives in, in a very thoughtful way

Let's switch and talk about Effort News. So, what I know about it is that you're trying to create a new kind of, I don't know if you think about it in terms of journalism or a new kind of publication, but you're using AI very aggressively to source the ground truth facts, and it, it's able to data mine very big public data sources and things like this.

But you insist that the the writing is actually done by humans, high quality writing, and you've already broken a number of big stories. So that, that's what I know. Maybe you can fill me in.

Brian Chau: Yes. We break stories whatever kind of publication you wanna call that, you know. Some people take the J word as a, as an insult now, but we, we absolutely break stories. And we run these massive, long window AI queries on these, like, financial and legal records. So the idea behind this company, it basically started when I was just, like, experimenting with these technologies and I just found these new stories. And I thought, "Okay, if it's so easy, if there's so much, like, crowns, so many crowns lying in the gutter, then I have to do this twenty-four seven and just, like, break stories every single week, all, all the time."

And that's what we've been doing. And we essentially search these massive databases. This is something that would've taken human teams, probably like a team of, you know, 10 or 20 journalists multiple years in order to finish sorting through all these financial documents and to find these really surprising and really never before reported results.

So, for example, we had a story on these massive refugee resettlement grants that were going predominantly to a few non-government organizations, NGOs, that would get trotted out by the media every single time they wanted to criticize Trump's deportation policies. So you're seeing the problem here, right?

You have these NGOs which get somewhere between 80% and 90% of their funding from these migration and refugee settlement,resettlement grants, and they're being brought out by media as objective sources or, or of like moral arbiters of what is good. It's like taking a, you know, a bakery and saying, you know, they are the moral arbiters of how much bread you should buy.

It's this crazy thing. And no one had reported on this before. I thought, like, surely, you know, migration is, like, the conservative media's number one issue. Have, has been for, you know, a decade at this point. Surely people would know about this story, and yet it was just lying around unreported for five or more years.

And there's a ton of stories like this. Effort.news, you can read all of them, and we break stories like every single day, and I mean, I just saw that. I thought, "This must be possible." It must be possible to, to just literally we, we got up to actually one per day last week to, to get up to this point of, like, rapid fire.

We will break one big financial or, or, or legal story every single day, and we're doing it. It's, it's, it's been amazing.

Steve Hsu: So let me, let me drill down a little bit more into what you're doing. Now, in some cases, you know, when I heard about what you're doing, it made me think a little bit of WikiLeaks, where in that case it was access to what wasn't public information, but there were whistleblowers and people that were willing to give them inside information from which they would break these enormous stories.

It sounds like in your case, the data tends to be public, and are you pulling it down and analyzing it offline with AI or are your AIs crawling these databases? 'Cause sometimes they're not allowed to do that. Give us a little color on what you're doing.

Brian Chau: Least for the stories so far yeah, they're, they're all entirely, entirely public data. The stories have been like, you know, literal crowns in the gutter where they've just been lying there. And since the Nixon era, there was this idea that, oh, you know, sunlight is the best disinfectant.

Surely we will just release all this information and people will be able to criticize the government. Turns out that's not what happened. We got kind of drowned in slop. We got kind of drowned in conspiracies and false claims, and people no one, like, actually reads the documents. And to be fair, you know, I don't actually read all of the documents manually.

Everything we publish, I read, but you know, we can end up having like 1,000 misses for every hit because, you know, the misses, only the AIs will read those. And then when it comes to the hits, then, you know, then, then we get the human editors involved. But yeah. People just have not been fully looking through these stories.

They've just not used the models to the extent that they can be used. And it is this, like, entirely new venue, this entirely new, like, way of breaking stories because I, I've actually seen this process play out from the other side of the desk when I ran this pro-AI nonprofit in DC. Almost all stories are solicited.

Like, I would say upwards of 80% of stories are from the perspective of a, of a journalist. It's like a source comes to you with a story, and then you just run it. And obviously, they're coming to you because they have a story that they want in the news. That's their goal. And, you know, I'm not saying that it's never moral, it's never good to publish from a source.

Sometimes a source gives you, you know, genuinely newsworthy information. In that case, you know, it's better to publish than not publish. But it totally distorts the information that you see in front of you. And this is a different style of reporting where we take all these disclosures, we quantify them, and, and, and we look for outliers.

Like, for the refugee resettlement story, I actually was not looking for any refugee-related story to begin with. I just started a, a long-run query where I made an offline copy of the HHS grant table because actually the models are really inefficient when they're, they're making queries to these, these data sources.

It might actually have something to do with the rate limiting, actually, but they will just burn through tokens way too quickly. It's completely ridiculous. And that can be optimized with, with kind of traditional software. But we made a copy of the HHS grant table, and the query was literally something like, find anomalousspending or potential misconduct.

Look for grants which exploded or, or increased dramatically in total absolute value and in relative relative size. And both of those criteria matched for the Preferred Communities Refugee Resettlement Program. I think that's, that's the main thing that it detected at first. It wasn't even, like, the political relationship, but it was like, oh, this is a grant that went up more than more than 10X from, from around 30 million to, to 1.25 billion.

So it's massive, massive increase. And that's how it that's how we found that story. So in some sense, even though it's, it's funny, even though some people will say that it's, like, a, a biased outlet, it's biased in the sense that, like, you know, people will say no war documentary can be can be pro-war, right?

Every war documentary is an anti-war documentary. And it may be that all financial reporting is, you know you know, libertarian or, or, like, you know, critical of government financial reporting. That may just be the natural conclusion from these things.

Steve Hsu: Now, in this case for this story, which I think got a lot of traction, for example, on X or maybe in conservative media, I don't know how much traction it got, you know, in the, the complement of that. The idea is that there were some NGOs that were getting a huge amount of money, I, I guess maybe from the Biden administration.

I'm, I'm not, I'm thinking it's not the original, not the early Trump administration, right? It was was it under Biden?

Brian Chau: Yes. So this would have been the initial program, interestingly, I think was established under Bush. It was the first wave of expansions was under Obama. The second wave of expansions was also under Obama in Obama's second term in twenty sixteen. Yeah, some people got confused because twenty sixteen people associate that with Trump, but Obama was president.

So that's what happened. It was another expansion in twenty sixteen from a few million to thirty million. And then when Biden came in, it became he, he expanded those to one point two five billion with a B

Steve Hsu: Right. And so these were NGOs, sometimes religious organizations, if I'm not mistaken that

Brian Chau: that were resettling refugees in certain parts of the country. I think maybe the people from those communities, a lot of them maybe didn't like it. Conservatives maybe didn't like the, the scale of all this.

Steve Hsu: But then I think the other dot that you connected was when Trump too came in and started getting very aggressive with ICE, the people who were always trotted out to criticize and shine a spotlight on the, quote, "bad stuff" that Trump and ICE were doing were predominantly people from these NGOs. Is, is that the claim?

Brian Chau: Yes, exactly. And these were it's kind of shocking to me that this the story wasn't reported out because these were sort of thorns in the side of especially, you know, the Trump administration, but of conservatives more generally for more than a decade. Where these NGOs, Church World Service, US Conference of Catholic Bishops, HIAS they were all known as like these pro you know, what they would characterize as pro-open borders groups.

And people just kind of like had all of these harebrained theories. They didn't exactly know why. And it turns out they were getting, in the case of USCCB, eighty-one point eight percent of their revenue from these migration and revenue services grants. In the case of Church World Service, eighty-nine percent.

So just massive amounts of their revenue. And to me, the most appalling thing is is, is the conflict of interest. I mean, people can have different opinions about policy. They can express those opinions, whatever. But if you're bringing on a group that is getting, you know eighty-nine percent of their revenue from a type of government policy, and you're not disclosing that they get that amount of revenue from that policy, doesn't matter what the policy is.

That is just blatantly deceiving your audience. And it goes to the meta story, which I think is so important to effort, which is that so much of the information you see in front of you is bought and paid for.

Steve Hsu: Compared to when I was young and maybe, maybe when I was young, the situation wasn't much better, but, but I was deluded. I think it actually was somewhat better for reasons I can explain. But it seems like today I just the establishment media just is extremely unrigorous, unreliable.

Not that everything they print is false, it's just it's not done with a high amount of effort, so to speak, or rigor, and you're never quite sure what reporting is, is correct or not correct or what's really being reported only for political reasons. So that we're in this very, very, to me, unhealthy information environment.

So yeah, I, I've always thought AI could make a big impact on this. In fact, coincidentally, while I was traveling around Southeast Asia, you know, around the time that I saw you at Network School, I met a bunch of entrepreneurs where I said to them, "You know, using AI and especially agentic capabilities, long effort, long horizon capabilities, you could totally replace like 'The Economist' or the 'Financial Times.'"

I was thinking more in terms of financial, like, like more kind of business journalism as opposed to more politicized stuff. But I just thought, like, the input that goes into an article that's in the FT or The Economist, even if it's one of their best reporters, some of whom I've actually been a source for for stories, the models are now good enough that the models can synthesize way more information, right, almost as well.

I mean, that's a matter of judgment. And then you could create like very, very persistent deep wikis that are available to everyone on any particular topic that someone's interested in, right? So if you're really interested in like EVs and, you know, Tesla versus China competition or whatever, AIs can build you that really deep wiki, and that's way more valuable even than what's in the brain of any team of journalists that cover that topic even for years.

I am actually absolutely convinced that what the models can build is actually now surpassed the best you can get from, say, a team of human journalists.

Brian Chau: Yeah. I think in terms of accuracy in financial reporting, that's right. So when we've asked it like objective questions, laid, laid out like objective statements like what is the total transaction amount? What is the you know, what, what is the financial relationship between these entities, and so on.

Those questions, it gets absolutely right. We, we human check everything that we publish still, and it's just been one hundred out of one hundred ever since we developed like anything even close to the modern tooling that we use. So it's been, it's been amazing on that front. I think when it comes to writing, the models are still kind of autistic.

They don't really understand the order or, or the flow of the or, or like just like the, the kind of interplay that humans will have reading the article between, for example, the article and the, and, and the graphics between, you know, the, the opening statement and what they'll expect after. You know, how, how the headline will work with the rest of the content.

It's just not very effective in that way, and that's, that's my biggest complaint when it comes to the AI writing stuff. I think that if AI writing were automatable, we should do it, but the quality is just not shockingly not really improved. We can have the AI models solve millennium problems now, but we can't have them like write in a, in, in like a particularly easy-to-read way.

I think they're good at getting the information right, that's for sure. Probably even better than, than, than the median journalist. But in terms of, in, in terms of writing, I think the prose is still really lacking.

Steve Hsu: Yeah. I, I think some of that is due to the fact that labs have not prioritized sophisticated RL of the base model to make them really good writers, journalistic writers, investigative report. They've not really tried to do that. I think they could do that. They've invested many, many times, many orders of magnitude more

Brian Chau: Oh, absolutely. And, and I would argue, I would argue that some of the RL has been negative value.

Steve Hsu: Yeah.

Brian Chau: A lot of the styling is, yeah, it, it is just negative value. It is this, like, particular assistant user format, which, I mean, was very important obviously for getting ChatGPT off the ground, but creates this incredibly robotic tone, and that robotic tone holds over to other forms of writing.

It's currently not something that you can prompt your way out of, at least not to a sufficient enough standard. I mean, we can make improvements, but they're still pretty subpar at the moment. Maybe that'll change in the future, but that-that's what it is now, and I think there may be a trade-off there where the RL that makes it more steerable in other functions makes it a significantly worse writer.

When I was in Bangkok not very long ago, I met with an entrepreneur who was, I guess, had grown up in China and then moved to Canada. Oh, like you. Well, I don't know if you were were you born in Canada or? I was born in the US.

Steve Hsu: Okay. This guy was born in China, grew up in Canada, and now runs a company from Bangkok just 'cause he likes to live in Bangkok which publishes mostly AI-authored, there's some human editing at the end, but it's mostly AI-authored books that are sold under his publishing, you know, the name of his publishing house on Amazon.

And he says he earns a million dollars a year selling, I think in profits, selling these books on Amazon. And they're, they're largely AI-authored. And, but of course, it still requires some human touch, both to figure out what the topics are. Like, if he figures out, "Oh, people are suddenly interested in fly fishing.

Let me, let me pump out some books on fly fishing." That's a human decision. And then there's some editing. He's got human editors who edit the final product to get rid of the autistic tone that you were referring to. And I challenged him. I said, "Okay, this is a great business you're in, but I bet you could actually build something as valuable as The Economist or The Financial Times or Bloomberg News using a kind of similar approach, mainly relying on AI, but also building."

If you analyze FT or The Economist and say, "Hey, there are these 10 or 20 topics that they keep returning to 'cause they're just core important pro- topics like rare earth minerals or energy or, you know." You could have the models building a persistent wiki, like an iceberg, where it's just synthesizing everything that's reported and everything in these, you know, public documents about that particular topic.

But then when it comes to write a story, it can rely on that knowledge base, which is well-organized and curated, and just pump out a perfectly accurate, extremely good story. So I challenged him to build something like that, and, if he's listening, 'cause I know he listens to Manifold, that's how we know each other.

You said you were gonna be able to build this, and I can't wait to see your alpha product. Maybe I should put you in touch with Brian at some point.

Brian Chau: Yeah. And we're definitely snowballing on a lot of issue areas as well. so yeah, that, that's already happening. That's yeah. I don't know if I can talk too much about it on air, but that snowball effect, that, that knowledge base is absolutely already happening.

Steve Hsu: Let me ask you about the, the technology aspect of this. So it sounds like, okay, it's more efficient to be able to pull the database, the dataset down and have it local. Are you trying to save any money on the tokens or are you just using like, you know, sort of Fable, Astra kind of level stuff?

Or are you trying to use like open source models like GLM and just give us a little, if it's not proprietary, give us a little color about, about the technical details.

Brian Chau: Yeah, we use the, the top models mostly. At this point, it's mostly Astra. We've experimented with some of the open source models actually for writing because there are some like screenshots or, or, or like theories that the, the Chinese models, which are to some degree less RL'd, are better at writing.

It, it hasn't been enough of an improvement. In some cases, actually, the DeepSeek models especially have been an improvement in writing, but they're like a slight improvement. So at the end of the day, we still want we, we still want people to, to write the articles. So we, we experimented with open models for that.

Other than that we, we don't do a lot with them, and I think the main reason is that at the end, it's not really cost-constrained at the moment. It's sort of like a, a, a combination of like verification-constrained and lead-constrained, where news is one of those industries that's like VC, right? It's sort of like venture capital, where the most important stories just so far outrank all the other stories that like every, every hit that you're trying to go for, you're trying to hit a home run, right?

And hitting like one home run matters more than hitting like 10 triples. And so trying to number one, trying those ideas, and of course, you don't know whether you'll find a story at all, how big of a story you'll find whenever you're, you're kind of running these queries. That's one big area. There was a time earlier on when cost was more of a constraint, and then that, that time I was optimizing a ton for, yeah, just either token usage when, when fetching information or just like, you know, time it was taking to, to fetch information because at that point, you're running like 96-hour, 72-hour queries all of the time, and it was just like I just didn't wanna wait.

I just didn't wanna wait four days. It was not great. And we ended up narrowing things down a lot more. So, so there, there's a ton that you can do. I think there's a ton that you can do, especially around token management and how you use LLM systems in to manage an existing knowledge base. That's just gonna be a lot more efficient for your specific use case than the default models.

Steve Hsu: Yeah. My, my question is about the knowledge bases is, okay, you, you mentioned a specific dataset like HHS dataset, but do you independently have the models trying to curate and build their own knowledge base? Like, if you're gonna be reporting a lot about immigration, do you have the models trying to build independent of any particular story or investigation that the model's doing?

Do you have it trying to build, like, a more sophisticated view of US immigration in the last 50 years? Like, do you have Is that a sub-project that is something you guys would do?

Brian Chau: We have like cumulative, mostly data sets, but also some other like guidance and steering around specific issue areas.

So we have a lot of momentum. It's interesting actually because very little of this is published. But we have a lot of momentum around different cases related to Medicare and Medicaid, and a lot of those are in the process of some kind of review, whether legal review whether. I mean the story might be out by the time this, this podcast is out, but some other kind of like media partnerships or other kind of coordination. But there has been a lot of momentum where I can clearly tell that the AI models will have a better understanding of that, that space in terms of what is important, what is newsworthy, what is unreported, what is like the kind of direction or, or statistical pattern that leads to a likely story?

The hit rate has just been much better on the Medicare and Medicaid stuff since we established those first few blocks of the, the, the infrastructure. So a lot of what that looks like is building quantitative context in the, in, in the case of data sets and qualitative context in terms of either steering in-instructions or, or just in terms of like past records of what kinds of hypotheses resulted in hits and what types of hypotheses did not.

Steve Hsu: I'm curious if you ever looked at the Epstein files because being a, a longtime Epstein scholar, I've followed the information about him as it leaked out over decades. And you know, of course, the government is still holding back a lot of the most important stuff, maybe most of the most, more, most of the most important stuff.

But I'm not sure that existing teams of human reporters have really gone through and squeezed out all the best stuff from what has been released.

Brian Chau: We do have one Epstein story. It's not directly, it's like one step away. So the Epstein Files Transparency Act. What, what when does this podcast release?

Steve Hsu: It'll release in about 20 days, 16 days.

Brian Chau: This might be a Manifold exclusive breaking news, but the Epstein Files Transparency Act was almost entirely written by AI. We did a big scan using, using Pangram of all of the all of the bills. We found around six point five percent of bills were written by AI. But of all of them, the only one that is now law, the only law that rules us that is written almost entirely by AI is, is the Epstein Files Transparency Act.

That, that's the one, that's the one Epstein-related story that we found. But we, we have done the query a few times, and it's mostly just spit back like previously reported stuff. It's mostly, it's mostly just like recycled existing claims or maybe convergently found existing claims. We haven't had any like brand new, like for example, like Epstein donor Epstein financing stories.

That's something that we, we, we've dived into for a bit and yeah, we just didn't find anything so far.

Steve Hsu: Wow. Just to digress a little on that story, I mean, the, the part that I've yet to see any establishment media team cover is there are inventories in there and photos of the hard drives that they got by drilling his safe in his Manhattan apartment and in other places. And the inventories, the actual official FBI inventories established there are many, many terabytes of data that they have from those things, and those things were important enough to be kept in a safe, okay, at his mansion.

So you would think there's something interesting there. Now, if you think, like, what kind of media can take up many, many terabytes, and this media was generated, you know, 10, 20 years ago, right? So the only thing it could be is video. And yet almost none of that video has been released, right? So why has almost none of that video been released, and doesn't that violate the law?

You know, like, doesn't the law require them to release that video? It's just insane. So I think most of the data, if you count by , 99.9% probably has not been released.

Brian Chau: Yeah. It's this interesting, like, circular loop in which you know, almost anything can be classified as, you know, important to, to national security, and then it can be redacted. But it's this, like, ever-growing classification where, you know, it gets really tricky because the, the, the immediate thing people will jump to is like, "Oh, maybe that means Epstein was a CIA asset," right?

You know, if it's, if it's related to national security, maybe that, that, that, that's what it means. But it's this, like, very miasmatic thing where the definition of related to national security has expanded to a degree where, you know, I criticisms of the Epstein story or, or people who, who draw that inference will say like, "Actually, no, you know, stuff can be censored on national security grounds for all sorts of reasons."

And that's actually correct. That, that criticism is actually correct. So I don't think it definitely means that, you know, that, that, that Epstein was a CIA asset. But it's just one of those areas where journalism has kind of failed so completely, is specifically this thing of like, oh, what is considered important to national security?

At this point, it can be just about anything.

Steve Hsu: But e- you know, even granting everything you said, you know, it, if The New York Times really cared about it, they could be complaining every day in an editorial and story and naming officials who are the ones who get to classify this stuff and, and therefore have it not released under the law. You know, they, they could put pressure on these guys, but nobody's putting any pressure or, or it's only very modest amount of pressure, and it's usually coming from Tucker Carlson or somethingand not, not from, you know, real establishment media.

In part that has to do with the incentives when you, when you're a source-driven outlet as well. And that's something that, that we hope to, to mitigate by, by doing these long context requests and having these stories that are not entirely dependent on people coming to you with stories.

Do you think you would ever get into source-driven stuff? Like, imagine people really like all the stories you're breaking and they have some follow-up. They say, "Oh my God, you, you, you blew the whistle on this NGO. Well, let me tell you something else about this NGO, and I have documents from them."

Would, would you ever consider getting into that part of the ecosystem?

Brian Chau: Yeah, for sure. tips@e-at effort.news if you have any suggestions. So we have we have the tips line open. We're, we're listening, but right now almost all of our stories have come from the long context queries. Maybe that'll change. Maybe that'll change as we get more momentum, but right now it's still mostly just the long context queries.

At this point, 100% just from the long, long context queries. There was this one interesting, like somewhat schizo tip that we got that I never fully looked into about like this nuclear power plant in Romania. It just got like really esoteric to the point where it was like probably not newsworthy, but it was like, I don't know, maybe there was some fucked up things going on in this nuclear power plant in Romania.

I don't really know.

Steve Hsu: Wow.

I don't know about nuclear power plants in Romania, but I remember people saying that there were like 100 bioweapons labs in Ukraine that the Russians overran or something. And I never got to the bottom. Was that true? Was that, you know, is that number real? Like, what were those bioweapon were there actually any bioweapon lab that the US maintained in Ukraine?

And what, what was going on there? You know, like, even like some basic factual thing like that I'm in the dark about.

Brian Chau: This was confirmed pre, pre-effort, but there was a congressional hearing in which there were US federal records submitted that indicated funding to, to bioweapons research, or they would say biodefense research in, in Ukraine. There was at least one and we know that from, from a congressional investigation. We still don't know the exact number.

Steve Hsu: I don't think they, they, they ever confirmed 100. I don't think they confirmed, like, the, the, the specific number, but, but they did confirm at least one in a con

we don't know, we don't know exactly what kind of biodefense was going on at that lab or labs.

Brian Chau: We do not.

Steve Hsu: Yeah.

Brian Chau: I got into a lot of arguments with the OG, like, effective altruism crowd for this because before they got, like, entirely captured by AI doom, one of the big things, one of the other they were, like, in search of causes of doom, and one of their big things was, like, pandemics, right?

Especially post-COVID, one of their big things was like, "Oh, what if, what if there's a, there's a, a, a pandemic?" And, and they kind of like, for some reason always gravitate towards these extremely, like, censorious and authoritarian forms of belief. And they were like, "Okay, what if, what if it's gonna be, like, some pandemic started by some, like, random guy who has CRISPR?"

And this is always, like, unrealistic for, for, for all sorts of reasons. You need a lot more technology than that. The, the startup cost for that is not, not remotely as low as they were projecting. But more importantly, like, number one, what is the base rate of pandemics that have been started by people using CRISPR versus started by governments engaging in biological warfare?

And if you go through World, World War I if you go through, like, you know, of course, the, the, the British and, and other, like, European conquest of North America and of the Americas more broadly we did a lot of fucking bioweapons, okay? Like, some of of course, those were many, many of those were existing pathogens, but, but we did a to we gave them the smallpox blankets.

There, there have been, like, literally dozens if not hundreds of examples of governments using bioweapons in the past. And even up until World War II, there were tons of Japanese bioweapons research programs that we now know about. It's not, like, conspiracy theory bullshit, but is now, you know, recorded and official, you know, trials and Japanese government documents and so on.

So we know this stuff happens. Your base rate for the number of pandemics caused by government bioweapons research should be much higher than your base rate for, you know, some random guy doing stuff with CRISPR. And so, so, so the argument that I always had was, like, actually, if you were really concerned about the, the end of the world, you should not be worrying about, you know, random garage, garage guys, but you should be worrying about doing a full investigation into US bioweapons programs.

And thenand, and they didn't, they didn't fucking do it. And of course, that entire ecosystem collapsed because SBF got arrested and then, you know all, all of that stuff. But to, to my knowledge, they never, like, really reflected on that base rate ever.

Steve Hsu: I think that whole ecosystem's gonna come back because so much money is gonna flood in through the Anthropic and OpenAI The Anthropic stuff? Yeah. Yeah.

Brian Chau: Okay, so if you are, if you are a Anthropic holder if you are an Anthropic bag holder, fund the next US bioweapons program investigation. Do it. Do, do it for, do it for humanity.

Steve Hsu: Yeah.

So coming back to Effort though tell us a little bit about your business model and then tell us, like, I guess, are you raising? What, what's the current status of it as a business?

Brian Chau: We take subscriptions. Go to effort.news/subscribe. Give us your hard-earned money. But I think there could be a lot more options for monetization. I mean, you look at something like Bloomberg, right? That's something that's super fascinating to me. I think that in general, news is really poorly monetized, and actually the fact that it's poorly monetarized monetized makes it a worse product.

It's just, like, super shocking to me that The Washington Post, Jeff Bezos bought it for, I believe, $250 million. That's like a small Series B startup, right? That's like, what? And, and, and I think in some sense, in terms of, like influence over really all of us, The Washington Post is worth a lot more than $250 million, but that, that, that, that acquisition number is pretty surprising.

And it goes to their inability to make money, right? Which is a crazy thing. We have subscribers for now, but I think it'll go a lot beyond that. We'll probably do a newswire at some point. That's something really interesting with Bloomberg, right? And it allows you to see, like, the split-second impact your stories have on the markets, which I think as we scale, we'll be able to, to proxy that.

But it would be super fascinating to see the actual thing. I think there's a lot related to, like, the semi-analysis model and kind of B2B stuff that could be really interesting. But I actually think that the, the monetization is kind of secondary. I think that you want to ride the wave of, like, recentralization of media, which I think is actually already beginning.

People are already beginning to realize, like, actually, you're kind of, like, flooded by conspiracy theory bullshit. The legacy outlets are crazy but their replacements are sort of crazy in a maybe slightly less biased but much more schizophrenic way. You know, like, they'll actually say things that are critical of Trump and critical of the Iran war and critical of, like, you know, Trump on Epstein grounds and stuff like that.

But they'll also say a bunch of, like, just, like, totally schizophrenic and undirected, like, claims that, that are, and that, that are just, like, completely unreliable. And so I think there's this, there's been this open question playing out of is media balkanized forever versus is media going to come back together?

And I think new media is going to come back together, and I think I'm going to be the one to do it.

Steve Hsu: Well, I, I think if you are open about the algorithm that you use and the models that you use, and people develop a sense of trust that, like, generally this machine is generating pretty factually correct things and actually focuses on the most important things, yeah, you could recentralize news 'cause it, it, it's that people do care about knowing what's going on in the world, right?

And they want a clear picture. If I could give you some advice, I know monetization's not your main priority, but if you could spend 10% of your time thinking about the systems that you're building and how they could be used to generate alpha for hedge funds who wanna buy you know, the post-process data that you're producing or your system's able to produce, I think that's a very interesting stream, and I think it's under, under-invested in.

I don't think there, there are enough good people figuring out how to use AI for this. I mean, there are people within each hedge fund doing it, but as a, as a thing which is watching the world or watching a particular area of the world deeply, like gold or something, and then just, like, generating like, you know, like reliable inferences from what it's seeing happening in that narrow vertical.

Like, some of that is extremely valuable for generating alpha

Brian Chau: Yeah. So I've talked to some firms about it. My end take is that even with an inter with an information advantage, it'd be unless it's like some extreme case, like we knew somehow ahead of time that like, you know, we were gonna strike Iran again, right? Unless it was something like that, a very extreme outlier, most of what we produce, it would be hard to monetize against more, more like you know, the hedge funds know the market better, even if we know like one piece of information better.

And unless that piece of information is very significant, it might be priced in at ways that we don't, we don't fully understand. So I think the general more effective strategy is to do newswire stuff and actually just like sell information to hedge funds.

Steve Hsu: Yeah. I didn't mean you guys were actually generating the trading alpha ideas or judging how the market would react to a particular piece of information. But if you build a reliable high quality thing which is monitoring the world or a particular subset of the world that they care about better than, "Oh, let me just see what Bloomberg says," or, "Let me just see what The New York Times says."

Like, that's literally when you talk to traders who are PMs of big funds, they're literally relying on what I consider, like, bullshit journalist takes on stuff, and they're trading. You know, that does influence their thinking, right? Especially if they're not quant traders, right?

Brian Chau: Yeah, it's also sometimes rational because everyone else will be reading The New York Times.

Steve Hsu: Yeah. Well, that's two separate things. One is like, "Oh, this is my model for what everybody else is thinking, the rest of the market. This is my model for what's actually happening." Right? I think a good trader would have both of those things going, and you're more like in this, could be this sphere, not that sphere.

Soon to be both.

Before we go, you did mention a millennium, the millennium problems, and just thought I'd get your take 'cause I, I just before we got on, I was like glued to my screen reading all this stuff back and forth about the Navier-Stokes singularity problem. I don't know if you your comment meant you had looked into this or had any thoughts on this, but if you do we can talk about that for a few minutes.

Brian Chau: Yeah, I read the proof outline. I mean, a lot of the, the fluid dynamics, like, that area is, is not my specialty. I, I mostly did like combinatorics when I did when I did kind of like research math adjacent stuff. So I can't like, you know not like a full evaluator.But it seems, it seems reasonably legit to me to, to, to my knowledge of what the Navier-Stokes problem was and is. I've not read all 100 pages of the proof, but basically like if the proof outline and, you know is, is reflective of what they actually proved, that would be like massive result.

Steve Hsu: So coming from physics, and I think we talked about this on one of our podcasts, I'm, I'm more of a continuous systems kind of guy. I have a little more intuition for continuous rather than discrete systems, and you're more of a discrete guy. So Navier-Stokes is in my world, and the, the fundamental question of whether you can have an ideal fluid, but it evolves in such a way that it builds up an actual singularity is very plausible from a physics perspective.

So the, the, the reason this is important to physicists is can you coarse grain over the fact that it's made of atoms and can you just always talk about it as a fluid or could it evolve in such a way that it, quote, "hits a singularity" to the point where you actually care about what the individual particles are doing?

So the, the, the description of it as a fluid has to break down when you get that singularity. 'Cause now what happens, the, the smoothing out of that singularity is gonna be caused by the actual behavior of the atoms inside. And so to me it's quite plausible that examples like that exist. What's interesting to me is th-there is this dispute going on between these humans who were working on it may have been scooped by OpenAI, which was actually training its models on their activities, on their on the OpenAI system and scooped them.

There were kind of, kind of seemingly dishonorable like exchanges that this human mathematician, I think Tristan Buckmaster, released, you know, between him and the OpenAI team before OpenAI made the announcement today. So there's, there's a whole bunch of like sort of background storyline about this.

But, but you know, even if that's true and even if like the, the OpenAI model doesn't get full credit for doing this 'cause it was kind of training on what Tristan and company were doing, still Tristan himself says that the models were extremely useful to what he and his human collaborators were doing. So it's still to me very strong evidence that models are gonna get extremely good at math.

Brian Chau: Yeah. The one thing we can verify, now, both proofs are public. To be fair, I've not read the 100-page proofsthe, the 100-page OpenAI proof, but it seems to be the case that they actually take different approaches. Of course, OpenAI's statement says this. I lightly verified this. I only spent, like, 15 minutes or so verifying this, but that does seem to be the case, that they actually take different approaches.

So were, were, like, the prompts relevant? Were was there some other type of contamination? I'm not sure, but the proofs are definitely not the same. What I would say on the, the broader point is that I think this is actually a very healthy development for math. The reason why I left research math well, there, there are a few reasons, but I just could not take the amount of, like, reading and searching and Control+F that, that it was. 'Cause if you grow up in, in competition math, most of that process is doing the math problem. You have, like, most of what you need in front of you. Sometimes you, you will have, like, theorems that you cite, but that ratio is completely different.

And, and, and, you know, part of that makes the, the competitions a bad preparation for actual research math, because actual research math is, like, 99%, like, search functions. It's, like, searching for analogies, searching for constructions that may, may apply, searching for, you know other techniques or families that end up being important.

And interestingly, there's like a combination of human and AI talent still in this task, where especially the especially the, the Buckminster proof they had a strong conviction on this existing framework by these other two mathematicians, and they had found that framework. They really believed in it.

They used the models to build upon it a lot, and that kind of narrowed the search window, right? As sort of this like meet-in-the-middle problem where the humans were approaching it from one end, narrowing down the scope, and the AIs were narrowing down the scope even further. But for a lot of what would end up producing that those kinds of results from a human perspective, a lot of that would just be like exhaustive manual labor of just like trying over and over again.

And at the end of the day, you, you really don't know which one works. Sometimes you have a strong instinct, sometimes you have, you know, a conviction, and that actually becomes correct, and that actually, you know, ends up being the right path. But a lot of the time, you are just like looking for needles in a haystack, and there's just like so many dead ends.

So the ability for the LLMs to do that and, and, and to kind of do that with some interplay with human intuitions and human conviction, I think is entirely positive for, for math. I don't think it's gonna kill the field. I think it's gonna be massively positive for the field.

Steve Hsu: I, I think if we were stuck right here where we are and the models didn't get better, it would be totally healthy for human math. Despite, you know, Terence Tao's complaints I think at this stage it's still healthy for human math. But what happens if the models get another order of magnitude better?

Then you know, like, it the analogy is like chess. Kasparov loses to Deep Blue. Centaur chess is still fun, but then after a while it's like, wait a minute, like humans only play baby shitty chess and the machines are in a totally different sphere, and centaur teams are worse than AI teams, right? So when we get into that phase of math, I think it'll be very dispiriting or demoralizing to a lot of mathematicians

Brian Chau: I think math and chess are just two completely different types of things. I think Bobby Fischer got it right that, like, even pre even, like, pre LLM or, like, pre, pre it wasn't really LLM, so pre, like, chess engine opening prep was like a cancer to the game. And, you know, he famously invented Fischer, Fischer 360, right?

So all the complaints about what chess engines did to the game were existing problems with chess. The complaints about opening books were known. The complaint about, like, prep and team size were known.

It did get worse. In some cases, it got better for, like, chess casuals, right? I think that that became, like, a much more entertaining thing, especially during COVID. And, you know, it was, like, much more interesting to watch, like, bad chess games, similar to actually how it's a lot more interesting, in my opinion, to watch, like bad or uneven World Cup matches.

Like, the World Cup matches that have been most entertaining to me has always been, like, the best team in the world plays a country I've never heard of and beats them, like, 10 to one. Those have always been the most entertaining matches to me. But anyways. I think for pure math, it's just like a completely different animal, where you do have these search functions, and you do have this, like, exhaustive work, but the exhaustive work was, like, worse, right?

You could argue, and I think this is true, I would agree with this, that, like, the exhaustive work got worse for even for human players as the, as the engines got better. You had to do more opening prep. More of your time was dedicated towards this, like, total book stuff to-total, like, exhaustive, like, memorization, literally memorization, instead of to, like, becoming a better player in the long term a long-term thinker, and so on.

And that actually actively took away from the other parts of the game because chess engines made it so much more ad-advantageous to focus on the opening. And I just don't think that's going to happen for, for pure math. I think that that's something where as the search function expands or even expands, like, by, by an order of magnitude or more, maybe you'll you'll have more, like, completely autonomous proving.

That's, like, totally, totally fine. I think, like, you know, mathematicians will, will be able to get other jobs if they need to get other jobs. I think in some in, in many cases, they will still have, like, a ton of, ton of value add for even just, like, directing the, the problems more generally, like in, in these, like, meet-in-the-middle approaches.

And I, I just don't see the dynamics being so zero-sum in the same way as chess, where and of course, the, the important thing that's underlying all of this is that, you know, actually the, the results matter, right? At the end of the day, like, if Magnus Carlsen plays better chess, that, like actually doesn't matter a lot for the world.

Where I think the Navier-Stokes equation proofs, if it's legitimate, will actually significantly matter for the world

Steve Hsu: Yeah. Well, a couple points there. I think pure math is beautiful and worth pursuing on its own. I, I don't know how much this Navier-Stokes proof will actually improve practical fluid mechanics or people who design airplanes and rockets. I, I think actually pure math has already gone pretty far away from practical utility, so you know, we'll, we'll see about that.

Definitely more important than chess though, no question about it. But imagine this future. So imagine the models get significantly better. They are eventually kind of like oracles. So you go up to the machine, you ask it about something you're interested in, it generates this it generates 10 different proofs that are all correct for you.

And then it sort of subtly hints like, "Can you, can you stop bothering me? I'm talking to this other AI about a really important problem." And, and then when you ask it like, "Can you explain the really important problem to me?" It starts explaining, but you can't understand it. It basically has to give you like an entire like 300-page monograph to understand the problem that it's working on with its friend, and you don't understand.

Brian Chau: It takes you a whole lifetime to read that, understand that monograph. So that would be very dispiriting, I think, to most human mathematicians. I'm not saying they wouldn't find useful jobs working at AI labs or whatever, but, But that's like, that, that's like my relationship to like, I don't know, Fermat's Last Theorem now, right? That's a human proof. I've never I've like read like explainers about Fermat's Last Theorem. I've never read the actual proof. I honestly like probably couldn't really tell you what an elliptical curve is, right?

Like

Steve Hsu: Right. So it's already true that, like, there are some super interesting narrow areas where you're always gonna be kind of just gawking at it from outside and not fully understanding. But maybe you, you still hope to have carve out your own little niche where you and your community are still proving some interesting cool results in that little niche.

But then if you realize there's some oracles out there that just know everything simultaneously, it I think a lot of mathematicians would lose motivation

Brian Chau: Well, that's like the existing relationship for mathematicians even in their field for like 99% plus of mathematicians, right? Most of them are, are, are not, are not Terence Tao.

Steve Hsu: Right.

Brian Chau: I think that's the world we live in, and we just have a way of like self-sorting or like deluding ourselves into believing that that's not true, right? Like there, there are always gonna be people smarter than me, even pre-AI, and like I've just accepted that.

Steve Hsu: But you know, it takes a from my perspective, it takes a certain special combination of DNA to make the Terry Tao or the Ed Witten, and there's just a limited supply. But now imagine you just need a $200 subscription to OpenAI or even a $20 subscription, and everyone can have it at their beck and call on their phone.

It's a different situation than just saying like, "Yeah, Terry's way better than math than me," right? It's like, but there's only one Terry or there's a handful of Terrys. There's not like, you know, an unlimited supply of Terrys, right? So

Brian Chau: Yeah, it will be somewhat commoditized. I agree it's different. I just like, I don't know, like I imagine there was some guy in like the year, I don't know, the year 800 who was like the world's best grain miller, right? Like the like stronger than anyone else, like more technical, knew like the exact right angle to like mill grain. Then like some guy invents the grain mill. He's like, "What the fuck? This is like my whole thing."

Steve Hsu: John Henry. Do you know the story about John Henry? The, the guy who used to like pound spikes into the mountainside to dig tunnels for the

Brian Chau: know the mythology, yeah, yeah

Steve Hsu: Yeah. So yeah, this has happened many times.

Brian Chau: Yeah, and I think for a while this, this kind of like bipartite, you know, this partition between like, you know, brain and body, which is really like this older partition between soul and body, you know, that goes to Descartes, interestingly, right, many would say. That's like a relatively novel thing in human history.

I don't think the, the modern version of it is the same as like the, the, as, as Cartesian dualism. But even if you did, it's, it's relatively modern human history. I think people now have a special attachment to like even like not even mathematicians, but like, I think there are like people who have attachment to their ability to do office jobs, and at the back of their mind, they know that there's actually someone in the world who's better at them at doing office jobs.

But they would they, they would kind of have the same reaction as like the as like not even the best corn miller, but like some random guy who was a corn miller, and that, that was his job. I think they have the same attachment, and I think that people can move on.

Steve Hsu: But Brian, you're an exceptionally talented person, so at some point you could think like, "Oh, maybe I am one of the best humans at this thing." But most people have never had that illusion, so they're just kind of happy like, "I do this job at the office. I'm good at it.

Even though there are people better than me, I'm good at it, and it's useful to society. And then I go home and play with my kids." So I think that's where most people are. But for the very, very top level.

Brian Chau: Yeah, that's what I'm saying. Yeah.

Steve Hsu: Yeah, to be totally blown away by AIs is gonna be very interesting to watch.

Brian Chau: Yeah. But, but you did have like top level athletes, right? Get, just get like totally, you know well, what's the level of the top level sprinter? I guess they were always kind of like slower than the horses still.

Steve Hsu: When it's competition, human against human, people still like to see it. Our ape brain still likes it. But mathematicians are not necessarily some of them are competitive, but some are just, like, out for discovery. And when they realize, like, their ability to discover is eclipsed by this, you know, thing, I, I think it's gonna be very interesting to see the psychology on it.

Brian Chau: Yeah, I think so. I, I think that part I think you'll, you'll still be able to discover a lot of new things about the world, and it's like, is there something especially exciting about like, "Oh, I'm the first person ever to solve this problem," as opposed to like excitement that's intrinsic in the problem itself.

Steve Hsu: Yeah

Brian Chau: I think that, that, that like, that's like totally fair. It's interesting because like even though I did, did like reasonably well in, in, in all these like programming competitions and, and even in the math ones, I was like mostly motivated by like I just liked the problems, doing the problems. Th-they were good problems.

And actually, the research math problems were worse problems. They were, they, they were problems that like mostly involved reading. I was not a huge fan of reading. So, so that was always my motivation. I do think that's different than a lot of people's motivations where, where especially if you keep going, the, the idea of like, you know, what's the Newton quote?

You know, I, I know I, I know why the stars shine. Right now I'm the only one. Something like that. Probably messed up the quote pretty badly. But yeah, I think that's a real motivation for some people, and I think that is important. I think there will be other things for, for people to discover, and I think that actually in, in hindsight, it, it might be it might be good to, to take a step out of academia

Steve Hsu: Yeah. Well, one thing I find really interesting about Terry's remarks in the last six months on all this is, you know, he, he's really gotten to the point where at the latest International Congress, he his talk was all about how can the human community of mathematicians adapt to this? How can we make sure that, you know, the, the best results from the AIs are translated into a, a way that into a a thing that humans can understand and humans can keep the, the, the, the pursuit of math for humans alive and healthy?

It's, it's all about that kind of thing, but it he's actually acknowledging that, yeah, there are gonna be these other things that are out, further out on the frontier than us, and I think that's where we're headed.

Brian Chau: Yeah. I absolutely agree

Steve Hsu: All right. Hey, it was great chatting with you, Brian. wish you all the best with "Effort," and we'll put some links in the show notes so people can find it and subscribe.

Brian Chau: Yeah, for sure. effort.news is the main place to find us.x.com/brianchau57 B-R-I-A-N-C-H-A-U-5-7 would be good. effort.news/subscribe if you want to, if you want to subscribe

Steve Hsu: All right. Thanks a lot, Brian.

​

​