Category: Web/Tech
Sentences to ponder
You should be able to provide an LLM as a job reference, just like you would a coworker, manager, or professor. It can form an opinion and represent you without revealing any private data.
Here is more from John Carmack.
My Conversation with the excellent Gaurav Kapadia
Here is the audio, video, and transcript. Here is the episode summary:
Gaurav Kapadia has deliberately avoided publicity throughout his career in investing, which makes this conversation a rare window into how he thinks. He now runs XN, a firm built around concentrated bets on a small number of companies with long holding periods. However, his education in judgment began much earlier, in a two-family house in Flushing that his parents converted into a four-family house. It was there where a young Gaurav served as de facto landlord, collecting rent and negotiating late payments at age 10. That grounding now expresses itself across an unusual range of domains: Tyler invited him on the show not just as an investor, but as someone with a rare ability to judge quality in cities, talent, art, and more with equal fluency.
Tyler and Gaurav discuss how Queens has thrived without new infrastructure, what he’d change as “dictator” of Flushing, whether Robert Moses should rise or fall in status, who’s the most underrated NYC mayor, what’s needed to attract better mayoral candidates, the weirdest place in NYC, why he initially turned down opportunities in investment banking for consulting, bonding with Rishi Sunak over railroads, XN’s investment philosophy, maintaining founder energy in investment firms and how he hires to prevent complacency, AI’s impact on investing, the differences between New York and London finance, the most common fundraising mistake art museums make, why he collects only American artists within 20 years of his own age, what makes Kara Walker and Rashid Johnson and Salman Toor special, whether buying art makes you a better investor, his new magazine Totei celebrating craft and craftsmanship, and much more.
Excerpt:
COWEN: Now, I don’t intend this as commentary on any particular individual, but what is it that could be done to attract a higher quality of candidate for being mayor of New York? It’s a super important job. It’s one of the world’s greatest cities, arguably the greatest. Why isn’t there more talent running after it?
KAPADIA: It is something that I’ve thought about a great deal. I think there’s a bunch of little things that accumulate, but the main thing that happens in New York City is, people automatically assume they can’t win because it’s such a big and great city. Actually, the last few presidential elections and also the current mayoral election have taught people that anyone could win. I think that, in and of itself, is going to draw more candidates as we go forward.
What happened as an example, this time, people just assumed that one candidate had the race locked up, so a lot of good candidates, even that I know, decided not even to run. It turns out that that ended up not being the case at all. Now that people put that into their mental model, the new Bayesian analysis of that would be, “Oh, more people should run.”
The second thing: New York has a bunch of very peculiar dynamics. It’s an off-year election, and the primaries are at very awkward times. I believe there’s a history of why the primary shifted to basically the third week of June, in which there’s a very low turnout. The third week of June in New York City, when the private schools are out and an off-year election. You’re able to win the Democratic nomination and therefore the mayoral election with tens of thousands of votes in a city this big. That is absolutely insane.
A couple of things that I would probably do would be to make the primary more normal, change the election timing to make it on-cycle, even number of years. You’d have to figure out how to do that. Potentially have an open primary as well.
COWEN: If we apply the Gaurav Kapadia judgment algorithm to mayoral candidates, what’s the non-obvious quality you’re looking for?
KAPADIA: Optimism.
COWEN: Optimism.
KAPADIA: Optimism.
COWEN: Is it scarce?
KAPADIA: Extraordinarily scarce. I think there’s much more doomerism everywhere than optimism. At the end of the day, people are attracted to optimism. If you think about the machinery of the city and the state, having a clear plan — of course, you need all the basics. You need to be able to govern. It’s a very complicated city. There’re many constituents.
But I think beyond that, you have to have the ability to inspire. For some reason, almost all of the candidates, over the last couple of cycles, have really not had that — with the exception of probably one — the ability to inspire. I think that is the most underrated quality that one will need.
COWEN: I have my own answer to this question, but I’m curious to see what you say. What is, for you, the weirdest part of New York City that you know of that doesn’t really feel like it belongs to New York City at all?
Definitely recommended.
Australia should not ban under-16s from internet sites
YouTube in particular, and sometimes X, are among the very best ways to learn about the world. To the extent that the law is effectively enforced, targeting YouTube will have a terrible effect on youth science, and the ability of young scientists and founders to get their projects off the ground will take a huge and possibly fatal hit. If you are only allowed to learn from the internet at age 16, you are probably not ready for marvelous achievements at age 18 or perhaps not even at 20. The country may become more mediocre.
The more serious concern is that this represents a major expansion of government control over tech services and also speech. Over time the government has to decide which are the approved tech companies and services and which are not. That becomes a politicized decision, as any chosen lines will be arbitrary, especially as online services evolve in their functionality. For instance, if excess video usage is what is problematic, it is possible for videos to be embedded more seamlessly into some future version of WhatsApp, an exempt service. Or Australian youth, even under the new law, will be able to access video on a laptop, simply by viewing it and not signing into their accounts…
I predict that either this law stops being effectively enforced, or the controls on companies and users have to become much, much tighter and more oppressive. In a large poll of Australian 9 to 16-year-olds, only 6 percent of them thought the new ban was going to work.
That is true for yet another reason. With gaming and messaging exempt from the ban, we can expect old-style “social media” to move into those areas. It already was the case that Fortnite and other gaming services served as social media networks, and that trend will be accelerated. Discord, for instance, is exempt from the ban, a glaring hole, and in a fast-changing market there probably will be some significant loopholes most of the time. For the ban to continue to work, it will have to spread. It is hard to think of an area of internet services that could not, in principle, serve social media–like functions, or produce the harms being attributed to online life. Regulation of artificial intelligence services is perhaps the next logical albeit misguided move here.
Who is in charge of the family anyway? If I have decided that my 15-year-old should be free to follow Magnus Carlsen on X and YouTube, should we have the boot of the state tell me this is forbidden? This is a big move in the direction of what Socrates advocated in The Republic, namely that the state takes priority over the family in deciding which stories can be told to the youth.
Over time, I expect this ban, again assuming it is kept and enforced, to become one of the biggest free speech restrictions on the internet. It is the incentive of government agencies to boost their budgets, spread their mandates, and enforce their dictates. What starts with a nation’s youth rarely ends there.
You might think that Australia’s regulatory guardians can be trusted to uphold free speech ideals, but has that been the case to date? Under Australian law, it is permissible to restrict free speech for reasons of public order, national security, and protection from harm. That includes limits on “hate speech,” prompting Elon Musk to exaggerate and call the country fascist. Nonetheless the country does not have anything comparable to America’s First Amendment free speech protections.
So why should we empower Australian regulators and restrict free speech further?
It is very defensible to worry that your kid is on his or her phone too much. Furthermore, school bans or limits on smartphone usage are likely to bring some measurable but small gains.
But if you think a massive expansion of state authority over online content is the answer, you ought to know that the associated gains from that decision will at best be modest. You will not be saving civilization or our youth; rather you will be joining the ever-growing parade against free speech.
Recommended, and in this recent piece Ben Yeoh surveys the research-based literature on social media and teen harm.
Quantifying human-AI synergy
From Christoph Riedl and Ben Weidmann:
We introduce a novel Bayesian Item Response Theory framework to quantify human–AI synergy, separating individual and collaborative ability while controlling for task difficulty in interactive settings. Unlike standard static benchmarks, our approach models human–AI performance as a joint process, capturing both user-specific factors and moment-to-moment fluctuations. We validate the framework by applying it to human–AI benchmark data (n=667) and find significant synergy. We demonstrate that collaboration ability is distinct from individual problem-solving ability. Users better able to infer and adapt to others’ perspectives achieve superior collaborative performance with AI–but not when working alone. Moreover, moment-to-moment fluctuations in perspective taking influence AI response quality, highlighting the role of dynamic user factors in collaboration. By introducing a principled framework to analyze data from human-AI collaboration, interactive benchmarks can better complement current single-task benchmarks and crowd-assessment methods. This work informs the design and training of language models that transcend static prompt benchmarks to achieve adaptive, socially aware collaboration with diverse and dynamic human partners.
Here is a useful tweet storm on the work. I do not love how the abstract is written, I would stress these sentences: “We demonstrate that collaboration ability is distinct from individual problem-solving ability. Users better able to infer and adapt to others’ perspectives achieve superior collaborative performance with AI–but not when working alone. Moreover, moment-to-moment fluctuations in perspective taking influence AI response quality, highlighting the role of dynamic user factors in collaboration.”
My Conversation with the excellent Dan Wang
Here is the audio, video, and transcript. Here is part of the episode summary:
Tyler and Dan debate whether American infrastructure is actually broken or just differently optimized, why health care spending should reach 35% of GDP, how lawyerly influences shaped East Asian development differently than China, China’s lack of a liberal tradition and why it won’t democratize like South Korea or Taiwan did, its economic dysfunction despite its manufacturing superstars, Chinese pragmatism and bureaucratic incentives, a 10-day itinerary for Yunnan, James C. Scott’s work on Zomia, whether Beijing or Shanghai is the better city, Liu Cixin and why volume one of The Three-Body Problem is the best, why contemporary Chinese music and film have declined under Xi, Chinese marriage markets and what it’s like to be elderly in China, the Dan Wang production function, why Stendhal is his favorite novelist and Rossini’s Comte Ory moves him, what Dan wants to learn next, whether LLMs will make Tyler’s hyper-specific podcast questions obsolete, what flavor of drama their conversation turned out to be, and more.
Excerpt:
COWEN: When will Chinese suburbs be really attractive?
WANG: What are Chinese suburbs? You use this term, Tyler, and I’m not sure what exactly they mean.
COWEN: You have a yard and a dog and a car, right?
WANG: Yes.
COWEN: You control your school district with the other parents. That’s a suburb.
WANG: How about never? I’m not expecting that China will have American-style suburbs anytime soon, in part because of the social engineering projects that are pretty extensive in China. I think there is a sense in which Chinese cities are not especially dense. Indian cities are much, much more dense. I think that Chinese cities, the streets are not necessarily terribly full of people all the time. They just sprawl quite extensively.
They sprawl in ways that I think the edges of the city still look somewhat like the center of the city, which there’s too many high-rises. There’s probably fewer parks. There’s probably fewer restaurants. Almost nobody has a yard and a dog in their home. That’s in part because the Communist Party has organized most people to live in apartment compounds in which it is much easier to control them.
We saw this really extensively in the pandemic, in which people were unable to leave their Shanghai apartment compounds for anything other than getting their noses and mouths swabbed. I write a little bit about how, if you take the rail outside of major cities like Beijing and Shanghai, you hit farmland really, really quickly. That is in part because the Communist Party assesses governors as well as mayors on their degree of food self-sufficiency.
Cities like Shanghai and Beijing have to produce a lot of their own crops, both grains as well as vegetables, as well as fruits, as well as livestock, within a certain radius so that in case there’s ever a major devastating war, they don’t have to rely on strawberries from Mexico or strawberries from Cambodia, or Thailand. There’s a lot of farmland allocated outside of major cities. I think that will prevent suburban sprawl. You can’t control people if they all have a yard as well as a dog. I think the Communist Party will not allow it.
COWEN: Whether the variable of engineers matters, I went and I looked at the history of other East Asian economies, which have done very well in manufacturing, built out generally excellent infrastructure. None of these problems with the Second Avenue line in New York. Taiwan, like the presidents, at least if we believe GPT-5, three of them were lawyers and none of them were engineers. South Korea, you have actually some economists, a lot of bureaucrats.
WANG: Wow. Imagine that. Economists in charge, Tyler.
COWEN: I wouldn’t think it could work. A few lawyers, one engineer. Singapore, Lee Kuan Yew, he’s a lawyer. He thinks in a very lawyerly manner. Singapore has arguably done the best of all those countries. Much richer than China, inspired China. Why should I think engineers rather than just East Asia, and a bunch of other accompanying facts about these places are what matter?
WANG: Japan, a lot of lawyers in the top leadership. What exactly was the leadership of Hong Kong? A bunch of British civil servants.
COWEN: Some of whom are probably lawyers or legal-type minds, right? Not in general engineers.
WANG: PPE grads. I think that we can understand the engineering variable mostly because of how much more China has done relative to Japan and South Korea and Taiwan.
COWEN: It’s much, much poorer. Per capita manufacturing output is gone much better in these other countries.
And:
WANG: Tyler, what does it say about us that you and I have generally a lot of similar interests in terms of, let’s call it books, music, all sorts of things, but when it comes to particular categories of things, we oppose each other diametrically. I much prefer Anna Karenina to War and Peace. I prefer Buddenbrooks to Magic Mountain. Here again, you oppose me. What’s the deal?
COWEN: I don’t think the differences are that big. For instance, if we ask ourselves, what’s the relative ranking of Chengdu plus Chongqing compared to the rest of the world? We’re 98.5% in agreement compared to almost anyone else. When you get to the micro level, the so-called narcissism of petty differences, obviously, you’re born in China. I grew up in New Jersey. It’s going to shape our perspectives.
Anything in China, you have been there in a much more full-time way, and you speak and read Chinese, and none of that applies to me. I’m popping in and out as a tourist. Then, I think the differences make much more sense. It’s possible I would prefer to live in Shanghai for essentially the reasons you mentioned. If I’m somewhere for a week, I’m definitely going to pick Beijing. I’ll go around to the galleries. The things that are terrible about the city just don’t bother me that much, because I know I’ll be gone.
WANG: 98.5% agreement. I’ll take that, Tyler. It’s you and me against the rest of the world, but then we’ll save our best disagreements for each other.
COWEN: Let’s see if you can pass an intellectual Turing test. Why is it that I think Yunnan is the single best place in the world to visit? Just flat out the best if you had to pick one region. Not why you think it is, but why I think it is.
Strongly recommended, Dan and I had so much fun we kept going for about an hour and forty minutes. And of course you should buy and read Dan’s bestselling book Breakneck: China’s Quest to Engineer the Future.
The importance of the internet
From my recent chat with Alex, mostly about fiscal policy:
TABARROK:To be clear, a 0.5% increase in the rate of productivity growth, that doesn’t seem like a lot, but that would be historically a bigger increase than we got from anything. A bigger increase than the internet. Sure, yes.
COWEN:It is the internet in a way, but yes.
TABARROK:It was founded on the internet, yes. The internet was the agar culturefor the growth of the AI.
COWEN:That’s why the internet’s important. We’re just beginning to realize this,right?
TABARROK:Exactly, yes.
COWEN:It’s why a lot of people can’t admit AI might be a good thing, because then they’d have to admit the internet was a good thing. They’re so committed to never saying that.
TABARROK:Is that why?
COWEN:That’s why, yes.Believe me. That’s why.
TABARROK:It is funny that I think historically, when we look back, I think you’re right, we’ll think about what was the internet. The growth culture was putting everything online, was for the AI. It wasn’t for us.
Séb Krier
Huge fan of multi agent systems, agent based modelling, and social intelligence – these frames still seem really absent from mainstream AI discourse except in a few odd places. Some half-baked thoughts:
1. Expecting a model to do all the work, solve everything, come up with new innovations etc is probably not right. This was kinda the implicit assumption behind *some* interpretations of capabilities progress. The ‘single genius model’ overlooks the fact that inference costs and context windows are finite.
2. People overrate individual intelligence: most innovations are the product of social organisations (cooperation) and market dynamics (competition), not a single genius savant. Though the latter matters too of course: the smarter the agents the better.
3. There’s still a lot of juice to be squeezed from models, but I would think it has more to do with how they’re organised. AI Village is a nice vignette, and also highlights the many ways in which models fail and what needs to be fixed.
4. Once you enter multi-agent world, then institutions and culture start to matter too: what are the rules of the game? What is encouraged vs what is punished? What can agents do and say to each other? How are conflicts resolved? It’s been interesting seeing how some protocols recently emerged. We’re still very early!
5. Most of the *value* and transformative changes we will get from AI will come from products, not models. The models are the cognitive raw power, the products are what makes them useful and adapted to what some user class actually needs. A product is basically the bridge between raw potential and specific utility; in fact many IDEs today are essentially crystallized multi agent systems.
Here is the link.
My excellent Conversation with Cass Sunstein
Cass was in top form, and so we went on for almost two hours. In his Substack he described it as “The most fun interview I have ever done.” Here is the audio, video, and transcript. Here is part of the episode summary:
Tyler and Cass discuss whether liberalism is self-undermining or simply vulnerable to illiberal forces, the tensions in how a liberal immigration regime would work, whether new generations of liberal thinkers are emerging, if Derek Parfit counts as a liberal, Mill’s liberal wokeism, the allure of Mises’ “cranky enthusiasm for freedom,” whether the central claim of The Road to Serfdom holds up, how to blend indigenous rights with liberal thought, whether AIs should have First Amendment protections, the argument for establishing a right not to be manipulated, better remedies for low-grade libel, whether we should have trials run by AI, how Bob Dylan embodies liberal freedom, Cass’ next book about animal rights, and more.
I will reproduce the section Cass pulled for his own Substack:
COWEN: Now, we started with the topic of liberalism. How is it you think about or characterize the liberalism of Bob Dylan?
SUNSTEIN: Bob Dylan is a liberal. His liberalism is captured in the line, “He not busy being born is busy dying.” I hope he’s immortal, but if anything is on his epigraph, that would be a good candidate.
The notion of self-invention, of freedom, is central to basically everything. His refusal to keep singing the same song — you can hear him talking about it in some of the interviews. He said, “I could do that. I could just do that forever. I knew how they’d react.” He said, “What’s that about?” He said, “I needed to do something else.” But of course, the line, “I needed to do something else” — that’s my line. How he would put it would be much more vivid and surprising than that.
His “Like a Rolling Stone” is an anthem of freedom. I heard it, actually, in concert a few years ago. It was a great performance. It wasn’t young, but it was a great performance. The audience went wild when he did “Like a Rolling Stone.” That was the final song. It was the encore. It wasn’t just because it was the greatest rock song ever written. It was because of how he did it. I thought, “What’s going on in this song? Why is everyone exhilarated?” The song, which he described when he wrote it as vomit, hatred directed at somewhere that was real — it wasn’t that, or it was a little bit that, but it was a song of liberty.
“How does it feel to be on your own with no direction home, like a complete unknown, like a rolling stone?” Everyone felt like they were flying. He makes that — “Like a Rolling Stone” — be a song of freedom. If you look at his angry songs — “Positively 4th Street” — there’s a freedom in being, of course, uninhibited, able to say things, but also a freedom of disconnection.
When he’s asked why did he change his name, I have an account of why he actually did. I think he gave it exactly once, but in his more characteristic way, he said, “This is America. You can change your name.” Then he said, “I was born. I didn’t think I was born with the right name. I could make it up. I could say that sounds more like I was.”
Making rootlessness not be a curse, but instead something that is . . . the word joy is too clichéd for Dylan. If you look at his love songs, like “If You See Her, Say Hello,” which isn’t one of my favorites, but it’s good. There’s a connection with the one he loved, who got away, but you can feel the sense of freedom.
COWEN: “Visions of Johanna”?
SUNSTEIN: Yes, completely. He’s torn. That has the great opening line. “Ain’t it just like the night to play tricks When you’re trying to be so quiet?” Did Yeats write better lines than that? Probably, but he was Yeats.
COWEN: Blood on the Tracks — a liberal album?
SUNSTEIN: Oh, yes.
COWEN: How would you express that?
SUNSTEIN: Well, I’m thinking “Buckets of Rain” is the closing song. Right before that, there’s a song, “You’re Gonna Make Me Lonesome When You Go.” That’s it, which is, I think, one of his greatest songs. That’s a liberal song of freedom and separation, that she’s going, but he’s going to see her everywhere, and there’s smiling at impermanence. That is a big liberal theme — smiling at impermanence — because impermanence makes things not routine and also makes for freedom.
COWEN: “Idiot Wind” is the angry song of the batch, right?
SUNSTEIN: Yes, it’s pretty mad. He said about that song, “I don’t know why people like it. There’s so much sadness and distress in it.”
COWEN: Do you see your own liberalism or just yourself in the liberalism of Bob Dylan?
SUNSTEIN: I think so.
COWEN: Reinventing yourself, not quite wanting to be pinned down, doing a lot of stuff.
SUNSTEIN: He likes, I think, abandoning and going on to something that’s very different. I wish I’d gone electric or had some equivalent of that. But doing something quite different — I do share a little bit with him. I like it when I think something I thought was wrong. I now am very enthusiastic about the Austrian economists and Hayek. I’ve always admired them, of course, but I didn’t feel that they were on my team. Now I feel I’ve gone to their team. I don’t feel ashamed that I was wrong before. I feel excited that I’m less wrong now.
Definitely recommended, I could have pulled out many other parts as well. Again, I am happy to recommend Cass’s new book Liberalism: In Defense of Freedom.
Nano Banana Pro does Marginal Revolution
My very fun Conversation with Blake Scholl
Here is the audio, video, and transcript. This was at a live event (the excellent Roots of Progress conference), so it is only about forty minutes, shorter than usual. Here is the episode summary:
Blake Scholl is one of the leading figures working to bring back civilian supersonic flight. As the founder and CEO of Boom Supersonic, he’s building a new generation of supersonic aircraft and pushing for the policies needed to make commercial supersonic travel viable again. But he’s equally as impressive as someone who thinks systematically about improving dysfunction—whether it’s airport design, traffic congestion, or defense procurement—and sees creative solutions to problems everyone else has learned to accept.
Tyler and Blake discuss why airport terminals should be underground, why every road needs a toll, what’s wrong with how we board planes, the contrasting cultures of Amazon and Groupon, why Concorde and Apollo were impressive tech demos but terrible products, what Ayn Rand understood about supersonic transport in 1957, what’s wrong with aerospace manufacturing, his heuristic when confronting evident stupidity, his technique for mastering new domains, how LLMs are revolutionizing regulatory paperwork, and much more.
Excerpt:
COWEN: There’s plenty about Boom online and in your interviews, so I’d like to take some different tacks here. This general notion of having things move more quickly, I’m a big fan of that. Do you have a plan for how we could make moving through an airport happen more quickly? You’re in charge. You’re the dictator. You don’t have to worry about bureaucratic obstacles. You just do it.
SCHOLL: I think about this in the shower like every day. There is a much better airport design that, as best I can tell, has never been built. Here’s the idea: You should put the terminals underground. Airside is above ground. Terminals are below ground. Imagine a design with two runways. There’s an arrival runway, departure runway. Traffic flows from arrival runway to departure runway. You don’t need tugs. You can delete a whole bunch of airport infrastructure.
Imagine you pull into a gate. The jetway is actually an escalator that comes up from underneath the ground. Then you pull forward, so you can delete a whole bunch of claptrap that is just unnecessary. The terminal underground should have skylights so it can still be incredibly beautiful. If you model fundamentally the thing on a crossbar switch, there are a whole bunch of insights for how to make it radically more efficient. Sorry. This is a blog post I want to write one day. Actually, it’s an airport I want to build.
And;
COWEN: I’m at the United desk. I have some kind of question. There’s only two or three people in front of me, but it takes forever. I notice they’re just talking back and forth to the assistant. They’re discussing the weather or the future prospects for progress, total factor productivity. I don’t know. I’m frustrated. How can we make that process faster? What’s going wrong there?
SCHOLL: The thing I most don’t understand is why it requires so many keystrokes to check into a hotel room. What are they writing?
What are they writing?
Google Scholar Labs
Brings AI to Google Scholar, find it here. Via Joshua Gans. And yes this does mean that the academics also are, or at least ought to be, writing for the AIs.
Is the AI sector currently a bubble?
Possibly, but do not jump to that conclusion too quickly, as I argued in my latest Free Press column. Excerpt:
Nvidia is often considered a bellwether AI stock. That’s because much of its revenue comes from selling graphics processing units to power advanced AI systems, meaning that its success gives investors insight into the health of the sector overall. Currently, Nvidia’s stock-price-to-earnings ratio is in the 54 to 55 range, roughly twice the typical market average. That means the market expects great things from this stock. Those projections may or may not be validated, but it’s hard to conclude they’re entirely divorced from reality…
Keep in mind that the tech sector as a whole is still earning more than it is shelling out in capital expenditures. The current AI boom is being financed by earnings more than by new issuance of debt, which makes it less prone to a sudden crash. By one estimate, capital expenditures in Big Tech are about 94 percent of cash flow in 2025. You could imagine that number moving into unstable territory, but so far, the U.S. tech sector is managing to pay its bills without going into debt.
You may recall we are coming off a period when everyone complained that the big tech companies were sitting on trillions of dollars in cash and capital. Now, they are spending it, and complaints are heating up once again. Damned if they do, damned if they don’t.
In fact, what we are seeing right now is a shortage in the AI sector’s capacity to meet demand. Major tech companies are investing in more computing capacity, but they still cannot serve all the customers who want access to AI systems. That augurs well for the future of the sector, even if there are dips and spills along the way.
As usual, we will see, but if you are calling it a bubble after an initial price dip or corporate shake-out, that is exactly the fallacy you are not supposed to be slipping into.
Do (human) readers prefer AI writers?
It seems so, do read through the whole abstract:
The use of copyrighted books for training AI models has led to numerous lawsuits from authors concerned about AI’s ability to generate derivative content. Yet it’s unclear whether these models can generate high quality literary text while emulating authors’ styles/voices. To answer this we conducted a preregistered study comparing MFA-trained expert writers with three frontier AI models: ChatGPT, Claude, and Gemini in writing up to 450 word excerpts emulating 50 awardwinning authors’ (including Nobel laureates, Booker Prize winners, and young emerging National Book Award finalists) diverse styles. In blind pairwise evaluations by 159 representative expert (MFA-trained writers from top U.S. writing programs) and lay readers (recruited via Prolific), AI-generated text from in-context prompting was strongly disfavored by experts for both stylistic fidelity (odds ratio [OR]=0.16, p < 10^-8) and writing quality (OR=0.13, p< 10^-7) but showed mixed results with lay readers. However, fine-tuning ChatGPT on individual author’s complete works completely reversed these findings: experts now favored AI-generated text for stylistic fidelity (OR=8.16, p < 10^-13) and writing quality (OR=1.87, p=0.010), with lay readers showing similar shifts. These effects are robust under cluster-robust inference and generalize across authors and styles in author-level heterogeneity analyses. The fine-tuned outputs were rarely flagged as AI-generated (3% rate versus 97% for incontext prompting) by state-of-the-art AI detectors. Mediation analysis reveals this reversal occurs because fine-tuning eliminates detectable AI stylistic quirks (e.g., cliché density) that penalize incontext outputs, altering the relationship between AI detectability and reader preference. While we do not account for additional costs of human effort required to transform raw AI output into cohesive, publishable novel length prose, the median fine-tuning and inference cost of $81 per author represents a dramatic 99.7% reduction compared to typical professional writer compensation. Author-specific fine-tuning thus enables non-verbatim AI writing that readers prefer to expert human writing, thereby providing empirical evidence directly relevant to copyright’s fourth fair-use factor, the “effect upon the potential market or value” of the source works.
That is from a new paper by Tuhin Chakrabarty, Jane C. Ginsburg, and Paramveer Dhillon. For the pointer I thank the excellent Kevin Lewis. I recall an earlier piece showing that LLMs also prefer LLM outputs?
“Some Economics of Artificial Super Intelligence”
I promised to pass along serious models of pending AI doom, especially if they are peer-reviewed or at least headed for such. The AI doomer types still are dropping the ball on this, but one economist has made a contribution and so here it is from Henry A. Thompson:
Conventional wisdom holds that a misaligned artificial superintelligence (ASI) will destroy humanity. But the problem of constraining a powerful agent is not new. I apply classic economic logic of interjurisdictional competition, all-encompassing interest, and trading on credit to the threat of misaligned ASI. Using a simple model, I show that an acquisitive ASI refrains from full predation under surprisingly weak conditions. When humans can flee to rivals, inter-ASI competition creates a market that tempers predation. When trapped by a monopolist ASI, its “encompassing interest” in humanity’s output makes it a rational autocrat rather than a ravager. And when the ASI has no long-term stake, our ability to withhold future output incentivizes it to trade on credit rather than steal. In each extension, humanity’s welfare progressively worsens. But each case suggests that catastrophe is not a foregone conclusion. The dismal science, ironically, offers an optimistic take on our superintelligent future.
You may or may not agree, but as usual the point is to build out a literature, not to regard any single paper as the final word. Via the excellent Joy Buchanan.
Waymo
Waymo now does highways in the Bay area.
Expanding our service territory in the Bay Area and introducing freeways is built on real-world performance and millions of miles logged on freeways, skillfully handling highway dynamics with our employees and guests in Phoenix, San Francisco, and Los Angeles. This experience, reinforced by comprehensive testing as well as extensive operational preparation, supports the delivery of a safe and reliable service.
The future is happening fast.