Category: Web/Tech
My excellent Conversation with Brendan Foody
Here is the audio, video, and transcript. Here is the episode summary:
At 22, Brendan Foody is both the youngest Conversations with Tyler guest ever and the youngest unicorn founder on record. His company Mercor hires the experts who train frontier AI models—from poets grading verse to economists building evaluation frameworks—and has become one of the fastest-growing startups in history.
Tyler and Brendan discuss why Mercor pays poets $150 an hour, why AI labs need rubrics more than raw text, whether we should enshrine the aesthetic standards of past eras rather than current ones, how quickly models are improving at economically valuable tasks, how long until AI can stump Cass Sunstein, the coming shift toward knowledge workers building RL environments instead of doing repetitive analysis, how to interview without falling for vibes, why nepotism might make a comeback as AI optimizes everyone’s cover letters, scaling the Thiel Fellowship 100,000X, what his 8th-grade donut empire taught him about driving out competition, the link between dyslexia and entrepreneurship, dining out and dating in San Francisco, Mercor’s next steps, and more.
And an excerpt:
COWEN: Now, I saw an ad online not too long ago from Mercor, and it said $150 an hour for a poet. Why would you pay a poet $150 an hour?
FOODY: That’s a phenomenal place to start. For background on what the company does — we hire all of the experts that teach the leading AI models. When one of the AI labs wants to teach their models how to be better at poetry, we’ll find some of the best poets in the world that can help to measure success via creating evals and examples of how the model should behave.
One of the reasons that we’re able to pay so well to attract the best talent is that when we have these phenomenal poets that teach the models how to do things once, they’re then able to apply those skills and that knowledge across billions of users, hence allowing us to pay $150 an hour for some of the best poets in the world.
COWEN: The poets grade the poetry of the models or they grade the writing? What is it they’re grading?
FOODY: It could be some combination depending on the project. An example might be similar to how a professor in English class would create a rubric to grade an essay or a poem that they might have for the students. We could have a poet that creates a rubric to grade how well is the model creating whatever poetry you would like, and a response that would be desirable to a given user.
COWEN: How do you know when you have a good poet, or a great poet?
FOODY: That’s so much of the challenge of it, especially with these very subjective domains in the liberal arts. So much of it is this question of taste, where you want some degree of consensus of different exceptional people believing that they’re each doing a good job, but you probably don’t want too much consensus because you also want to get all of these edge case scenarios of what are the models doing that might deviate a little bit from what the norm is.
COWEN: So, you want your poet graders to disagree with each other some amount.
FOODY: Some amount, exactly, but still a response that is conducive with what most users would want to see in their model responses.
COWEN: Are you ever tempted to ask the AI models, “How good are the poet graders?”
[laughter]
FOODY: We often are. We do a lot of this. It’s where we’ll have the humans create a rubric or some eval to measure success, and then have the models say their perspective. You actually can get a little bit of signal from that, especially if you have an expert — we have tens of thousands of people that are working on our platform at any given time. Oftentimes, there’ll be someone that is tired or not putting a lot of effort into their work, and the models are able to help us with catching that.
And:
COWEN: Let’s say it’s poetry. Let’s say you can get it for free, grab what you want from the known universe. What’s the data that’s going to make the models, working through your company, better at poetry?
FOODY: I think that it’s people that have phenomenal taste of what would users of the end products, users of these frontier models want to see. Someone that understands that when a prompt is given to the model, what is the type of response that people are going to be amazed with? How we define the characteristics of those responses is imperative.
Probably more than just poets that have spent a lot of time in school, we would want people that know how to write work that gets a lot of traction from readers, that gains broad popularity and interest, drives the impact, so to speak, in whatever dimension that we define it within poetry.
COWEN: But what’s the data you want concretely? Is it a tape of them sitting around a table, students come, bring their poems, the person says, “I like this one, here’s why, here’s why not.” Is it that tape or is it written reports? What’s the thing that would come in the mail when you get your wish?
FOODY: The best analog is a rubric. If you have some —
COWEN: A rubric for how to grade?
FOODY: A rubric for how to grade. If the poem evokes this idea that is inevitably going to come up in this prompt or is a characteristic of a really good response, we’ll reward the model a certain amount. If it says this thing, we’ll penalize the model. If it styles the response in this way, we’ll reward it. Those are the types of things, in many ways, very similar to the way that a professor might create a rubric to grade an essay or a poem.
Poetry is definitely a more difficult one because I feel like it’s very unbounded. With a lot of essays that you might grade from your students, it’s a relatively well-scoped prompt where you can probably create a rubric that’s easy to apply to all of them, versus I can only imagine in poetry classes how difficult it is to both create an accurate rubric as well as apply it. The people that are able to do that the best are certainly extremely valuable and exciting.
COWEN: To get all nerdy here, Immanuel Kant in his third critique, Critique of Judgment, said, in essence, taste is that which cannot be captured in a rubric. If the data you want is a rubric and taste is really important, maybe Kant was wrong, but how do I square that whole picture? Is it, by invoking taste, you’re being circular and wishing for a free lunch that comes from outside the model, in a sense?
FOODY: There are other kinds of data they could do if it can’t be captured in a rubric. Another kind is RLHF, where you could have the model generate two responses similar to what you might see in ChatGPT, and then have these people with a lot of taste choose which response they prefer, and do that many times until the model is able to understand their preferences. That could be one way of going about it as well.
Interesting throughout, and definitely recommended. Note the conversation was recorded in October (we have had a long queue), so a few parts of it sound slightly out of date. And here is Hollis Robbins on LLMs and poetry.
The Molly Cantillon manifesto, A Personal Panopticon
I find this piece significant, and think it is likely to be one of the most important essays of the year:
A few months ago, I started running my life out of Claude Code. Not out of intention to do so, it was just the place where everything met. And it just kept working. Empires are won by conquest. What keeps them standing is something much quieter. Before a king can tax, he must count. Before he can conscript, he must locate. Before he can rule, he must see. Legibility is the precondition for governance…
The first thing Claude solved was product blindness. NOX now runs on a cron job: pulling Amplitude, cross-referencing GitHub, and pointing me to what needs building. It handles A/B testing, generates winning copy, and has turned customer support into a fully autonomous department.
Once I saw this was possible, I chased it everywhere. Email, hitting inbox zero for the first time ever, with auto-drafted replies for everything inbound. Workouts, accommodating horrendously erratic travel schedules. Sleep, built a projector wired to my WHOOP after exactly six hours that wakes me with my favorite phrases. Subscriptions, found and returned $2000 I didn’t know I was paying. The dozen SFMTA citations I’d ignored, the action items I’d procrastinated into oblivion. People are using it to, I discovered, run vending machines, home automation systems, and keep plants alive.
The feeling is hard to name. It is the violent gap between how blind you were and how obvious everything feels now with an observer that reads all the feeds, catches what you’ve unconsciously dropped, notices patterns across domains you’d kept stubbornly separate, and—crucially—tells you what to do about it.
My personal finances are now managed in the terminal. Overnight it picks the locks of brokerages that refuse to talk to each other, pulls congressional and hedge fund disclosures, Polymarket odds, X sentiment, headlines and 10-Ks from my watchlist. Every morning, a brief gets added in ~/𝚝𝚛𝚊𝚍𝚎𝚜. Last month it flagged Rep. Fields buying NFLX shares. Three weeks later, the Warner Bros deal. I don’t always trade, sometimes I argue with the thesis. But I’m never tracking fifteen tabs at 6am anymore.
It feels borderline unfair seeing around corners, being in ten places at once, surveilling yourself with the attention span of a thousand clones.
A panopticon still, but the tower belongs to you.
There is more at the link, or this link, and yes she is related to the 18th century Irish economist Richard Cantillon.
The US Leads the World in Robots (Once You Count Correctly)
If you search for data on robots you will quickly find data from the International Federation of Robotics which places South Korea in the lead with ~818 robots per 10,000 manufacturing workers, followed by China, Japan, Germany and finally at 10th place the US at ~304 robots per 10,000. The IFR, however, misses the most sophisticated, impressive and versatile robots, namely Teslas with FSD capability. Teslas see the world, navigate complex environments, move tons of metal at high speeds and must perform at very high levels of tolerance and safety. If you included Teslas as robots, as you should, the US leaps to the top.
Moreover, once you understand Teslas as robots, Optimus, Tesla’s humanoid robot division, stops being a quixotic Elon side-project and becomes the obvious continuation of Tesla’s core work.

O-Ring Automation
We study automation when tasks are quality complements rather than separable. Production requires numerous tasks whose qualities multiply as in an O-ring technology. A worker allocates a fixed endowment of time across the tasks performed; machines can replace tasks with given quality, and time is allocated across the remaining manual tasks. This “focus” mechanism generates three results. First, task-by-task substitution logic is incomplete because automating one task changes the return to automating others. Second, automation decisions are discrete and can require bundled adoption even when automation quality improves smoothly. Third, labour income can rise under partial automation because automation scales the value of remaining bottleneck tasks. These results imply that widely-used exposure indices, which aggregate task-level automation risk using linear formulas, will overstate displacement when tasks are complements. The relevant object is not average task exposure but the structure of bottlenecks and how automation reshapes worker time around them.
That is from a new paper by Joshua S. Gans and Avi Goldfarb. Once again people, the share of labor is unlikely to collapse…
Luis Garicano career advice
Take the messy job:
The other option is to go for a messy job, where the output is the product of many different tasks, many of which affect each other.
The head of engineering at a manufacturing plant I know well must decide who to hire, which machines to buy, how to lay them down in the plant, negotiate with the workers and the higher ups the solutions proposed, and mobilise the resources to implement them. That task is extraordinarily hard to automate. Artificial intelligence commoditizes codified knowledge: textbooks, proofs, syntax. But it does not interface in a meaningful way with local knowledge, where a much larger share of the value of messy jobs is created. Even if artificial intelligence excelled at most of the single tasks that make up her job, it could not walk the factory floor to cajole a manager to redesign a production process.
A management consultant whose job consists entirely of producing slide decks is exposed. A consultant who spends half of her time reading the room, building client relationships, and navigating organizational politics has a bundle AI cannot replicate.
Here is the full letter.
Dan Wang 2025 letter
Self-recommending, here is the link, here is one excerpt:
People like to make fun of San Francisco for not drinking; well, that works pretty well for me. I enjoy board games and appreciate that it’s easier to find other players. I like SF house parties, where people take off their shoes at the entrance and enter a space in which speech can be heard over music, which feels so much more civilized than descending into a loud bar in New York. It’s easy to fall into a nerdy conversation almost immediately with someone young and earnest. The Bay Area has converged on Asian-American modes of socializing (though it lacks the emphasis on food). I find it charming that a San Francisco home that is poorly furnished and strewn with pizza boxes could be owned by a billionaire who can’t get around to setting up a bed for his mattress.
And:
One of the things I like about the finance industry is that it might be better at encouraging diverse opinions. Portfolio managers want to be right on average, but everyone is wrong three times a day before breakfast. So they relentlessly seek new information sources; consensus is rare, since there are always contrarians betting against the rest of the market. Tech cares less for dissent. Its movements are more herdlike, in which companies and startups chase one big technology at a time. Startups don’t need dissent; they want workers who can grind until the network effects kick in. VCs don’t like dissent, showing again and again that many have thin skins. That contributes to a culture I think of as Silicon Valley’s soft Leninism. When political winds shift, most people fall in line, most prominently this year as many tech voices embraced the right.
Interesting throughout, plus Dan writes about the most memorable books he read in 2025.
Taxation in a strong AI world
Here is Dwarkesh’s tweet, based on his recent paper with Trammell, raising the issue of whether wealth taxes will become desirable in the future. A few points:
1. I think quality homes in good locations will be extremely valuable. Those could be taxed more. You could call that a wealth tax, but arguably it is closer to a “housing services tax.”
2. You could put higher consumption taxes on items the wealthy purchase to a disproportionate degree. Paintings and yachts, and so on. Tom Holden argues: “In a world in which capital is essentially the only input to production, taxing capital reduces the growth rate of the economy. Whereas at present capital taxes have only level effects. So if anything, capital taxes will become less desirable as the labour share falls.”
3. I think the amount of money spent on health care will go up a lot. And people will live much longer, which will further boost the amount spent on health care. Taxing health care more is the natural way to address fiscal problems. Some people will fly abroad for their knee surgeries, but for a long time most health care will be consumed nearby, even in a strong AI world. If the way we keep the budget sane is to have people die at 95 instead of 97, there may be some positive social externalities from the shorter life spans. We also could use some of that money for birth subsidies.
4. As a more general point, capital will not be a perfect substitute for labor, or anything close to that, anytime soon.
5. Final incidence of the AI revolution is not just about the degree of substability of capital for labor. It is also about supply and demand elasticities in goods and services markets. For instance, to the extent AI makes various services much cheaper, real wages are rising not falling. That may or may not be the dominant effect, but do not assume too quickly that wages simply fall.
5b. It is not an equilibrium for capital to simply “have all the goodies.” Let’s say that Simon Legree, using advanced AI, can produce all the world’s output using a single watt of energy. And no one else with an AI company can produce anything to compete with that (this already sounds implausible, right?). If Simon simply hoards all that output, he has no profit, though I guess he can cure his own case of the common cold. The prices for that output have to fall so it can be purchased by someone else. The nature of the final equilibrium here is unclear, but again do not assume all or even most of the returns will stay with capital. That is almost certainly not the case.
Addendum: Here is some follow-up from Dwarkesh. I think he is talking about a world very different from our own, as there is talk of ownership of galaxies. That said, many other people wish to implement his ideas sooner than that.
Existential Risk and Growth
By Philip Trammell and Leopold Aschenbrenner, a new paper:
Technological development raises consumption but may pose existential risk. A growing literature studies this tradeoff in static settings where stagnation is perfectly safe. But if any risky technology already exists, technological development can also lower risk indirectly in two ways: by speeding (1) technological solutions and/or (2) a “Kuznets curve” in which wealth increases a planner’s willingness to
pay for safety. The risk-minimizing technology growth rate, in light of these dynamics, is typically positive and may easily be high. Below this rate, technological development poses no tradeoff between consumption and cumulative risk.
Self-recommending…
A model of girl happiness, a compensatory-use study
A statistical model was used to examine these relationships simultaneously by predicting the likelihood that a girl reports being very happy.
The model includes socioeconomic status, parent–child communication, screen-time limits, and an interaction between limits and communication.
The results reinforce the patterns in the figures. Parent–child communication dominates the model. Girls who report strong communication are about three to four times more likely to report being very happy than those who report none. Socioeconomic status shows a smaller independent association. Screen-time limits contribute little on their own and matter modestly only when strong communication is already present.
If phones were the central problem, limits would emerge as a robust solution across contexts. They do not…
What the compensatory-use model rejects is a stronger claim. It rejects the idea that smartphone exposure itself is the primary driver of youth distress and that prohibition is therefore the central remedy. If that causal story were correct, limits would show large and consistent benefits across households, including among those with the weakest communication and highest distress. They do not.
And to close:
The most reliable way to improve youth well-being is to meet individual needs through connection instead of control.
That work depends on cooperation, not compliance.
Here is the full essay by Owen Kellogg, of course this is only a single study.
Will AI Improve Undergraduate Economics Education?
For decades, undergraduate economics educators have followed a well-worn playbook featuring textbooks, lectures and problem sets. Students have passively listened, taken notes and studied for exams. AI disrupts this educational process. Some students are using this tool as a substitute for their own precious time. What is our best response? This paper provides a prospective analysis of how to restructure every phase of the undergraduate economics experience to improve the major and better prepare students for their uncertain future. Departments face a principal/agent issue in implementing major curricular reforms. I discuss the incentive problems that arise both within economics departments and across departments. If we win this competition to reimagine the undergraduate experience, will the Deans reward us?
TC again: No.
Séb Krier, continued
Or more specifically Nenad Tomašev, Matija Franklin, Julian Jacobs, Sébastien Krier, and Simon Osindero:
AI safety and alignment research has predominantly been focused on methods for safeguarding individual AI systems, resting on the assumption of an eventual emergence of a monolithic Artificial General Intelligence (AGI). The alternative AGI emergence hypothesis, where general capability levels are first manifested through coordination in groups of sub-AGI individual agents with complementary skills and affordances, has received far less attention. Here we argue that this patchwork AGI hypothesis needs to be given serious consideration, and should inform the development of corresponding safeguards and mitigations. The rapid deployment of advanced AI agents with tool-use capabilities and the ability to communicate and coordinate makes this an urgent safety consideration. We therefore propose a framework for distributional AGI safety that moves beyond evaluating and aligning individual agents. This framework centers on the design and implementation of virtual agentic sandbox economies (impermeable or semi-permeable), where agent-to-agent transactions are governed by robust market mechanisms, coupled with appropriate auditability, reputation management, and oversight to mitigate collective risks.
Here is the link, this is some of the most important work of our time. Here is the previous MR post on Krier.
Markets in everything
People Are Paying to Get Their Chatbots High on ‘Drugs’
An online marketplace is selling code modules that simulate the effects of cannabis, ketamine, cocaine, ayahuasca, and alcohol when they are uploaded to ChatGPT.
Here is the full Wired article. Via David.
Falling costs
Unbelievable progress that even I underestimated! Gemini 3 Flash has practically beaten ARC-AGI-1 [an AI evaluation] at cost/score parity! It achieved the same score at more than 500x lower cost than the o3 model from a year ago & 6x lower than the just-released GPT-5.2!
Here is the link.
My Conversation with Alison Gopnik
Here is the audio, video, and transcript. Here is part of the episode summary:
Tyler and Alison cover how children systematically experiment on the world and what study she’d run with $100 million, why babies are more conscious than adults and what consciousness even means, episodic memory and aphantasia, whether Freud got anything right about childhood and what’s held up best from Piaget, how we should teach young children versus school-age kids, how AI should change K-12 education and Gopnik’s case that it’s a cultural technology rather than intelligence, whether the enterprise of twin studies makes sense and why she sees nature versus nurture as the wrong framework entirely, autism and ADHD as diagnostic categories, whether the success of her siblings belies her skepticism about genetic inheritance, her new project on the economics and philosophy of caregiving, and more.
Excerpt:
COWEN: If it’s something like height, where there is clearly an environmental component, especially if the child is not well-fed, but it seems perfectly fine to say above a certain dietary level, it’s mostly genetic, right? No one says that’s ambiguous, and more and more traits will become like that.
GOPNIK: Well, first of all, I’m not sure that’s true. To a striking degree, the traits that people have looked at, like educational attainment, for example — we haven’t found consistent relationships to genetics. I think the reason for that is exactly because there’s this very complicated developmental process that goes from the genetics to the outcome.
Even if you think about fruit flies, for example. I have some geneticist colleagues who work on this — fruit fly sex determination. You’d think, “Well, that has to be just the result of genes.” It turns out that there’s this long developmental — long by fruit fly standards — developmental process that goes from the genetics to the proteins to the morphology, and there’s lots of possibility of variation throughout that. I think that hasn’t turned out to be a scientifically helpful way of understanding what’s going on in development.
The other thing, of course, is, from my perspective, the common features of, say, what kids are doing are much more interesting than the variations. What I really want to know is how is it that anyone could have a brain that enables them to accomplish these amazing capacities? Thinking about, is this child smarter than the other one, given how unbelievably smart all of them are to begin with, I just think it’s not an interesting question.
COWEN: But say, what you would call the lay belief that smarter parents give birth to smarter children, at least above subsistence — surely you would accept that, right?
GOPNIK: Again, what does smarter mean?
COWEN: How you would do on an IQ test.
GOPNIK: What does genetics mean? It’s interesting, Tyler, that IQ tests, for example — they have their own scholarly and scientific universe, but they’re not something that we would teach about or think about in a developmental psychology class, and there’s a good principled reason for that. The good principled reason — this has come up a lot in AI recently. There’s this idea in AI of artificial general intelligence, and that is assuming that there’s something called general intelligence.
Again, I think, a lot like consciousness or life, it’s one of these lay ideas about how people work. When you actually look at children, for example, what you see is not just that there isn’t a single thing that’s general intelligence. You actually see different cognitive capacities that are in tension with one another. You mentioned one about the scientist who’s trying to think of some new idea versus the scientist who’s looking at a more specific idea, right? A classic example of this tension that I’ve talked about and studied is in computer sciences: exploration versus exploitation.
What do you count as IQ? In fact, most of what IQ is about is how well do you do in school? How well do you do on school tests? That’s actually, in many respects, in tension with how good are you at exploring the world around you? The kinds of things that you need to do to have particular goals, to accomplish them, the kinds of things that we emphasize a lot, say, in a school context, are actually in tension. This gets back to the point about babies being more conscious than we are — are actually in tension with the kinds of things that will let you explore.
Think about the Bayesian example. If you have a flatter prior, and you pay more attention to evidence, you are probably not going to do as well on an IQ test…
COWEN: There’s you — you’re tenured at Berkeley, you’re famous. There’s Blake, The Definitive Warhol Biography, and Adam, who’s amazing, writes for the New Yorker, and you don’t believe inheritability and IQ being very concrete things? I just don’t get it. I think you’re in denial.
GOPNIK: Actually, I think that example is maybe partly why I don’t believe in that. In fact, what I do believe is that the effect of caregiving is to increase variability, is to increase variation. Our family, our care — there were six of us in 11 years. My parents were graduate students, and even before they were graduate students, they were that great generation of immigrant kids.
We had this combination of a great deal of warmth, a great deal of love, an enormous amount of stuff that was around us — books and ideas. We got taken to the Guggenheim, when Adam was three and I was four, for the opening of the Guggenheim. We both remember this vividly. But we were also completely free. We were just in regular public schools. As was true in those days, in general, we came home after school, and we basically did whatever it was that we wanted. I was involved. The kids were taking care of each other a lot of the time.
The result is that you get a lot of variation. It’s an interesting example in our family where we have six kids who presumably all have somewhat similar genetics, all in that 11 years grow up in the same context, and they come out completely differently. They come out with really different strengths, really different weaknesses, things that they’re good at, things that they’re not good at. Even if you think about what Blake and Adam and I are like as thinkers, we’re all foxes instead of hedgehogs. We’re all people who have done lots of different things and thought about lots of different things.
So, my view is that what nurture will do is let you have variability. That’s the thing that, in a sense, is heritable. That’s contradictory, the idea that what’s heritable is the standard deviation instead of the mean, but that’s my view about that. I think my childhood did have the effect of making me suspicious of those simple nature-nurture oppositions.
Here are the books of Alison Gopnik.
An RCT on AI and mental health
Young adults today face unprecedented mental health challenges, yet many hesitate to seek support due to barriers such as accessibility, stigma, and time constraints. Bite-sized well-being interventions offer a promising solution to preventing mental distress before it escalates to clinical levels, but have not yet been delivered through personalized, interactive, and scalable technology. We conducted the first multi-institutional, longitudinal, preregistered randomized controlled trial of a generative AI-powered mobile app (“Flourish”) designed to address this gap. Over six weeks in Fall 2024, 486 undergraduate students from three U.S. institutions were randomized to receive app access or waitlist control. Participants in the treatment condition reported significantly greater positive affect, resilience, and social well-being (i.e., increased belonging, closeness to community, and reduced loneliness) and were buffered against declines in mindfulness and flourishing. These findings suggest that, with purposeful and ethical design, generative AI can deliver proactive, population-level well-being interventions that produce measurable benefits.
That is from a new paper by Julie Y.A. Cachia, et.al. A single paper or study is hardly dispositive, even when it is an RCT. But you should beware of those, such as Jon Haidt and Jean Twenge, who are conducting an evidence-less jihad against AI for younger people.
Via the excellent Kevin Lewis.