Notes — Brendan Foody on Teaching AI and the Future of Knowledge Work

Notes on Brendan Foody in conversation with Tyler Cowen — Conversations with Tyler #267, recorded 16 October 2025, published 7 January 2026.

Four questions — Adler’s reading frame

Q1 — What is it about as a whole? Foody, CEO of Mercor — the marketplace that supplies AI labs with the domain experts who write the rubrics and evals models are trained and measured against — walks Cowen through four connected threads: how frontier labs now measure real progress on economically valuable tasks rather than academic benchmarks (Mercor’s APEX index); what kind of human-supplied data is actually scarce (a graded rubric, not raw text, and RLHF preference-ranking where no rubric can be written); what happens to knowledge work once teaching a model to do something becomes a one-time, fixed-cost investment rather than a recurring task (Foody’s ‘society as one giant RL machine’); and, in a long biographical closing stretch, how Foody’s own hiring philosophy, dyslexia, and adolescent donut-arbitrage business inform his view of talent and markets.

Q2 — How is it argued? A rapid-fire conversational interview, not an essay. Cowen presses for concrete numbers (‘Can you put a number on that?’), stages named hypotheticals (Larry Summers and Cass Sunstein as Mercor-hired evaluators; a pendant that records every conversation; scaling the Thiel Fellowship 100,000x), and repeatedly supplies his own competing forecast against Foody’s, inviting Foody to agree or push back rather than simply asserting his own view. The philosophical high point is Cowen invoking Kant’s Critique of Judgment — that taste is precisely what cannot be captured in a rule or rubric — against Foody’s claim that a rubric is the data Mercor most wants; Foody’s answer is to fall back on RLHF preference-ranking as the mechanism for the residue a rubric cannot capture.

Q3 — Is it true, in whole or part? Foody’s specific figures — $150/hour poet pay, GPT-5 scoring 64% on APEX, a 25–30%-per-year rate of improvement on economically valuable tasks, $400M ARR reached faster than any company in history — are self-reported by the company’s own CEO and unverified in the conversation [?]. His forecasts are explicit personal guesses, several hedged in the moment: ‘that might be about two or three years, it wouldn’t surprise me’ (models stumping Cass Sunstein in law), ‘less than a year’ (matching the median Neruda poem), open uncertainty on matching Neruda’s best poems. Cowen visibly disagrees with some of these (he guesses six months, not two-to-three years, for the Sunstein case) and the disagreement is left unresolved, which is more intellectually honest than a monologue would be. The central economic thesis — that knowledge work will restructure around one-time ‘teach the model’ investments the way software restructured around zero marginal cost of copying — is a coherent analogy rather than a demonstrated mechanism; it assumes tacit, undocumented expert judgement (‘taste’) can eventually be captured in a rubric plus RLHF, which is exactly the point Cowen’s Kant objection contests and which the conversation leaves open rather than settles.

Q4 — What of it? The episode gives the wiki’s Evals concept page a genuinely new measurement methodology to sit alongside its existing accounts: rather than a single rubric-writing exercise, APEX starts from surveys of how professionals spend their time (consulting, medicine, law) and derives evals weighted by that economic allocation, explicitly rejecting academic benchmarks (GPQA, IMO) as disconnected from customer-perceived value. It also sharpens the wiki’s existing aesthetic-evals debate (first raised via Sam Altman on Trust, Persuasion, and the Future of Intelligence) by adding Kant as an explicit philosophical objection to rubric-based training of taste, and by proposing RLHF preference-ranking as Foody’s answer to the objection — extending Reinforcement Learning from Human Feedback. Separately, Foody’s ‘society as one giant RL machine’ argument gives that page’s existing RL-environments material (via Edwin Chen on AI Data Quality) an economic mechanism — the fixed-cost-investment analogy to software — and a specific labour-market prediction: a new job category of people who find where a model fails and turn that failure into a trainable RL environment, replacing rather than merely displacing today’s analyst-style knowledge work.

Glossary

APEX (AI Productivity Index) — Mercor’s benchmark for model capability on economically valuable tasks, built from surveys of hundreds of domain experts (ex-McKinsey, Bain, BCG consultants; lawyers; doctors) about how they actually spend their working time, then converted into prompts and rubrics for each time-allocation bucket. Deliberately contrasts with academic evals (GPQA, IMO) that Foody says are disconnected from real economic value. [§ On measuring real-world AI progress]

Rubric — a structured grading scheme (Foody’s analogy: how a professor grades an essay or poem) that specifies what a good model response looks like and by how much to reward or penalise deviations. Foody’s central claim: a rubric — a way of measuring success — is more valuable training data than more raw text to learn from, because it lets a model attempt a problem many times and learn from graded feedback. [§ On why rubrics are the new oil]

RLHF (reinforcement learning from human feedback) — see Reinforcement Learning from Human Feedback. Foody’s fallback for domains where taste resists an explicit rubric: generate two candidate responses, have a high-taste human choose the preferred one, and repeat until the model has absorbed the preference pattern. [§ On why rubrics are the new oil]

RL environment — see Reinforcement Learning from Human Feedback. In Foody’s usage, the artefact an expert builds once — encoding a task, its tools, and how to judge success — so that an agent can then repeat that task indefinitely at near-zero marginal cost, the mechanism behind his ‘society as one giant RL machine’ claim. [§ On turning society into one giant RL machine]

Y-intercept vs slope (hiring) — Foody’s own metaphor for a limitation in talent assessment: it is far easier to measure someone’s current ability (the Y-intercept) than their rate of improvement on the job over months (the slope). [§ On why vibes-based hiring fails]

Price elasticity of demand (applied to jobs) — Foody’s heuristic for which knowledge-work roles survive AI-driven productivity gains: if making a role 10x more efficient causes 10x more of that work to be demanded (his example: software engineering), the job grows rather than shrinks; if demand is roughly fixed regardless of efficiency (his examples: accounting, some customer support, arguably teaching), the same productivity gain instead reduces headcount. [§ On AI and employment]

Key claims by section

On measuring real-world AI progress [§ On measuring real-world AI progress]

Foody’s starting complaint about AI research: the field over-indexes on academic evals (GPQA for PhD-level reasoning, IMO for Olympiad maths) that are disconnected from what customers actually want automated — a medical diagnosis, a legal draft, a financial analysis. APEX corrects this by surveying experts on how they spend their working time and deriving evals weighted by that allocation, using time spent as a proxy for economic value. On his account, GPT-5 scores 64% on APEX, up from a much lower GPT-4o baseline a year earlier — a 25–30% per-year improvement rate on economically valuable tasks [?]. He is candid about a mapping problem this leaves open: in medicine a 30% failure rate is unacceptable (near-perfection required, as with autonomous vehicles), while in law or consulting the same score already represents real, usable value.

On why rubrics are the new oil [§ On why rubrics are the new oil]

Foody distinguishes two kinds of training data: curriculum (text a model reads and learns from) and measurement (a rubric, a test question with a known answer, a unit test) — and argues the second is far more valuable because it lets a model attempt a problem repeatedly and learn from graded outcomes, rather than absorbing it once as text. Pressed by Cowen with Kant’s claim that taste is exactly what a rubric cannot capture, Foody concedes some domains (poetry especially, more ‘unbounded’ than a scoped essay prompt) resist rubrics and require RLHF preference-ranking instead — a genuinely different data type from a rubric, not a patch on it.

On enshrining taste in LLMs [§ On enshrining taste in LLMs]

Asked whose taste a model should learn when studies show ordinary readers often prefer AI-generated poems that experts judge worse, Foody frames it as a design choice for labs and researchers, not a fact to be discovered — ‘what do they think good looks like?’ Cowen raises the sharper version: the very best poetry, by most people’s own standard, is old (Milton, Wordsworth, Blake), so training only on present-day taste may enshrine a taste inferior to the past’s, and there is no way to resurrect the greatest historical judges to grade against. Foody’s guess is that, given enough time, models will be able to draw on the taste of every era and personalise which one they invoke, rather than the field settling on one.

On turning society into one giant RL machine [§ On turning society into one giant RL machine]

Foody’s central economic thesis: a large share of knowledge work will shift from recurring effort to one-time teaching. Instead of an investment banker re-analysing a data room from scratch on every deal, or a customer-support agent handling the same ticket type daily, the expert instead builds an RL environment once — codifying the task and how to judge success — after which an agent performs it indefinitely. He draws the explicit analogy to software’s zero marginal cost of copying: ‘they’ll be able to use that many times as they use their agent,’ the same shift that made building software once and distributing it at scale replace bespoke repeated labour.

On when AI will stump experts [§ On when AI will stump experts]

Using Cass Sunstein as the test case, Foody guesses it will be two to three years before Sunstein struggles to find a model error in law, because so much of legal judgement is undocumented tacit taste rather than codified rule; Cowen disagrees, guessing six months, and the two do not resolve the gap. On poetry specifically, Foody expects a model to match the median Pablo Neruda poem within a year but is not confident enough to forecast when a model matches Neruda’s best work, agreeing with Cowen’s intuition that the long tail of quality is the hardest part to close. His general heuristic: models are advancing fastest on 50–75% of what an expert does, and will struggle for a long time with the last 25% — the tacit, undocumented residue.

On why vibes-based hiring fails [§ On why vibes-based hiring fails]

Foody’s diagnosis of conventional hiring: interviewers substitute an unstructured, personality-driven conversation for direct measurement of the actual skill the job requires, and Mercor’s founding insight was to automate resumé review, interviewing, and skills assessment around measured task performance instead. His caveat on AI-assisted assessment: rather than banning tools during a test (as some academic settings tried with ChatGPT), Mercor now asks candidates to use every available tool, since performance-with-tools is the better predictor of on-the-job impact. He names the harder, unsolved half of talent assessment as measuring someone’s slope — how fast they improve on the job — versus their easily-measured current ability (their Y-intercept).

On solving labor market matching problems [§ On solving labor market matching problems]

Foody’s account of the core labour-market inefficiency: hiring is radically disaggregated — a jobseeker applies to a few dozen roles, a company considers a fraction of a percent of the workforce — and no aggregator plays the role for jobs that (say) Amazon plays for retail. He argues the missing piece is not distribution (LinkedIn already has that) but matching — reliably predicting performance from a profile — and expects this to improve as full-time roles increasingly fragment into remote, fractional RL-environment-building work. On the risk that AI-optimised applications (perfect cover letters, AI-coached interviews) push hiring back toward nepotism and referrals, Foody partly agrees this will happen in some firms, but hopes model-driven, evaluated-against-outcomes hiring decisions can discount unearned nepotistic signal rather than reward it.

On scaling the Thiel Fellowship [§ On scaling the Thiel Fellowship]

Foody, a Thiel Fellow himself who dropped out of college, argues the Fellowship faces the identical matching-and-scale problem as hiring generally — it can interview only a tiny fraction of plausible candidates — and says Mercor has already worked with the Fellowship on AI-assisted interviews and transcript analysis to select fellows better. Pressed on whether the Fellowship’s value depends on staying a small, self-selecting, ‘ornery’ pool, Foody grants the referral effect is real but argues broadening access matters for candidates who would otherwise never reach a venture capitalist; he still expects the process could scale 100x or more with the right technology, while conceding that an aggregate panel of domain experts will likely out-predict any single interviewer, including Peter Thiel, for a long time.

On donuts, debates, and dyslexia [§ On donuts, debates, and dyslexia]

Biographical stretch, source-grounded but lightly summarised here: Foody ran an unlicensed doughnut resale business in eighth grade (‘Donut Dynasty’), was shut down by his school, relocated off-campus, and undercut a rival on price before he ‘had learned anything about anti-competitive laws.’ He met his Mercor co-founders on a champion high-school policy-debate team at age 14 and is dyslexic — a trait he links, via a cited (but unnamed) correlation in the entrepreneurship literature, to comparative-advantage thinking: early, forced practice at delegating tasks he could not do as fast as others (reading evidence quickly in debate), and a bias toward unconventional problem-approaches.

On Mercor’s next steps [§ On Mercor’s next steps]

Foody’s stated next priority is scaling ‘super-realistic’ evaluations of tool use over long horizons (tasks that would take a human days or weeks), aimed specifically at enterprise adoption — his framing is that the last two years of AI progress chased intelligence, while the bottleneck has shifted to usefulness: whether enterprises can actually deploy what the models can now do.