What this covers

Brendan Foody and Tyler Cowen discuss what actually limits AI progress in automating knowledge work. Foody, a researcher focused on AI evaluation, argues that the binding constraint is not model capability but the ability to measure what "good" looks like—establishing evals and rubrics that track economically valuable outcomes rather than academic benchmarks. The conversation moves through how models will automate 50–75% of expert work but will struggle with the final quarter that depends on undocumented taste and judgment, and how this creates a new labor category: people who train agents rather than repeatedly performing tasks themselves. The core insight is that expert time as economic value proxy should guide evaluation design, not standardized tests disconnected from what customers actually pay for.

The two cover distinct domains to illustrate the principle: medicine (where near-perfect performance is mandatory), law (where tacit knowledge dominates), poetry (where taste cannot be rubricized), and hiring (where skills-based hiring beats vibes). Foody distinguishes between two kinds of data—the content models learn from versus measurement data that lets them iterate—and explains how price elasticity determines which jobs AI will displace most (elastic domains like software expand; inelastic ones like accounting shrink). He introduces APEX, a benchmark built to measure economically valuable tasks by surveying expert time allocation, and describes the shift toward software economics of knowledge work: a single investment in building an RL environment replaces repeated manual labor. The discussion reaches into how Kant argued taste resists codification, and what role consensus among exceptional graders should play in subjective evaluation.

Sharpest takeaway

Foody argues that the scarcest resource in AI progress is knowing what 'good' looks like—measuring economically valuable performance via expert-built evals and rubrics—and that human expertise will remain essential for the hardest last 25% of knowledge work, creating a new job category of people who train AI agents.

  • Academic evals (GPQA, IMO math) are disconnected from the economically valuable outcomes customers care about
  • Models can automate 50-75% of expert work but will struggle with the final 25% that depends on undocumented taste
  • Knowledge work will shift to a fixed-cost model of teaching agents once and reusing them, like software

The argument · threads10 threads · 42 claims
0.71

Better AI assessment and matching technology will dramatically improve labor-market efficiency despite inherent confounding variables.

10 pointscentrality 4/5
  • The Thiel Fellowship is constrained by the same matching problem as labor markets—it can only interview a tiny fraction of people who might be a good fit—so Mercor built AI interviews and used models to analyze transcripts to find signals that better select Thiel Fellows.

    “theThiel Fellowshipis constrained by that exact matching problem... they can only consider, and interview a fraction of a percent of the people in the world... So, we’ve worked with them on building out AI interviews that are able to better assess Thiel Fellows”

  • Body language in an interview can be predictive but is also often a false signal—Foody has over-indexed on a candidate seeming awkward who then did phenomenal work—so it is important to be cautious about which signals are actually correlated with job performance.

    “I think it can be, but I also think it can be a false signal because I’ve definitely had cases where I over-index on, “Oh, this person feels a little bit awkward,”... but they do a phenomenal job at the actual work”

  • The largest hiring mistake people make is not measuring the actual skills and capabilities they want exhibited on the job—instead they have a vibes-based conversation about where someone grew up, how similar they are, and whether they'd enjoy hanging out—over-indexing on cultural fit relative to demonstrated skills.

    “one of the largest problems that people make is that they don’t measure the actual skills and capabilities that they want someone to exhibit on the job. Instead of focusing on how do we measure how well this person does their investment analysis... they have this vibes-based conversation of where did the person grow up, how similar are they”

  • An AI-run, more efficient labor market will likely have many second chances and accommodate late bloomers, because the models will recognize that second chances are often effective—identifying when someone wasn't the right fit for a first role but would excel in another—since many jobs exist that almost everyone could do well given the right match.

    “there will be a lot of second chances, and the reason that there will be is that oftentimes they’re effective. The models will identify that, and realize that maybe someone wasn’t the right fit for that first role. There’s another role that they could be a really good fit for.”

  • While the best models will be better than any single best individual interviewer, the aggregate sum of all experts across every domain will likely remain better than the models for a long time.

    “Certainly, the best models will be better than any single best individual, but I would expect that the aggregate sum of all experts in each domain will likely remain better than the models for a long time.”

  • Rather than fighting AI tools in assessment (like forcing essays on paper to block ChatGPT), the right approach is to let candidates use all the tools—e.g., record their screen building a product with Codegen tools over an hour—because what they can do with the tools is a far better predictor of real impact than performance without them.

    “the right way of approaching it is seeing what people can do when using all of those tools... that’s a far better predictor of this person’s ability to actually deliver impact than it is to say, “Don’t use the tools at all.””

  • AI predicting job performance from taped interviews will be superhuman because humans are bad at it, but variance will remain due to confounding variables—family issues, illness during the interview—though with enough contextual data (manager notes about the person's life during interviewing and on the job), it could become phenomenal over roughly a 10-year horizon.

    “it will be certainly superhuman because humans aren’t very good at it, but it’s still such a difficult problem that there’s going to be variance... there’re a lot of confounding variables. Did the person have an issue in their family... Did they get sick during the interview process”

  • It is much harder to measure a candidate's slope—how they will develop on the job over a six-month horizon—than their Y-intercept, their current level of ability; for non-project roles you have to use prior similar roles as the best proxy and drill into details and references.

    “It definitely is more difficult to measure someone’s slope and how they’ll develop on the job over a six-month time horizon than it is to measure their Y-intercept.”

  • The largest labor-market inefficiency is disaggregation—candidates apply to a couple dozen jobs and companies consider a tiny fraction of people—but the binding constraint is a difficult matching problem, not distribution: LinkedIn already has distribution to every company and candidate, yet it is incredibly hard to predict from a LinkedIn profile whether someone will perform well at a given job.

    “LinkedIn has all the distribution to pretty much every company and every candidate, but at the very same time, it’s incredibly difficult to understand, based on someone’s LinkedIn profile, whether they’ll actually perform well at a given job... it’s very much a matching problem — less so a distribution and aggregation problem”

  • As candidates all submit perfect cover letters and AI-optimized profiles, hiring may revert to nepotism and referrals, but Foody hopes models running companies will hold an eval set of all employees' performance reviews and make accurate predictions about whether a given reference or piece of nepotism should be weighted positively or treated as a countersignal.

    “the models have an eval set of all the performance reviews of people in that given company, and they’re able to make an accurate prediction over whether this reference, or that piece of nepotism should actually be considered, or maybe as a countersignal”

0.70

Measurement data and expert-built evaluations are the most valuable training signal for models.

5 pointscentrality 5/5
  • There are two kinds of data: output/curriculum data the model reads and learns from, and measurement data (rubrics, test questions and answers, unit tests) that lets models attempt a problem many times, score responses, and learn from them—and the second kind is the most valuable.

    “The second kind of data is some way of measuring success, where you have the rubric for the response, you have the test question and answer, you have the unit test, and code. That second kind of data is the most valuable, where we’re able to have the models attempt the problem many, many times, score those responses, and learn from them”

  • A good heuristic for the most valuable data is: if you could build a model that, without seeing the test, could max it out, how much economic impact would that add? Whatever test best measures that economically valuable capability is the most helpful.

    “Maybe a good heuristic is, if we could build a model that, without seeing this test and reading through it, could max out the test, how much economic impact would that add? Whatever test is able to measure that the best is most helpful.”

  • APEX measures economic value by surveying hundreds of experts within each industry about how they spend their time, then having them write prompts and rubrics for each activity bucket; expert time allocation is the best proxy available for economic value because it maps to what customers are willing to pay for.

    “Using their time is the best proxy we have for the economic value associated with their salary, or what customers are willing to pay for”

  • The largest way deep domain experts can contribute to AI advancement is by defining the evals—building phenomenal tests for model capabilities—because once researchers have these tests, they can hill climb and optimize around them remarkably fast.

    “the largest way that these deep domain experts can help to contribute to the advancement of AI is defining the evals. When we have these phenomenal tests for model capabilities... it’s amazing how fast the researchers can hill climb them and optimize around them”

  • The largest disconnect in AI research is that everyone focused on academic evals like GPQA for PhD-level reasoning or IMO for Olympiad math, which were wholly disconnected from the outcomes customers actually care about—automating a medical diagnosis, a legal draft, or a financial analysis—which is why APEX was built to measure economically valuable tasks.

    “everyone was focused on academic evals, likeGPQAfor PhD-level reasoning orIMO for Olympiad math, which werewhollydisconnected from the outcomes that customers actually care about”

0.67

Models will automate 50–75% of expert work but struggle with the last 25%, making human expertise a persistent bottleneck.

4 pointscentrality 5/5
  • Models struggle at two things humans excel at: longer-horizon tasks that take 50 to 100 hours rather than a few hours, and integrating multiple tools (including interacting with people) into a response—but Foody expects enormously capable models on both dimensions within six to twelve months.

    “The first is these longer horizon tasks of not just something that we could do in a few hours, but something that might take us 50 or 100 hours to do. Then the second thing is integrating multiple tools”

  • Models will struggle in law for a long time because much of the right way to approach legal problems is not written down or codified—it exists in the heads of experts—and without those tokens in pre-training data or expert-created post-training data, the model will inevitably struggle.

    “There’re interestingly a lot of areas in law where the right way of approaching something is not written down or codified. It exists more in the heads of experts”

  • Models are advancing very quickly toward automating 50 to 75 percent of what humans and experts can do, but will really struggle with the last 25 percent; human expertise will for a very long time be the imperative bottleneck to that final 25 percent and thus to greater economic prosperity.

    “models are advancing very, very quickly in being able to automate — call it 50 percent or 75 percent of what humans and experts are able to do — but will really struggle with that last 25 percent. I think that for a very long time, human expertise will be imperative to help accomplish that last 25 percent as the ultimate bottleneck”

  • Models will likely match the median Pablo Neruda poem within about a year, but matching the very best Neruda poems is much further out, because the longer tail of advancement is generally the most difficult.

    “I think that’s probably not too far off.”

0.53

Domain experts contribute most by defining evaluations that capture economically valuable capabilities and edge-case behaviors.

4 pointscentrality 4/5
  • Mapping test scores to economic impact is complex because in high-stakes domains like medicine you cannot tolerate a 30 percent failure rate—you need near-perfect performance like driverless cars—whereas in domains like an initial legal draft or consulting analysis a 64 percent score already has profound, accelerating impact.

    “in certain industries like medicine, you can’t have a 30 percent failure rate. You need to have near-perfects or similar to driverless cars in some ways. In other industries, like an initial legal draft or a consulting analysis, this technology is already starting to have a profound impact”

  • Mercor can pay top experts like poets $150/hour because once these experts teach the models how to do something once, that knowledge is applied across billions of users, making the per-expert investment highly leveraged.

    “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”

  • Nonprofit institutions don't freely send their evaluative data to AI labs because of two things: lack of awareness about why evals matter outside Silicon Valley, and fear about how AI will impact their jobs, work, and ability to contribute meaningfully to the economy.

    “everyone worries about how is AI going to impact their jobs, their work, their ability to contribute to the economy, and be meaningful. I think that that’s always top of mind, even for nonprofit organizations that want to contribute and preach this world of abundance”

  • For subjective domains like poetry, you want some degree of consensus among exceptional graders that they are each doing a good job, but you do not want too much consensus, because disagreement captures edge-case model behaviors that deviate from the norm.

    “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”

0.50

Knowledge work will shift from task repetition to training agents and building RL environments as a fixed-cost investment.

3 pointscentrality 4/5
  • Knowledge work will shift to be like software: instead of repeatedly doing tasks (an investment banker reanalyzing a data room every few weeks, a support rep answering tickets daily), workers will teach an agent or build an RL environment once as a fixed-cost investment, then use the agent many times—so a huge portion of the economy will become an RL environment machine.

    “we’ll move towards these fixed-cost investments of teaching an agent how to do something, building an RL environment for something, and then being able to use agents as many times as we want to perform that activity. That’s why I believe that a huge portion of the economy will become an RL environment machine.”

  • Rather than automating jobs away, AI will create a new job category of people training agents and building RL environments; within five years a majority of high-end knowledge workers may be training models—whether in full-time jobs or through marketplaces—instead of doing the underlying analysis themselves.

    “we’ll move towards these fixed-cost investments of teaching an agent how to do something, building an RL environment for something, and then being able to use agents as many times as we want to perform that activity. That’s why I believe that a huge portion of the economy will become an RL environment machine.”

  • To hold the new agent-training jobs, a person needs to know the domain thing, not technical AI—the only technical skill required is the ability to find where the model makes a mistake and understand the frontier of its capabilities, after which it is relatively easy to create criteria so the model can learn.

    “They just need to know about the thing. The only element of technical AI that they’ll need is to find where the model makes a mistake... then it’s relatively easy to create some criteria or way of measuring that mistake so that the model can learn from it.”

0.43

AI personalization and expert-driven reputation will reshape trust and education even as systems match or exceed human expertise.

5 pointscentrality 3/5
  • Dating apps are generally good because better technology to solve matching problems enables people to be happy in their lives, contrary to the widespread negative view of them.

    “I think they’re good. I’m very much a proponent of better technology to solve these matching problems and enable people to be happy in their lives.”

  • Most people will continue to value privacy and won't be comfortable adding their personal data to broad model weights, so companies with strong privacy brands like Apple—even if not at the AI frontier—will gain a trust advantage that lets them collect personalized information for individual agents.

    “some of the companies that have done a very good job around their brand of privacy are going to have an advantage in it. Apple — while maybe not totally at the frontier of AI yet — has done such a good job in theirbrand around privacy. That’s going to allow them to have a lot of trust from users”

  • Education will be transformed by everyone having a Sal Khan–like personal tutor available 24/7 to teach any topic, making it easier to self-motivate and providing better access to and explanation of information; teaching roles may shrink in part, but teachers will remain important for guiding students' emotional development and personal relationships in smaller class settings.

    “a good heuristic is if everyone hasSal Khanas their personal tutor, available 24/7 to teach them whatever topic they want to learn”

  • There is a dating crisis for smart young men in San Francisco, likely driven by gender imbalance in certain industries, but not in New York; dating apps generally help drive efficiency in solving the matching problem in society.

    “I think certainly in San Francisco. Not in New York... I think it’s probably gender imbalance in San Francisco, especially in certain industries.”

  • When top models clearly surpass human experts, the reputation of expertise may actually rise rather than fall, because the model delivers an impersonal distillation of expert knowledge without an annoying personality, and not having a face causes people to place high trust in it.

    “There’s some element of it not having a face to it that causes us to place this high trust.”

0.36

Subjective domains like poetry require taste-driven training methods when rubrics cannot fully capture quality.

3 pointscentrality 3/5
  • There is a real tension between optimizing models for what power users of AI products want to see and be happy with, versus what the top 1 percent of poetry experts would call a broadly good poem; striking that balance is up to the labs' researchers and product leaders defining what good looks like.

    “that sometimes deviates from the types of responses that the top 1 percent of experts in poetry might say is a broadly good poem. So, striking that balance is really up to a lot of the researchers and product leaders at the labs”

  • Over a long enough horizon, models will enshrine taste from every decade and era, and then learn what taste an individual user has and pull on each of those knowledge bases to personalize responses to their preferences.

    “in a long enough time horizon, we’ll enshrine taste from every different decade and every different era. Then the model will be able to learn what taste doyouhave and how does it pull on each of those knowledge bases to best personalize it to your preferences”

  • When taste cannot be captured in a rubric—as Kant argued—an alternative is RLHF: have the model generate two responses, have people with taste choose which they prefer, and repeat many times until the model understands their preferences.

    “Another kind isRLHF, 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”

0.34

Price elasticity of demand determines whether AI-driven efficiency expands or displaces jobs in each sector.

2 pointscentrality 3/5
  • Building businesses, product and distribution of software, capital allocation, and resource allocation within companies are all price-elastic domains with enormous remaining inefficiency, so better AI investing and analysis will produce much more of these activities.

    “there’s still enormous inefficiency with respect to how we allocate capital in the economy... there’s so much elasticity with respect to how we build more products, how we distribute those products, and how we allocate resources within companies more effectively”

  • Price elasticity of demand is the right thing to focus on for how job displacement evolves: if AI makes software engineers 10 times more efficient, we may get 10 times as many engineers building 100 times as much software, whereas inelastic domains like accounting or customer support won't expand because we only need so much of them.

    “I think the elasticity is the exact right thing to home in on with respect to how job displacement will evolve in these domains. I think if we make software engineers 10 times more efficient, we’ll have even more software engineers... Versus other domains... we only need so much accounting in the world, or we only need so much customer support”

0.23

Success depends more on identifying and leveraging comparative advantages than on fixing weaknesses.

3 pointscentrality 2/5
  • Hiring has so much overlap with investing that imagining Peter Thiel interviewing everyone in the world at 18-20 and deciding whether to give them a $100K check would be powerful for economic mobility and company creation; Foody believes this will eventually happen given the right technology and focus.

    “hiring has so much overlap with investing. Imagine if we could have Peter interview everyone in the world when they’re 18 or 20... and make a decision around whether he wants to give them a 100K check. That would probably be very powerful with respect to economic mobility”

  • So much of finding success in your career is understanding what your strengths are and how to leverage them—much less about your weaknesses—which is how Foody approaches Mercor and encourages employees to think about their comparative advantages within the business.

    “So much of finding success in your career is just understanding what are your strengths, and how do you leverage those, and much less about what are your weaknesses.”

  • Beyond intelligence, being good at extemporaneous speaking depends on high clarity of thought (which correlates strongly with intelligence), confidence and willingness to iterate through practice, and speed of thought—a dimension distinct from aptitude, since some people have high aptitude but think slowly while others are quick on their feet.

    “high clarity of thought often correlates very strongly with people that speak very well... a second thing is confidence... maybe a third one is, more than just intelligence, it’s also the speed of thought”

0.20

Model improvement in economically valuable tasks sustains roughly 25–30% annual growth, as shown by APEX benchmarks.

2 pointscentrality 2/5
  • The rate of model improvement at economically valuable tasks is roughly 25-30 percent per year; GPT-5 now scores 64 percent on APEX, up from GPT-4o a year ago, though sustaining that rate would be challenging.

    “call it a 25 percent–, 30 percent–improvement. Per year, exactly. Now, GPT-5 is at 64 percent, so maintaining that would definitely be challenging.”

  • Mercor's next goal is scaling super-realistic evaluations that measure how models use diverse tools on trajectories taking days or weeks, bridging the gap between the intelligence focus of the last two years and what enterprises actually want to use.

    “scaling up a lot of these super-realistic evaluations... of how do we measure the ways that models use all sorts of different tools on trajectories that would take someone days or weeks to do... For the last two years, people have been very focused on the idea of intelligence rather than the idea of models being useful.”