Longreads + Open Thread

Longreads

• Dwarkesh Patel interviews Ilya Sutskver on AI, scaling, taste, and much more. He makes an interesting point early on about models being excellent at narrowly-scoped tasks that have discrete, measurable goals, and worse at broader ones. One way to think about this is that for regular people, there's a correlation between skill on narrow tasks and whatever general trait is being measured; if you add another plate to your max deadlift, you're probably getting stronger in general, and if you get really good at LeetCode, you've probably gotten good at figuring out which algorithm to apply where and at implementing it effectively. But even though AI models have a counterintuitive level of skill transfer (train a big enough model tofor produce prose, and it'll also learn to write poetry and Python), they have uneven training data across different domains and meta-domains. So the more you optimize for specific evals, the more you're training a family of idiot savants. Towards the end, Sustkever talks about taste, and says that one of his guiding principles is that results should be beautiful: he liked building things that replicated stylized neurons because we already have empirical evidence that you can stack lots of neurons together and, with the right starting parameters and inputs, achieve intelligence. But that's the result of a lot of evolutionary brute force, and while evolution can produce beauty, it's perfectly satisfied with functional ugliness—the same process that gave us human brains as a local optimum for one environment produced the naked mole rat as the optimum for another. And it might be the case that intelligence has some level of irreducible kludginess to it.
• Will Dunn in The New Statesman on how bond traders have become an increasingly important force in UK politics. In a sense, it's perfectly legitimate to say that if millions of voters support one thing, and a few dozen very well-paid hedge fund traders oppose it, markets are interfering with a democratic process. But what markets are doing in this case is ringing up the order and telling the government what their plans will cost. If bond prices react slowly to spending plans, those plans will tend to get locked in before their real cost is known. The piece mentions one fund that was putting on trades ahead of last year's general election—they started shorting once they were confident Labour would win, because they were also confident Labour would mess things up while in power. That at least smooths out the impact and means that governments have to adjust to higher borrowing costs in advance.
• Nicholas Hune-Brown on accidentally commissioning a freelance article from an LLM-wielding fabulist who had invented both fake quotes for stories and fake jobs to pad out her résumé. It's hard to have a sustainable setup where the same freelance writing commission can require interviews, research, and hours of writing, or just typing an outline of the article into an LLM. But this also raises the question of how much value the average article really adds to the world—when the editor investigates some previous pieces written by the same writer, she talks to someone who'd allegedly been interviewed, who says the interview never happened but "The quotation attributed to me is the sort of thing I might say." As models get better, their hallucinations will better-approximate reality—ask an LLM to answer something in a public figure's voice, and it'll give you a fake quote, but probably not a quote that that public figure would be alarmed to be associated with. So in some categories, models are pretty competitive with human labor even if they make a convincing fake rather than the real thing. (Maybe at some point the model router will see that you're writing an article for a real publication, give you a list of potential sources, and offer a confirmation screen where you can send your cold email to those sources and see if any of them will give a real quote.)
• Brian Potter asks how Extreme Ultraviolet lithography, building on decades of US government- and industry-funded research, ended up being dominated by a Dutch company. As it turns out, the answer is that ASML was the only company in the venn diagram of having the technical resources necessary, having an interest in actually building the devices, and not being Japanese. Before US chip policy was dominated by concerns about China catching up, the biggest worry was that Japan was already far ahead. US companies didn't feel the need to own the process themselves, but did worry that Japanese companies might develop it and keep the results to themselves. It's a fun story on both the technical and economic fronts (on the technical side, I'd known that one step of the EUV process was generating superheated droplets of tin, but this piece explains more about why).
• This brutal Richard Posner takedown of Supreme Courtjustice William O. Douglas has been circulating recently. If you like reading articulate haters, you'll love it. But one thing the piece illustrates is that people who are very flawed in one domain (or, in Douglas' case, in many different domains) can still make some good calls. Douglas faced impeachment threats three separate times in his tenure, and one of them was because he'd called for the US to cultivate ties with Communist China during the Korean War, as a counterbalance against the Soviets. That did turn out to be an important part of US policy, and it was perceptive to see that possibility so early. The piece also has some good thoughts on what the Supreme Court is actually for; he makes the Straussian case that judges basically legislate, but that they can only do so legitimately when they can come up with a legal justification for their actions. If you're trying to come up with a coherent ruling that fits in with precedents, it constrains the scope of judicial activism, but means that it can still happen—they can basically legislate downhill, by codifying implicit assumptions behind existing laws and rulings, rather than uphill by creating entirely new ones.
• One question a user shared from the Diff AI chat was about how to start researching stocks to buy or short. It's worth taking a step back to ask how to think about whether or not this is a good idea. Investing can be a great education, but it comes with tuition—and you learn different things if you're speculating with a tiny bit of fun money as opposed to betting a substantial fraction of your bankroll. So the first question to ask is what you're trying to get out of it. If you want superior risk-adjusted returns, your best option is almost certainly an index fund; if you're pursuing a career as an investor, you need to be aware that the techniques and information resources available to professionals are very different from what everyday investors can access (and that these professionals also operate at a scale where some strategies that work fine for individuals just can't work). Probably the healthiest attitude is to treat investing as a calibration tool: we all have views about what will happen in the future, and about what matters now. Turning those views into trades is a test of whether your bet on the future is right, whether it matters, and whether or not everybody else saw it coming.
• In Capital Gains this week, we look at the viral $140k is the poverty line piece. Some of the numbers are hazy—if you think through the assumptions, the hypothetical average family is probably at the peak of their non-discretionary expenditures, and to the extent that they're spending in a way that makes sense, they're implicitly assuming that they'll make more money in the future. But the piece also raises the broader question of how to think about poverty in a country that has so much aggregate material abundance, but that also has cost inflation for some unavoidable expenses like having somewhere to live and not dying of a preventable illness. If you look at how people's spending evolves as they get richer, you'll see that even in an economy with pretty optimal policy, you'd expect to hear complaints about the high cost of housing, education, and healthcare, because those are the sectors that can absorb the most spending when every other need has been satisfied.

Books

Peak: Secrets from the New Science of Expertise: There's a fun thing you can do when you write about human nature, where you take our messy middle-of-the-road intuitions, exaggerate them in one direction, and then take the other side of the bet. If the smug, self-satisfied middle class demands a message that they deserve everything they've got because they worked so gosh-darn hard, Nickel and Dimed is an essential wake-up call. If we've forgotten America's roots as a meritocratic, industrious country where anyone can get ahead, to the point that books like Nickel and Dimed can be bestsellers, what we really need is a book like Scratch Beginnings to counter that narrative. And if America is a very rich country, but wealth has a component of inheritance (in the form of direct financial transfers from parents, and the transmission of class norms and behaviors), well, it's tough to write a book whose thesis is "the average person thinks this is a 50/50 issue, but you'll actually find that it's closer to 45/55, though some of that's a matter of which specific studies you read and how you interpret them."

So: Peak, as a book about the concept of deliberate practice and the various ways it can be implemented, is good. Peak, as a book about how world-class performance in a variety of domains is accessible to basically anyone who practices effectively, is not so good. Fortunately, 1) most of the meat of the book is about the first point, and 2) given that all of us have more control over ourselves than over all of society, the good parts of the book are the ones that matter.

Deliberate practice is basically a more structured and serious way to make the point that if you do something over and over again, you'll get better at it. But, you might find that the longer you practice, the slower your rate of improvement. Deliberate practice aims to deconstruct this process, by identifying discrete, measurable outcomes, and targeting a difficulty level where what you're working on is challenging but not impossible.

This discipline is very well-developed in some fields, like music and sports. If you follow a standardized workout plan ("Attempt X sets of Y squats, and if there's ever a session where you finish and feel like you could do more, add five pounds to the bar. Repeat indefinitely.") you're following a deliberate practice protocol: that's just a simple algorithm for finding your point of maximum effort and spending time there. In both physical fitness and music, you can break down and restructure whatever it is that you're working on. A musician who's struggling with a particular section of a piece can just keep setting the metronome to a slower pace until they get everything right, do that a few times, and then start speeding up again. Or they can look at the specific sub-problem—maybe it's hard to produce a particular syncopated rhythm, or there's a tricky series of arpeggios, so you do some exercises focused on that general skill. 

But these are stylized domains where it's easy to narrow things down. If what you're practicing is small talk, it's pretty hard to arrange for someone to engage in small talk with you but to talk at one third of a normal speaking speed so you get in some solid reps of "figure out a normal-human way to keep this conversation going" and then to slowly move uptempo. (Though, come to think of it, maybe an LLM social skills tutor could turn out to be Ozempic for awkwardness—a low-stakes way to rehearse social interactions and get feedback leading to improvement. There are some real-world implementations of this: if you wanted to design some training system where people got used to striking up conversations, developed emotional immunity to rejection, and then were swiftly removed from the long-term social consequences of being the weird person everyone rejects, you might reverse-engineer LDS mission trips. A surprising amount of enterprise software is sold from places like Provo, by people who view hitting their quota as one of the lower-stakes cases times they’ve had to contact a bunch of strangers and persuade them to make a significant decision.)

But it's too simplistic to divide the world into domains where deliberate practice works and is thus table stakes, and domains where it doesn't. There are practice-like things that you can do in a variety of fields. The book talks about Benjamin Franklin's writing exercises, where he'd try to do things like turn an article he read into a poem, then return to the poem in a few weeks and try to reproduce the original article. There are some investors who will have their analysts look at financial statements for some mystery company, with the dates removed, and ask them to assess the business before revealing that they're looking at financials for Coca-Cola in 1988 or Worldcom in 2002 or Netflix in 2011 or whatever. Programmers have LeetCode, Project Euler, and the like.

But the messier your field is, the more you have to be careful about what it is that you're deliberately practicing. You could aim to become a good writer by following Anthony Trollope's pomodoro-esque technique of doing a dozen 250-word sprints a day. But if you're not really careful, you'll find that you're actually practicing the art of padding out ideas until they hit the quota, and of writing about topics that will never get resolved. You can definitely work around that by trying to measure quality, but there's a whole sub-industry within media devoted to debating which novels, movies, songs, etc. have the most artistic merit; the mere existence of critics means that quality is a noisy metric. You could use popularity as a proxy, too, but that's a moving target: a so-so work produced by someone who has an established fan base will probably do better, at least at first, than a masterpiece from a complete unknown. And anyway, gunning for pure popularity leads to the same infinite-torrent-of-slop attractor as targeting a specific word count.

And yet, those constraints really do discipline people. Shakespeare needed seats filled night after night if he was going to get paid to write the next play. I, Claudius was written for rent money, and Balzac literally recycled his personal anecdotes about being a free-spending writer who was behind on his deadlines and had publishers breathing down his neck about returning the advance into plots, like some weird literary parody of Say’s Law. Setting a concrete, explicit goal for quantity and then having too much ego-plus-talent to compromise on quality does seem to work.

At least, for some people. One of the frustrating themes in the book is that the authors see that deliberate practice is associated with achievement in several fields, but they'd really prefer that deliberate practice be the main determinant of success in every field. So they come up with some elaborate, loopy arguments against the concept of talent. For example, one narrative they have is that having a high IQ helps chess beginners learn the basic rules of chess, at which point their teachers notice that they're making more progress and lavish them with attention, which makes them more likely to form habits of deliberate practice, and so on. So the mechanism is that raw talent drives social feedback which drives practice which produces output.

But if you're learning chess, the part where you think really hard does not stop once you've memorized the rules for when you can move a rook. It's still very helpful to recognize patterns, reason by analogy, discover core similarities between seemingly different situations, etc. Being smart drives more efficient memorization! You can make up for that with spending more time memorizing things, and the authors point to an example of a study of young chess players that found that 1) the best of them had slightly higher IQs than the rest, but 2) within that group, there was a negative correlation between IQ and skill, and that the factor that made up the difference was total practice time. The mechanism they proposed is that IQ once again drives propensity to practice, but this time in the opposite direction (suspicious!): the smarter chess players coasted, and their peers hustled to keep up and wound up getting ahead. But any time you find yourself saying "I wish I were less naturally talented so I'd be motivated to spend so much more time practicing that I overcame this disadvantage and then some," you want to ask yourself if you've run into an example of range restriction. If you take a narrow slice at some skill level, that slice will include some people who had natural talent offset by laziness, and others who had less natural talent but made up for it through effective practice. The top chess player in their sample had an 1835 ELO (the group's average was 1603), which is both good for 11-year-olds and not the absolute elite for that age group. And these chess players were the members of chess clubs at a group of local schools that were studied. Which is important because young chess players who are impossible-to-ignore talented will be more likely to be selected out of a sample of students who are in chess clubs at school. To the extent that that happens, the grim outcome of all of this is that practice is a substitute for talent and that the top performers in a given field are the ones who had a head start from the beginning and then proceeded to out-practice everyone, too. (It's important to note that this is speculative; the study's authors don't have good data on the practice habits of students who weren't selected for the original study, and I obviously have even less information on this matter. But any time you see a counterintuitive negative correlation in a narrow slice of a broader sample, this should be your first thought.)

The same chapter talks down some other famous examples of prodigies, but usually taking the form of: there's a crazy legendary story about this person being able to produce masterpieces when they were kids, but that's wrong. Instead, they were better than the average adult professional in their domain as kids, and didn't start producing immortal artistic masterpieces until young adulthood. (So, while writing this review, I wanted to use an example of a random famous musician, so I could say something to the effect of "If you could play Chopin as well as Martha Argerich did when she was ten years old, you're probably both the best piano player in your social circle and not remotely qualified to be a professional. But it occurred to me that I didn’t know exactly when Argerich started. From Wikipedia: "A precocious child, Argerich began kindergarten at the age of two years and eight months, where she was the youngest child. A five-year-old boy, who was a friend, teased her that she would not be able to play the piano, and Argerich responded by playing perfectly, by ear, a piece their teacher played them." That's a pretty big head start! When Tyler Cowen interviewed conductor David Robertson, Robertson referred to the precision with which some composers recognize pitch as "uncanny." And this is someone whose entire career has been in the field of paying incredibly close attention to the nuances of performing classical music.)

All of this is very important in two situations:

1. You've read about deliberate practice and decided to become the best in the world at something for which you haven't previously demonstrated signs of natural talent, or
2. You want deliberate practice to be a bigger part of education policy specifically to address inequality.

If you're just interested in improving your skill at something you enjoy, absolutely none of this matters! If you like painting watercolors because it's meditative, or because they look nice, or because your paintings impress your friends, then by all means take a rigorous look at your work, identify weak points, come up with exercises that address them, and evaluate the results. One of the important points the book makes is that elite performers do not enjoy practice, just the output thereof, and it's a very good idea to develop the habit of doing unpleasant-in-the-moment things because of the eventual rewards. But don't quit your job to take up some practice-driven skill with the expectation that you can be #1 in the world at that. The amount of practice you can put in is the amount that someone with more natural ability in that domain can use to retain their lead.

The science of expertise is hard because we don't have good controlled experiments. The authors talk about a study they ran where they coached someone on memorizing long strings of randomly-selected numbers, and were able to hit world record-level performance. And one reason they were able to do that was because there weren't many people trying! As memory contests have gotten more popular, they've gotten much more competitive. As the idea of deliberate practice spreads, differences make a smaller difference in how well people do, in the same way that differences in childhood nutrition have a smaller impact on adult health outcomes in parts of the world that are rich enough that everyone can afford to eat. If everyone optimizes the things under their control, the outcome is determined by whatever wasn't under their control, which is some variety of luck. That's a positive externality for the rest of the world, because it means that we get to see incredible elite-level performance in so many different domains. But it means that if you're going to win through being more willing to grind than everyone else, the main criterion you need to use is the ratio between how much you care about winning that competition and how much the rest of the world does, because when enough people care, maximum grind is table stakes.

There are eight billion of us in the world, and I don't think we could collectively come up with a list of eight billion separate things-one-could-be-good-at. So we can't all be the best. And "best" status will be some combination of raw ability and effort, with effort playing a bigger role in absolute outcomes over time (we learn more about training) but a smaller role in relative outcomes at the tails (you can go from practicing for three hours a day to nine hours a day, but you can't triple it again after that). So, if you want to feel better, remember that even if we can't all be the best, we can all get better—with the exception of people who are already elite-level performers and have put in the maximum possible amount of practice, but have reached the point where the skill accumulation of practice is offset by skill depletion of aging. So there’s that.

Open Thread

• Drop in any links or comments of interest to Diff readers.
• What are some versions of deliberate practice that work well in your field? 

Diff Jobs

Companies in the Diff network are actively looking for talent. See a sampling of current open roles below:

• A startup is automating the highest tier of scientific evidence and building the HuggingFace for humans + machines to read/write scientific research to. They’re hiring engineers and academics to help index the world’s scientific corpus, design interfaces at the right level of abstraction for users to verify results, and launch new initiatives to grow into academia and the pharma industry. A background in systematic reviews or medicine/biology is a plus, along with a strong interest in LLMs, EU4, Factorio, and the humanities.
• Series-A defense tech company that’s redefining logistics superiority with AI is looking for a MLE to build and deploy models that eliminate weeks of Excel work for the Special Forces. If you want to turn complex logistics systems into parametric models, fit them using Bayesian inference, and optimize logistics decision-making with gradient descent, this is for you. Python, PyTorch/TensorFlow, MLOps (Kubernetes, MLflow), and cloud infrastructure experience preferred. (Salt Lake City or NYC)
• A hyper-growth startup that’s turning the fastest growing unicorns’ sales and marketing data into revenue (driven $XXXM incremental customer revenue the last year alone) is looking for a senior/staff-level software engineer with a track record of building large, performant distributed systems and owning customer delivery at high velocity. Experience with AI agents, orchestration frameworks, and contributing to open source AI a plus. (NYC)
• Well funded, Ex-Stripe founders are building the agentic back-office automation platform that turns business processes into self-directed, self-improving workflows which know when to ask humans for input. They are initially focused on making ERP workflows (invoice management, accounting, financial close, etc.) in the enterprise more accurate/complete and are looking for FDEs and Platform Engineers. If you enjoy working with the C-suite at some of the largest enterprises to drive operational efficiency with AI and have 3+ YOE as a SWE, this is for you. (Remote)
• Ex-Bridgewater, Worldcoin founders using LLMs to generate investment signals, systematize fundamental analysis, and power the superintelligence for investing are looking for machine learning and full-stack software engineers (Typescript/React + Python) who want to build highly-scalable infrastructure that enables previously impossible machine learning results. Experience with large scale data pipelines, applied machine learning, etc. preferred. If you’re a sharp generalist with strong technical skills, please reach out.

Even if you don't see an exact match for your skills and interests right now, we're happy to talk early so we can let you know if a good opportunity comes up.

If you’re at a company that's looking for talent, we should talk! Diff Jobs works with companies across fintech, hard tech, consumer software, enterprise software, and other areas—any company where finding unusually effective people is a top priority.