A very particular set of skills

If you're really good at something, do that thing?

If you ran a mid-sized coffee chain, and a million Yale-educated lawyers showed up at your corporate headquarters offering you their services for $5.92 an hour, you would have four options:

1. Ask them to leave. You already have an in-house lawyer that you probably like, and you don’t need a million more. 

2. Replace your current lawyer. Even if you like them, they almost certainly cost a lot more than $5.92 an hour, and you probably don’t like them that much. You might even replace your lawyer with ten lawyers: Perhaps you’re constantly negotiating new leases and navigating complex employment regulations and fighting lawsuits about not putting any pineapple in your Pineapple Passionfruit Refresher.¹ For $5.92 an hour, you might as well hire an entire team of lawyers. 

3. Hire the lawyers to do other jobs. Yale-educated lawyers are smart, and can probably learn how to do other useful things at your friendly neighborhood coffee conglomerate. They might be good baristas, coffee bean procurement managers, and customer service representatives. This is a less obvious thing to do with a million lawyers, but it’s a pretty obvious thing to do with a million talented people who are willing to work for less than six dollars an hour.

4. Close the coffee shop and become a law firm? And this is clearly the right choice? Of course, maybe you really like making coffee, and you’re in it for the love of the game. But more likely, you’re a business person who doesn’t care if you’re making coffee or making burritos, so long as you’re making money. And while a $5.92-an-hour-Yale-educated lawyer might be a passable barista who costs 60 percent less than the average barista, a $5.92-an-hour-Yale-educated lawyer both makes a better lawyer and costs 95 percent less than the normal Yale-educated lawyer. For a coffee shop, an infinite army of cheap lawyers offers a way to save a little bit of money. But for a law firm, an infinite army of cheap lawyers offers a way to make tons of money.² 

On one hand, this fourth option is clearly crazy. You run a coffee shop! You sell pastries! It would be insane to try to pivot a cafe into a corporate law firm. On the other hand, it would also be insane not to? You have a million lawyers ready to work for you! For next to nothing! How could a change in circumstances that dramatic not change everything about the business that you’re running?

In other words, if you run a business and suddenly you have access to a new resource that has a very particular set of skills, it’s tempting to shoehorn those skills into whatever you’re already doing. You make coffee and need a few lawyers; use the resource to make more coffee or replace your handful of lawyers. But the correct thing to do—or at least the thing that would make you more money—is probably to shoehorn a new business around the resource. Rather than trying to fit the skill to your current problems, find some new problems that match the skill.

Anyway. We can all now hire a million lawyers³ for $5.92 an hour,⁴ and seem to be having trouble figuring out what to do with them. From Jim Covello, the Head of Global Equity Research at Goldman Sachs:

Many people seem to believe that AI will be the most important technological invention of their lifetime, but I don’t agree given the extent to which the internet, cell phones, and laptops have fundamentally transformed our daily lives, enabling us to do things never before possible, like make calls, compute and shop from anywhere. Currently, AI has shown the most promise in making existing processes—like coding—more efficient, although estimates of even these efficiency improvements have declined, and the cost of utilizing the technology to solve tasks is much higher than existing methods. For example, we’ve found that AI can update historical data in our company models more quickly than doing so manually, but at six times the cost.

More broadly, people generally substantially overestimate what the technology is capable of today.⁵ In our experience, even basic summarization tasks often yield illegible and nonsensical results. This is not a matter of just some tweaks being required here and there; despite its expensive price tag, the technology is nowhere near where it needs to be in order to be useful for even such basic tasks. 

Though I love a good rant about how some popular thing fails—and I said something similar about AI back in March—this seems like an incomplete argument about its limitations. Almost two years into the generative AI hype cycle, most companies have tried to apply it to existing problems: Help us write code; help us create marketing collateral; help us cheat on our homework. Help us sell coffee. And as Covello says, AI hasn’t been especially good at very many of these things. 

But perhaps that’s because we’re approaching the problem from the wrong side. Perhaps the way AI breaks through—or maybe more concretely, the way that the companies that figure out how to make lots of money with AI end up using it—isn’t by solving some existing problem with an LLM; it’s by changing the problem to make it easy for AI to solve. 

There is some precedent for this sort of thing. Decades ago, big investment firms made money by picking stocks. Though they were quantitatively rigorous about how they made trading decisions—they looked at charts and numbers and most didn’t exclusively invest on hunches⁶—the problem was fundamentally an analytical one. Experts did some math, stared at their numbers and charts, and tried to outsmart the market. 

And then, computers got really fast. People created early artificial intelligence programs; one of them “solved” chess, in ways that made no sense to human players. And so, financial firms had two choices: They could either use this new computing technology to make them better stock analysts—that is, they could apply the new resource to an existing problem—or they change the problem to fit the resource, and turn stock trading into a game that could be played by machines that were really good at chess. 

Renaissance Technologies—potentially the best performing investment firm of all time—took the second approach:⁷ 

Early on, [ Renaissance founder Jim ] Simons traded like most others, reading the news and betting on gold prices and other investments. He scored gains but found it hard to endure the market’s volatility. …

Simons set out to build an automated trading system that could discover fleeting and overlooked patterns in prices.

“I want models that will make money while I sleep,” Simons told a friend at the time. “A pure system without humans interfering.”

Simons hired a bunch of computer scientists and physicists who knew nothing about finance, and found a problem that could be solved by computers and people who were really good at math:

It’s clear that the way [ Renaissance Technologies’ executives ] Brown and Mercer approached programming was fundamentally different from the way other hedge-fund programmers thought about it. At Tudor, for example, Sushil Wadhwani trained a machine to approach markets in a way that made sense for human traders. By contrast, Brown and Mercer trained themselves to approach problems in a manner that made sense for computers. … Presented with apparently random data and no further clues, they sift it repeatedly for patterns, exploiting the power of computers to hunt for ghosts that to the human eye would be invisible. …

Robert Mercer … [says]: “If somebody came up with a theory about how the phases of Venus influence markets, we would want a lot of evidence.” But he adds that “some signals that make no intuitive sense do indeed work.” Indeed, it is the nonintuitive signals that often prove the most lucrative for Renaissance. “The signals that we have been trading without interruption for fifteen years make no sense,” Mercer explains. “Otherwise someone else would have found them.”⁸

Though Renaissance dropping their old approach and becoming an automated trading firm wasn’t as extreme of a change as a coffee shop becoming a law firm, the two pivots are different in degree more than different in kind. Simons, who taught math at Harvard and MIT, was well connected to lots of mathematicians and scientists. He initially tried to make them financial analysts, and use their skills to help him pick stocks. He discovered they weren’t especially good at that, but they were very good at math. And he realized that, if he could make financial markets a math problem rather than an analytical problem, he could—and then, did—make a lot of money. 

If you have friends who are mathematicians, don’t ask them to be accountants; come up with a way to make solving math problems very lucrative. If you can hire a million lawyers, don’t turn them into baristas; become a law firm. And if you have access to today’s state-of-the-art generative AI models, don’t ask them to send emails and schedule appointments and be digital clones of ourselves; instead, find a problem that can be well solved by a machine that can consume enormous amounts of information, remember nearly all of it, and write remarkably creative, if not always entirely accurate, summaries of what it read.⁹ 

Of course, I don’t know what that problem is. If I did, I wouldn’t be writing a blog; I would have 31 billion dollars. But we won’t find these sorts of problems until we stop spending most of our time trying to teach LLMs how to be a better barista.

And at least in the narrow world of the data industry, we seem stuck in the coffee shop. Over the last two years, countless startups have tried to use AI to make us more effective at solving the same problems that we wanted to solve three years ago. AI agents write SQL; they produce documentation; they alert us about anomalies and do root cause analysis; they find “insights.”

They are not particularly good at any of this. But, startups and big companies alike insist on plowing ahead, trying, as Tudor’s Sushil Wadhwani did, to train the machine to approach data in a way that makes sense for human analysts. And when it not useful, we often “trust that use cases will proliferate as the technology evolves.” It doesn’t work now, but something something Moore’s law, and it will.

Maybe, but I'm not convinced. AI, as remarkable as it is, may simply not fit the analytical problem. 

Take Amazon. Amazon is widely considered one of the most effective “data-driven” organizations in the world. Cedric Chin recently wrote an essay examining the important pillar of Amazon’s data practice—the weekly business review—and concluded that doing what Amazon does is a very hard human concern:

The WBR is an extremely heavy lift. It assumes you are willing to work at a certain operational cadence, and that you are willing to get at the truth. …

The problem with the WBR isn’t the mechanical details of doing the meeting. That is easy to articulate, as I have just done in this essay. It’s not even the software that is difficult—Amazon ran their WBRs using Excel in the early years, and printed every goddamn deck on paper.

No, the problem with the WBR is organisational discipline. It takes a certain amount of operational rigour to compile, examine, and then test a causal model of your business each week. It takes rigour to hold yourself to a high bar of excellence, facing uncomfortable truths early on by looking, unflinchingly, at your numbers, and iterating—always iterating—on your business.

It’s not at all obvious how today’s flavors of AI help with this problem. A robot that can read really fast won’t make us more disciplined. A creative writer won’t hold us to a tight operational cadence. Cold fusion and room-temperature superconductors would be amazing new technologies too, but we wouldn’t try to contort them into a better BI tool. So why should we assume that’s the right thing to do with chatbot?

Better, it seems, to do what Jim Simons did, and try to imagine how we can contort our problem—understanding the world, so we can make better decisions about it—to fit what the technology is good at. 

To me, that means giving up the goal of being “data-driven,” at least as it’s currently understood. It’s dropping our attachment to numbers and statistical rigor, and realizing that good decisions don’t need to trace their origins to spreadsheets. They can also come from careful observation and research—which are things that LLMs are good at. We don’t need to quantify something to be smart about it.

Of course, not every company should do that, just as not every financial firm should turn into a quantitative trading outlet like Renaissance. But if this current generation of AI technologies is to produce its Renaissance, that seems like the sort of company it’ll be—one that all but abandons quantification entirely, so that it can take full advantage of the million researchers and writers who are offering it their services for $5.92 an hour.

1. Apparently, none of the Starbucks fruit drinks contain any of the fruit that’s in their names. There is no mango in the Mango Dragonfruit Refresher; there is no açaí in the Strawberry Açaí Refresher; there is no pineapple in the Pineapple Passionfruit Refresher. Instead, according to a class action suit, Starbucks’ “hot chocolate contains cocoa, its matcha lattes contain matcha, and its honey mint tea contains honey and mint,” but its juices “are predominantly made with water, grape juice concentrate, and sugar.” Starbucks argued that this was fine, because the name “describe the flavors as opposed to the ingredients” of the drinks. 

There is, I have discovered, a long history of legal precedent about this. Do there need to be whole pieces of wheat in whole wheat bread? No. Do honey graham crackers need to be sweetened with honey? Probably, though it depends on the size of the font on the package (that’s actually real; see footnote four). Do Cap’n Crunch Crunch Berries need to be real berries? No—“no reasonable consumer would believe that Cap’n Crunch derives any nutritional value from berries” because “depictions of the Crunch Berries did not ‘even remotely resemble any naturally occurring fruit of any kind.’” 

Notably, there is no apparent precedent that says—and Starbucks certainly didn’t attempt to argue—that the Pineapple Passionfruit Refresher can be called a Pineapple Passionfruit Refresher because you’re supposed to drink it while eating a pineapple. If we call a food an ingredient, our government officially says, it should probably contain that ingredient, but at least needs to be flavored like that ingredient. Ahem.

2. The rough math: According to a blog post from Toast, an average coffee shop makes about $4,000 in profit a month, and spends about $14,000 a month on labor. Lowering labor costs by 60 percent would increase the coffee shop’s profit to $12,500 a month, or $150,000 a year. Law firms make about a million dollars a year per lawyer; at $5.92 an hour, you could hire a lawyer for less than $20,000 a year, netting you $980,000 of annual profit. Moreover, law firms could probably generate a lot of business by hiring more lawyers and offering their services at below market rates. It’s not clear that coffee shops could meaningfully increase their revenues by underpaying their baristas and offering pour overs for eight dollars instead of ten. (Obviously, the way to meaningfully increase your revenue as a coffee shop is to start selling Cinnamon Toast Crunch lattes.)

3. According to Meta’s AI team, Llama 3, Meta’a open source LLM, got 80 percent of the questions right on the LSAT (which converts to a score in the low 160s), scored a 162 on the quantitative section of the GRE, and a 166 on the verbal section. The median Yale Law student typically scores a 175 on the LSAT, scores 163 on the quantitative section of the GRE and 167 on the verbal section. So Llama isn’t quite a Yale-educated lawyer, but it’s not far off. It’s a UNC-educated lawyer, I guess.

4. I’m a personal investor in Modal.

5. For example, some idiots cavalierly compare ChatGPT to Yale-educated lawyers.

6. Some, by contrast, just liked the stock. 

7. The first quote is from The Man Who Solved the Market, by Gregory Zuckerman; the second is from More Money Than God, by Sebastian Mallaby. Both are via Matt Levine.

8. I also like this story because it fits my belief that there are patterns in the world that we can neither see nor comprehend. Technologies like AI can find those patterns, but “if you insist on the signals making sense to you, you will just get in the way.” When the computers tell Renaissance’s employees what to do, it’s theirs not to make reply, theirs not to reason why, theirs but to do and die (deliver incredible earnings).

9. People often cite LLMs’ inaccuracies as a big problem, which has always struck me as a little strange. That complaint makes sense if you think of them as bad computers, but makes less sense if you think of them as better people. As a person to read a book and tell you what happened, and they’ll get some stuff wrong. Ask an LLM to read a million books, and it’ll get less wrong.