I frequently wonder what the error bars on my life choices are. What are the chances I ended up a chemist? A scientist of any type? Having two children in graduate school?
If I had the ability, I would want to restart the World Simulator from the time I started high school, run a bunch of replicates, and see what happened to me in different simulations. And this wouldn’t just be useful for me personally—there are lots of things in the world that are just as contingent and path-dependent as one’s life choices. What would have happened if Charles the Bold hadn’t died in 1477 and Burgundy had preserved its independence? If the 1787 convention were rerun several times, how might the US Constitution differ?
Sadly, we’ll never know the answer to these questions. But what we can do instead is find cases in which analogous institutions evolved in parallel, and try to learn from the similarities and differences between them. It’s an imperfect substitute for rerunning the World Simulator, but it’s still pretty cool. (This is far from an original idea: see for instance Legal Systems Very Different From Ours.)
Lately, I’ve come to think about the tech/startup world as somewhat parallel to academic science in this manner. Why? For one, both tech and academia deal with hard problems that demand obscure/arcane domain-specific knowledge inaccessible to non-experts. (It’s true that the problems are typically scientific in academia and engineering-related in tech, but I’ve argued previously that this distinction is flimsier than it seems.) And in both fields, a few high performers vastly outperform the rest of the field, be it a “10x engineer” or a Nobel laureate.
Startups, like academic labs, are small and agile institutions which face the task of raising money, building a team, selecting a hard yet solvable problem, and finding a solution all within a few years. In both cases, too, there are nonlinear returns to success: moderate success is not much better than failure, pushing founders/assistant professors to be as ambitious as possible.
If we accept these two fields as vaguely analogous, what interesting differences can we observe?
I’ll quote from an essay by Paul Graham, founder of Y Combinator and noted startup sage:
Have you ever noticed how few successful startups were founded by just one person? Even companies you think of as having one founder, like Oracle, usually turn out to have more. It seems unlikely this is a coincidence.
What's wrong with having one founder? To start with, it's a vote of no confidence. It probably means the founder couldn't talk any of his friends into starting the company with him. That's pretty alarming, because his friends are the ones who know him best.
But even if the founder's friends were all wrong and the company is a good bet, he's still at a disadvantage. Starting a startup is too hard for one person. Even if you could do all the work yourself, you need colleagues to brainstorm with, to talk you out of stupid decisions, and to cheer you up when things go wrong.
Ever since I read this, I’ve wondered why no labs ever have multiple PIs. I guess this would mess with the semi-feudal organization of university bureaucracy, but it doesn’t seem intrinsically bad—after all, lots of startups seem to do just fine.
The VC strategy, as I understand it, is basically “fund a bunch of companies, and one or two of them will make it all worth our while.” This is a little bit different than how universities approach hiring assistant professors: each university will typically hire a small number of professors each year, after much deliberation, and they have a pretty high likelihood of giving them tenure, at least relative to the likelihood of any given startup succeeding. (Basically, startups are r-selected, whereas academic labs are K-selected.)
There are a lot of reasons why this might be. For one, faculty members aren’t just trying to pick a winner but also their future colleague, so personal considerations probably matter more. Failure in science seems more cruel, too: while a failed startup founder can often negotiate the “sale” of their company and parlay that into new jobs and the constant churn of tech means that there are always new openings for talented ex-startup employees, a lab that doesn’t get tenure takes a toll on professor and students alike.
A hypothesis for why the success rate for new labs is so much higher than the success rate for new businesses is that many labs only succeed a little bit. They don’t actually achieve what they dreamed about in their initial proposals, but they pivot and accrue enough publications and cachet to earn tenure nevertheless. In business, it seems harder to succeed a little bit—the market is a harsher critic than one’s peers.
Paul Graham again, this time talking about the dangers of fundraising:
Raising money is terribly distracting. You're lucky if your productivity is a third of what it was before. And it can last for months.
I didn't understand (or rather, remember) precisely why raising money was so distracting till earlier this year. I'd noticed that startups we funded would usually grind to a halt when they switched to raising money, but I didn't remember exactly why till YC raised money itself. We had a comparatively easy time of it; the first people I asked said yes; but it took months to work out the details, and during that time I got hardly any real work done. Why? Because I thought about it all the time.
The broader conclusion, from this and other essays, is that any distractions from the core mission of the startup are very dangerous, and should be avoided at all costs. This is very different from the lifestyle of new PIs, who are typically juggling departmental responsibilities, writing a curriculum, lecturing for the first time, and writing grants all while trying to get their lab up and running.
In tech, people obsess about recruiting the best people possible—I reviewed a whole book about this last year. Hiring bad programmers is #6 on PG’s list of mistakes that kill startups, and there seems to be a general consensus that a great company takes great engineers, no matter what.
In contrast, professors don’t have full control over whom they hire (for graduate students), making recruiting much harder. Graduate students are selected through a complex two-stage system involving admission to a school and then a subsequent group-joining process (and new assistant professors sometimes aren’t even around for the first of these stages). You can obviously try to coax talented students to work for you, but the pool of accepted students interested in your subfield might be tiny, and they might all prefer to work for an established group…
(Plus, there’s not a good way to reward top performers in academia. All graduate students are equal, at least on paper—you can’t give someone a year-end bonus, or a promotion.)
A nice concrete example of this is how professors struggle to hire competent programmers, even as research scientists—they aren’t allowed to pay enough to match market rates, even when the expense would be well worth the money. To quote Bret Devereaux: “academic hiring, to be frank, is not conducted seriously” (he’s discussing the humanities, but the point stands).
As a startup succeeds, it grows: while a seed-stage startup typically has <15 people, startups at Series A often have 20–40, and startups at Series B–C might have as many as 300 employees (one ref; rough numbers broadly consistent with other sources). Good companies grow, while bad ones die.
In contrast, it’s rare for even the most successful US academic labs to grow past 30 people (although it occasionally happens), limiting the reach of top-performing professors. While a huge proportion of tech employees work for the best companies (Google, Meta, Amazon, etc), only a very small number of students work for the best professors.
The imperfect nature of the analogy means that some of these points might not be useful in a normative sense: universities are not really optimized to produce research as efficiently as possible, and maybe that’s fine. Likewise, startups aren’t optimized to produce unprofitable research or train future scientists, even if these activities may in the long run be beneficial. (This is why basic science is considered a public good, and why the government funds it at all!)
Nevertheless, I think there’s a lot that scientists can learn from startups. There is a whole army of people working to solve challenging technical problems in the most efficient way, and it’d be prudent to study the wisdom that emerges.
Thanks to Ari Wagen and Jacob Thackston for reading drafts of this piece.