Cheng Tan

A Lab of One

The single-person research group, and the one thing it cannot produce

· researchaiphilosophy ·中文

The software industry has a running bet about the first one-person billion-dollar company. Sam Altman says his tech-CEO group chat keeps an actual betting pool on the year it happens—“unimaginable without AI, and now will happen.” What is the research equivalent of the one-person company? The single-person research group: one researcher with good taste, a fleet of agents, and no meetings. I think it is coming faster than my colleagues expect. I also think it has exactly one flaw—and the flaw is the interesting part.

Why groups got big

Research groups were never big for comfort; they were big because the artifacts demanded it. Systems research in particular has a brutal norm: describing a hypothetical system whose implementation is “just beginning” is unacceptable—the reader has a right to know whether the system is real. The value of a half-built system is not half the value of a built one; it is roughly zero. When the artifact is a cliff rather than a slope, a group too small to reach the top cannot afford to start climbing. The stakes priced small groups out.

So groups grew with the artifacts. The UNIX paper had two authors in 1974. Spanner had twenty-six in 2012. The Gemini technical report lists over nine hundred. Across SOSP, the average paper had fewer than two authors in 1973 and nearly eight in 2023 (per DBLP). For fifty years, the way to do bigger systems research was to be a bigger group.

That logic rested on one premise: implementation is expensive. The premise died recently. Agents write the code, run the experiments, produce the figures, and draft the prose; the results are no longer toys. Strip the labor out of the research group and what remains is its irreducible core: taste in, artifact out.

Brooks, inverted

Here is the part I find delicious. The most famous book in software management, The Mythical Man-Month, is one long meditation on the cost of headcount: Brooks’s Law says adding manpower to a late project makes it later, because newcomers must be trained and coordination effort grows as n(n-1)/2. Every group leader lives a mild version of it: the week dissolves into syncing, unblocking, re-explaining; the group’s n(n-1)/2 eats the group leader first.

The lab of one is the limit Brooks could never reach: n equals one, zero channels, the idea and the implementation living in the same head. Agents are manpower that does not attend meetings. This is why the single-person group will not merely match the big one—on speed, it will embarrass it. An idea can become a tested artifact in days, with no seminar in between. (One honest caveat: writing context for agents is the new meeting. The coordination tax shrank; it did not vanish.)

The reproduction problem

So far, the forecast is pure gain: the same taste, more artifacts, faster cycles, no meetings. I promised a flaw. Here it is.

A research group was never only a production unit. It is how the field makes more researchers. And the one input the lab of one still requires—taste—happens to be the one thing that cannot be downloaded. Polanyi said it precisely: “An art which cannot be specified in detail cannot be transmitted by prescription, since no prescription for it exists. It can be passed on only by example from master to apprentice.” That is a working definition of the PhD: sit close to someone’s judgment for five years and absorb the rules they could not tell you, because they do not fully know the rules themselves. Fields are lineages.

The single-person research group breaks the lineage. It produces papers and no people. A lab of one has no alumni.

Scale that to a big fraction of a field, and the arithmetic turns grim. The first casualties are the small groups’ lineages: no students means no heirs, and the tastes cultivated in a thousand small labs simply stop reproducing. What survives is the taste of the giant groups—and we already live in a preview of that world, where the biggest groups command a voice, amplified by social media, louder than their research alone would earn. Then the deeper worry: if the next generation learns judgment from AI instead of from people, their tastes may converge—every apprentice sitting beside the same master, absorbing the same silences. And the endgame: when a generation of unanimous taste becomes the only judges, thinking differently stops being an asset and becomes a disqualification—the researcher who thinks differently cannot survive review, precisely because they think differently.

Biology has a name for a population that reproduces quickly and identically: a monoculture. It flourishes, right up until the environment changes. A lab of one has no alumni; a field of labs-of-one has no descendants, only copies. The single-person research group wins every race except one: the relay.

References