For most of the history of academic work, co-authorship was not a virtue. It was a necessity.

A paper needed someone to collect the data, someone to clean it, someone who knew the statistics, someone who could write it up, and usually someone whose name on the front made an editor open the envelope. Five people on a paper was not five minds. It was one question and four kinds of labour.

The labour got cheap. The question did not.

What actually changed

I now treat research questions the way I treat product features: as experiments to run, not as commitments to make.

In software you do not argue about whether the green button converts better. You ship both and look. The cost of running the test collapsed, so the correct number of tests went up, and the skill moved from having the right opinion to designing the right test and reading the result honestly.

Research is going the same way. I can take a dataset, a method, and a hypothesis, and have a result in an afternoon. Not a publishable paper. A result. Most of them are nothing. That is the point. When testing a hypothesis costs an afternoon instead of a term, you stop choosing hypotheses by how defensible they sound in a meeting and start choosing them by how interesting they would be if true.

I have run this for a year now across my own papers. Blues improvisation across thirteen artists. Charlie Parker's harmonic content as a time series. High order interdependencies in cross platform attention. Mutual compensation across 1,182 jazz piano performances. Different corpora, overlapping machinery, all single author.

The machinery compounds, and that is the real unlock

Here is the thing nobody tells you about working alone: your second paper is cheaper than your first, and your sixth is almost free.

I keep a catalogue of methods rather than a folder of scripts. Interval vectors. O-information. Transfer entropy. The Gaussian copula transform. A directional causality score I had to name because nothing existing said quite what I meant. Each one was built for one paper and has since been used in three.

A transition matrix pipeline I wrote to study how improvisers choose their next note turns out to work, almost unchanged, for studying how a composer chooses the next chord. Same code, same comparability discipline, new object. That reuse is not available to a team that assembles for one grant and disperses.

What the AI cannot do yet, and it is not a small thing

I know more about music than the model does.

Not more facts. More of the thing that matters: I know which questions a jazz musician would find insulting, which would be obvious, and which would make someone put their drink down. I can look at a corpus of standards and see that Cole Porter wrote both words and music while the Gershwins split them, and that Richard Rodgers kept his harmony and swapped lyricists halfway through his career. That is not a fact retrieval problem. That is a research design, and it came from twenty years of playing this repertoire, not from reading about it.

Ask an agent to test a hypothesis and it will test it well. Ask it what hypothesis is worth testing and you get the median of everything ever written on the subject. The median is not where the papers are.

So the division of labour is clear enough. I bring the question, the domain judgement, and the decision about what counts as an interesting result. The machine brings tirelessness. It will run the thirtieth variant of an analysis at the same quality as the first, which no graduate student will, and no collaborator should be asked to.

You need a dataset, and that is the actual barrier

Everything above assumes you have data nobody else has.

This is where my day job turned out to be an unfair advantage. I build scrapers for a living. I have corpora sitting on disk that exist because a product needed them, and some of them are more interesting as research objects than as product inputs. The marginal cost of asking a research question of a dataset I already own and already understand is close to zero.

If you do not have that, the independent route is much harder, and I would not pretend otherwise. Methods are free now. Compute is cheap. Data that is both yours and interesting is still rare, and it is the one input that AI does not hand you.

The friction that is left is institutional, not intellectual

I want to be straight about the part that still does not work.

The research is doable alone. The publishing is not quite.

arXiv will not take a first submission in a category without an endorsement from someone who has recently published there. It is a reasonable policy and it exists for good reasons, and it also means that the moment you have a finished paper with a DOI and a journal submission behind it, you are emailing strangers to ask permission to post a preprint. An institutional email address does not fix this. A co-author does, permanently, which is a quiet argument for collaboration that has nothing to do with the research itself.

Peer review has a similar shape. Nothing in it requires an affiliation, and plenty of it assumes one.

So the honest version of this post is not that institutions are obsolete. It is that they have stopped being necessary for the work and remain necessary for the paperwork, and that gap is going to be uncomfortable for a while.

What I would tell someone starting

Pick a domain you genuinely know, deeply enough that you can tell a boring question from a live one. Get a dataset that is yours. Learn four or five methods properly rather than twenty badly, because the ones you know well are the ones that will compose into the next paper.

Then run a lot of tests and throw most of them away.

The skill is no longer execution. It is knowing what is worth executing, and being willing to bin the results that are merely true.