The Brain That Was a Tax Is Now an Asset

In an earlier version of this essay, this was the point where the article quietly became a different article. The first one was about data entry — about systems losing the shape of the work and making people put it back by hand. It ended at a thing called Thread. If you haven’t read it, the short version is: I have spent twenty years noticing the same thing across very different jobs, mostly without realising it was the same thing.

This one is about the person who kept noticing. About why he kept losing tools (in the garage, but also scripts I wrote), abandoning projects at 80%, and looking, on paper, scattered. And about what happened when prescribed ADHD medication and a swarm of cheap fast AI agents arrived more or less at the same time.

The thesis, stated up front so I don’t have to dance around it: ADHD plus medication plus AI agents changed the economics of being a cross-domain generalist. The same wiring that used to produce unfinished projects, dropped tools, and a CV that read as scattered can now be useful, for now, when paired with specialist agents that absorb the boring grind. I don’t know how long the for now lasts. I’ll get to that.

Worth flagging up front: the technical through-line in the previous article existed in my subconscious for twenty years. The conscious experience was chaos. The two facts coexisted, and I’d like to talk about how, because the way they coexisted as disparate entities is the bit that just stopped being permanent.

Less than one week ago I started on prescribed ADHD medication. It has been the most useful single intervention I’ve had in my professional life. (Yes, a Director of Flight Sciences blogging about ADHD meds is a deliberate exposure decision. It’s 2026.) The reason it earns a paragraph here, beyond the personal, is that the meds let me see why AI is a good fit for the way my brain works. The AI’s job is the boring grind I’m bad at; the meds gave me enough executive function to actually hand the grind over rather than abandoning the project at the same point I always did before.

The thing the meds clarified, which I hadn’t expected: I always thought ADHD meds would stop me jumping between things^*My wife and I have spoken about whether I have ADHD before.. What I learned is I was never that good at it. I’d drop tools when I was done with them like they ceased to exist. I’d start a thing, work on it intensely, get it to a state I considered “essentially done,” and never come back to harden, document, productionise, or in some cases even commit it. The boring grind at the end of a project wasn’t visible to me — not as in “I avoided it,” more as in “I didn’t perceive it as a thing that existed.”

The meds didn’t kill the through-line. Topology, refusing to flatten, network-shaped problems — that was always there. What the meds turned down was the chaos around it. I can now look at a half-finished thing, see that it’s half-finished, and finish it. That sounds embarrassingly basic. It is.

The flowchart is a two-state animation. State one is the meme everyone with an ADHD diagnosis has seen forty times: get new idea → start new project → tell everyone → loop, with “finish project” floating off to the side, disconnected. State two is my actual flowchart now: get new idea → recognise it’s connected to three previous ideas → hand the boring bits to a swarm of agents → stay at the connection level → actually finish. The original loop is preserved, because that bit didn’t go away and I don’t want to pretend it did. What’s new is the connection to the AI swarm that does the boring grind, and the connection from that swarm back to “finish project”, through me, now a green node instead of an orphan.

The brain doesn’t get fixed. It gets supplemented. I’m not going to extend that into self-help — half the internet is doing that already — but I need it on record for what comes next, because the configuration the meds enabled is the configuration the rest of this post is about.

The current AI configuration that works for me, in 2026, is one person with cross-domain taste running a swarm of fast cheap specialists.

The PhD supervisor analogy is the closest thing I have to an honest description of what the orchestrator role looks like day to day. A good supervisor doesn’t write the thesis. They don’t run the simulations either. They notice that the result you’re stuck on connects to a chapter from somebody else’s group last year, send you a paper, ask the right next question, kick you off the dead end you’d have stayed on for three more weeks. Their value is almost entirely about routing and shape-recognition, and almost none of it is about doing the bounded work. Anyone who’s had a good one will know what I mean (I had a good one, thanks Eric.) Anyone who’s had a bad one will know what I mean even more clearly.

I think there’s a tonne of self-aggrandising YouTube videos and posts about being neuroatypical and how it’s a superpower. I do not subscribe to that. But I do think that my ADHD-shaped brain finds coupled problems satisfying in a way that it does not find decoupled ones.

Aeroelasticity is the canonical example of a problem you cannot solve sequentially. The airflow loads the wing, the wing deforms, the deformation changes the airflow, the changed airflow changes the load, and if you pretend those are separate problems you get the wrong answer. Sometimes you miss flutter entirely, which is quite a bad way to learn that the coupling was real.

I loved it immediately.

Not because it was easy. It wasn’t. Because it felt like the first honest problem I had been given. The interesting answer lived in the refusal to split the thing into tidy pieces.

Once I saw that, I started seeing the same taste everywhere. Thread treats simulation, provenance, and the trajectory through design space as one thing rather than three separable concerns. The topology-vector acquisition function treats the design point and the path to it as one thing. The conventional move, in each case, is to decouple: run the simulation here, store the metadata there, reconstruct the history later; choose a point, score the point, ignore the path. The move that produces the interesting answer is to refuse.

That is the link between the engineering work and the ADHD bit. Not architecture. Taste. ADHD, at least in my case, does not naturally separate things in the first place. That is a problem when the task wants clean sequential execution — school, admin, most forms of adulthood, most forms of “can you just” — and an asset when the task is genuinely coupled and rewards the refusal to decouple. The brain finds its native problem class.

Which explains a lot of my life in retrospect.

I was a smart kid, which made me a nightmare student. ADHD plus capability is a beautifully efficient way to produce a child who will not engage with anything that fails to challenge him. Almost nothing in school did. I got mostly Bs at GCSE, including Maths, which is funny now in a way it was not funny then - my level of mathematical ability probably surpasses my secondary school teachers, and that’s OK because they weren’t teaching kids vector calculus. I got As in Science. College had the same shape until Nick Anderson, a physics teacher who seemed to recognise how my head worked when most teachers did not.

Then I had a Traumatic Brain Injury and did an extra year at college, and the bargain changed. Mathematics became genuinely difficult. Learning became difficult. I had to learn how to learn, because simply absorbing the material without ever building study muscle was not available anymore. After that, I got full marks in some maths exams, not because I had become a different person, but because the problem and the teaching had finally required the part of me that had been mostly idle.

University repeated the pattern. The first two years did not count, so I coasted. In the third and fourth years I found aeroelasticity, and Eric Gillies, and the engagement problem solved itself. Top marks in aeroelasticity. Graduated top of the class.

Two teachers across two decades read the architecture correctly, and got different output as a result.

This is where the internet usually says “ADHD is a superpower,” and I want to push back hard enough that the useful grain survives. There is a grain of truth. For coupled problems, where refusing to separate things is the whole game, an ADHD brain can have a real advantage - at least for me. But the viral framing overstates it badly.

ADHD is not a superpower. It is a cognitive architecture with a narrow band of problems it is exceptionally good at and a wide band of ordinary tasks it is bad at. The net outcome depends almost entirely on whether your environment lets you spend most of your time in the narrow band. School does not. Most jobs do not. Aerospace research at the right company does.

Calling it a superpower flattens the cost: the years of underperformance, the things that do not get done, the executive-function tax that does not disappear just because you found work that fits. The honest version is less viral and more useful. It is an asset for coupled problems and a tax for decoupled ones, and most of life is decoupled.

A brain wired the way mine is — refusing to flatten, jumping between problems, holding loose connections across communities, pathologically uninterested in the boring grind — was previously an engineering tax. You paid it in unfinished projects, dropped tools, and a CV that read as “scattered” to anyone whose filter rewarded coherence over connection. There was no operational answer to the grind. You either ground through it yourself, or you didn’t, and the work didn’t ship.

Two things changed at roughly the same time.

The meds made the unfinished loop visible to me. Before, “essentially done” felt like done. Now it doesn’t. I can see the gap between “the interesting bit works” and “this is a thing other humans can use,” and I can stay in the room long enough to close it.

The agents made closing it cheap. The 80% to 100% — tests, docstrings, edge cases, the boring data cleaning, the literature sweep that confirms you haven’t reinvented something from 1987 — used to be the bit that killed the project. Now it’s the bit a swarm picks up while I’m doing something else.

Neither alone would have done it. The meds without the agents would have given me visibility of the grind without much help getting through it. The agents without the meds would have given me help on a grind I still couldn’t see. The pair is what works.

A small social side-effect I wasn’t expecting: I used to vocalise hair-brained ideas at whoever was in earshot, because saying a thing out loud is how I worked out whether it was any good. Now I can talk to people about normal things and route the half-formed pre-ideas to an AI instead. My wife has noticed.

Use the previous article as the worked example. I told the LLM about Plotly, Paradigm, Postgres, Thread. It saw the surface pattern (refusing to flatten data) early and confidently. It did not see the deeper one (topology as first-class) until I named it^*That’s an excellent observation, Harry, and you’re right to call me out on it…. We had four beats fitted into a “data richness” frame before I noticed the frame was wrong.

Asked afterwards why it missed it, the LLM gave me a useful answer: almost no one is a specialist in two of those communities at once. DAGs in software engineering, graphs in OR, network topology in telecoms, provenance in databases and archaeology — same mathematical object, different folklore. The cross-walk papers don’t get written, because the people who’d write them are too busy doing the cross-walks.

Even a model that has read more about graph theory and CFD provenance than I ever will didn’t bridge them, because the bridge isn’t in the corpus. The bridge was in my head. The bridges live in the heads of generalists and there isn’t currently a substitute - which, for now, is rare and valuable.

I want to be careful not to oversell that. AI may develop something taste-shaped — something that holds problem shape across communities and notices when their vocabularies are pointing at the same underlying thing. If it does, the orchestrator role compresses, possibly hard. The honest version: right now, in 2026, this configuration is real and rare. The window depends on how fast taste-shaped AI arrives, and I don’t know how fast that is.

What I’ll say with more confidence is that the configuration didn’t exist before. Now the same wiring that used to be a tax is the bit the orchestrator role rewards, because the swarm picks up the grind tax that used to compound on you alone. Unfinished projects are something the swarm finishes. Scattered interests are something the swarm follows up on. The connection-noticing — the bit I was always actually doing, even when I didn’t know it — is the bit no specialist in the swarm can do for me, because the connection requires having been all of those specialists, badly, at one point and thinking about them together.

I should be transparent about something. This article was written this way.

I wrote prompts. A swarm drafted. I edited. The LLM had a copy of Writing Style - Harry Smith.md and tried, with mixed success, to sound like me. I rewrote the bits where it didn’t. The voice you’ve just read is mine, not because the LLM wrote in my voice, but because I rewrote until it did. The orchestrator-and-swarm configuration in this article isn’t a thesis I’m asserting at you in the abstract — it’s the thing that produced the article you’re reading. Form follows content one more time. That’s the post.

The bit the swarm couldn’t do, here, was the same bit it couldn’t do with Paradigm and Thread. It could draft any individual section in roughly the right voice. It could not, on its own, notice that two articles were tangled together inside one and decide where to cut. The shape decision was the human bit. The grind around the shape decision wasn’t, and I’m grateful for that, because I would otherwise have abandoned this draft at the same point I usually abandon drafts.

One more thing, by way of P.S. I made a video version of an adjacent argument a while back — Where did that plot come from? — same instinct, done for fun rather than for prose. Thread, but louder.

So here’s where this lands.

The brain didn’t get fixed. It got supplemented. The meds made the unfinished loop visible. The agents pick up the grind that used to kill the project at 80%. What remains for me is the part I was apparently doing all along: holding the shape, noticing the bridge, deciding which specialist answer belongs where.

Maybe AI compresses that role later. I don’t know. In 2026, it is real.

The brain that was a tax is now an asset.