// writing

2026-07-03 // 19 min // ai, cognition, work

three kinds of nothing

I’m a machine learning engineer. A fair chunk of my job is figuring out what AI is actually good for, and this series is me doing it in public. Writing code is one of the settled cases: the LLM is good at it, I hand it code every day, and I’m not here to relitigate that.

I’m here because of two kinds of day. On my heaviest orchestration days, agents running, diffs to review, threads to steer, I do real work the whole time. Decisions, priorities, architecture calls, the parts that matter most. And by evening I’m more tired than any day of writing code ever left me, with the thing I used to call being deep in it nowhere in sight. Then there are the hours I block off to write code myself, no LLM, on purpose. Three of those go by like twenty minutes.

The part of my job that got automated is the only part where that state ever lived. That’s what’s been bugging me, and working it loose turned up something bigger: the LLM doesn’t just optimize away tasks. It optimizes away states.

three kinds of nothing

Work runs through three states that produce nothing you can point at while they run.

Idle is the gap between tasks: the queue, the shower, the ten minutes between closing one thing and opening the next. Stuck is the impasse inside a task: blocked, circling, nothing coming out. Absorbed is the state from my no-LLM hours: deep in the doing, self forgotten, time gone strange.

To an optimizer these are the same thing. Time without output. And every tool that buys one of them gets priced the same way: by the gains. The task goes faster, the wait feels shorter, the blocker dissolves. What the state was doing while it ran never makes the ledger, because a state produces nothing to point at, and nobody invoices the removal of something that felt like nothing. The bill arrives anyway, just not on the task. It shows up later, in you.

This accounting has already run to completion once. The feed bought the idle gap years ago; almost nobody stands in a queue with their own thoughts anymore. Fifteen years on, the consequences are arriving in order, first the obvious ones, then the ones nobody priced. That finished purchase is the best preview available for the two the LLM is running right now.

the completed experiment

The case for idle is better than it has any right to be. An idle brain isn’t off. Resting cognition has its own organized mode, mapped twenty-five years ago, and mind-wandering recruits the default network and executive control regions together.¹ Left alone, a mind spends most of its drift on the future, doing the unglamorous planning of a life: next week, the conversation you owe someone, what you actually want.²

Ten minutes of wakeful rest after learning something and you remember it better a week later than someone whose ten minutes got filled.³ And the causal anchor: block mobile internet on people’s phones for two weeks, in a randomized trial of 467 adults, and sustained attention, mental health, and well-being all improve.⁴ So that’s the first-order loss, the one you could have predicted: the work the state was doing stops, and the ledger records only the reclaimed minutes.

The second-order loss took a decade to surface, and it’s the one nobody priced. It hides under fake numbers: the goldfish attention span traces to a marketing deck citing a firm with no such study, and the “MIT and Stanford” study of 45,000 people watching attention collapse has no authors, no DOI, and no existence. Both circulate anyway, keynote to keynote.⁵

What actually moved is quieter and worse. Attention capacity didn’t decline; a meta-analysis across 32 countries finds children flat and adults slightly improved. What moved is the threshold: how long a moment has to stay empty before it becomes intolerable.⁶ And it didn’t move on its own. It’s the dislike of boredom, more than boredom itself, that drives phone use, and the arrow runs harder in reverse: problematic phone use predicts later chronic boredom more than boredom predicts later use.⁷ The feed didn’t just occupy the idle gap. It manufactures the restlessness it sells relief for, and every pass through the loop raises the price of an empty moment.

Something else got bought with the gap: the message. Boredom is information, the report that what you’re doing has stopped engaging you, and its whole job is to push you toward something that does.⁸ Answer it with a feed and the message never gets delivered. The restlessness still fires, more than it used to. What’s gone is your ability to hear what it’s about.

Third-order, nobody knows. What happens to people who can no longer hear their own boredom is not a literature yet, because nobody was measuring when the purchase went through. That’s what a precedent is for: by the time the losses were visible enough to study, the default had already flipped.

the chat box

The LLM is running the same purchase on the stuck gap right now. The moment a problem pushes back, relief is one keystroke away. What that buys off is incubation, the shower-thought effect: real on aggregate, and it pays under two conditions, after you’ve genuinely struggled, and when the break is lightly occupied rather than demanding.⁹ For well-specified problems the skeptics are right and the relief is pure gain: the bug with a clean repro, the lookup, the boilerplate, there was nothing in that struggle for you.¹⁰ The ill-structured impasse is different, the one where what blocks you is your own framing. Getting unstuck there runs on forgetting your bad leads, and forgetting needs the one thing the instant unblocker removes: time away, with a loaded head.

And one loss here isn’t cognitive bookkeeping at all. The click, insight as a felt event, only exists downstream of an impasse. Route every impasse to the chat box and answers keep arriving, on time, forever, and nothing ever comes apart in your hands again.

the habitat

Then there’s the state I actually miss, the one the feed never touched because it doesn’t live in the gaps. It lives inside the work, and it has requirements.

Look at where those conditions hold in software work. Scoping and architecture are real thinking, but the feedback loop runs in weeks. Meetings are meetings. Review is judgment applied in bursts. Writing the code is the one layer where everything lines up at once. The problem pushes back at exactly your level, the goal is concrete, the compiler and the tests answer in seconds, and nothing about the activity demands you stop. Flow in this job never lived everywhere. It lived there.

That’s the layer the LLM automated.

This is where the precedent stops being reassuring. The feed took a state every human had; queues and showers are universally distributed. Flow had a habitat requirement, and only some work ever met it. Orchestration is taking a state only some jobs offered, and taking it from exactly the people whose jobs offered it. A smaller blast radius, and a total loss inside it.

And the work that remains fails the entry conditions one by one. Orchestration is interrupt-shaped, so sustained attention goes first. When METR ran its second developer study, time measurement partly broke because participants “would often work an unrelated task while waiting for the agent to complete its work.”¹¹ A study that wasn’t asking about flow recorded the workflow’s shape by accident. Feedback arrives in diff-sized lumps instead of continuously. And a reviewer’s challenge-skill match is undefined: the agent’s output isn’t pitched at your level, it isn’t pitched at all.

This also squares the one result that looks like a contradiction. GitHub’s own survey found 73% of Copilot users saying the tool helped them stay in flow.¹² That’s a vendor survey, and it’s about autocomplete, which lives inside the writing-code layer: it shortens the keystrokes and leaves the conditions standing. Orchestration replaces the layer. Different regime, different result. The practitioners describing the loss are describing the second one.¹³

Flagged as a feeling, not a finding: orchestration gives me no absorption at all. Not even the compulsive slot-machine kind Thomas calls dark flow; I read those accounts the way you read letters from a country you haven’t visited.¹⁴ What I have instead is a policy. Some hours are LLM-off, code written from scratch, because I wasn’t willing to lose the capacity. Those hours are where flow still shows up, reliably, like it had been waiting.

the control room

None of this is a new shape. Bainbridge wrote “Ironies of Automation” in 1983, about control rooms. Automate the doing and you leave the human the parts that are hardest to do well, monitoring and exception handling, while the skills those parts depend on erode underneath you.¹⁵ Orchestration is her control room with a chat box. The hard parts stayed. The layer the conditions lived in left.

What four decades of that literature never measured is what the transition feels like from inside. And as I write this, nobody else has either. I’ve checked three times. The closest study tracks one survey item called “flow state” across users of every kind of AI assistant mixed together; the item erodes while people’s sense of their own productivity holds flat. The studies that look at orchestration interview developers about oversight and control and never measure states. An IBM survey says it straight: there’s “a notable absence of analysis targeted to agentic workflows.”¹⁶

So the evidence here is me, and that requires swallowing something first. METR ran the experiment that keeps me honest: experienced developers, their own repositories, randomized. They were 19% slower with AI while believing they were 20% faster.¹⁷ A 39-point gap between the feeling and the fact, and it discredits everybody’s self-reports about the LLM, mine included. Which is why every firsthand claim here is about states and never about effects. The absorption is gone; what it costs, I can’t measure from in here, and neither can anyone else yet.

one economy

The three purchases share a mechanism. The brain economizes on effort: Kool and colleagues ran six experiments where the harder option carried no penalty, and people still picked the easier one; effort avoidance behaves like a primitive preference, not a calculation. Barr’s group then watched that preference meet the smartphone: the less analytically people think, the more they let the phone do it for them.¹⁸ Any tool that lowers the price of not-thinking gets used, and it gets used by default, not by decision.

Absorption sits at the expensive end of that economy, because it has an entry cost. The state doesn’t start when you sit down; it starts twenty or forty minutes into sustained, conditions-met work. An environment where the cheap exit is always available keeps collecting you before the run-up pays. My LLM-off hours are this mechanism read backwards: remove the exit, rebuild the conditions, and the state comes back like it never left.

The threshold from the feed shows up at the workbench too. In METR’s second study, developers would no longer tolerate the unassisted condition: 30 to 50% withheld tasks they didn’t want to do without AI, and one participant said it exactly: “my head’s going to explode if I try to do too much the old fashioned way because it’s like trying to get across the city walking when all of a sudden I was more used to taking an Uber.”¹¹ The feed raises the boredom threshold. The agent raises the unassisted-work threshold. Same trade, different aisle. None of these effects runs large at population scale, and populations aren’t the subject. The subject is what the defaults do to one person who doesn’t push back, and I’m reporting from inside the push.

the asymmetry

Two kinds of claim are in play here, and I don’t want them confused. For flow, the claim is about how it felt: absorption was part of the pay, part of why the work was worth doing. Nobody disputes the feeling; the argument is only ever about what produces it. For idle and stuck, I concede the feeling completely: mind-wandering predicts unhappiness across thousands of people pinged mid-moment, and people fled the empty moment long before the phone showed up to help.¹⁹ These states were doing work while feeling like nothing, and feeling like nothing is exactly why nobody invoices their removal.

the countermove

One practice survives all of this, and I stole it from the incubation literature. Take the problem for a walk before you take it to the chat box. Walking reliably boosts divergent thinking, the generating kind, and does nothing for convergent, focused work; the anchor study ran on a treadmill facing a blank wall, so it’s the walking and not the scenery, and a recent meta-analysis pools a large effect across 23 studies.²⁰ The practice has a shape: walk to generate, sit to close. The bounds ride in a footnote; I’m claiming the practice, not a theory of legs.²¹

And the practice pays. Earlier this week I’d been stuck for days on how to measure counterfactuals in a domain where randomized experiments aren’t possible. The whole thing came apart on a thirty-minute walk I hadn’t taken to think about it, and two experiment streams I’m running now date from that half hour. One anecdote, worth exactly one anecdote. But look at its shape: loaded by days of struggle, lightly occupied foreground, not even deliberate. That’s the incubation pattern, rebuilt by accident. The walk is the empty moment with its conditions restored.

the states I keep

So, the practices, plainly, and I’m not selling a system. The LLM-off hours guard the habitat. Hard problems get a walk before they get a chat box. And some gaps stay unfilled on purpose, in full knowledge that they’ll feel like nothing, because feeling like nothing is what a working state looks like from the inside.

And I won’t pretend any of it scales. The defaults all run the other way, the environment now bills you for thinking, and boredom’s push resolves badly when there’s nothing good to resolve into.²² What I’m describing is a swim upstream. I’m describing it anyway, because the current isn’t going to reverse, and the swimming turns out to be worth it.

The question I can’t close: what happens to the craft, to the satisfaction that kept people in it, and to whatever consolidates skill, when the layer of work that produced all three is the layer that got automated. As of this writing, nobody has even started measuring it. I wrote in the reps you stop taking that the skills go quietly, with disuse, feeling like nothing while they go. The states go first, and they go the same way.

The no-LLM hours are mine. I’m keeping them.