A few days ago, a swyx newsletter on Latent.Space happened to catch my eye. It’s titled “Humanity’s last gasp,” the last breath of humanity. It’s not a shouted title—quite the opposite. It’s written in that calm, slightly oblique tone that, precisely because it doesn’t raise its voice, forces you to listen more closely.
Inside there are three images that stuck with me. Aaron Levie says that in Silicon Valley nobody is working less—if anything, the opposite. Tyler Cowen, as an economist, argues that you should work much harder now , both if you think AI will devalue your work and if you think it will make it more valuable. And then there’s Simon Last from Notion talking about sleepless nights and a new anxiety, “token anxiety,” something that two years ago didn’t even exist as a concept.
The paradox swyx puts at the center is simple and, for that very reason, a bit unsettling. Agents do more work than ever, benchmarks saturate, models outperform human experts by percentages that until recently we would have called science fiction. And yet the people building and managing these systems have never worked this much.
And here a question lights up for me that I can’t easily turn off: if the promise was “more productivity,” why is the widespread feeling “more pressure”?
The convenient explanation: it’s just a phase
The most reasonable objection is also the one that comes to me naturally, if I try to be honest. Maybe it’s a transitional phase. The usual disruption dynamic: first it breaks things, then it recomposes them. We’ve already seen it with the personal computer, with the internet, with smartphones. Jobs change, some disappear, others are born, productivity grows, and in the end—at least for those who manage the transition well—average well-being improves.
It’s a serious objection. Ignoring it risks sliding into a kind of keyboard luddism, that thing where you convince yourself the problem is technology itself, rather than the way we’re inserting it into economic and social systems.
But there’s a point where this objection stops working. And that point, for me, is the body.
The body doesn’t scale
The problem isn’t just whether AI will replace knowledge workers or empower them. The problem is that the human body doesn’t scale.
It doesn’t scale the way tokens scale. It doesn’t scale the way a cluster of tens of gigawatts scales. It doesn’t scale the way a model’s learning curve scales when it gulps down an enormous amount of knowledge in weeks.
The body has a biological clock that doesn’t get reset with an update. It needs sleep, and if you take it away for months you might keep going anyway—at least on the surface—until the moment something breaks. And often it breaks in ways you didn’t anticipate and that aren’t so easy to reverse.
It needs movement, natural light, rhythms. It wasn’t designed for 3 a.m. spent debugging an agent that generated eight hundred lines of code that “work,” but that you don’t really understand.
These things aren’t new. Occupational medicine says them, neuroscience says them, any primary care doctor with a minimum of intellectual honesty says them. And yet, in the dominant discourse on AI, the body seems like an implementation detail. A legacy constraint to work around with enough caffeine and enough willpower.
Where I’m looking from
I’m a partner and technical lead at a small ICT company in Pescara. We’re about ten people. For a while now I haven’t been writing production code anymore, or at least not like before. My work has become research, strategic spikes, architecture evaluation, and also regulatory compliance, which is an unsexy word but a very real one.
I spend my days figuring out where the market is going and deciding what makes sense to do and what doesn’t, for a company that doesn’t have the luxury of an investment round to absorb mistakes. And that has direct responsibility for the people who work there—people who go home in the evening, who have families, bodies, limits.
From this position I see something that, maybe, when you’re inside the daily flow you have a harder time bringing into focus. AI hasn’t reduced anyone’s workload on my team. It has transformed it.
It shifted the weight from producing code to supervising code produced by others—where “others,” in this case, is a statistical model that doesn’t sleep, doesn’t get tired, doesn’t get sick. But it makes mistakes. And it makes them in creative, unpredictable ways.
And here there’s a detail that seems important to me, even if I still can’t explain well how important it is. Reading code written by an alien intelligence is a different job than writing it. It requires constant vigilance. It’s a form of attention that keeps you always a bit on alert.
And this alertness is incompatible with something that, in cognitive work, is almost a natural form of protection: flow, the state of deep focus where you aren’t fragmented, you aren’t continuously interrupted, you aren’t forced to make micro-judgments every thirty seconds.
Demand goes up, control goes down
There’s a concept in occupational medicine called “job strain,” in Karasek’s model. In very simple terms it describes the most toxic combination: high job demand and low control over your own activity.
Well, I have the impression that the massive introduction of AI agents is doing exactly this: it’s changing the relationship between demand and control.
Demand increases because, if the tools are faster, the expectation of output grows along with them. “If the agent can do it in an hour, why are we taking a day?” is a question that arrives almost automatically, often without malice, just out of inertia.
Control decreases because you delegate an increasing part of the process to systems you don’t fully understand and that change every few weeks. Even when you do understand them, you understand them in a different way than you understand a colleague or a traditional software component. It’s a more probabilistic, more fragile understanding.
In this sense “token anxiety” doesn’t seem to me like a linguistic fad. It seems like a symptom. Anxiety, after all, is an adaptive function. If it shows up at industrial scale, maybe it’s not an individual problem. Maybe it’s a system signal.
“If we slow down, China won’t slow down”
At this point another objection almost always comes up, and this one too has a core of truth. If we slow down, others won’t. If we set limits, others push. And in a global race, whoever gets there first wins.
But I wonder whether there isn’t a category error hidden inside that sentence.
Competition between nations and between companies is competition between systems, not between individuals. A system that burns the people who make it up is not a competitive system. It’s a system that is consuming its most precious capital—namely the intelligence and creativity of human beings who function well only when they’re healthy, rested, and motivated by something other than fear.
The history of software is full of companies that “won” in the short term with crunch and lost in the medium term. Because the best people leave, or get sick, or simply stop having ideas. An exhausted brain isn’t an engine to squeeze—it’s an ecosystem that collapses.
Maybe the first problem is inside the chest
If the point isn’t only the race for benchmarks, then what is the first problem?
I believe it’s right under our nose—actually, inside our chest. The most valuable problem is figuring out how to integrate tools of extraordinary power into working life without working life devouring life.
It’s not a technical problem, or at least it’s not only a technical problem. It’s a problem of organizational design, managerial culture, labor policy. And in the end it’s also a philosophical problem, because it forces you to decide what is worth doing with the time we have. Which is finite in a way no technology can change.
Something personal, but it weighs
I have a son. It’s not a logical argument, it’s a biographical fact. But biographical facts, sometimes, weigh more than arguments.
When in the evening I close the laptop and look at him, I think that all the code written during the day—even the code “multiplied” by agents—is not worth an hour of that time. Not because work doesn’t matter. It matters a lot. But because that time is unrecoverable in a way code is not.
Code can be rewritten. A childhood can’t.
I know that in certain environments this sounds like weakness. Like a lack of ambition. Like sentimentality that won’t survive the next market evaluation. But maybe it’s the opposite. Maybe the real ambition is to build systems, companies, and lives that last.
And lasting requires taking care of the medium through which everything else happens. The human body, with its non-negotiable need for rest, relationship, and time that isn’t measured in tokens per second.
The alignment we’re not looking at
The AI industry loves to talk about alignment, about how to make models do what we want.
But I wonder if the first alignment problem doesn’t concern the models. It concerns us.
We’re building an economy of attention and productivity in which incentives are often misaligned with what we know is necessary for human health. We sleep little, we move little, we pause little. And then we try to compensate with technology—with meditation apps, sleep trackers, supplements.
It’s as if we created a productive system that no longer produces well-being as a side effect of work, and then we’re surprised an industry of well-being pops up to repair the damage.
A direction, not a solution
I don’t have a clean solution, and I distrust those who offer one. But I do have a direction.
Take seriously that human health and human time are the primary constraint. Not model efficiency, not deployment speed, not the number of lines of code generated in an hour.
Every organizational decision, every sprint planning, every system architecture should also be evaluated with a very concrete question: how much life does it cost?
And if the answer is “too much,” then maybe it’s not the right problem. Or it’s not the right time. Or it’s not the right way.
This doesn’t mean “working less” in the banal sense of the term. It means working knowing that the most sophisticated system in the known universe isn’t an llm. It’s the human brain that designed it. And that brain works according to rules we didn’t write and can’t rewrite at will.
That last gasp, maybe, isn’t humanity’s last gasp. Maybe it’s the last gasp of a way of working that treats people as fungible resources inside an optimization pipeline.
If that’s the case, maybe we shouldn’t try to prolong it. Maybe we should let it go and learn to breathe in a different way. A way that takes into account that we’re made of flesh, time, and relationships. And that no agent, however powerful, will ever be able to live in our place.
Key takeaways
The body doesn’t scale like tokens or gigawatt clusters — sleep, attention, and finite time are the primary constraint, not model efficiency or deployment speed.
Reading code written by an alien intelligence is a different job than writing it: constant vigilance that’s incompatible with flow, the natural protection of cognitive work.
Karasek’s job strain model describes what AI agents are doing to knowledge work: demand goes up because tools are faster, control goes down because you delegate to systems you don’t fully understand.
Every sprint planning and every system architecture should also be judged by a concrete question: how much life does it cost?
A system that burns the people who make it up is not a competitive system — it’s one consuming its most precious capital under the banner of a race against China.
Questions & answers
If AI agents do more work, why are people working more?
Because the system’s productive capacity grew in one place — the machine — while pressure spread across the whole system. More output means more output to review, more decisions to make, more coordination. The bottleneck moved from production to orchestration, and orchestration still needs human bodies with physiological limits.
What is the 'token anxiety' Simon Last from Notion talks about?
A form of anxiety that didn’t exist as a concept two years ago: the sense of not having enough tokens, compute budget, or mental speed to keep pace with what an AI agent produces. It’s the anxiety of someone who has to validate and correct output arriving faster than they can reason about it.
Isn't this just a transition phase, like PC, internet, or smartphones?
Partly yes, but the “it’s just a phase” objection is true historically and useless operationally for anyone with a body right now. Previous transitions took decades and none closed on timescales compatible with an individual’s health. AI’s cycle opens faster than previous ones: the risk isn’t whether it will recompose, it’s how many burnouts it produces along the way.
Is the answer to use AI even more to handle the load?
That’s the worst trap. Increasing the machine’s output makes the problem worse, not better: more artifacts to judge, more decisions, less slack. The bottleneck is the human body — sleep, attention, finite time — and no amount of compute routes around it. Work has to be reorganized around the human limit, not using the machine to deny it.
The author
