Around two in the afternoon, the machines are still ready to go.
I am not.
This mismatch is becoming one of the defining features of my work with AI. I can move faster than I ever could before. I can tab between six terminals, keep several agents working, review a draft, chase down a UI bug, and start something new before the old thing has fully cooled (or maybe even finished).
For a while, this feels like abundance. The distance between an idea and a working artifact collapses. The queue keeps moving. It becomes easy to believe that my capacity has expanded along with the system’s capacity.
But I still have a limited context window in my own head.
By mid-afternoon, the warning signs start to appear. I read a diff and realize I have stopped asking why it even exists. I find myself making excuses for hallucinated features. A solution feels slightly forced, but I want to merge it anyway — enough of the surrounding machinery is green, right? I begin to overlook details I would have caught that morning, when I was fresh.
I start more things because starting still feels exciting, even though my ability to integrate them is falling. Even though the visceral anxiety downstream of all these little indiscretions is beginning to mount.
That first ping of worry in my forehead or gut is a warning sign I ignore at my peril.
The AI is not tired. That does not mean the system is not tired. I am part of the system.
I am toying with calling my response an AI Siesta. I don’t actually nap — sleep is not the point. The point is to stop increasing cognitive throughput when my capacity to judge the output has begun to contract. Close the terminals. Stop prompting.
Walk, sit, stare at something and blank my mind. Just look. Preferably I look at something far away that let’s me use my whole peripheral vision. Something green. Something that does not generate a response.
Let the work become quiet enough that I can hear whether it still makes sense. This is not a productivity hack. It is an offramp.
It’s the only way I can come back a few hours later with gas in the tank.
The work expands faster than the worker
Siddhant Khare recently described the same paradox from inside AI engineering: AI makes each task faster, so we do more tasks.
Production gets cheaper while coordination, review, and decision-making get more expensive. Those costs land on the human. That distinction matters. Much of the AI conversation treats generated output as completed work. But generation is often the cheap part.
Someone still has to hold the purpose of the system, compare the output against reality, notice the edge case, and decide whether this is the right thing to ship at all.
Recent research on what workers have started calling “AI brain fry” points in the same direction. In a survey of nearly 1,500 full-time workers, heavy use of multiple AI systems was associated with acute cognitive fatigue: mental fog, slower decisions, and a sense of crowded thinking.
AI did not merely shrink workloads. It expanded what the researchers called the “sphere of accountability.” People became responsible for producing and monitoring more in the same amount of time. That phrase lands for me. My tools give me a larger sphere of action without automatically giving me a larger sphere of attention.
This creates a dangerous period late in an otherwise productive run. I still have enough energy to act, but not enough clarity to discriminate.
The system continues offering novelty and apparent progress. One more prompt. One more agent. One more fix. Each action is plausible on its own. Together they can carry me past the point where I am actually steering.
Withdrawal is information
The Archetypal Wavelength is my attempt to describe a recurring rhythm underneath very different systems: Rising, Peaking, Withdrawal, Diminishing, Bottoming Out, and Restoration. Everything that oscillates can be mapped to the sine curve of these six phases.
Withdrawal is the turn just past the peak. Energy is still relatively high, but it has changed direction. In its healthier form, Withdrawal is discernment, introspection, collecting data, and stepping back. In its unhealthy form, it looks like burnout, anxiety, overlooking details, or trying to preserve the peak by force.
That is the moment I am trying to awaken to at two in the afternoon. Awaken to so I can rest.
Our culture tends to treat contraction as failure. If expansion feels good, more expansion should feel better. When the energy begins to recede, the reflex is to override it. Add stimulation. Open another tab. Ask the model to try again. Keep the productive feeling alive.
But the withdrawal phase is not an error in the wave. It is information about the condition of the system.
Ignored, it can become an ugly reinforcing loop. Reduced attention creates weaker work. Weaker work creates more cleanup, more uncertainty, and more things to hold in mind. That extra load reduces attention further. Eventually the system finds an offramp anyway, often through a mistake, conflict, failed release, or painful collapse.
What could have been a deliberate contraction becomes a nosedive.
The siesta interrupts that loop while the choice is still mine.
A stock called energy
Donella Meadows taught me to look for stocks, flows, feedback loops, and delays rather than blaming individual events. Here the stock is energy: not merely physical energy, but the capacity to attend, judge, integrate, and care. AI increases the flow of possible work.
It does not replenish that stock at the same rate. Early in the day, a reinforcing loop dominates. A good output creates momentum. Momentum makes it easier to frame the next task. The next result arrives quickly, which creates more momentum. For a while, energy seems to rise with output. But growth systems meet constraints. At some point another loop strengthens. Each new branch creates review work. Each decision leaves residue. Each context switch draws from the stock. Because there is a delay between expenditure and felt depletion, I can continue accelerating after I have crossed the sustainable limit. This is classic overshoot behavior. The signal arrives late, so the actor keeps consuming the stock as if the earlier conditions still hold.
Meadows warned that complex systems cannot be controlled from the posture of an omniscient conqueror.
Her alternative was more humble: get the beat, listen to the system, expose our mental models, and expand the time horizon.
An AI Siesta is a small way to do that. It is a feedback policy for the feedback system. When judgment starts thinning, stop adding throughput.
The same wave, one scale up
I do not think this is only a problem for tired software engineers. It mirrors our predicament in the polycrisis.
AI tells a familiar story: intelligence is becoming abundant, productivity can compound, and growth can continue because the newest technology will dissolve the previous limit.
Oil told a version of this story.
So did the broader civilizational project Charles Eisenstein calls the ascent of humanity: progress as increasing mastery over nature, matter, and constraint.
When I am flowing between six terminals, the story feels true. I can understand its seduction from the inside. Brendan Graham Dempsey describes modern civilization as a lopsided spring. Our cognitive and technological capacities wind tighter and faster than our ethical, relational, and ecological capacities develop. We become more capable of acting without becoming equally capable of sensing the whole in which we act. Capacity outweighs insight.
AI accelerates the winding.
The issue is not that acceleration is evil, or that we should retreat from these tools. I am deeply optimistic about what they make possible. The issue is that a reinforcing loop has no wisdom about when to stop reinforcing.
A system organized around expansion will interpret every new unit of capacity as permission for another unit of demand. At the scale of a workday, that means the saved hour becomes three more tasks. At the scale of a civilization, it means efficiency becomes more extraction, abundance becomes more consumption, and every constraint becomes a technical problem to overpower.
The balancing signals—fatigue, ecological damage, social fragmentation, loss of meaning—arrive as inconveniences to suppress rather than information to receive.
Then the system overshoots.
Inviting contraction
The AI Siesta is not my attempt to personally solve the polycrisis by taking a nap. That would be an impressive amount of thought-leader bravado, even for Substack.
It is a practice at the scale where I actually have agency.
I am learning to recognize the moment when speed stops serving the work. I am learning that the first flicker of doubt after a creative peak does not always need another prompt.
Sometimes it needs distance.
I am learning to invite contraction before contraction has to force itself upon me.
This changes what rest means. Rest is not fuel purchased in order to maximize the next round of output. It is one of the phases by which a living system remains whole.
Withdrawal makes room for discernment. Diminishing releases what cannot be sustained. Bottoming Out admits the limit. Restoration gathers energy without demanding that it immediately prove its worth.
The machine may have another thousand context windows available. I do not. The wiser system is not the one that pretends otherwise.
It is the one designed around that truth.