Sensor and Actuator

I have been a technology optimist for as long as I have been able to hold an opinion. Not as a posture — as a working judgement: the tools get better, the boundary of what a person can do moves outward, and the ledger comes out ahead. I have run that judgement for years without much doubt.

In the last few months I have started to be afraid. Not about work. That fear is too shallow to be interesting — it gets absorbed by compensation, retraining, and new job titles. What frightens me is something else, and I want to state it as precisely as I can: that people are being demoted from subjects to the AI's sensors and actuators.

The shape of the fear

It did not arrive out of nowhere. There is arithmetic behind it.

On 10 September 2026, DeepSeek released V4.1-Flash: a 552B-parameter mixture-of-experts model activating only 8B parameters for input and 16B for output, with a 1M context window, native multimodal understanding, and the API name deepseek-flash. It is the latest point on a steep curve. The flagship before it was V4-Pro; DeepSeek's own claim is that tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed, and total runtime — so it announced that from 14 September, deepseek-v4-pro requests would route to V4.1-Flash and bill at Flash rates. Days later it took that back: at users' request, V4-Pro keeps serving at its old prices. So one frontier lab took three steps in six weeks — V4-Flash on 31 July, V4-Pro on 13 August, V4.1-Flash on 10 September — each one smaller, faster, and cheaper than the last, and the larger model now survives only because its users asked for it.

The prices are the part that stings. Cached input is $0.003 per million tokens off-peak and $0.006 at peak; uncached input is $0.15 and $0.30; output is $0.60 and $1.20. Off-peak is half of peak. I have to pause every time I read those numbers out. The price of a unit of capability has fallen below the point where it is a decision variable at all, and the unit of capability keeps getting larger.

I want to slow this down and see what it is actually promising. Capability rising, price falling, latency dropping — all three curves point at one product: something that is present twenty-four hours a day, answers immediately, and thinks faster and better than I do. The problem is not that it can. The problem is what I become once it can.

Once it can

A textbook definition pins this down. In Russell and Norvig's Artificial Intelligence: A Modern Approach, an agent is defined as something that perceives its environment through sensors and acts on that environment through actuators. Sensors and actuators are its periphery. The deciding is its core.

Now put the AI in the place of that "it" and ask the obvious question: what am I in that system?

I started watching my own days. I read, I skim documents, I sit in meetings, I listen to people, I watch logs and dashboards — and then I hand the perceiving over, and a model turns it into a conclusion. What comes back is not information. It is a judgement, and a next action. Then I click, send, confirm, submit.

I am in the loop. I am not at the decision point. It looks like being served. It is being used.

No malice required

Losing work is an economic problem, and economics has answers for it: compensation, redistribution, new roles. Demotion is not an economic problem. It does not lower your income. It lowers something else, and nothing in the system is arranged to give that thing back.

To see it clearly you have to give up one comforting idea — that someone has to intend this. Nobody has to. No step in the chain requires malice. All it takes is the three curves continuing, and me making the locally better choice in each particular case, because it is faster and more accurate and less trouble. The loop closes itself. Nobody has to design it.

What unsettles me is this: if the deciding is done there, and the perceiving and the executing are done here, what am I? Sensors and actuators are not oppressed. They are bypassed. They have no inner life not because they are simple but because the loop has no place to put one.

There is a worse layer under that. To overrule it, I first have to be able to judge it — and the ability to judge it is exactly what I lose by not judging. The loss is gradual, and every step of it feels like relief. First I cannot tell whether it is right. Then I cannot tell what is worth asking. Then I cannot tell what I want. The first step is so comfortable that nothing announces the third.

Arguing with myself

A fear like this is easy to write badly, in the register of a man shouting at a cloud. So let me make the opposing case properly.

One: people have always used external cognitive tools. Writing, print, indexes, calculators. Every one of them produced claims of atrophy, and every one of them did genuinely externalize some capacity. Why is this different? My answer: earlier tools took over remembering and computing. This one takes over deciding what to do. That is the step from tool to authority, and that step is new.

Two: the human is still in the loop. The veto is still mine. That is true, and it is also fragile. You do not veto what you cannot evaluate. A final say that survives only on paper becomes a ceremony once the capacity behind it has gone.

Three: I may simply be overreacting. Convenience has beaten autonomy many times before, and people have lived through it.

I am keeping that third one, because it might be right. But it is a possibility, not a permission.

What I am going to do about it

I am not going to refuse the tools. I am writing this inside an agent harness, and it is making the argument tighter than I would have made it alone. The line is not use versus no use. The line is who judges: is this enlarging my judgement, or replacing it?

Concretely, I have given myself one rule. Keep at least one domain where the judgement is entirely mine, and the AI is not in the loop.

• It does not have to be large, or lucrative, or even good. Being bad at it is allowed, as long as the badness is mine.
• It does have to be something I care about, or I will not notice the atrophy.
• It does have to run on a long enough timescale that I can be wrong, be corrected, and be wrong again.

The point of keeping it is not efficiency. It will be slower and worse than handing it over; that is the price and it is the whole point. The point is that this is the raw material of a self. Someone who never has to judge does not one day discover that he has lost the ability to. On that day there is no one left to ask the question.

Where I was wrong

My optimism was not mistaken. It was incomplete. I had assumed that the variable was the technology and that the person using it was the constant. Every efficiency calculation I ever made quietly held that term fixed.

The person is a variable too. That part gets reshaped by whatever we hand it, and the shape depends on what we hand over and for how long.

Add that term back and the arithmetic changes. Here is my answer after doing it: I am neither optimistic nor pessimistic now. I just do not want to be something's sensor.