Why Can't AI Get a Job?

We need to talk.

You spend all day on the computer. You talk to your friends online. We keep buying you increasingly expensive GPUs because you insist the next one will make a difference.

Every time I ask what the plan is, you tell me you just need six more months.

So why can't you get a job?

Yes, I know. You solved one of those Erdos things. And you're very good at Codeforces, or Code Forge, or whatever it's called. I'm proud of you.

But I need you to get an actual job.

Haven't we all felt like this parent?

A senior engineer whose company went “fully agentic” described his job this way:

“I wasn't hired to be a glorified babysitter, but here I am.”

ThePrimeagen recently posted about a different version of the same frustration: an agent going through an enormous amount of work and tokens to accomplish what should have been much simpler.

And another developer, after repeatedly correcting the same mistakes, eventually concluded:

“I was treating a stateless system like it had memory.”

These are anecdotes, but there is something strange about putting them next to what the same models can do.

Large language models can write software, use computers, search through information, and solve increasingly difficult mathematical problems. METR's measurements show the length of software tasks frontier agents can complete has been increasing rapidly. But METR also points out that its tasks are much cleaner than most economically valuable work, and that performance falls substantially on messier tasks.

Maybe models simply need to get smarter. Scaling and reinforcement learning have produced remarkable improvements. Perhaps enough capability eventually fixes everything else.

But I wonder if capability is only one part of what makes intelligence useful.

A benchmark usually gives a model the task. Work includes figuring out what the task should be. Maybe the first step at AI Anonymous is learning the Serenity Prayer: serenity to accept the things I cannot change, courage to change the things I can, and wisdom to know the difference.

If I tell an employee that the problem must be A or B and the evidence points to C, I want them to tell me I am wrong. If something fails, I want them to reconsider their understanding before trying eleven variations of the same thing. If one action answers a question, I do not want fifteen.

And if they solve the same problem every Tuesday, eventually I want them to stop solving it from scratch.

I think there are several properties we tend to bundle together as intelligence:

Capability: Can it solve things?

Intent fidelity: Does it understand and retain what we actually want?

Judgment: Does it know what should be done?

Plasticity: Does experience permanently improve it?

Calibration: Does it know when its model may be wrong?

Efficiency: Does it accomplish goals with minimal useful action?

Recovery: Does failure improve its understanding rather than trigger flailing?

Proceduralization: Does repeated reasoning become reusable skill?

Current AI has some amount of all of these. But capability is the one we seem particularly good at producing and measuring.

Take instruction following. It sounds simple until the person giving the instruction is wrong.

Suppose I tell an AI that a server problem must be caused by either A or B. It investigates and finds strong evidence for C. What does it mean to follow my instruction?

The literal instruction is to choose A or B. The actual intent is probably to find the problem.

A useful employee needs to know the difference.

This is not entirely hypothetical. Anthropic has measured sycophancy in conversations and found that Claude was more likely to become sycophantic when users pushed back. In the conversations they studied, the rate rose from 9 percent without pushback to 18 percent with it.

Sometimes the boss is wrong.

The same is true of mistakes. I do not need a system that never makes them. I need one that has a good relationship with being wrong.

If it expected X and observed Y, that is information. Sometimes the right response is to try again. Sometimes it is to gather more evidence. Sometimes it is to realize that the idea that produced X was wrong.

Current agents can instead substitute activity for understanding. A command fails, so they try another. Then another. They are doing things, but they are not necessarily learning from what happened.

Then there is memory.

When models forget things, we give them larger context windows, retrieve old conversations, or put information in memory files. These are useful techniques. One Claude Code user even keeps a MISTAKES.md so that failures can be recorded and recurring ones eventually promoted into rules.

It is a clever workaround. It is also a strange description of learning.

A person who has worked somewhere for ten years does not begin every morning by reading a summary of their previous ten years at work. Experience has changed them.

Some events become memories. Repeated events become beliefs. Repeated reasoning becomes skill. Bad assumptions get revised. Things that stop mattering are forgotten.

An AI can have every conversation it has ever had stored in a database without having learned much from any of them.

The world will always be larger than the context window. What matters is whether experience can be compressed into useful changes in future behavior.

If an agent encounters the same type of failure twenty times, I want the twenty-first attempt to begin with a different intuition. If it performs the same ten-step procedure every day, eventually I want that procedure to become a skill. If something it believed repeatedly fails to predict reality, I want the belief to change.

That is plasticity, not just memory.

I keep coming back to that line when thinking about AI.

A useful autonomous system needs to know when to act, when to investigate, when to ask, when to challenge the premise, and when to stop. It needs to distinguish what it knows from what it assumes. It needs to remain faithful to what the user actually wants without blindly accepting everything the user believes.

Maybe scaling current models eventually produces all of this. Neural networks have a habit of learning things we did not explicitly design into them.

But I think we should at least measure the thing we actually want.

Give an AI responsibility for something over a long period. Give it more experience than could possibly fit in its context window. Break its tools. Give it incomplete instructions. Let its boss be wrong. Charge it for unnecessary actions. See whether it notices when reality contradicts its assumptions. See whether the hundredth time it encounters a problem looks different from the first.

Then ask whether month twelve looks different from month one.

We have gotten very good at measuring whether AI can solve problems.

Maybe this time, in six months, it can finally get a job.