Why the labs keep hiring

Every frontier lab has the best coding model in the world installed on every engineer's desk.

Nobody there is waiting on procurement, a license, or budget validation. If AI could replace a good engineer, these are the people who would find out first, and they have every commercial reason on earth to show us.

Right now OpenAI is advertising 818 open roles. More than 400 of them are engineering or research. Anthropic is advertising 609, with 256 in engineering. In March, OpenAI told the Financial Times it intends to roughly double its headcount by the end of this year.

They have the best tools anyone has ever had, and they are hiring humans as fast as they can find good ones.

That tells you something, and I don't think it's what the discourse says it is.

 

So here is my take, and I have been saying a version of it for three years.

The best AI-generated software comes from the best engineers. The best AI-generated art comes from the best artists. The best AI-generated marketing comes from the best marketers.

The best AI-generated anything comes from the best anything.

Not usually. Every time.

And yes, AI lifts the floor. A beginner produces better work today than a beginner produced three years ago.

(But owns nothing, learns nothing, and barely understands its own work.)

It is also exactly why the floor stopped being worth paying for. Everyone has the floor now.

(Sorry, but coding a "good enough" app or product is worth nothing now.)

What I want to do this week is explain the why, properly, because I am not sure I ever have.

GPS Coordinates

I have told you before that your expertise works like GPS coordinates. Without them, the model drifts to the middle of its own probability space and hands you the average of everything it has ever seen.

Here is the mechanism.

A model is a prediction engine. It reads what you gave it and produces what most plausibly comes next. Every word you write changes what "plausible" means for the word after it. That is the entire machine.

So the question was never whether your prompt was long, or polite, or well structured. The question is how much YOUR WORDS (and sequences of words) narrow the field.

If you don't know the right vocabulary, the right questions, the right concerns, and you don't provide the relevant context (sequence of words) about the situation. Your work is going to be "average".

A word is an address

Say you ask a model to make your code faster.

That request exists ten million times in the training data, attached to every level of skill on earth, from student homework to production systems at scale. You will get something reasonable back. You will (most likely) get the middle of that enormous pile.

Now say: this is an N+1 query.

You have done something completely different. That phrase is rare, and it shows up almost exclusively in places where somebody knew what they were talking about: code reviews, post-mortems, things written by people who got burned once and made sure to write it down.

When you use it, you are not just narrowing the request.

You are summoning the neighborhood that phrase lives in.

The model now has a much better idea of who is asking. And it writes for that person.

That's the GPS coordinate metaphor.

The word does the work of the paragraph you did not write.

It is why length does not save you. You can write a thousand words of enthusiasm and land in exactly the same lobby. One correct term takes you further than a page of adjectives, because the term drags its whole world in behind it and the adjectives arrive empty.

And it is why you cannot fake it. You can only name what you know.

It works the same way outside engineering. Ask a model to review a client's portfolio, and you will get something that could sit in any bank's brochure. Ask it whether the gap between your internal ratings and the agency ratings on the three tranches maturing next year matters less than the covenant triggers sitting in the subordinated notes, and you are somewhere else entirely.

Same model. Same subscription. But different context.

The expertise is in the question, the context, the follow-up. That's what allows you to sample from the "EXCELLENCE sub-space" of possible answers.

Nobody tells you the request was thin

Now what makes this more complex and somewhat dangerous is this:

There is no error message for a vague request.

(Or worse, a request that puts you in the "mediocrity sub-space" or "hack-your-company sub-space". Labs implement guardrails to prevent some of that, but as you know, no system is perfect.)

Think about every other tool you have ever used. Bad code throws an exception, a broken formula turns red, a dead link gives you a 404. They all tell you when you have handed them something they cannot work with.

The model never does.

Ask it something thin, and it answers with exactly the same fluency and confidence, the same tidy paragraphs it gives the person who knew precisely what to ask for. It does not say you have not told me enough.

That's not how LLMs work.

(This might btw be included in future harnesses. Cf. my previous letters if you don't know what a harness is.)

It has no way of knowing what you wanted and failed to say.

So you get the average, and it arrives looking like an answer.

That is why this gap is invisible from the inside. Nobody standing in the lobby knows they are standing in the lobby. You find out you were getting the median only once you know enough to see what is missing, which is to say once you no longer needed to be told.

I watch this every term with students. They are not lazy and they are certainly not stupid. They just have no way of knowing that what came back was the cheapest thing the model could have given them.

My theory is that most people get only a tiny fraction of what today's most capable LLMs can do.

Why a better model makes this worse

The standard reply is that all of this is temporary. Today's models need steering. Tomorrow's will not.

I think that is exactly backward, and there are three reasons.

The map gets bigger. A more capable model reaches more places. That is what more capable means. It does not make the address easier to know. It multiplies the number of places you could have gone and did not. Every jump in capability raises the cost of not knowing where you were headed.

That's probably why most people don't "notice" the gap when switching to the most recent, capable models. More capability doesn't mean most people can reach or extract that capability from models.

The default gets more average, not less. This one surprises people. Models are tuned on human preferences, and when several answers are equally acceptable, what that tuning quietly rewards is typicality. So the more finely a model is aligned to what most people approve of, the harder it pulls toward the middle whenever you have not said where to go. Alignment does not rescue you from a vague request. It raises the price of one.

So you pay more for the same outcome.

The mistakes get harder to see. When models were bad, anyone could catch them. The errors were obvious and often funny. Now the output is polished and well argued and usually right, so the errors that survive are the subtle ones. Catching those takes more domain knowledge, not less.

Put the three together, and you get the opposite of the comfortable story. Capability does not reduce the human requirement. It concentrates it.

What I would actually do on Monday

Stop collecting prompt templates.

Start collecting words, intuition, first principles, context, deep understanding, and knowledge.

Take the thing you are genuinely good at and write down what you know about it. Write terms that a competent outsider would not use correctly. The words that carry real distinctions. Write the questions, what can go wrong, and how you read the context around your work. Not only the work itself, but also the humans around your work, what motivates them and what you want to achieve.

Put that in a folder. That's an important part of your AI Operating System.

That list is your coordinates. Most people have never written theirs down, which is why they cannot use them under pressure.

Then do the harder half. Go and read something written by somebody better than you in your own field, understand and collect the insights you did not have. That is the only way the list ever grows, and it is the same way it grew before any of this existed.

But if you collect only the words without understanding them, you won't know how to follow up, and you can't own your own work.

The only way to be great at using AI for your work (not average), is to be very good at what you do. There's no workaround.

If you manage people, this is also the cheapest diagnostic you have. Ask a junior to tell you what good looks like on the thing they are about to build, before they open anything. If the answer comes back in adjectives, you have found the problem, and it was never going to be solved by a better prompt template or a harness or Copilot training...

Back to the labs

The labs are not hiring engineers because their models are weak. Their models are extraordinary. They are hiring engineers because an extraordinary model in the hands of somebody who cannot tell it where to go produces confident, fluent, beautifully formatted average work.

That is an expensive way to average.

The people worth hiring are the ones who can say the specific thing. They were always the people worth hiring. AI did not change that. It just raised the price of the gap, and it will raise it again with the next model, and the one after that.

You can only ask for what you can name.

AI is only as good as the human operating it.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)