Everyone gets faster. Not everyone gets more valuable.

#ai #engineering #careers
Wilbur the Pit Bull Terrier dressed as a distinguished professor in a white lab coat, red bow tie, and round glasses, standing at a whiteboard and pointing a stick at a hand-drawn U-shaped graph titled 'The Canine Value Curve, according to Wilbur, PhD, D.B.A. (Dog Barking Analytics).' The curve rises green and labeled 'AI amplifies your value,' dips orange through 'AI erodes your value,' and crosses a dashed 'net zero' line, with its three regions marked 'newly enabled: outsiders, generalists,' 'the squeezed middle: routine execution (ouch),' and 'amplified: judgment and taste.' The lab around him is packed with jokes: a desk nameplate reading 'Wilbur, PhD, Paws. In. Dy-namics,' a stack of books titled 'Canine Econometrics,' 'Good Boy Strategy,' and 'The Art of Sniffing Signals,' mugs that read 'Caffeine + Chaos' and 'Best. Boy. Ever.,' bubbling beakers labeled 'Ideas' and 'Data is delicious,' a glowing crystal ball, an 'E = MC squared (Bark squared)' scrawl, and a smug gold cat statue tagged 'stop dog innovator.'

You’ve seen the story by now. Somebody with no technical background, no business writing a line of code, sits down with an AI chatbot and ships a real, working website over a weekend. A year earlier it would have been impossible. The floor came up. People who could never do the thing can suddenly do the thing.

That is the story everyone wants to tell about AI right now, and the research seems to back it up. There’s a study of a few thousand customer support reps where the rookies got dramatically more productive while the veterans barely moved. There’s a controlled experiment where developers built a little web server and the least experienced ones got the biggest speedup. There’s one on professional writing where the weakest writers gained the most. Same shape every time. AI helps the bottom more than the top. The gap closes.

So the comfortable conclusion is that AI is the great equalizer. The tide comes in, every boat rises, and the little boats rise fastest.

I think that conclusion is wrong. Not the data. The data is fine. The conclusion. And it goes wrong in a specific, instructive way. It takes one slice of the picture and mistakes it for the whole thing.

Two true stories that disagree

There are two competing intuitions about what AI does to a workforce, and the funny part is that both of them are correct. They just describe different parts of the same picture.

The first is the steady multiplier. AI gives everyone some constant boost, a flat percentage on top of whatever you already produce. Under that story the strong pull ahead, because a percentage of a big number is a bigger number. The gap widens.

The second is compression. The bottom rises faster than the top, the floor comes up, and the distance between the worst and the best shrinks. That is the equalizer story, and it’s the one the headline studies found.

Here’s the trick. If you stand at one end of the picture and squint, you see the multiplier. If you stand at the other end and squint, you see compression. Each camp is looking at a real region and reporting honestly what they see in it. The mistake, every time, is assuming the slice you happen to be standing in is the whole graph. It isn’t. Step back far enough to take in all of it at once and the straight lines bend into something more complicated, and a good deal less comfortable.

The compression is real, and it’s bait

Start with the compression studies, since they’re the ones people wave around. Look at what they actually measured. A support ticket with a known good resolution. A standalone coding kata with a clean spec. A short writing prompt. Every one of them is bounded, single skill, clearly specified, and cheap to check. You know exactly what done looks like before you start.

In a world like that, of course the novice gains the most. AI hands them a floor of competence they didn’t have and they sprint to it. The expert was already near the ceiling, so there isn’t much headroom left to gain. Compression isn’t a deep truth about intelligence. It’s arithmetic. Give everyone a floor and the people standing on the floor benefit while the people near the ceiling don’t.

I lift. Now imagine there was a machine at the gym that lifted the weight for you. Not the kind you lift against, an actual machine that does the lifting while you stand there and watch. It would not make you strong. Showing up and moving the load yourself is the only thing that makes you strong, and nothing shortcuts that. A floor of competence is that imaginary machine. It is not the same thing as being able to carry the load yourself when the machine isn’t there.

The trouble is that almost no real software work looks like the exercise. Real work is unbounded, underspecified, and expensive to verify. The spec is wrong. The requirements contradict each other. The client is an asshole who needs to get smacked down. The thing has to live inside a system that is already on fire in three places. Done is a judgment call. Sometimes done is a lengthy and expensive legal debate. And that is where the floor quietly stops helping you.

Speed is not value

Here’s the move almost everyone misses. Going faster on the task is not the same as being worth more.

There’s an old idea from manufacturing. Walk the factory floor and look for where the inventory piles up. The pile is sitting in front of your slowest machine, and that slowest machine is the only thing actually setting the pace of the whole line. Speed up any other station and you don’t ship one more unit. You just grow the pile in front of the bottleneck faster.

AI is a very fast station bolted onto the front of a line whose bottleneck was never typing. The constraint was always review, integration, knowing whether the thing is correct, and deciding whether it should exist at all. So you flood the front of the pipe and the pile in front of the slow part gets enormous. The big annual DevOps survey landed on exactly this and called AI an amplifier. Individual output goes way up, more pull requests, more tasks closed, while actual delivery at the org level stays flat. Everyone is producing more code. The same amount of value comes out the other end.

Then there’s QA, which was the underfunded, unloved part of the process long before any of this started. We cannot test and verify as fast as a model can generate. The model will produce more plausible code in an afternoon than an honest review and test pipeline can vet in a week. So the pile in front of the bottleneck isn’t just features waiting their turn. It’s unreviewed risk, sitting there, waiting to surface at the worst possible moment.

There’s a widely shared essay that named this the 70% problem, later updated to the 80% problem. AI gets you most of the way there, fast. The last chunk, the edge cases, the security, the production integration, the part where it has to be right and not merely plausible, does not shrink. If anything it gets harder, because now you’re debugging code you didn’t write and don’t fully understand. That last stretch is pure judgment.

I cofounded a company called StriveDB. It’s software for victim service organizations, domestic violence shelters and the like. AI can write a lot of that code. What AI cannot do is own the consequence. A data leak there is not a bad week and a blog apology. It’s a safety incident for someone hiding from a person who wants to hurt them. The part of that job that matters has nothing to do with how fast the code gets typed. It’s the judgment about what could go wrong and who pays when it does. That part does not speed up. It only gets more important.

What the market is actually doing

If you want to know who AI is really helping, don’t read the productivity studies. Read the hiring.

A payroll analysis covering millions of workers found that since AI showed up in force, employment for people early in their careers dropped sharply in the jobs most exposed to it. For young software developers specifically, the drop was steeper still. Older, more experienced workers in the same jobs were basically untouched. A separate talent report covering a huge slice of the workforce found that new graduate hiring at the largest tech companies has fallen off a cliff in recent years, even as demand for senior people kept climbing.

So sit with the contradiction. The studies say AI helps beginners most. The market is firing beginners and bidding up veterans. Both are true at the same time, and reconciling them is the entire point of this article.

Here’s the reconciliation. The person who builds a website over a weekend with no background is an outsider. They have no skin in the wage game, so a new capability lands as pure gift. The credentialed junior developer, two years out of school, is in a completely different spot. They are paid like a professional, and the work that justified that paycheck, turning a clear ticket into working code, is exactly the work the model now does in thirty seconds. Same tool. For the outsider it lands as a gift. For the new grad it lands as a layoff. The difference was never the tool. It was where you were standing when the tool arrived.

The whole curve

Step back and look at all of it at once and the two straight lines bend into a U.

The AI market-value curve A U-shaped curve plotting the effect on your market value against experience, competence, and talent. It starts slightly above the net-zero line for the newly enabled (outsiders and generalists), where tools broaden access; drops well below zero through the routine middle, the most exposed group, where routine execution is compressed; then climbs steeply above zero for judgment-led work, the most amplified group, where leverage compounds with judgment. The AI market-value curve Positive Negative Net zero Low Developing Competent Advanced Exceptional Experience / competence / talent Effect on your market value Newly enabled outsiders & generalists Judgment-led work most amplified Routine middle most exposed Tools broaden access Leverage compounds with judgment
Net effect on your value across the curve: a small lift at the newly enabled end, a dip through the compressed routine middle, and the steepest gains where judgment compounds.

On the left, the newly enabled. Outsiders and generalists who get a real, new capability. The lift is genuine but small and a little fragile. They hit the 70% wall without the mental models to climb the rest, and the thing they built is brittle in ways they can’t see.

In the middle, the squeezed. People whose value was reliable execution of clearly specified work. They are caught from both sides, more expensive than an outsider with AI for routine output, and not yet differentiated by the judgment the model can’t do. That is the bottom of the U.

On the right, the amplified. People whose value is nonroutine judgment. Architecture, framing the problem, knowing what is worth building, verifying what comes back. AI scales exactly those things, and the returns compound.

One honest caveat, because someone will throw it at me. This U only works if the horizontal axis is the kind of value you bring, not your years on the job. Plot it against seniority and it isn’t a clean U, it’s closer to a cliff at the entry level. The reason the new grad falls in the hole isn’t that they’re young. It’s that they’re paid as a professional for work that now sits at the commodity floor. Mistake the axis and the whole picture falls apart.

It’s not seniority. It’s fundamentals.

Here’s the part I can’t prove, so I’ll just say it and you can argue with me.

AI is a task efficiency multiplier. On any single bounded task, it makes you faster, and that much is probably just true. But how that little truism plays out across a whole market, against your knowledge of fundamentals, against the limits of the language you are even able to think in, is not a flat multiplier and it isn’t a rising tide. It’s the curve. And as with very nearly everything else in human history, once you let it run across all of those dimensions at once, the benefits accumulate at the top.

The reason is unglamorous. The people who get the most out of these tools are the ones who understand what is actually happening underneath them. Not the people with the best titles. The people who know how the computer works. How memory works, how the network works, why the database is slow, what the abstraction is hiding. The people who, if you peeled back three layers, would still know exactly where they are. The tool is ruthless about this. It rewards you in almost exact proportion to how well you understand what it’s doing for you.

The people who get the least are the ones who thought the job was typing the right magic words into a terminal. They had a ritual, the ritual produced results, and they never needed to know why. AI is fantastic at handing those people a faster ritual. It is no good at making them more effective, because they can’t tell when the output is wrong, can’t steer it when it drifts, and can’t ask it for the thing they don’t have a word for.

My friend Craig, the most thought-provoking conversationalist I know, says this better than I do. We go through life owning the ability of our team and of our whole society as if it were our own, and it isn’t. You don’t find the edge of what you actually know until the moment you are standing in front of the problem alone. His favorite proof: everyone is certain they know how a toilet works. They do not. Or ask someone to draw the real working mechanism of a bicycle, how it’s powered, how the motion actually happens. Most people cannot do it, people who have ridden one their entire lives. They know what a bike is. They have no real idea how it works.

AI is the most convincing borrowed competence we have ever had. It lets you operate way out past the edge of what you actually understand, right up until you’re alone with the piece it got wrong, and then the bill comes due. The floor it hands you is real. It is also somebody else’s understanding, rented by the hour, and you don’t find that out until the moment you have to debug it yourself.

That last part about the word is the whole game, and it’s older than computers. We think in language. The words you have are the thoughts you are allowed to have. If “race condition” isn’t in your head, you cannot ask the model about one, and worse, you won’t recognize it when the model hands you one, smiling, in code that looks perfectly fine. Your mental model is the resolution limit on every prompt you will ever write. The tool reaches exactly as far as your language can point, and no further.

I learned to code in a bedroom in Albuquerque, grinding, with no floor under me and no AI to catch me. It was slow and it was often miserable and there were no shortcuts. But that grind is where the mental models came from, and the mental models are the thing I actually get paid for now. The struggle wasn’t in the way of the learning. The struggle was the learning. Which brings me to the thing that bothers me most.

I don’t want to do anything anymore

Let me tell you how this article got made, because it’s a little funny and a little grim.

I’m not typing this. I’m talking. I’ve got a local model running on my own machine turning my voice into text, because somewhere in the last year I decided typing was too much friction. I’m talking to Claude, which is going to take what I say, write it up, drop it into the personal site Claude built for me, and deploy it to infrastructure Claude configured. And the whole thing started from a research paper I had a different AI write, under my direction, with my argument, chasing down the studies I wanted, so I’d have something rigorous to push against.

So here’s the honest stack. My opinions. My judgment. My taste. My argument. And underneath all of it, AI, the whole way down. The thinking is mine. Almost nothing else is.

And I am not even slightly embarrassed about it. That is the part that should give you pause, because it means the seduction works on me too, and I am supposed to know better.

There’s a randomized trial I keep coming back to. They took experienced open source developers, people who had maintained their own codebases for years and knew every corner of them, and gave them AI tools. They got slower. Not a little, either. And here’s the genuinely sinister, shitty part. They were certain they’d gotten faster. They predicted a speedup going in and reported one coming out. They were wrong about their own experience while it was happening to them.

I feel the same pull in myself. I used to hold a whole schema in my head. Now I ask. I reach for the model before I reach for the thought, and every time I do, the muscle that would have had the thought gets a little weaker. I built a little agent called Kessler that runs my inbox, my calendar, my follow-ups, all the daily admin a human shouldn’t have to touch. It even sends my mom a daily affirmation. I automated being a good son. Sit with that one. The bot is a more reliable source of warmth in my mother’s morning than I am.

Wilbur, my dog, looks at me perplexed while I scream into a microphone at a robot that can’t hear me. I’m doing this instead of putting on some jazz or a little classical, typing peacefully, being mildly contemplative, sliding into something like a flow state. The dog is confused. I cannot honestly pretend I’m booking a win for my mental health here.

That’s the trap, and it closes on the experts too, which is the part nobody warns you about. The fundamentals are the engine of your leverage. Laziness, letting the thing do your thinking because it’s right there and it’s so easy, is what quietly takes the engine apart. You can sit on the right side of that curve and then atrophy your way down it, one offloaded thought at a time, feeling more productive the whole way down.

The tool only works in the hands of someone who still knows what’s happening. The moment you stop knowing, it stops being a multiplier and becomes a very confident way to be wrong faster.

So

AI is not a tide, and it isn’t a flat multiplier either. Both of those are just what you see when you only look at one slice of it. Take in the whole thing and it’s a curve, and the benefits pile up at the top, the way they have for pretty much every powerful tool we have ever invented.

As the work moves from bounded, clearly specified implementation toward unbounded judgment, figuring out what to build, whether it’s right, what happens when it’s wrong, the value flows to the people whose contribution was always judgment, and it drains away from the work that was always just execution. Everyone gets faster on the isolated task. Not everyone gets more valuable. Those are different axes, and confusing them is how you end up cheering for the thing that’s quietly eating your job.

Understand what is actually happening underneath you. Keep doing some of the reps yourself, even when the machine offers to move the weight. That is the whole edge. It might be the only one that lasts.

For the pedants and nerds.