AI will change everything, and also not that much

Originally published on christianvismara.com.

There are a lot of people on the internet making the most absurd claims, that's why when you find a well-balanced one nowadays it almost feels more shocking than the exaggerated ones I grew numb to. This episode of Lenny's Podcast with guest Benedict Evans is a great example: he went on to lay out where he thinks AI is all going, and the best line IMO was this: AI is as big a deal as the internet or mobile, and only as big a deal as the internet or mobile.

The first half is for anyone who still thinks this is a normal software cycle, a slightly better autocomplete you can wait out. It isn't, and "as big as the internet" is already an enormous claim. The second half is for the people who think it's the last thing humans invent, the gap that swallows every job by Christmas. The internet was as big as the internet, and it didn't end work, end truth, or end us. You don't have to inflate the thing into a religion to take it seriously.

Evans's read is that we're in 1997. Most of it doesn't work yet, most of what people will build hasn't been invented, and if you'd tried to pick the internet's winners from 1997 you'd have missed almost all of them. You'd have backed the portals, you wouldn't have guessed that a nearly-bankrupt computer company from Cupertino would own the next era, or that a search box with a primary-colored logo would quietly take the whole thing. The lesson there is narrower and more useful than "nobody knows anything": you can be completely right that something is enormous and completely wrong about how it pays off and who captures it.

You can watch that gap in the one corner of this I work in every day, coding tools. For months Claude Code was the clear leader, on the benchmarks that matter and on the raw share of code getting written. OpenAI was falling behind despite initially being the first mover in the space, at least until they released Codex, through which they're silently catching up to Claude Code on downloads. The next release could flip it back: when the lead changes hands that fast, "model moat" sounds like the wrong phrase for it. The models don't seem to have the network effects that let one pull ahead and stay there; they look less like Windows, which locked you in, and more like cloud capacity, where there's always another supplier a click away. Evans puts the sharp version of it this way: blind-test the same prompt across the big models, and most people couldn't reliably tell you which one wrote which.

The jobs conversation, where most of the noise lives, calms down the same way. Evans separates a task from a job. A task gets automated; a job is a bundle of tasks with judgment and accountability wrapped around it, and the bundle holds together better than any single task in it. You can't look at a senior partner at a law firm and say "17% of her work could be automated, therefore," because the sentence has nowhere to go. Also look at the labs everyone names as the job-killers: they are hiring as fast as they can fill the seats, the big consultancies are booming on AI work, and some of those same firms are cutting elsewhere at the same time. Work is moving around, the way it moved when the spreadsheet arrived. The spreadsheet didn't end accounting; the number of accountants rose through fifty years of software built to replace them.

I get to watch the move from inside it. The product studio I run is busier than it has ever been, and the reason is AI, companies that need someone to actually build the thing they've decided they want. That work didn't exist in this shape three years ago.

The posture under all of it is the part I've actually kept. Evans calls it presuming radical uncertainty, and pairs it with the line he gives as his motto: probably going to be okay, not for sure. It lets you take the technology completely seriously, use it every day, and still refuse to pretend you know which bets land. The alternatives are both cheap: You can put your head down and feel morally superior about how much you hate all of it, which costs nothing and helps no one, or you can pick a date for the singularity and post through it. Neither asks you to do any work.

Where I'd push past Evans a little: I think the next year resolves less of this than people expect, not more. The labs going public will drag out numbers we've never had, and my guess is they'll come back messier and more boring than either camp wants, enormous real usage sitting on top of economics nobody has worked out yet. I'd rather be early to that messy reality than to the next confident prediction everyone lines up behind.