Crossing
the Threshold

If you want to guess what the next seven years of AI will look like, you really only have to answer two questions. Do AI agents get reliable enough that you'd trust them to do things on their own? And does AI start to make AI better? Nearly everything people argue about — which companies win, which jobs go, how fast any of it happens — turns out to hang on those two.

I'm not going to try to predict one future, because I don't think anyone can. What you can do is map the few futures we might actually end up in, work out the order they'll separate in, and notice the signs that tell you which one is coming. The near future is fairly easy to see. The far one forks hard. A map is useful mostly because it shows you where the roads split before you get there.

1. The next few years are easy to predict.

   Agents go from impressive demos to things you can actually rely on, at least for narrow tasks. A lot of thin software gets absorbed into bigger platforms. Companies settle on owning their data and swapping the model underneath. None of this needs a breakthrough. It's just the spread of things that already work.

2. The models improve faster than companies can use them.

   New abilities show up in months. Actually putting them to work takes years. That gap — not how good the models are — is what decides how much changes soon. “AI is moving incredibly fast” and “nothing really feels different yet” are both true, and for the same reason.

3. Two questions split the whole thing open.

   Whether agents get reliable enough to trust, and whether AI starts improving AI. Put those two together and you get four fairly distinct worlds. Only one of them makes the rest impossible to predict.

4. Compute and power are the real moat, and they end up in few hands.

   Whoever can train the biggest models and buy the electricity to run them has an edge that compounds, and not many players can do it. But open models, which can't be taken back once released, keep a floor under everyone else. The interesting fights happen in the space between those two facts.

5. Jobs change task by task, long before they disappear.

   Roles get hollowed out from the inside as pieces of them get automated. The first place you see it is entry-level hiring, not layoffs. What's left moves toward judgment, taste, and being the person who's accountable. The averages can hide a lot of concentrated pain.

6. The far future is unpredictable for one specific reason.

   If AI starts making AI better, everything speeds up and the questions in this report stop being the right ones. The first sign won't be a benchmark score. It'll be in reliability curves, and in how close open models stay to the best ones. Watch those.

A good map doesn't need many lines on it. Lots of things about AI are uncertain, but most of them depend on just two. Nail those two down and the number of futures you have to worry about drops from infinite to four.

The first question — can you trust it?

Can you trust an agent to finish a job without someone checking its work? This isn't really about how smart the model is. A model that's brilliant but wrong five times in a hundred is something you have to supervise. One that's wrong once in a thousand is something you can hand things to and walk away from. Those aren't two points on the same line. They're different in kind — the difference between a tool and a coworker. So the thing that matters most in the near term is an error rate, not an IQ.

The second question — does it improve itself?

Does AI make the job of building better AI go faster? Right now that work is done by human researchers using AI as a helper. But if that changes — if the models take over a real chunk of the research itself — then progress stops being a straight line and starts to compound. This is the one thing that makes predicting the long term almost pointless, because anything that compounds quickly outruns whatever map you drew in front of it.

The two questions don't get answered at the same time. Reliability comes first — it's closer, and the signal is clearer. Whether AI improves AI comes later. So if you read the next seven years left to right, they look like a tree: one fork early, two forks late.

The seven years aren't all the same. They come in three parts: one you can mostly see coming, one where the paths pull apart, and one where you finally know which path you're on.

I. Things spread and settle  ·  2026–2028

This part is easy to predict, because it's mostly the spread of things that already exist. Agents get more reliable at narrow, valuable tasks. The thin kind of software — basically a nice screen on top of a database — gets swallowed as its job becomes a feature of something bigger. Companies settle on owning their data and swapping the model beneath it. And quietly, the thing that limits progress stops being how good the models are and starts being how much power and how many chips you can get.

II. The paths pull apart  ·  2029–2031

Now the fork gets visible. Whichever way reliability goes, and then whichever way the AI-improves-AI question goes, the four worlds start to separate in ways you can actually measure instead of just argue about. The job effects show up first in hiring numbers for entry-level work, not in unemployment headlines. This is the stretch where the map starts to pay off, because the signs you were watching finally point somewhere.

III. You know where you are  ·  2032–2033

By now you can tell which world you're in. The reliability question has an answer. The AI-improving-AI question at least has a direction. The bets you'd been hedging across all four scenarios can finally be placed. What used to be foresight is just the situation you're in.

Each of the four is a world that makes sense on its own — its own way of working, its own winners, its own story about jobs, and its own tell. I've put them in order from the slowest to the most disruptive.

A few things stay true no matter which branch you end up on. Think of them as the load-bearing walls — they hold up whichever room you're standing in.

Compute and power end up in few hands

Training the biggest models, and getting the electricity to run them, is the one real moat — and it belongs to a handful of players almost by physics, not just strategy. This is the strongest reason to worry about everything getting concentrated. It's not that one company out-designs everyone else. It's that the raw materials of frontier AI naturally sit in few hands. In every one of the four worlds, the power grid and the price of a gigawatt matter more than any clever product decision.

Open models keep a floor under everyone

Once a model's weights are out, you can't take them back. Open models stay maybe a generation behind the best ones, but they're good enough for a huge share of what people actually need — and just by existing they drag the price of the commodity parts toward zero and keep buyers from getting locked to one vendor. This is the thing pushing against a winner-take-all ending, and the reason there's a real fight in the middle at all.

The value moves from the app to the data

As the models turn into a commodity, the thin app layer stops being defensible and the value drains toward the parts that don't: your own data, the retrieval that respects who's allowed to see what, the orchestration, the evaluation, and the people accountable for all of it. “Own your data, rent the model” isn't a slogan. It's just where the money ends up.

You won't figure out which world is coming from product launches or benchmark records. You'll figure it out from a short list of signs, most of them kind of boring. These are the ones worth watching.

Here's the whole thing in a sentence. The next three years are mostly things you can see coming — stuff spreading, companies merging. Years five through seven split on two questions: whether agents get trustworthy enough to act, and whether AI starts improving AI. And it's that second one that makes predicting the long term almost hopeless.

My best guess is the Agentic Buildout: a big, fast change you can still make sense of, where the value slides away from thin software toward data, orchestration, and human accountability, and where jobs change task by task well before they change wholesale. The two extremes bound the range — a Long Diffusion that just takes decades, and a Compression that squeezes it all into a few years. Which one you're heading for gets written down early and quietly, in reliability curves and in how close open models stay to the best ones, long before it ever reaches the news.

So plan for the middle, keep some room for the extremes, and watch the signs, not the headlines.