here is a structural reason the frightening version of AI is the one you keep reading.

A risk story is complete on the day it is written. Something might go wrong, here is who it would hurt, here is what we should do about it. It has a beginning, a middle, and a call to action, and it takes ninety seconds to read.

A capability story is not complete for ten or fifteen years. A method works in a lab. Then it works reliably. Then it works cheaply. Then somebody builds a business on it, and then, eventually, something in your life is different and you cannot name the paper that caused it. There is no day on which that becomes a headline. So it never does.

I don't think this is a conspiracy. It's a shape. Fear fits in a news cycle and progress doesn't.

But if you go looking at what's actually moving, and you stay away from the science fiction — no moon colonies, no uploaded consciousness, nothing that requires a physics we don't have — the next five, ten, and twenty years look substantially better than the feed suggests. And the reasons are almost all in layers nobody writes about, because the layers are boring.

Here are four of them.

In February 2025, the European Centre for Medium-Range Weather Forecasts put a machine learning model into operational service alongside the physics-based system it had spent decades building. Not as an experiment. As the forecast. Its ensemble version went operational that July. DeepMind's GenCast beat the ECMWF ensemble — the global gold standard — on about 97% of the targets it was tested against, and produced better results on extremes. Tropical cyclone track accuracy improved by up to a fifth.

The part that matters isn't the accuracy. It's that these models run in minutes on a handful of GPUs instead of hours on a supercomputer, at something like a thousandth of the energy.

Which changes who gets a forecast.

Weather prediction has been, for its entire history, a capability of rich countries. It required a supercomputer, which required a national meteorological budget. A subsistence farmer in a country without one has been planting on the basis of intuition and memory, in a climate where the memory is no longer reliable.

Skilled forecasting is about to become something you can run without a national budget. Nobody is going to write about the year that happened. It will be one of the largest transfers of practical capability in modern history, and it will arrive as an app that tells a farmer when to plant.

This is the one I'd point at if I could only pick one, and it is the one with the least public attention by a distance.

Enzymes are how biology does chemistry — at body temperature, in water, without the heat and pressure and toxic solvents industrial chemistry requires. Designing a new one from scratch has been, until very recently, close to impossible. The success rates on computationally designed enzymes were embarrassing for over a decade.

That broke in the last couple of years. Generative models are now producing functional, de novo enzymes at rates that make it an engineering discipline rather than a lottery. A method published this year builds the protein structure around the reactive center directly, instead of rummaging through databases for something close enough.

Follow that out. The chemical industry is roughly a tenth of global energy demand, and much of that energy is spent forcing reactions that an enzyme could do at room temperature. Manufacturing a drug becomes something you can do at smaller scale, closer to where it's needed. Compounds we have no way to break down — the ones sitting in soil and water with no exit path — become things that something can eat.

You won't wake up one morning and find all of that has happened. It shows up one reaction at a time, in industries that don't announce things, over about fifteen years. But each one is permanent once it lands.

In 2025, results were published from a phase 2a trial of rentosertib, a drug for idiopathic pulmonary fibrosis. Both the target and the molecule came out of generative AI. Over twelve weeks, patients on the higher dose gained lung function while the placebo group lost it.

Idiopathic pulmonary fibrosis is scarring of the lungs that gets worse over time for reasons nobody has identified — that's what idiopathic means, we don't know why. The scarred tissue stiffens, and breathing gets progressively harder. Most people diagnosed with it live three to five years. The two approved drugs slow the decline. Neither stops it, and nothing reverses it.

So gaining lung function, in a disease defined by losing it, is not a small thing.

Phase 2a is a small, early trial, and it is not approval. Drugs look good at this stage and fail later all the time. But there are now on the order of 170 AI-discovered programs in clinical development, with early-phase success rates running meaningfully above the historical base rate. Some of that is selection effects and some of it is real, and we will not know the split for years.

The reason to pay attention isn't any single molecule. It's that the front of the pipeline stopped being the bottleneck.

Berkeley's A-Lab runs materials synthesis with robots directed by a model. Carnegie Mellon has a cloud lab doing the same thing for alloys. The pattern is the same in both: the model proposes, the robot runs the experiment, the result goes back into the model, and the cycle repeats without a person in the middle.

Take one thing these labs are working on: a battery that stores grid-scale power using sodium instead of lithium.

Sodium is table salt. It is everywhere, it costs almost nothing, and it doesn't require a mine in a politically fragile country or a supply chain three governments are fighting over. The catch has always been that sodium batteries hold less energy for their size, and the fix is a better electrolyte — the chemistry the charge moves through. There are millions of plausible candidates. Testing them one at a time is how the field has spent the last thirty years.

That is exactly the kind of problem this loop eats. Propose ten thousand candidates, have the robot make and test the promising ones overnight, feed the failures back in, run it again tomorrow.

Now picture what a cheap, safe, boring sodium battery does. Solar power stops being a daytime resource. A village that has never been on a grid gets panels and a battery shed and has power at night — not as a development project, as the cheapest available option. A hospital refrigerator holds vaccines through an outage. The intermittency problem that every argument about renewables eventually collapses into stops being the argument.

None of that requires a breakthrough in solar panels. It requires a better electrolyte, which is a chemistry problem, which is the kind of problem that used to take a career and now takes a season.

And the same loop is pointed at membranes that pull salt out of seawater, magnets that work without rare earths, and concrete that doesn't emit carbon when you make it. Not one of those is a moonshot. Every one of them is a search through a very large space of possible materials, which is the thing that just got dramatically faster.

That's why I think the twenty-year picture is better than the five-year one. We aren't only getting better answers. We're getting faster at the process that produces answers, and that speedup applies to every field at once.

Underneath all of it is a price.

An answer that required the most expensive model available in 2023 now comes out of a cheap one. Hold the quality steady and ask what it costs to produce, and the price has fallen something like 98% in two to three years. A dollar of work then is about two cents of work now.

Inside companies that mostly turned into bigger invoices, because organizations always find more to spend it on. That's the corporate version of the story and it's the one that gets written about.

The version I care about is what a collapse like that does for people who were never customers in the first place.

Think about what expertise has actually cost, and who has therefore had it. Not knowledge — knowledge has been free-ish since the encyclopedia and effectively free since Wikipedia. I mean expertise: someone who knows your specific situation and tells you what it means. A doctor who looks at your symptoms and says what to do about them. A lawyer who reads the lease before you sign it and points at clause fourteen. A tutor who explains fractions a fourth time, in your language, without sighing. An accountant who tells you which of these two options costs less.

Every one of those has been rationed the same way, and the rationing has almost nothing to do with how much humanity knows. It's that the knowing was stored inside individual expensive people, who trained for a decade, charged accordingly, and mostly lived in wealthy cities. If you were born somewhere else, or born poor somewhere wealthy, the knowledge existed and simply wasn't available to you. It has been that way for all of recorded history, and we mostly stopped noticing because it seemed like a fact about the world rather than a fact about price.

That's the constraint that's coming apart. Not perfectly, not evenly, and not without real failure modes — bad advice at scale is its own problem and I don't want to pretend otherwise. But the direction is unmistakable, and the size of it is hard to hold in your head. Most of the people who have ever lived have made the biggest decisions of their lives — medical, legal, financial, educational — with no access to anyone who actually knew the answer. Within a decade that stops being the default condition of being human.

That's the one I'd bet most on. It's also the least covered of anything here, and I think the reason is uncomfortable: it doesn't change much for the people who write the coverage. If you already have a doctor you can text, an expert becoming free reads as a minor convenience. If you don't, it's the whole thing.

I'm not saying the disruption is imaginary.

A lot of it lands in the next five years, on people who did nothing wrong and get no vote. Someone who chose a field at nineteen because it was stable, trained for it, was good at it, and did everything the culture told them to do. The fact that the aggregate outcome is positive is not an argument that person needs to hear, and offering it to them is a kind of cruelty dressed up as perspective.

I've watched smaller versions of this from close range. Every enterprise rollout I've been part of has a moment where someone realizes the thing they were the expert in is about to be handled differently. That's a bad afternoon for a person, and multiplying it by millions doesn't make it less bad — it just makes it harder to see any individual instance of it.

So let me separate two things that keep getting answered as one question.

Where this ends up is one question. How people get through the middle of it is a completely different one, and the second is mostly not a technology question at all. It's about whether retraining is real or a press release. Whether the gains show up in wages or only in margins. Whether a fifty-two-year-old gets a genuine second act or gets managed toward the exit. Those are choices, made by people, in rooms that already exist. The technology doesn't decide them and won't.

The doom stories and the boosterism make the same mistake from opposite directions. Both treat the outcome as though it's already determined by the technology — one says the machine ruins us, the other says the machine saves us, and neither leaves room for the part where it depends enormously on what we do. The fear also has a specific cost that's easy to miss: a country convinced this ends badly makes different decisions than one that thinks the destination is worth reaching. Fear doesn't produce caution, mostly. It produces paralysis, or it produces rules written by whoever is loudest in a bad week.

What I'd want someone to take from all this isn't optimism exactly. It's proportion.

The fear is legible and the progress isn't. A risk is a story; an enzyme is a footnote. That asymmetry is doing more to shape how people feel about the next twenty years than any actual look at what's coming, and it's worth knowing that's what's happening to you when you read the feed.

Meanwhile the forecast model went operational and almost nobody noticed. The enzymes are being designed right now. The robot in Berkeley ran experiments last night. In fifteen years somebody will write the history of this period, and the boring parts will be the plot.

— Isaac