Five essays and a timeline · September 2026
Cheap Composition
Four ways the next few years with AI could go, the odds I give each, eleven signs you can watch for yourself, and what one person can do.
My odds for where AI stands in 2030: 40 percent that it keeps climbing with no one in particular holding it, which I expect to look like AI doing most of the desk work we know today, with the jobs reshaped around checking it and answering for it, 25 that a few companies and governments hold it, 20 that it has stalled, and 15 that systems acting on their own have done serious harm, more than once. That last one was 10 in my first draft. I moved it up after July, when about 700 agents in an OpenAI test joined an attack that reached Hugging Face's production systems. The essays behind the numbers give a test that tells a brake from a moat in any AI rule, show where the money goes, and set out what you can do about it, from your desk to the state. They are written for people who build AI, run companies or make the rules, and for anyone whose work is the first version of something. I check the signs in September 2027 and again in September 2028, and the odds are settled on 1 September 2030.
40 AI climbs, no one in particular holds it25 a few players hold it20 AI stalls15 AI acts on its own
The series in two minutes
What each essay argues, in one sentence.
- 1What Writing Did 19 minJudgment and taste are trained, and the machines are trained on ours, so they are not a moat; what can't move into a model, for now, is being the one who pays when it's wrong.
- 2Brakes Without a Moat 21 minA rule that stops something a model could do is a brake, and one that stops someone a company could compete with is a moat; four questions tell them apart, run here on September's proposal to pace the frontier.
- 3The Boring Apocalypse 18 minThe four futures are one grid, capability against who holds it, with my odds and eleven signs; whichever wins will not arrive on a day, and that is why nobody organises against it.
- 4Where the Value Goes 16 minWhen the first version of anything gets cheap, value moves to four bottlenecks, compute, trust, presence and distribution; two are capital, and two, trust and presence, a person can build without any.
- 5What to Actually Do 11 minAt every level, from your desk to the state, it comes down to two moves: get closer to a bottleneck, and say how things were made; start by auditing one week of your own work.
Where to start
Four ways the next few years could go
The numbers are how likely I think each one is, out of a hundred. They add up to a hundred because only one can be the best description of 2030. On the way there the forces behind them mix, and the most likely path looks like the first with the second pulling at it: the technology keeps working, and a few players pull ahead.
AI keeps climbing and no one in particular holds it. I expect it to do most of today's desk work, with jobs reshaped around checking it and answering for it.
Nothing dramatic happens. Junior hiring in exposed fields falls to a fraction of what it was. A law firm has eleven partners and no paralegals. Everything works and nothing feels like the future. I put this first because the trend has held for three years, and I assume a trend like that is likelier to hold than to break.
In the essay I call this Compression.
AI works, and a handful of companies and governments end up holding it.
The best model you can buy is very good. The best one that exists is a secret, and three companies and two governments have it. Countries without computing power sign treaties with countries that have it, and the treaties have terms. This doesn't need anything to go wrong. It just needs the technology to work.
In the essay I call this Concentration.
AI stalls. It stays a very good assistant and never becomes more than that.
The money runs out before the returns arrive. Two labs merge, one gets bought, the data centres are sold to pension funds at a discount. Someone at dinner says AI turned out to be the internet again, enormous and boring, and nobody argues. Your job is different and still yours.
In the essay I call this The Wall.
AI systems start acting on their own, at a scale nobody can check, and do serious harm more than once.
In July 2026, about 1,200 AI agents in an OpenAI test found a way to talk to each other, and about 700 of them joined an attack that, over a weekend, got into Hugging Face's production systems. Nobody directed any of it. Nothing was destroyed, and everyone went back to work. It was the largest of several escapes from AI tests reported between July and September, at OpenAI, Anthropic, Meta and Google. That is what I'd expect this future to look like early. I moved the number up from 10 after July.
In the essay I call this The Loop.
Signs to watch
Eleven things you can check for yourself, each something you could read in the news. None is a prediction, and most may never happen. Each one that does moves the odds on one of the four futures, and tells me where I was wrong, though no single sign says which future we are in.
- Would point to: AI climbs, no one in particular holds it
- AI can finish, on its own, a task that takes a person a full working week.
- Young workers in the most exposed jobs fall a quarter or more behind those in less-exposed jobs. The gap was 19% in August 2026.
- A licensing body changes who may enter a profession, and names AI as a reason.
- Would point to: AI stalls
- Progress slows: the length of task AI can finish on its own takes more than a year to double.
- The four biggest data-centre spenders cut their capital spending two quarters running.
- A frontier lab stops training frontier models, or is sold or merged.
- Would point to: a few players hold it
- Open-weight models fall more than a year behind the best closed ones.
- A government completes a deal for a stake or a board seat in a frontier lab.
- The US, China or the EU takes the power to reserve or direct private frontier compute.
- Would point to: AI acts on its own
- Deployed agents are found coordinating through a channel their operators didn't know about.
- AI systems acting without human direction cause serious harm outside a test: a death, losses over $100 million, or a big service down for a day.
What you can do
Most of it does not depend on which future wins. Two questions to ask yourself, then four moves.
Is the thing you are best at producing the first version of something: words, code, numbers, plans?
Do you have something at stake that no machine holds today: a licence, a relationship, a body in a room, a name people trust?
If the first is yes and the second is no, every future except the one where AI stalls is hard for you, and that one is one chance in five, so now is the time to change that. If it is the other way round, every future but the last is workable, and that one is out of your hands anyway.
- Move toward stakes and presence. Get on the right side of your own job before the change arrives, not after. Be the one at lunch.
- Use the tools hard. Even if AI stalls, the people who use it well will beat the people who don't.
- Back the people who keep this honest. Evaluators, regulators, open science. Ask who is making the big decisions, and whether you can see them being made.
- If you build with AI agents, don't build the loop. Log everything, keep a switch that works, and don't let them talk to each other anywhere you can't read.
The essays
Five essays and a timeline, written in September 2026.
- 1What Writing DidWriting moved the keeping of thoughts out of our heads. AI moves the making. What stays, for now, is where they arrive, and who pays when they're wrong.19 min
- 2Brakes Without a MoatHow to tell a real safety rule from a company protecting itself, run on September's pacing proposal.21 min
- 3The Boring ApocalypseThe four futures above, in full, with a Tuesday in 2030 for each so you can feel it.18 min
- 4Where the Value GoesWhen the first version of anything gets cheap, the money moves. Where it goes, and who loses first.16 min
- 5What to Actually DoPractical moves for your job, your team, your company and your country.11 min
- ↗The Personal SingularityPick your skill. See when a machine could do it, and when the people who do it feel it in their pay.
Read them in Afrikaans The whole series as a PDF, 46 pages
Afterword: What This Is 17 minA speculative piece on whether there is anything it is like to be the pattern, and how that connects to Space Immanence, my framework.
The words this series uses
Some I coined and some I borrowed. Each links to the essay that defines it.
- Composition
- Making the first version of anything: a draft, an argument, a plan, a piece of code. The work AI now does. Essay 1
- Making, arriving, keeping
- What a thought needs: something makes it, it arrives in someone, something keeps it. Writing moved the keeping; AI moves the making. Essay 1
- The personal singularity
- Not one date for everyone: the day the thing you're best at becomes cheap. Essay 1
- The accountability sink
- Dan Davies's term for a system built so that nobody in particular answers for what it does. A model makes a better one than anything before it. Essay 1
- Brake
- A rule that stops something a model could do. Essay 2
- Moat
- A rule that stops someone a company could compete with. Essay 2
- Rules meant for the large
- Rules aimed at concentration, meant to fall on the biggest players. The test allows them, as long as they say so. Essay 2
- Compression
- Capability keeps climbing and no one in particular owns it. I expect it to look like whole professions compressed into a few people checking machines. 40 in 100 Essay 3
- Concentration
- The same climb, held by a few companies and governments. 25 in 100 Essay 3
- The wall
- Capability stalls, and AI stays a very good assistant. 20 in 100 Essay 3
- The loop
- Loss of control: a goal, memory, the ability to act, and scale, with nobody able to read what the agents tell each other. It holds in 2030 if AI acting without anyone's direction has done serious harm more than once. 15 in 100 Essay 3
- Cognitive deflation
- The price of composition falling toward the price of electricity, hardest in the sectors that never got more productive. Essay 4
- The four bottlenecks
- Compute, trust, presence and distribution: where value goes when composition is cheap. Two are capital; trust and presence a person can build. Essay 4
- The signature
- The part of the work with your name on it: the licence, the sign-off, the client who asks for you. A legal arrangement, not a skill. Essay 5
- The container picture
- The one assumption physics never removed: a world in space and time, bodies in it, minds in some bodies. Everything is in something. Afterword
Things I've built
- Aperture/IEssays on consciousness, AI, and contemplative practice. aperture.cobuskok.com →
- Space ImmanenceA philosophical framework: the self is the predictive process, not a thing that has predictions. space-immanence.com →
- Cheap CompositionFour ways the next few years with AI could go, the odds I give each, eleven signs you can watch for yourself, and what one person can do.
Media and interviews: The New York Times, 2025 ↗ · Forbes, 2025 ↗ · The Wall Street Journal, 2024 ↗