AskGib Pictures presents
WhenRivalsAgree
The frontier labs have started saying the same thing about slowing down. Companies that cannot stand each other do not converge by accident. History has a word for what it looks like, and it has never ended well for the public.
Written by Gibril BachouchiSeptember 14, 2026Running time 8 min
When Rivals Agree
Something odd is happening at the top of the AI industry.
The frontier labs have started talking about slowing down. Longer gaps between releases. More testing before a model ships. Less race, more pause. And, in the same breath, a request that governments get on the same page.
These are companies that poach each other's researchers, sue each other, and leak against each other in the press. They agree on almost nothing. And now, within a few months of each other, they have all landed on the same idea.
That is the part I find interesting. Not the slowdown. The agreement.
Rivals do not converge by accident
When people who dislike each other suddenly say the same thing, there are usually only two explanations.
The first is that they know something the rest of us do not.
The second is that their interests have lined up. Nobody coordinated. The ground shifted under all of them at once, and the same move became the best move for each of them. Even enemies can read a board. When one direction is clearly the way to keep gaining, they all walk that way, and they do not need to like each other to do it.
Most of the time it is a bit of both. So here are the two stories I can tell, and why I think both are true.
Story one: the plateau
The frontier is hitting diminishing returns.
The labs are still improving their models. Just not at the pace the market has priced in. Training runs cost billions. Investors expect each generation to be a leap over the last, and the leaps are getting harder to show. A model that scores a few points higher on a benchmark does not justify a valuation that doubled.
In that world, "we are slowing down for safety" is a very useful sentence. It turns a technical ceiling into a moral choice. It protects the valuation, preserves cash, and keeps the company away from public markets until it knows whether the curve is still a curve.
If this is what is happening, the slowdown is not a decision. It is a description, dressed up as a decision.
Story two: the drawer
The opposite. The labs have something significantly better sitting internally, and they do not want to release it.
Better at autonomous reasoning. Better at coding. Better at improving itself, or at coordinating with copies of itself. Release that openly and the open-weight ecosystem and the Chinese labs study it, reproduce it, and ship it for a tenth of the price within a year. I wrote a few weeks ago about how fast that loop now runs. The gap between the closed frontier and the open one is measured in months.
So the incentive is to keep the best thing in the drawer. Use it internally. Let it write the next model. Sell access to the layer below it. And say, in public, that the world is not ready.
Dario Amodei's recent essay leans this way. He names recursive self-improvement and swarms of autonomous agents as the risks he is watching. That does not mean anyone has superintelligence in a basement. It might only mean they have seen enough strange behaviour in testing to notice that their safety work is not keeping up with their capability work. But it tells you where their attention is, and it is not on the plateau.
Both are true
These two stories are not in competition.
The public frontier can plateau while the private one keeps moving. The model you can buy today improves slowly. The internal system that trains it improves fast. The labs are not lying when they say they are slowing down. They are just not saying which frontier.
Put those together and the alignment needs no conspiracy. Every lab faces the same cost curve, the same copy risk, the same investors. The same move became the best move for all of them. People who would never share a stage end up reading from the same script, because the script is simply what the board rewards.
The pivot
There is a third thing the slowdown hides, and I think it is the biggest.
Selling the model does not work. Not any more. The Chinese labs sell something close for a tenth of the price and hand out the weights. You cannot charge frontier prices for a product that has an open-weight twin a few months behind it. The labs know this. Their investors know this. So the money has to come from somewhere else.
And they know exactly where. For three years, every company on earth has been feeding them its work. Drug discovery pipelines. Factory schedules. Legal briefs. Supply chains. Robot telemetry. Sales calls. Code. Maybe not the raw data, the terms say so. But the shape of the work: which questions get asked, which answers get accepted, where the bottleneck is, what a good result looks like. That is the most valuable training set ever assembled, and the customers assembled it for free.
So the next move is obvious. Stop selling intelligence by the token and start selling the thing the token was used for. A lab that has watched ten thousand biotech teams work does not need to sell to biotech. It can be biotech. Same for manufacturing, robotics, law, finance, logistics. Each one is a market larger than the API business, and each one is defended by knowledge the lab already has.
That is what "growing internally" looks like from the outside. Fewer public releases. More vertical products. Partnerships that turn into acquisitions that turn into competitors. The customers of 2024 become the targets of 2027.
Nobody slows down for no reason. You slow down before a turn.
And a group of rivals who all take the same turn at the same time, and then ask the government to steer with them? History has a word for that. It is not "safety".
History has a word for this
I am not saying anyone signed anything. I am saying the shape is familiar, and every time it has appeared, it has been sold with the same three words: quality, safety, order.
1557, printing. England gives the Stationers' Company a royal charter. The printers get a monopoly. The Crown gets a guild that polices every press in the country and reports the unlicensed ones. Both sides call it order. The public gets licensed books for 138 years, until the licensing law lapses in 1695. "The world is not ready" is the oldest sentence in the book.
1815, Europe. Russia, Austria and Prussia, who had spent a generation at war, form the Holy Alliance. In 1819 the Carlsbad Decrees put censors in the universities and the newspapers. Rival empires agree on one thing: nobody below them should be able to organise. The people who disagree are exiled, jailed or pushed to the side. It holds until 1848, when the whole continent goes up at once.
1892, railroads. The railway barons have just been handed a regulator, the Interstate Commerce Commission. Richard Olney, a railroad lawyer about to become Attorney General, writes to a railroad president that the commission "satisfies the popular clamor" for supervision while that supervision stays "almost entirely nominal". His advice: do not destroy the commission. Use it. That letter is the founding document of every industry that has ever asked to be regulated by rules it helped write.
1924, light bulbs. Osram, Philips and General Electric, rivals across three continents, form the Phoebus cartel. They agree that no bulb should last more than a thousand hours, and they fine each other when one does. It is sold as standardisation. Bulbs had lasted 2,500 hours before. Consumers pay for the difference for fifteen years, until the war breaks the cartel.
1954, tobacco. The big cigarette makers publish a joint "Frank Statement" and fund a shared research committee. Rivals, together, promise to study the health question seriously. Forty years of "more research is needed" follow. In 1994 seven chief executives sit in a row before Congress and swear, one after another, that nicotine is not addictive. Shared safety research was the delay.
1969, cars. The US Justice Department sues General Motors, Ford, Chrysler and their trade association for agreeing, since 1953, to slow the development and rollout of pollution controls. It is settled by consent decree, no admission, no trial. Three rivals, one pace, sixteen years of smog.
Every case has the same five parts. Rivals agree. They ask the state to agree with them. The reason given is safety, quality or order. The bill goes to the public. And the people who disagree are treated as the problem.
It always ends. It always ends late, and the public always pays for the gap.
An executive at Archer Daniels Midland, caught on tape fixing the price of lysine in the 1990s, said the quiet part out loud: the competitor is our friend, the customer is our enemy. I do not think anyone at a frontier lab would say that. I also do not think it matters what they would say. Cartels are not made of villains. They are made of incentives, and the incentives here are all pointing the same way.
Trust less, own more
So what should a normal person, or a normal company, take from this?
Not that the labs are evil. That they are labs. They will act like every concentrated industry before them, and the correct response is the same one it has always been: reduce your dependence on them.
The good news is that this is the first cartel in history whose product is already leaking out the sides. Open weights exist. Chinese labs ship frontier-adjacent models for a tenth of the price and put the weights on the internet. A model that matches what the frontier sold last year runs on a laptop this year. A cartel can slow down what it sells. It cannot slow down what it does not own.
There is a precedent for this too. In the 1990s, governments and a handful of companies wanted to control strong encryption. Export bans, key escrow, a chip in every phone with a spare key for the state. Then one programmer published his encryption tool as open source, the code got printed in a book, and the argument was over. Today everyone has strong encryption and nobody asks permission. Open weights are the same move, one technology later.
Independence here is concrete. Run a model you can download. Keep your data where you can reach it. Build on a layer you can swap. Treat every closed API as a rented room, not a home.
Why this is still good for builders
I build on top of these models, so I am not a neutral observer. But I think this is good news for everyone who builds, and I would say it even if I did not.
If progress is slowing, the tools built on today's models get to mature. An application, an agent, a harness has time to get good, to find its users, to become a habit before the ground moves again. For three years, every product built on a model got rewritten every six months. A slower cadence is the first time builders get to finish something.
If progress is not slowing, only hidden, then the scarce thing is not intelligence. It is trust. Evaluation, provenance, confidence, control. Ways to know why a system reached an answer and whether to believe it. The more capable the model in the drawer, the more that layer is worth, because nobody will run a system they cannot check.
And if the frontier is behaving like a cartel, then independence is a feature people will pay for. Not as an ideology. As insurance.
What a product should look like now
If I take my own reading seriously, it changes what I would build.
Treat the model as a part, not the product. It will be swapped. Twice. Build so that swapping it is a config change and not a rewrite, and so that the thing users pay for survives the swap.
Keep an open-weight exit. Every feature should have a version that runs on a model you can download, even if it is a bit worse. The day a closed provider changes its terms, its prices or its rules, you should already know what you lose and what you keep.
Do not teach your supplier your business. What you send to a closed API is what it learns. Keep the parts that make you you, your data, your workflows, your evals, on infrastructure you control, and send the generic work outside. Your supplier is a future competitor with a head start you paid for.
Sell the check, not just the answer. Any model can produce an answer. Show where it came from, how sure the system is, and what it looked at. The answer is the commodity. The receipt is the product.
Make the tool a habit before the next wave. A pause at the frontier is a window. Use it to get boring: reliable, fast, integrated into the places people already work. Boring is what lasts.
Assume the best model is not the one you can buy. Design for the day a much stronger system lands with little warning. The products that survive that day are the ones that already know how to verify, limit and explain whatever sits underneath them.
The labs will keep saying they are slowing down. AI is not stopping. Open weights will keep catching up. The frontier will keep moving, in the light or in the dark. When rivals agree, the right response is not to trust them more. It is to need them less.