Two documents, same week

In June 2026, Anthropic put two texts on the table within a few days of each other. One is an economic policy framework describing how the United States should prepare for AI-driven labour-market disruption. The other is a research report on how far the company has come toward recursive self-improvement: AI building AI.

Read separately, both are well written, nuanced and unusually candid for a commercial company. Read side by side, they describe two different futures. The space between them is the interesting part.

The economic document: the staircase

The economic framework is built like a staircase with three steps. The first step is a world with roughly five percent unemployment. There, Anthropic proposes capital accounts that can hold AI equity, retraining grants, occupational-licensing reform and wage insurance. The second step is ten percent unemployment. There, the proposal shifts toward expanded unemployment insurance, sector-specific transition support and relief for basic needs. The third step is called unprecedented unemployment. There the framework says, with unusual frankness, that it is less certain about the right answers and names basic income, sovereign wealth models and equity sharing as possible directions.

Notice the pattern: the level of detail is inversely proportional to the severity. The five-percent scenario gets a toolbox. The ten-percent scenario gets an insurance solution. The scenario that would alter society at the root gets a shrug and a referral to future research.

The framework is accompanied by two concrete commitments: a 200 million dollar research fund and a 150 million dollar fellowship programme. In total, 350 million dollars.

The second document: the curve

The research report on recursive self-improvement contains figures Anthropic had not previously published. More than 80 percent of the code merged into Anthropic’s own codebase is now written by Claude. A typical engineer merges eight times as many lines of code per day as two years ago. The length of tasks models can handle independently is doubling roughly every four months. Two years ago those were tasks measured in minutes; today they are twelve- to sixteen-hour tasks, with full workdays within reach this year. On the most open-ended problems, those without a clear specification, the model’s success rate rose by fifty percentage points in six months. In a recurring internal experiment where the model tries to speed up training code, the result went from a threefold improvement to a fifty-twofold improvement in eleven months. A skilled human researcher needs four to eight hours to reach fourfold.

The report dutifully discusses the possibility that the curve will flatten, that the exponential is really an S-curve. It then treats that as unlikely. Every capability Anthropic can measure has so far followed the same trajectory, and the company has not yet seen the curve bend.

Capital rarely lies

So much for the words. Then there are the actions.

In May, Anthropic signed an agreement with SpaceX for the full capacity of the Colossus 1 data centre: more than 300 megawatts and over 220,000 Nvidia GPUs. According to Axios and SpaceX listing documents, the financial terms amount to 1.25 billion dollars per month through May 2029, or about 15 billion dollars per year, with contractual off-ramps. That comes on top of an already staggering list: up to five gigawatts with Amazon, five gigawatts with Google and Broadcom, 30 billion dollars of Azure capacity and 50 billion dollars in American AI infrastructure. The same month, Anthropic recruited Andrej Karpathy, one of OpenAI’s founders and Tesla’s former head of AI, not for a product role, but to build a team whose job is to use Claude to accelerate pre-training for the next generation of Claude. In other words, exactly the loop the research report describes.

Now put the two sums next to each other. Three hundred and fifty million dollars to soften the social shockwave. Fifteen billion dollars per year, for a single compute agreement, to build the thing causing it. If one counts the whole infrastructure stack, the ratio lands somewhere between one to a hundred and one to five hundred.

This is not hypocrisy in the banal sense. It is something more revealing: revealed preference. A company can phrase its policy documents as cautiously as it likes, but capital allocation shows which future it is actually planning for. You do not buy compute for fifteen billion dollars a year for a technology you expect to flatten. You buy it for an exponential.

And then there is the epistemic asymmetry, almost elegant in its neatness. When the argument concerns safety regulation, where the conclusion is that the state needs the power to stop dangerous models, Anthropic trusts the curve: it has not seen it bend. When the same curve would imply the third labour-market scenario, the tone suddenly changes: it is difficult to predict confidently how and how fast. The same company, the same week, the same exponential. Uncertainty is invoked selectively, depending on which conclusion it is asked to carry.

The view from the floor

I have worked thirty-three years in industry, the last twelve as an independent consultant in manufacturing companies around Europe. I have seen automation waves before: robotisation that took the heavy lifts, ERP systems that took the paper flows, improvement work that took the wasted time. Each wave moved people. The person who stood at the machine yesterday programmed it tomorrow, or moved into maintenance, quality or planning. The toolbox in the five-percent scenario, retraining, licensing reform and wage insurance, is built for exactly that movement: moving labour between occupations, assuming demand appears somewhere else, often one step higher in the value chain.

This wave moves in the other direction. It does not start at the machine but at the screen: in analysis, calculation, reporting, planning and code. These are the tasks people used to be retrained into. I can see it in my own work. A large and growing share of what I do behind a screen can already be done by today’s models, and done faster. That is not a forecast. It is an observation from someone who invoices for precisely that kind of work. The question no transition model answers is simple: when substitution hits from above on the skills ladder, where is the arrow of retraining supposed to point?

To be fair to Anthropic, a policy framework calibrated for thirty percent unemployment would not be read by a single decision-maker. It would be dismissed as science fiction and take the whole document down with it. The five, ten, unprecedented staircase may be deliberate rhetoric: meeting politics where it stands. And the report’s own numbers should be read with scepticism in the other direction. A company in a talent war and a feverish capital market has strong reasons to overstate its acceleration, something the report partly acknowledges when it notes that lines of code overstate real productivity. The framework also contains a sentence remarkable to find in a policy document: the pace of development may render some proposals inadequate before they can be implemented.

But that is exactly the point. They know.

The Swedish question

All of this is written for the United States. What is Sweden’s answer to the third scenario?

On paper, our model is better equipped than the American one: unemployment insurance, transition organisations, collective-security agreements and a labour-market model that has handled structural change for seventy years. But that whole machine is calibrated for business cycles and structural transformation: situations where demand for work disappears here and arises there, and where the task is to bridge the journey between them. Nothing in the system is built for general substitution, where demand for human labour falls broadly and simultaneously, and where there is no there to travel to. Transition organisations cannot retrain away an exponential.

If the American answer to the third scenario is “we don’t know”, Sweden has not yet even asked the question. And the two documents from June 2026 suggest that anyone who waits for the curve to bend before beginning will begin too late.

Sources and checks

The sources below cover the core factual claims about Anthropic's policy framework, recursive self-improvement, compute agreements and the Karpathy hire. The conclusion about capital allocation and Swedish preparedness is the author's.

Rolf Skogling writes AI-skiftet from an industry-near and practical perspective, grounded in work with AI in real organisations.