The Fault Line
On AI acceleration, European committees and the ground shifting beneath our feet
Touching Some Grass
This morning, woken at the crack of dawn by my little daughter, I felt a wild urge to fly to San Francisco. To be in the place where unprecedented things are happening. Then I checked my Linkedin and wanted to cry.
But first things first.
In my head, I imagined getting a coffee from Tartine and strolling Dolores Park afterwards. I would feel close to the epicentre of it all. I would watch dogs chase one another and the occasional ball. I would ask myself whether dogs are to us what we will be to superintelligence. Then I would stop, because this type of thinking defies all rational thought. A dog cannot understand what a human thinks. Its limitations are part of what makes it a dog, just as ours may be part of what makes us human.
Eventually, I would lie down on a lush patch of green and seriously touch some grass, letting the thoughts rest rather than chase each other around yet another corner.
The Cracks
San Francisco is the epicentre of what may be the defining shift of humankind. Over the past few years, those who were looking found the first cracks. Initially so hairline thin and shallow that you would barely notice them, they started to appear everywhere, and yet the surface still looked solid.
But something in the way sound travelled on the pavement had changed. A subtle hollowness beneath each step, some faint resonance where there used to be none. Had earth developed an echo?
Then, the cracks started finding each other, as if they had been searching for one another all along. They connected and branched, covering every stone and ground surface.
The ground shifted for the first time. Just once, and only slightly. The tremor was so brief you immediately question whether it happened at all. Whether it was you, walking too quickly and stumbling over your own feet, or indeed the earth and your body finally registering what your eyes had been trying to tell you.
The cracks widen. You can see depth in them now. Darkness lies beneath. It is then that you realise the ground you are walking on is not really ground, but a ceiling. And somewhere below, something has been pushing up against it for a long time.
You see it now and once you do, you know that at some point not too far from now, everything will happen, all at once.
How ironic that all of this should be happening in the city built over the San Andreas fault, the tectonic boundary between the Pacific Plate and the North American Plate.
In the Bay Area, those who have seen the cracks are responding with the full force of their intelligence and resources. And somewhere else, in that glorious land I call home, committees are forming.
The Committee
The EU is launching yet another expert forum on frontier AI, complete with a fancy public application form. Mind you, this forum is not for just anyone. They are looking for “senior-level participants” after all. The forum will map existing efforts in frontier AI, address challenges, and identify opportunities for Europe’s competitiveness, sovereignty and security.
When somebody sent me this initiative on Linkedin, asking me to apply, it made me want to cry. In moments like these, I feel a sense of desperation in every fibre of my body.
It is March 2026. We are seeing the first recursive self-improvement loops working. The frontier is being shaped by AGI-near models, world models and embodied intelligence. And here comes the EU, calling for senior-level participants to map and identify opportunities…two months down the line. Two months in AI in 2026 is like a whole career at a German automaker.
The Acceleration
The developments of the past few days have been nothing short of astounding. The gap between the frontier and everyone else is widening, and it pains me that the community driving this forward is largely concentrated in the 18,000 square kilometres of California’s Bay Area. Most people outside it have so much catching up to do that entire lifetimes may not suffice anymore.
Recursive self-improvement is the idea that AI can conduct part of its own research, analysing, optimising and upgrading its code, architecture or training processes to produce a better next-generation model.
The most advanced American labs are beginning to turn AI inward, using it not just to build products but to help build the next generation of itself. What looks today like a few thousand researchers and engineers at AI labs may soon function like a workforce many times that size. As these systems take over larger parts of research and engineering, the effective scale of each frontier lab could swell into the tens of thousands, then hundreds of thousands, long before their headcount does.
“This means that soon, the vast majority of frontier AI lab staff will neither sleep nor eat nor use the bathroom.”
Dean W. Ball, On Recursive Self-Improvement
Anthropic’s key developers no longer write code. They are now routinely building Claude with Claude. AI researcher Andrej Karpathy released a repository where an AI agent autonomously iterates on LLM training code in a loop, committing improvements every five minutes, with the human only writing the initial prompt. It works well enough that the agent deserves an entry on the leaderboard for fastest time to GPT-2 performance. It also caused a major AI acceleration psychosis on X.
Shopify Founder Tobi Lutke built on Karpathy’s repo and let his AI agent run overnight. He woke up to find they were able to significantly improve the quality of his model, leading to the casual realisation that “OK, this thing is totally insane.”
Recursive self-improvement loops, in their earliest innings, are staring us blankly in the face.
Last week, my team and I locked ourselves in for two days to play around with Claude Cowork. We pointed Claude at a folder of due diligence material, like financials, product information, and customer calls. It one-shotted a memo that is at least 90% of the sophisticated analysis we do as investors. In many places, it was better than what we produce. In other places, it still got things wrong, forgot data and made incorrect assumptions. But I feel confident with some tweaks we will get Claude to 95%+ investment memo quality over the next few days. What takes experienced investors days and often all-nighters, Claude does in roughly fifteen minutes. The signs of acceleration are everywhere. As AI founder Yuchen Jin said: “AGI is billions of AI agents doing autonomous research together.”
The Openings that Remain
There are still optimistic takes on Europe’s strengths in green tech and pockets of deep tech, and they are not entirely wrong. “Between 2015 and 2025, China’s share of global VC spending fell from 30% to 10%. Europe’s grew from 12% to 16%” (The Economist).
Two things are true simultaneously. There are incredible European entrepreneurs working like crazy every day to advance the frontier. They give me hope and plenty of energy. But there is also what increasingly feels like a black hole beneath our feet, about to turn the many hairline cracks into vast crevices.
So what are our options?
Competing at the forefront of LLMs may take nothing short of a miracle. But miracles happen and open endedness research, which focuses on developing systems that can continually generate new, increasingly complex, and diverse solutions, is still in its early innings. Some very promising European teams are now tackling this challenge.
Then there is the fact that AI models do commoditise over time. DeepSeek R1 matched o1-level reasoning for $5.6 million in training costs. Kimi K2.5, open-sourced in January 2026, outperforms closed frontier models on multiple benchmarks and costs 100 times less to run. Yesterday’s moat will become today’s baseline. I can see a future in which we won’t have to be at the very top of the leaderboard in every category, just close enough to still matter.
On image, video and world models, the race is still more open than on LLMs. Several European companies are competing seriously and I desperately hope more will over the coming months.
Nvidia CEO Jensen Huang has explicitly called out Europe’s dominance in robotics as a key ingredient to future economic leadership.
“You can now fuse your industrial capability, your manufacturing capability, with artificial intelligence, and that brings you into the world of physical AI. Robotics is a once-in-a-generation opportunity for the European nations.” Jensen Huang, World Economic Forum, Davos
German suppliers are already signing contracts to manufacture Tesla Optimus robots and supplying leading robotics companies including Boston Dynamics and Figure AI.
World models are the connective tissue here. They are AI systems that build an internal simulation of physical reality, learning not just to predict the next word, but the next state of the world: how objects move, how actions lead to consequences. Where LLMs learn patterns in language, world models learn the causal structure of physics itself. They provide the missing link between robotic capability and machine intelligence, and they represent a path where Europe can genuinely compete.
Europe Has to Earn the End Game
The road ahead still holds real opportunity. But getting there requires a scale of transformation Europe has not yet reckoned with. We are not facing one transition. We are facing three, layered on top of each other, each running at different speeds.
First, sprint.
In the short term (2026 & 2027), the task is brutal and simple: sprint. Sprint like our lives depend on it. That means working harder and faster, maximising AI adoption across the economy, building competitive companies in hardware, foundation models and applications, and investing aggressively in energy.
I would call this phase “transient AI hypercapitalism.” It means accepting temporarily higher inequality in order to build an AI asset base, and prioritising speed over procedural comfort in regulation and labour rules. None of this feels good. This phase will not feel egalitarian. It will, in that sense, feel deeply un-European. Ownership of AI assets will concentrate, at least temporarily, as the system builds the productive base of the new economy. Most of this is politically toxic. But the alternative is to fall so far behind that whatever toxicity a radical AI strategy brought about will look like cough syrup compared to the treatment we will otherwise need.
Next, reorganise.
The mid term (2027+) is where things get interesting. The sprint cannot last, and it should not. The goal of the hypercapitalist phase is to fund a transformation. Once the infrastructure and leading companies are established, the system begins to reorganise around them. Workers who learn to use AI well become dramatically more productive. Other workers require earlier forms of protection. Companies consolidate around the leaders while weaker firms release capital back into the system. The tax and welfare state gets rebuilt around AI profits rather than wage income. Inequality peaks in this phase, but it is also the moment when new mechanisms for redistribution can be built. Ownership becomes the central question. Pension funds, sovereign funds and public investment vehicles need to hold meaningful stakes in AI infrastructure so that a broader share of society participates in the upside.
Eventually, steward.
The long term end game (2040+) is genuinely new territory. The companies that dominate AI and energy may begin to resemble utilities: stable, cash-generating infrastructure underpinning the entire economy. When energy and compute are abundant and cheap, when most people no longer work primarily for economic survival, the logic of the system can truly change. Ownership structures shift toward public and broad private models.
The focus of politics shifts from redistribution to stewardship: responsibly managing the people in its care. Europe’s economic leverage, regulatory standards, value system and narrative power then become tools to shape global rules. Labour shifts toward care, community, art and politics.
The tricky thing is that we cannot skip to the end game. As investor and writer Will Manidis put it, we have to earn the right to a good end game by playing the opening and middle game well.
Unfortunately, we have already played a range of catastrophic moves in the opening. We played it like someone only half-remembering the rules. Europe may now have to pass through a hypercapitalist phase, losing part of itself in order to earn the chance to find itself again. That is the mindset with which we need to enter the middle game. It is messy and may last far longer than the cracks on the floor suggest. But every move still counts.
Running Towards the Storm
This morning, my eyes still heavy with sleep, I felt like a storm chaser. Watching a violent tornado from afar, mesmerised by its tremendous beauty and lethal force in equal measure. Rather than run from it, as any sane person should, I found myself taking a few steps closer, mouth slightly open, eyes wide.
Then my daughter jumped on me in full force, and even the fiercest tornado would have to wait.



Poetically written and full of painful truths!
Frontier Artificial Intelligence: Capabilities, Risks, and the Future of Artificial Intelligence, https://promptengineer-1.weebly.com/ai-developments/frontier-ai-capabilities-risks-and-the-future-of-artificial-intelligence
Also, Frontier AI: https://promptengineer-1.weebly.com/frontier-ai.html