Progress Is a Stack. And Europe Is Missing Most of It.

The cultural layer: ambition before policy

Grant Mulligan, writing about his fellowship at the Roots of Progress Institute, names something that policy documents never do: progress begins as a shift in self-conception. Before you build anything, you have to believe you're allowed to. Before you declare an idea, you have to become legible to yourself.

I recognize this because I've watched it happen in real time, hundreds of times.

In 2019, a woman named Lucía showed up to a Saturdays.AI cohort in Málaga. She was a social worker. Mid-forties. She told me on the first Saturday that she'd registered by accident — she thought it was a digital literacy workshop. She'd never written a line of code. She didn't own a laptop; she borrowed her daughter's.

Twelve weeks later, Lucía presented an NLP model that classified domestic violence reports by urgency level, trained on anonymized case data from her own department. It wasn't state-of-the-art. The accuracy was modest. But the room went silent when she explained it, because everyone understood: this wasn't a technology demo. It was a woman who had reclassified herself. She walked in as someone who used tools. She walked out as someone who built them.

That transformation — not the model, not the accuracy score — is what Grant is pointing at. Europe's AI conversation almost never starts here. It starts with regulation, or with compute, or with sovereignty. But regulation without ambition produces compliance. Compute without builders produces infrastructure for someone else's models.

The industrial layer: sovereignty is necessary but not sufficient

Mistral's European AI playbook is, on paper, impressive. Talent visas, university-industry links, applied AI institutes, student compute access, procurement reform, SME adoption, energy-linked infrastructure. It reads like someone finally took the European bottleneck seriously.

But here's my problem with it: sovereignty arguments tend to assume that once you control the stack, adoption follows. It doesn't. I've watched this from the inside, and the gap between "we have the model" and "someone actually uses it" is where most European AI ambition goes to die.

Let me give you a concrete example. At PresidencIA, we built CiudadanIA — a system to summarize citizen letters to the Prime Minister of Spain using a locally hosted LLM. Sovereignty mattered. We deployed an open-weight model on-premise precisely because these letters contain personal data, sensitive complaints, real people writing about real pain. A US-hosted API was not an option.

The model worked. Linguistically, it was solid. And then, in the third week of testing, it invented a ministry.

A citizen had written about a problem with their pension. The model summarized the letter accurately — tone, content, urgency, all correct — and then attributed the issue to the Ministerio de Coordinación Institucional y Transparencia. A ministry that does not exist. The name sounded plausible. It had the right bureaucratic cadence. A junior analyst might have let it pass. But the civil servant reviewing the summaries caught it. She circled it in red pen, walked to my desk, and said: "If this tool invents institutions, how do I trust it with anything?"

She was right. And the question she was asking wasn't about hallucination rates or temperature settings. It was about institutional legibility. The model understood Spanish. It did not understand Spain — the map of ministries, the chain of competencies, the difference between what sounds like a government body and what actually is one. We spent the next two months rewriting system prompts, building entity guardrails, and — most importantly — redesigning the workflow so that she and her colleagues weren't asked to trust the model. They were asked to verify it. That distinction changed everything.

But here's the part Mistral's playbook doesn't cover: the procurement story. Even buying eight AI licenses for the cabinet was a project. In a private company, that's a credit card and ten minutes. In the Spanish government, it required a formal justification document explaining what a large language model is, why it differs from enterprise search, why the licenses already in place were insufficient, and how we'd ensure compliance with the Esquema Nacional de Seguridad. The document was eleven pages long. The approval took six weeks. Six weeks to buy a tool that costs less than the office coffee budget.

Mistral is right that Europe needs its own infrastructure. But infrastructure without institutional capacity is a highway with no on-ramps. The missing piece isn't compute or models. It's the ability of actual organizations — ministries, city halls, schools, hospitals — to evaluate, procure, deploy, and govern AI systems. That's not a technology problem. It's an organizational design problem, and it's where most European AI strategies go silent.

The social contract layer: who benefits?

OpenAI's industrial policy paper — "Industrial Policy for the Intelligence Age" — is ambitious and comprehensive. Worker voice in deployment. A "right to AI." Public wealth mechanisms. Portable benefits. Trust stacks. I don't doubt the sincerity. But I question the framing.

This is, at its core, a lobbying document. Not because everything in it is wrong — much of it is directionally right — but because it positions OpenAI as the architect of the social contract for a technology it dominates. The paper proposes auditing regimes and incident reporting, but does so from the perspective of a company that decides what gets audited. The section on democratic input is thoughtful, but written by the entity that would be subject to that input. These aren't contradictions that invalidate the paper. They're tensions that anyone reading it should name openly.

What I take from it is more structural: the intelligence age needs an explicit social contract, and that contract cannot be designed by labs alone.

I learned this the hard way during a Saturdays.AI cohort in Bogotá. A team of four fellows — two engineers, a journalist, and a policy analyst — built a prototype for a local government: an AI tool that would help a municipal social services office prioritize housing applications. Technically, it was one of the strongest projects in the cohort. Clean data pipeline, interpretable model, well-documented code.

The municipal office rejected it.

Not because it didn't work. Because they'd never been asked what problem they were trying to solve. The team had assumed the bottleneck was prioritization — too many applications, too few case workers, let the model rank urgency. But when the social workers finally sat down with the team (after the prototype was already built), they explained: the bottleneck was trust. Applicants didn't believe the process was fair. What they needed wasn't faster sorting. They needed a system that could explain, in plain language, why one application moved forward and another didn't. The team had solved the wrong problem — brilliantly.

That failure taught me more about AI deployment than any policy paper. The social contract isn't about what AI can do. It's about who defines the problem, who participates in the design, and who can challenge the output. OpenAI is right that we need shared upside and democratic input. But those phrases remain slogans until someone sits in a room in Bogotá and listens to a social worker explain why a technically correct system is socially illegitimate.

The legitimacy layer: what do people actually want?

Anthropic recently launched the Anthropic Institute, led by co-founder Jack Clark, to study the societal consequences of frontier AI from inside the lab. It brings together red-teaming, economic research, and societal impact work into a single unit that promises to report candidly on what they find.

This matters because it formalizes something most AI organizations do informally or not at all: studying how AI systems behave in the real world, not just in benchmarks. Their previous 81,000-user study across 159 countries showed something obvious but underappreciated — people simultaneously feel hope and alarm about AI, and their concerns are concrete, local, and specific.

This resonates with me because of what CiudadanIA revealed about Spain — not as a technology output, but as a mirror.

When we started processing citizen letters, the policy team expected the usual: complaints about healthcare, pensions, housing. The macro issues. And those were there. But the letters that stayed with me were different. Loneliness described not as a feeling but as an infrastructure problem: no bus, no pharmacy, no connection fast enough for a video call with family. People asking, in careful formal Spanish, whether public programs were meant for people like them. Small business owners wondering whether AI would make them obsolete.

These letters didn't map to any policy dashboard. They didn't fit into the categories the model was trained to sort. They were specific, local, and full of a kind of intelligence that no benchmark captures — an understanding of what life actually feels like in a particular place, at a particular moment, for a particular person.

"Translation localizes language. Representation localizes intelligence. And you can't build that representation from inside a lab in San Francisco, no matter how many countries your survey covers."

I'll add a challenge to Anthropic's approach: studying consequences from inside the lab has an inherent limitation. You learn what your models do to users. You don't learn what happens when people lack access, when they distrust the system, when their government can't procure it, or when their language is well-represented in the training data but their institutional reality isn't. The legitimacy layer needs voices from outside the lab as much as data from inside it.

Where Europe actually stands

If you map Europe's readiness against the four layers of the stack, the imbalance becomes stark. We've invested heavily in the top — regulation, ethics frameworks, governance structures — while leaving the foundation dangerously thin.

The investment picture tells a similar story. Europe's private AI investment is a fraction of what the US deploys — and that gap compounds across every layer of the stack.

What I actually believe

Here's the thesis I've arrived at, not from reading these four texts but from the friction of doing this work — from watching a reviewer circle a fake ministry in red pen, from watching Lucía reclassify herself, from watching a brilliant prototype get rejected by the people it was meant to serve, from reading letters from citizens who weren't worried about AGI but about whether there was a bus.

That means talent pipelines that don't just train engineers but activate builders from excluded communities — Lucía in Málaga, the fellows in Bogotá, the 30,000 people who've come through Saturdays.AI across ten countries. That means procurement systems that can evaluate AI tools without an eleven-page justification for an eight-seat license. That means verification infrastructure — evaluation, auditing, incident reporting — treated as core public capability, not an afterthought. That means local alignment: not just translating models into Spanish, but building systems that understand what a citizen actually needs when they write to their government.

Progress, it turns out, is a stack. Culture at the bottom — the belief that you're allowed to build. Infrastructure above it — the compute, the models, the sovereign capacity. Social contracts above that — the agreements about who benefits and who decides. And legitimacy at the top — the lived experience of people who encounter AI not as a benchmark but as a bus schedule, a pension letter, a question about whether a course is "for people like me."

Each layer depends on the one below. None of them is sufficient on its own.