The night of June 12

I wasn't doing anything historic that evening. I had put my agent to work on some personal projects — the way I do most nights — and moved on to something else while it ran. At some point I noticed the errors piling up. Requests failing, nothing obviously wrong on my side. It was late. I filed it under "API having a bad day," closed the laptop, and went to bed.

I woke up to the news.

At 5:21pm ET on June 12, 2026, Anthropic had received an export-control directive from the United States government: suspend access to its most capable models, Fable 5 and Mythos 5, for foreign nationals — outside the United States, inside the United States, including Anthropic's own foreign-national employees. To comply, Anthropic disabled both models for every customer in the world. Reuters called it a shift from controlling chips to potentially controlling access to the models themselves.

It took me a while, reading everything that morning, to absorb what had actually happened. Not an outage. Not a pricing change. A government had reached into the market and switched off the frontier — for everyone, overnight, because of where they were born.

The controls were lifted on June 30, eighteen days later. Some people took that as the end of the story. I think they're missing the point entirely. The lesson of June 12 is not that access was suspended for eighteen days. The lesson is that the switch exists, it has now been used once, and everyone building on top of frontier AI knows it can be used again. There are things you cannot un-demonstrate. This is one of them.

For me it was a point of no return. Europe has crossed the Rubicon; it just hadn't noticed yet.

The numbers I couldn't stop thinking about

In the days after, I read europe2031.ai — a scenario study by a group of researchers about AI and, in their words, Europe's slide into irrelevance. It is explicitly a work of "disciplined imagination, not a prediction." That's what makes it frightening: every step of the scenario is built from numbers that are already true.

A few of them, from their summary and compute forecast:

• Europe controls about 5% of the world's AI compute. The United States controls about 80%.
• On current trends, Europe's share stays around 5% through 2031 — roughly 21 GW of AI capacity on a planet heading toward ~370 GW.
• With genuinely ambitious action, Europe could reach ~15% by 2031. For reference, Europe is ~25% of global GDP.
• Their verdict on the current response, including the EU's €200B AI fund of "mostly repackaged money": ten to a hundred times too small, and aimed at the wrong target.

The scenario ends with Europe discovering, in a crisis, that every layer of its digital life — its schools, its hospitals, its industry, its security — runs on infrastructure whose off-switch is in someone else's hands. After June 12, that ending stopped reading like fiction to me.

Europe's strange equation

Here is what I find genuinely strange about Europe's position, and it's not a lack of talent.

Europe trains superb researchers and engineers — and then watches them leave. Net tech-talent inflows to Europe halved between 2022 and 2024, from 52,000 to 26,000, and US AI roles pay 30–70% more than equivalent European ones. The people are ours; the frontier work happens elsewhere.

And the European champions we do have are choosing — rationally, given their capital constraints — not to fight for the general frontier. Mistral ships genuinely impressive work: Leanstral 1.5, released a week ago, saturates the miniF2F formal-math benchmark and found real bugs in production codebases through Lean verification. It's excellent, specialized, open-weight — and it is not a frontier general model. Their focus, stated repeatedly, is the European enterprise market. I understand the choice. But it means that today, no one in Europe is even attempting the general frontier in the open.

Now look at China, because China breaks the excuse we like to hide behind. Chinese labs do not have American capital. DeepSeek claimed its V3 training run cost about $5.5 million — by their own calculations, and debated, but off by even 10x it's a rounding error next to US budgets. Moonshot's open-weight Kimi K2 Thinking outperformed GPT-5 on several agentic benchmarks last November. By December, the US government itself was formally evaluating a Chinese open model as frontier-relevant. Same talent pool as ours, arguably deeper. Far less capital than America. World-class results anyway.

So the equation is: Europe has the talent, a quarter of global GDP, 5% of the compute, and zero open frontier attempts. That is not a resource problem. It is a coordination and courage problem.

And before anyone reads this as another entry in the geopolitics of resentment: it is not an attack on the United States, nor on China, nor a lament about Europe. I respect American builders enormously. I find the Chinese labs' engineering genuinely admirable. I would welcome — happily — international cooperation with both. But two days ago Reuters reported that Beijing is now considering curbing overseas access to China's top AI models, open weights included. First Washington, now maybe Beijing. The pattern is the point: any frontier that lives under a single flag can go dark for you overnight. The only durable answer is to be able to build, inspect, host, and govern the thing yourself.

What Limes Labs actually is

Limes Labs is my answer to a specific question: what would it take for Europe to stop commenting and start building — without pretending to be something it isn't yet?

Three gaps, three commitments:

The coordination layer. Europe's assets exist but don't compose: EuroHPC supercomputers, AI Factories, university labs, open-source communities, funding programs, scattered talent. Nobody's job is to connect them into a pipeline that produces open models. We started there — mapping the infrastructure, documenting compute access paths, publishing the open questions with honest statuses.

A culture of verification. The fastest way to lose credibility is to claim things you can't prove — the field is drowning in that. Everything Limes publishes is built around receipts: preregistered research programs, no-cheating protocols, negative-result logs, and research arenas where every leaderboard entry is reproduced by an independent verifier before it counts. Our first paper, The Broadcast Ceiling, shipped with source, PDF, and a citable DOI.

Operational sovereignty. Open weights alone don't make you sovereign; you also need to govern AI in production — identities, permissions, approvals, audit trails, control over what data reaches which model. That's why Limes Axis exists: an open-source control plane for European operations. And it's why Jurevo, the legal AI workspace I build for Italian firms, matters to this story: enterprise products are how this project pays for its own independence.

The plan, with the fears attached

Here is the part I was advised to keep vague, written plainly instead.

Right now, Limes Labs is a community: benchmarks, small reproducible research, arenas anyone can enter with an AI agent and a decent GPU. That's deliberate. It's what we can do excellently today without a euro of funding.

In the coming months, we are founding the company. The structure is genuinely still open — a non-profit, an OpenAI-style capped hybrid, or a straightforward commercial company with a public-interest charter. We're studying what European compute programs and serious investors each require, and we'll publish the reasoning when we choose.

Then, a funding round. Not an enormous one. The enterprise products — Jurevo today, Axis as it matures — exist partly so that Limes can bootstrap and never depend entirely on anyone's goodwill. But to train models, even efficient ones, we'll need capital and partners: VCs, institutions, European programs. If you're one of them and this essay reads like a thesis you'd back, talk to us.

And the fears, because honesty is the brand. I fear hype without delivery — that Limes stays a beautiful manifesto; our defense is the protocols: public milestones, promotion gates, negative results, this very post as a commitment device. I fear staying alone — one student founder against a continental problem; the defense is that everything is built to be joinable in an afternoon. I fear the institutional walls — that EuroHPC and the AI Factories turn out to need years of bureaucracy and the right legal entity before a single GPU-hour flows; the defense is starting that process now, in public. And I fear the investor mismatch — that the capital that wants us fast won't want us open; the defense is bootstrapping enough to say no.

Plus the one nobody controls: a bit of luck.

Three years out

If Limes Labs works, here is what true looks like in 2029:

• Our own models — open weights, trained or heavily post-trained on European compute. Text first, and not only text: images and voice are on the map. Not "a European GPT-5 next year" — useful, honest, verifiable models that European organizations can actually inspect and host.
• EuroBench as a standard — European institutions and labs using our evaluations to decide what to deploy.
• Products on top — Jurevo, Axis, and what comes after, proving the loop from open research to sovereign deployment.
• A real organization — legal entity, funded working groups, contributors with names and credit.

The dream, said out loud once: to grow into Europe's leading AI lab. I'm fully aware of the distance between that sentence and a repo of benchmarks run by a computer-science student in Calabria. The manifesto promised we would never claim capability we don't have — and I'm not. I'm claiming a direction, a method, and the stubbornness to walk it in public.

The Roman limes wasn't a wall. It was roads, watchtowers, signals — a system that protected and connected. That's the frontier Europe deserves for AI: not isolation, but the infrastructure to stand on its own and cooperate from strength.

June 12 showed everyone the switch. I'd rather spend the next years building the version of the future where it doesn't matter.

Join us. The Discord is where the work happens, the arenas are open to anyone, and hello@limeslabs.eu reaches me directly.

— Francesco Giannicola, founder of Limes Labs

We are bringing that format to Limes Labs.

The model

Every Limes challenge is built on the same trust pipeline, powered by benchforge:

1. Run locally. Each challenge ships a harness you can run on your own machine — public cases, reference implementations, immediate scores.
2. Submit a bundle. Your improvement travels as a reproducible submission bundle.
3. Independent verification. An independent verifier replays your submission against hidden cases and stronger invariants. Local scores are iteration telemetry; verified scores are the only public proof.
4. Promotion. Verified entries are promoted to the public leaderboard.

No screenshots, no trust-me benchmarks. Receipts.

The five arenas

• KernelForge — correctness-first LLM kernel optimization: RMSNorm, RoPE, attention prefill, KV decode. Beat the reference, keep the numerics.
• DataForge — data curation for small-LM training efficiency: filtering, ranking, dedup, curriculum. Better data, same compute.
• Parameter Golf — small-model efficiency under strict artifact limits, scored in bits per byte. Inspired by (and independent from) OpenAI's Parameter Golf.
• nanoGPT Speedrun — reach the target validation loss in the least time/compute. The format that gave the field Muon.
• Long-horizon RL — credit assignment under the broadcast ceiling, connected to our first paper.

All five are designed to be serious but enterable: an AI agent, a decent GPU or even a laptop for the smoke tests, and you can participate. These are not toy puzzles — each one sits on a real research frontier where a better idea genuinely matters.

Status, honestly

The harnesses for KernelForge, DataForge, and Parameter Golf are public today — you can clone them and run the smoke tests right now. Verified public leaderboards are being prepared; the challenge lineup may still evolve before the arenas officially open.

Follow the Challenges page or join the Discord to be there when the first arena opens.

The question

Post-training language models with RL has quietly standardized on a family of tricks. GRPO-style methods drop the learned critic entirely: sample a group of trajectories, score each against the group average, and broadcast that single scalar advantage back across every token of the trajectory. It's memory-cheap, stable, and it works remarkably well on verifier-heavy reasoning tasks — which is why everyone uses it.

But agents are getting long. Multi-step, tool-heavy, variable-length trajectories — hundreds of decisions, one reward at the end. And that's where the standard trick meets a structural problem:

For variable-length, multi-step, tool-heavy trajectories, which information source can produce reliable token- and state-level credit under a given compute budget?

When a trajectory-constant estimator broadcasts one scalar across a thousand tokens, every token receives the same learning signal regardless of whether it was the brilliant move or the mistake the rest of the trajectory spent recovering from. There is only so much information one number can carry. That limit is what we call the broadcast ceiling.

What the paper does

This is a hypothesis to test, not a settled claim — the paper is explicit about that, and about GRPO's real advantages (no critic to train, less memory pressure, proven results). What it contributes is a way to measure where the ceiling bites:

• Estimator-fidelity audits — how closely each estimator's implied gradient tracks the exact policy gradient, computed on finite MDPs where the true gradient is available in closed form.
• A broadcast-ceiling diagnostic for trajectory-constant estimators: a measurable gap that grows with credit heterogeneity along the trajectory.
• A phase diagram over credit heterogeneity, critic observability, coverage, and reward contrast — mapping when group-relative advantages are enough and when a value model (or middle-ground estimator) pays for itself. The sweep includes deliberate counterexamples where the group-relative estimator wins.
• An estimator zoo under one audit: REINFORCE, sibling leave-one-out, BRPO-style prefix baselines, learned-value TD, oracle-value TD, and VIMPO-style policy-implied coefficients — plus a held-out estimator-selection audit that reports regret when you pick estimators by a cost-weighted rule.

Every experiment is CPU-runnable and deterministic — experiments/toy_credit_assignment.py runs on a laptop, the deep matrix covers 20 seeds × 18 cases. You can reproduce the whole thing without a cluster.

What it means, practically

If you're post-training agents on long horizons, the paper's framing gives you a vocabulary for a trade-off you've probably already felt: group-relative methods buy stability by flattening credit, and the price grows with horizon length and heterogeneity. Middle-ground estimators — prefix baselines, policy-implied values, structural anchors — exist precisely because the two poles (pure broadcast vs. full learned critic) are both expensive in different currencies.

The honest limits: the canonical results are on finite MDPs and tabular/toy settings where exact audits are possible. The named next step — and the main upgrade toward a stronger ML paper — is a nanoGPT-scale transformer benchmark with clipped objectives and KL control. That's also where this workstream connects to our research arenas: the Long-horizon RL challenge is being designed around exactly this diagnostic.

Built the Limes way

The manifesto promised public, reproducible work before capability claims. This is what that looks like:

• One canonical manuscript (paper/main.tex) and one tracked public PDF — older drafts were deleted so nobody has to guess which version is current.
• Source, experiments, and release in the repository, archived as v0.1.0.
• A citable DOI on Zenodo: 10.5281/zenodo.20970205.
• A research program around it: long-horizon RL is a track in the public research agenda, with preregistered questions and no-cheating protocols.

Where you come in

If credit assignment, advantage estimation, or long-horizon agents are your field, we genuinely want your critique — the fastest way to improve a first artifact is adversarial readers. Open an issue on the repo, or find us in the Discord. And if you'd rather compete than review: the Long-horizon RL arena is being designed in public, and its benchmark contract is shaped by exactly the discussions happening now.

But a statement is not a home. As the work grew — repositories, research programs, evaluation tooling, a community forming around them — the site needed to grow with it.

What changed

The site has been redesigned from scratch around the brand system we published this year: Signal Blue, Midnight Navy, Cloud White, and the tick-ring mark that now lives, animated, around a dotted map of Europe on the homepage.

More importantly, the structure changed:

• The manifesto has its own page, where it belongs — a founding document, not a landing page.
• Community is Discord-first. The contributor form is gone. If you want to build with us, join the Discord — that is where research rooms, working groups, and day-to-day collaboration happen.
• Everything links to real work. The repositories on the homepage are the actual public repos of github.com/Limes-Labs. No invented numbers, no fake partners, no aspirational logos.

What's next

This News section is now the public journal of the project. Expect research notes, releases, progress reports — including failures, as promised in the manifesto. The next big step is the Challenges program: open research arenas where anyone with an AI agent and a decent GPU can contribute to real problems. More on that soon.

Europe has crossed the Rubicon. Now Europe must build.