I've been turning two things over for a few weeks now. The first is that LLMs are quietly losing their claim to AGI โ not because they've failed, but because the people who built them are starting to say so out loud. The second is that the AI Alliance just launched Project Tapestry with Yann LeCun as Chief Science Advisor, and the framing is unmistakable: open weights are not enough. If you want sovereign AI, you have to shape pretraining, not just download the result.
These two threads converge. This note is about what that convergence means for Univrs, Sepahsalar, and Imaginarium, and what I think we should actually do about it on a 2027โ2028 horizon.
The shift LeCun is calling
LeCun's argument has gotten sharper. Predicting the next token in pixel or text space produces a "blurry average" โ it can never reach the abstract causal understanding that action in the world requires. JEPA inverts this: predict in latent representation space, force the model to learn the semantic structure of what's happening rather than its surface statistics. V-JEPA 2 hit 80% on zero-shot robot control. LeWorldModel is now training stably from raw pixels with two loss terms and ~15M parameters on a single GPU.
The deeper move โ the one that matters more than the architecture โ is the redefinition of good data. It's no longer about volume or label accuracy. It's about whether the data produces a representation space with the right causal geometry. Every assumption baked into Common Crawl pipelines breaks under this frame.
This is the future Tapestry is being built for. The federated-sovereignty story isn't really about today's LLM pretraining. It's about who gets to participate when the substrate becomes world models โ and world models are far more data-hungry in the embodied, multimodal, locally-meaningful sense than text ever was.
What I think Univrs actually has
I want to be honest about the assets, because the strategy depends on this. The three layers we've been building have a coherence I don't think I appreciated until I started writing this down:
- Thesis Univrs.io โ the public-infrastructure / commons argument. The Mazzucato mission-state framing. The Common/cloud project. The "full-stack commons" claim that ties compute, grid, and governance together.
- Research Sepahsalar Labs โ Nexus clusters as compute substrate, VUDO as agentic runtime, DOL as semantic interoperability. Plus the CRDT and distributed-systems lineage that's been quietly the actual moat the whole time.
- Community Imaginarium โ the builder studio. Students and early technologists shipping real products. Not just a logo on a coalition page โ an actual throughput of contribution.
The piece I keep coming back to is DOL. Federated training across heterogeneous institutional data โ national medical, regional language, industrial sensor โ is fundamentally a semantic alignment problem. Most coalition members can't solve it because the work requires linguistics-and-systems crossover, which is rare. We have exactly that intersection in the building. This is our wedge.
The horizon
The Paris founding workshop was 7โ8 May 2026. We missed it, and I think that's fine. Possibly better than fine. Founding membership comes with governance entanglement that would compromise the independent thesis work, and the real architecture decisions get made in 2026โ2027, lock in for years, and only crystallize in 2028. 2027โ2028 is the right horizon โ we're not racing for the photo, we're racing to be the obvious technical partner when specific problems get hot.
Position and publish. DOL paper to arXiv. Open-source the minimum reference implementation. Get indexed, get read.
First technical contact through AI Alliance working groups. Artifact in hand. Frame as architectural input on standards-in-formation, not partnership ask.
VUDO and Nexus built as Tapestry-compatible federated execution infrastructure. Imaginarium build cycles produce tooling and reference implementations. Conference and workshop track.
Univrs proposed as the reference governance pattern for non-state participation โ sub-national, cooperative, research-collective nodes. Imaginarium as the onboarding pipeline.
The decision test
The test I keep applying to this plan is: does it work in both futures?
If Tapestry succeeds โ Univrs is positioned as the obvious sub-national federation pattern, with two years of demonstrably interoperable infrastructure behind us.
If Tapestry stalls โ we've spent two years building federated infrastructure with independent commercial and research value, a semantic-alignment language that solves problems regardless of which coalition adopts it, and a contributor pipeline that compounds.
Both outcomes pass. That's the test of horizon-positioning rather than betting.