Codex Hackathon · Paris
The world model that learns your factory, your hospital, your city. Made smarter by the people who use it. Owned by no one who can take it away.
Fig. 1. Next-latent prediction error, recorded live
The question
Robotics world models are becoming infrastructure. The model that understands a factory, a hospital, a city is being built somewhere else, and it can be priced, throttled, or switched off on terms you never agreed to, by people you will never meet.
A handful of companies, or all of us?
The bargain
Hand over your data. Let it train a mind you will never see inside. Then pay to rent that intelligence back, on terms you do not set.
Are we really fine giving the world away to train something we are not allowed to own?
We can do better. So we built it.
The alternative
Thousands of people, each holding a sliver of the real world, make one shared model smarter. The data never leaves their devices. The gains are real and measured. The upside belongs to them.
But what kind of intelligence can hold a world?
The bet
A language model learns the world secondhand, from text. A child takes in more of physics by the age of four than every model trained on the whole internet. Words are a thin trace of a world they never touch.
No feel for gravity, for contact, for what stays true once you look away.
Every token is a guess. Over a long horizon the guesses drift into nonsense.
A stream of words is not a state you can search. Predicting text is not reasoning about consequences.
More data makes them remember more, not understand more. To act in the physical world, a machine needs a model of the world.
World models
ChatGPT was the moment a model understood language well enough to be useful to everyone. A world model is that leap for the physical world: it predicts what comes next in representation, not pixels, keeping what matters and ignoring the noise it could never predict. JEPA makes it stable, and small enough to run on the device in your hand.
Predict the representation, not the pixel. That one choice is the whole bet.
And you are about to watch one think.
It predicts what comes next, every frame. Surprise is the moment the world does something it did not expect, measured in latent space, not pixels. One number, and you can watch it spike.
About 6 ms per step on a plain CPU, over 80 frames a second.
surpriset = ‖ predicted next latent − actual next latent ‖²
A scalar prediction error in latent space, not a per-pixel heatmap. The model emits one 192-d vector per frame, so surprise is a recorded signal, the kind a journal once printed from an instrument.
Implemented in web/federated-demo/lewm_probe.mjs. The recorder runs it once per frame.
The agent sprints across the room. A thousand pixels change, but the model saw it coming, so the trace stays flat.
The agent barely moves but does something off-distribution. Almost nothing changes on screen, yet the trace spikes.
It measures prediction, not pixels changing. A live frame-difference trace runs beside it on stage to prove the point.
Minutes ago, people in this room each trained a small piece of this model on their own private trajectories. Nothing raw ever left a single device. No account, no upload, just a model that came back better than the one they walked in with.
Fig. 9. The same frozen backbone, one shared revision. The signal goes quiet
12,512 parameters, learned on your own local trajectories.
0.07% of the modelA clipped adapter delta is sent. The data itself stays put.
The offsets are averaged into a single update to the model.
The backbone stays frozen. Only the adapter moves. The data never leaves the device it was made on, and the model still gets better for everyone.
Held-out prediction error dropped 12.3%. Across five independent seeds it dropped every single time, by 16.8% on average, and never less than 5.4% in the worst draw.
Fig. 11. Held-out error, before and after federation. Every seed improved
Source: lewm_tworooms_system_probe.json and the five-seed sweep. Press down for the full table.
| seed | baseline MSE | adapted MSE | improvement |
|---|---|---|---|
| 20260612 | 0.0604 | 0.0530 | this run+12.3% |
| 1 | 0.0547 | 0.0400 | +26.8% |
| 2 | 0.0564 | 0.0534 | worst+5.4% |
| 3 | 0.0769 | 0.0717 | +6.8% |
| 4 | 0.0738 | 0.0497 | best+32.6% |
| mean · stdev · verdict | +16.8% mean · 11 pts stdev · every seed improved |
Even the worst of the five seeds still improved, by 5.4%.
The economy
EUR 1,000,000
What the improvement is worth to a buyer who needs it. They are paying for a measured result, not a promise.
The money flows to the people who made the model better, not to whoever happens to host it. Not the platform. Not a lab. Them.
The reward ledger
| contributor | weight | reward |
|---|---|---|
| L'Ensemble Labs · the network | 20% share | EUR 200,000.00 |
| Ada | 4122.1740 | EUR 363,136.75 |
| Tomás | 3123.0300 | EUR 275,118.65 |
| Priya | 1426.7360 | EUR 125,686.17 |
| Mateo | 409.3200 | EUR 36,058.43 |
| balanced to the cent | EUR 1,000,000.00 |
Ada trained on the most data, so Ada earns the most. Every share is earned: data trained on, rounds joined, how much each update actually helped. The ledger balances to the cent.
Powered by Mollie · test mode API key · simulated payments
A world model you can watch think. Made smarter by a crowd that kept its data. Paid for by those who need it. Owned by no one who can take it away.
This is what sovereign intelligence looks like.
Open. Come build it: github.com/AbdelStark/Lensemble
The back matter, for the breakout room and your questions.
Cartographer
The same model, turned into a rotating point cloud of its latent states, with planning trajectories igniting toward a goal. You can literally watch it imagine its next move.
A healthy model spreads across many dimensions. A collapsed one folds into a line.
browser join (scan a QR code) local trainer -> adapter delta -> submit_update -> aggregate -> model revision + shared offset headless probe -> docs/evidence/...system_probe.json
tworooms_env -> live trajectory lewm_runtime -> encode / predict (ONNX) surprise = MSE(pred_next, actual) lewm_adapter -> before / after offset -> recorder + perturbation + frame-diff
Runs on WebGPU with a WASM fallback. The seismograph recorder is pure 2-D canvas, so it never depends on a GPU.
lewm_tworooms_system_probe.json this run, +12.3%lewm_tworooms_probe_seedsweep.json five-seed mean and worstAGENTS.md § Claim Discipline the binding languageNothing on the wall is hand-typed. Each figure traces back to a file you can open.
Recorded demo
A captured run of the live demo, kept as the fallback so the proof shows even when the room's network does not.
