WAKE
World Agent Knowledge Engine
AI can help your business. Wake makes it work for yours — it acts in a messy world, on incomplete information, and every outcome sharpens what it does next.
World models are not large language models (LLMs) — they are the next step beyond LLMs, and they sit at the center of the Advanced Machine Intelligence architecture set out by Yann LeCun, a pioneer of modern AI.
Grow your own intelligence. The WAKE world-model platform learns your business the way a person learns a craft — by watching, remembering, imagining what comes next, and planning. It learns from raw video. Today four models read every swim video: V-JEPA learns its own representation of the motion, MoveNet tracks body pose, SAM 3D estimates body shape, and Gemini writes a plain-language read. You bring what you know. Wake turns it into an intelligence that is unmistakably yours.
You feed in what's happening in your business and the goal you're after; Wake feeds back guidance from a careful plan. (A faster reflex mode is designed but not built yet.) Each planned action carries a likely result and a bad-case estimate, weighed against what you care about — so you can trust it, or overrule it, on purpose. Tell it how things turn out, and it learns from the result.
Try it, you'll like it! Try the pilot on the sidelines first. It can run in parallel with your current system for comparison. The pages below explain what we can build together — for you.
Where it stands today
What runs, what is built but not serving yet, and what is not built.
Checked 2026-09-23 against the code and the live systems. Wake runs in two environments, staging and production; the beta of our swim app uses staging.
Running today
Swim video analysis. V-JEPA, MoveNet, SAM 3D and Gemini read every uploaded swim video. Each clip is matched to its three closest of 24 seeded swim-technique concepts, with a coaching cue for the 18 that have one written.
GPUs booked on demand. The V-JEPA video model (and, in staging, SAM 3D) runs on flex-start GPUs: they switch off after 15 idle minutes, and the next video books one from a queue.
Swim features. One tap adds a drill to a swimmer's month, building a month of workouts if there is none. When a video finishes, “Your swim is ready” appears in the swimmer's in-app chat. Every swimmer starts with one goal: move through the water more efficiently. Each swim gets one pace per 100 m, from Garmin laps first, then Apple Watch lengths, then recorded rep times (the app does not send Apple Watch lengths yet).
Planning. Deployed in both environments: /v1/propose in planning mode asks the world model for a careful plan. No outside caller has used it in the past 30 days.
End-to-end traces. The swim app sends its traces to a relay (POST /v1/traces), deployed in both environments, which passes them to that environment's trace collector, so a trace can follow a tap in the app into the backend.
Built, not serving yet
The perception doors. /v1/encode and /v1/perceive answer 503 in production. /v1/encode answers 503 in staging too, because the gateway image does not include the encoder yet. The reader /v1/perceive needs is part of the V3 backend change being merged now.
The swim-form readout. The app's World Model Intelligence tab gets its one takeaway from a backend change (V3) that is being merged now. Until it lands, the tab stays empty and swimmers see the Language Model Intelligence tab.
The swim_v7 pack. One swimmer pack with the 24 technique concepts, 21 drills and 56 drill-to-concept links is built but not yet published, so day plans still use the earlier swim pack.
Strict move enforcement. Rejecting a planned move outside the declared set is on in staging, not yet in production.
Typed SDKs. Python, TypeScript and Swift SDKs, including perception.perceive, are on the development branch of a private repository. They are not published to PyPI, npm or Swift Package Manager yet.
The robot-arm world model. Trained on 1,348 DROID robot videos spanning 12 kinds of moves, and promoted to the staging model registry on 2026-09-03. Production has no arm model.
Not built or not open yet
The quick reflex (Mode 1). A /v1/propose call that does not ask for planning mode fails today.
Outside access in production. Only the goal door (the Configurator Gateway) is reachable from outside. The other five production doors have no public front door yet; on staging, all six are reachable with an API key.
A robot-arm video reader, and moving the arm. The robot-arm demonstration asks the world model for plans, but those plans do not drive the arm yet, and no hosted copy of the demonstration is deployed.
Explore the platform