Language had it easy. One architecture won, and everything after was a scaling problem: more data, more compute, better answers. Computer vision never got that clean break, and it has ten competing recipes and no champion, not one of which can reliably walk into a scene it has never seen before and just work.
The model will not decide the winner. The founder of a robotics company I know argues the deciding constraint sits underneath it, in where the data has to be processed.
Bandwidth, not architecture, sets the limit
The problem starts with a hard physical limit. You cannot pipe video from every camera, all day, to a data center and run a large model over it. The bandwidth does not exist, the bill is absurd, and the lag kills anything that has to act in the moment. The model therefore has to live where the camera lives, and in practice that means something tiny, on the order of a 19MB model running on the device itself.
That constraint decides the design space: the question is not which architecture is most capable but which capable thing fits on a camera and answers in real time. Data does not just train the model; it dictates the architecture. The place where the data has to be processed decides the shape of everything you are allowed to build.
Most of the field treats architecture as the interesting decision and data as the input you shovel in. On-device vision inverts that order: the physics of where the data lives comes first, and the model is whatever fits inside it.
Two bets on world models
To understand a scene instead of just labeling it, a system needs a world model, some internal sense of what is happening and what comes next. Two camps disagree on how to build one:
- Predictive, in the JEPA lineage. It learns to anticipate the next state in a compressed, abstract space, never drawing the future, only seeing it coming.
- Generative. It predicts the world frame by frame, effectively rendering what happens next in pixels.
Generative demos are the ones that go viral, because you can watch them work. But for a model that has to run on a camera, the predictive approach is cheaper, and a lot more honest about what you can actually afford.
The "deliberate Frankenstein"
His system is not one elegant model but what he calls a deliberate Frankenstein: a distilled reasoning model, a conventional vision-language model, and a pure detection model, bolted together and running about ten custom primitives, each doing one narrow job.
No single approach is good enough yet, so shipping means assembling the parts that each earn their keep. In a field still waiting for its one big idea, combining what works today beats holding out for what might work tomorrow.
Why the data layer is the durable position
Models are getting commoditised; the durable position is the data. You do not buy your way into the right data, you earn it by deploying systems that work.
Access is won in the field, and then it compounds.
A team deploys a working system into a real environment, and that deployment earns it access to a camera network. The network supplies the exact data distribution its models need. Better data makes the next deployment easier to win, which earns more access, which produces more data. Whoever gets that flywheel spinning first builds a lead a better model, arriving late, cannot purchase.
Vision may still be waiting for its GPT moment, the single recipe that turns everything after it into a scaling exercise. The winners will not be whoever finds that recipe; they will be whoever already owns the data pipeline it would need, and that position is taken by shipping rather than waiting.