“For coding agents, language model, it's okay if it has latency, right? … But for robots, the inference latency has to be very short. The robot needs to act. You can't just hold a knife and not do anything for three seconds. That's going to freak people out.”
Latent Space · AI Builders · October 2026
Ma, of Dyna Robotics, answering whether robot models run on the device, in the cloud or both. He explains why physical agents can't tolerate the few-second pauses people accept from chatbots.
Transcript
…I'm just really excited about this reaction because I think even though we have scaled to a million hours of data, we are deploying a bunch of robots. The robots actually already starting to do very cool things. We're just at the beginning, just the tip of iceberg for what is there to come.
…20B is very fit for your average MacBook. Then there's a larger one that's fit for a single GPU. Do you work in the constraints of robot model on device in cloud, a hybrid? And then how does that affect your scaling law of data? A million hours is great, but do we just scale up parameters, data?
Yeah, no, that's a great question. So I think what's really different about robots versus language model and other applications is that for coding agents, language model, it's okay if it has latency, right? But like, you know, today, whenever I ask ChatGPT, it takes like a couple of seconds before it spits out anything. But for robots, like the inference latency has to be very short. The robot needs to act. You can't just hold a knife and not do anything for three seconds. That's going to freak people out.
We saw someone put their real hand under our
models pretty smart, so that's okay. But in general, if you think about very dynamic tasks, for example, you want your robot to hold a, I don't know, a tennis racket and juggle the ball. It's very dynamic, very reactive. You can't just like, the physics gravity doesn't wait for you. So for that…
you have
More from Latent Space
