Brad Porter bets general dexterity is 2 to 3 quarters away
The prediction rests on internal results nobody outside has seen, and John Carmack is making the case for the other side.
Brad Porter says he stood on stage at the MIT AI conference and made a very specific prediction. Fluid generalized physical AI for dexterous manipulation, he says, is 2 to 3 quarters away.
fluid generalized physical AI for dextrous manipulation is 2-3 quarters away
His reasons are all internal. Porter points to the scaling laws his team is seeing with their own pre-trained models, says their pre-training and post-training recipes are working, and says they know how to fuse sim and real world data, how much data they need for the next milestone, and the exact labeling formula that works. None of that has been published or shown outside the company.
The demo he offers
The one concrete thing in the post is a Friday afternoon test. An engineer put a coding agent inside the ROS container on a robot called Proxie, gave it access to the lidar and camera topics, and gave it a single prompt to find a cart and attach to it. Over about an hour, with some minor inputs, the agent iteratively wrote code to predict the position of the handle bar plus a controller to drive up and attach.
Porter is upfront that this is not new capability. They already have code that does this and have grabbed and moved hundreds of thousands of carts. The point is that frontier models are now good enough to drive continuous self-improvement. The engineer also noted the agent wrote a bunch of safety checks without being asked, accounting for control latency.
Porter also repeats, as a rumor, that OpenAI has moved its robotics work from a research endeavor to an applied program. He does not say where that came from and nobody has confirmed it.
Why it is still quarters, not weeks
Porter's own answer is generalization. Whether an AI writes the control code or humans train the models, the industry is still learning individual tasks one at a time, and that has proved elusive to generalize.
He cites John Carmack as articulating the gap well. Carmack's post is a blunt reality check on flashy robot demos. Changing the batteries in a jack-o-lantern means opening a battery package, getting a precision screwdriver out of its case, and unscrewing four tiny screws that require a hard press and twist. You could build and train a specialist robot for that today, Carmack writes, but there is a very long way to go to replace the full range of human capabilities.
Porter agrees the gap Carmack points to is real. He just thinks it is closing very fast. What is on the table so far is a founder's forecast built on unreleased results, plus a one hour cart attachment experiment on a task the company had already solved.
There is no robot in the world that can change the batteries in a jack-o-lantern.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.