Chris Paxton calls Machenta's two-minute factory the future
The praise comes with its own caveat, and a follow-up clip teaching one robot three different behaviours on the same objects.
Chris Paxton watched Machenta's three-robot packaging line and said it shows the future. His framing is the part worth reading twice. He described the goal as a guy who unpacks his robots, shows them a task, and gets them doing something useful basically right away. He also said that future is still a ways off, and that the robots cannot be fully trusted on their own yet.
That is a researcher's compliment and a researcher's caution in the same breath. The demo he was reacting to is the one Machenta's Edward Johns posted earlier, three different bimanual robots collaborating on a packaging task after two minutes of teaching, all running the same end to end model, all trained entirely in simulation on objects the model had never seen.
The steerability follow-up
Johns then posted a second clip aimed squarely at the obvious objection. If a pre-trained model does a task after you show it one demonstration, how do you know it is actually watching your demonstration rather than replaying something it already learned during pre-training? Johns calls the answer steerability, and the test is simple. Keep the objects the same, change the demonstration, see if the behaviour changes.
In the video he says the team teaches the robot three different interactions with the same set of objects, just by swapping the demonstration each time. He says Machenta's model is instantly steerable from a single demonstration.
Shown versus claimed
What is shown is a short clip of three interactions on one set of objects. What is claimed is that the model is genuinely conditioning on the teaching demonstration rather than pattern matching to pre-training. Nobody outside the company has run that test independently, and Johns is the one both running the demo and grading it.
Still, the fact that Johns bothered to post the steerability clip at all says something about where the argument has moved. One-shot learning demos are easy to doubt, because a model that has already seen a thousand pick and place tasks will look like it learned yours. Changing the demonstration and showing the behaviour follow is the cheapest way to answer that in public.
Paxton's line about a real world Claude is the one people will repeat, and it captures why this demo landed. The appeal is not the dexterity. It is the setup time. Weeks of system integration compressed into a couple of minutes of showing a robot what you want, with a person in the loop because the robots are not trustworthy alone yet.
this is what a "real world" claude should look like
we've built our model to be instantly steerable from a single demonstration
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.