I do wonder what “adoption” here measures. Chatbots are especially good at leading us down seductive, non-productive rabbit holes.

Big bet: Deploy robots → learn WM data process → solved. Even if we grant a perfect WM, turning it into general behavior isn’t solved—not even for modest toy domains. Glad to see resources going into WMs, but I wouldn’t put all eggs in that basket.

Companies that ditch entry-level roles for chatbots deserve the mediocrity they’ll get. The smart play: hire, train, and let juniors learn to wield chatbots under seasoned eyes.

So the claim is: if we can build enough robots, then every job disappears in 5 years. But don’t the robots also need to be able to do all those jobs? That’s nowhere close — and won’t be in 5 years. I sense that maybe my parody detector is malfunctioning.

Another nice demo of how LfD yields insect-like intelligence. “Keep folding” isn’t a plan—it’s just the state machine grinding away, oblivious to the broader context. Worth remembering: these controllers have no (or little) exteroceptive memory. They live only in the moment.

Great intro to vision-language models by Chris Paxton Vision-Language-Action Models and the Search for a Generalist Robot Policy open.substack.com/pub/itcanthink…

Amazing progress on the Agility AI stack.

A subtle thing here: many humanoids switch from manipulation → walking mode, which looks like clunky “robotic marching.” This is true whole-body control—moving however it needs to so the hands efficiently get where they must.

Gave ChatGPT 5.0 two weeks as my weekly/daily planner (even the specialized GPT). Felt like working with an absent-minded secretary with drunken confidence and endless encouragement. Fired. Back to my old system.

I’ve heard this "HRI is the moat" point many times. For robots that do physical work, I’m not so sure. Capability feels like the real moat—the hard part that will set companies apart. Interfaces matter a lot, but seem easier to copy and refine once the capability is there.

This is what I have been building for the last few months: a whole-body controller trained in Isaac Sim using RL. The ability to put the hands wherever they are needed and let the NN take steps and adjust stance as needed is a huge boost in capability for teleoperators + LfD.

If 'UFC of robots' means spectator sport, it’s far too early. Current battles get dull after 30 sec. As a spectator sport it would fall between e-sports and pro-wrestling—with no clear sweet spot. More likely profitable as arcade rental than arena show.

If the win condition is based on falls, then I predict the meta-game for this type of robot will not involve punching or kicking. A good whole-body controller can be robust enough to absorb the forces produced the robots punches or kicks.

The productivity feeling AI gives can be deceiving. Cal Newport’s warning: in certain uses, it erodes the deep focus that real progress requires. Something to watch for as we lean on these tools. youtu.be/_aFd7ROyrsU?si… via YouTube

55% in reading, 43% in math, 42% in science. And we’re supposed to celebrate because it’s “better than statewide”? We managed this while also eliminating accelerated math opportunities. Stop putting a positive spin on failure -- we need a societal wakeup call.

My first impressions of the Figure 03: it continues figures incredible progress towards a general-purpose humanoid with a few clear improvements over 02, without fundamentally addressing the capability gaps we saw in the previous versions, most notably around whole body control

Impressive demo! Whole-body control with object interaction is the next step after free-space agility. Most of these demos still rely on motion capture—just a small step from proprioception. The next leap: ego-centric perception.