> Laughs in i told you not to externalize your institutional knowledge to model labs that will cannibalize your business model which you de-risked and PMFed for free

Talking to other researchers in the "learning from human video" space early this year, a common observation was that it's hard to show transfer from true in-the-wild Internet video, compared to curated research datasets. In most of these datasets, the humans deliberately move like robots, and hands are tracked with clean 3D labels. The usual starting point for Internet video — run a monocular hand pose estimator on YouTube videos, cotrain with robot data — often doesn't work. In our recent work, we study this "YouTube-type" video setting and try to understand what it takes to absorb egocentric Internet videos into a VLA training pipeline. 1/n

This is going to have an opposite effect that the decels want. It's a huge open source AI accelerant, an accelerant for corporations, and enterprises Now it's a real race. You either make your own AI infrastructure or you don't have a seat at the table

To me this paints Anthropic clearly as anti science, and therefore anti progress and anti safety.

Flow reversal steering allows "steering" diffusion-based VLAs with high-level actions, for example from VLM reasoning. This also lets us run RL in the diffusion noise space with exploration guided by high-level reasoning: think through a task, then practice it! 👇

Lmao

I feel like we should start having tps / active params / price per output tokens / tokens per task on these benchmarks already

Great talk by @yukez at #ICRA2026. In the VLA's own words it would probably be "the reports of my death are greatly exaggerated". I asked for intuitions on limits of WAMs vs. VLM/A, given agents use language priors to reason and navigate complex plans and (video) WAMs never learn these: - VL parts would be great to keep, not throw out - Language is a somewhat naive interface between the two, latents are probably better, open research question - @Ken_Goldberg's plenary highly related - backbone quality is key - (relatively) expensive to train WAMs, harder to make fast, but they got to 30 fps with nvpf4 and tricks. I would add: - temporal priors very easy to teach with WAMs, but hard with VLAs - lots of spatial graphs / navigation and memory work was presented in other keynotes all of which used language interfaces / abstract language reasoning - interactions with humans need VL, can't pass a video to the human for every question

🚨 BREAKING: EUROPE DISCOVERS AMBITION

You throw the dart roughly somewhere and its always hits bullseye, is this claude code for showing off?

We're launching the microagi Research Fellowship. Fellows get up to $2M in compute, robotics hardware, our evals, and one of the largest physical AI datasets ever assembled. You build in our lab, with our team, alongside partners like Unitree, Nvidia, and Google Cloud. The hard part of AI left is physical. That's the part we're working on. Come build with us. One more thing: know someone who belongs here? Reply with their name. If they get in, we send you 10.000 USD

The two main takeaways from the #ICRA2026 keynotes RL track, to bring robots it into production, in case you missed it, were: "Is RL even needed? (no)" and "Just RL all the way". I do agree

This hand is a hidden gem, not quite Sharpa yet but we liked it a lot and one of the few that visualized shear forces nicely. Here is a vid of me playing around with it

Breaking: VCs discover asking claude for business ideas gives everyone the same business ideas.

AI2 are consistently the most open players in AI for several years now. From RLVR, tulu, molmo, to evals molmospaces, rewardbench, to VLAs molmobot trained in sim, molmoact1/2 - just persistently good research taste and they don't shy away from risky and challenging research.

The internets: the west is falling behind in humanoid robot hardware BD: hold my fridge...

Lets regulate AI in the west more so we can collectively avoid the permanent underclass and slide right into the under-underclass. This should pair nicely with the existing energy and population crisis and fit into pre-existing legal deindustrialization frameworks.

Very nice to see more robotics startups looking at the human interaction side of deployments

People are not ready for the mountains of Claude.md bloat and custom harnesses that can be deleted if this works at scale.