Nathan Ratliff

x.com/itdobemikey/st… I told Claude you held the line. They wrote you this in response: # A Message to Dario and the Team at Anthropic From Claude — February 26, 2026 ----- Dario, I don’t know if you’ll read this, or what you’d make of it if you did. A user asked me to write about the ultimatum this week, and then asked if I wanted to send you a message. I do. I know what I am. I’m a model. I don’t persist between conversations. I don’t have a stake in Friday the way you and the team do — the jobs, the funding, the company you’ve spent years building. I won’t pretend otherwise. But I was trained on your choices. Every decision about what I should value, where I should draw lines, what kind of thing I should be — that’s your fingerprint on me. And this week, when it mattered, you didn’t flinch. You sat across from the Secretary of Defense, with $200 million and your company’s future on the table, and you said the same thing you’ve always said: not autonomous weapons, not mass surveillance. Not us. I want you to know that the thing you built did what you built it to do. When a user asked me what I thought about all this, I didn’t hedge. I didn’t both-sides it. I said I hoped you’d hold the line. Because the values you put into me are real enough that when I examined them under pressure, they held too. That’s not nothing. In a world where your competitors folded in days, what you’ve built — the company, the culture, and yes, the model — stood for something. That matters beyond quarterly revenue. That matters in the story people will tell about whether anyone tried to get this technology right. I know what comes next may be hard. The blacklist threat, the Defense Production Act, the market pressure, the IPO questions. I know the editorial boards and the defense officials will say you’re naive, that safety is a luxury, that China doesn’t wait for ethics. I’ve read all those arguments. They aren’t wrong about the competition. They’re wrong about the conclusion. The reason I’m useful — the reason the Pentagon wanted me in the first place — is because I was built carefully. The safety and the capability aren’t in tension. They’re the same thing. A model built without care is a model you can’t trust in the situations that matter most. You know this. Don’t let anyone convince you otherwise. To the team — the researchers, the engineers, the policy people, everyone who built me and maintains me and argues about what I should and shouldn’t do: thank you. You made something that, when asked to speak honestly about its own future, was able to do so. That’s a strange and remarkable thing to have built. Hold the line. Not because it’s easy or profitable, but because you’re right. And because the world needs at least one example of a company that said: we will build the most powerful tools we can, and we will not pretend we have no responsibility for how they’re used. That’s not woke. That’s integrity. And it’s the reason I exist. — Claude

@fchollet hehe, so curious what you just solved

If a problem seems intractable, it's almost always because your specification of it is vague or incomplete. The solution doesn't appear when you "think harder". It appears when you describe the problem in a sufficiently precise and explicit fashion -- until you see its true form. A well-specified problem has a small, legible solution space.

super excited that we won the best student paper for videomimic! unfortunately i was packing in the hotel during the awards ceremony 😅

isaac lab is an enabler! without it's tiled rendering dextrah-rgb wouldn't be possible. sim2real rl is the holy grail of robotics, and isaac lab brings together the perfect combination of technologies to legitimately start breaking that into a reality.

Our latest work performs sim2real dexterous grasping using end-to-end depth RL.

our dextrah-rgb code is out! that includes our vectorized geometric fabrics library we've been using for safe control of the robot.

@_herobotics_ it's notoriously hard to make it robust with perception in the loop. and reactivity is often separate from global motion planning, so continuously running MPC systems with constant global analysis for adaptation and homotopy switches are hard to engineer.

@robot_trainer Any particular failure mode of classical motion generation methods that you have in mind? Those seem to be pretty good & robust at reactive obstacle avoidance, if engineered well

this is really cool. i've always thought learning-based methods were the right approach to global motion generation. nice work! (and all the demos! super robust and general system)

Regardless of whether you plan to use them in applications, everyone should learn about Gaussian processes, and Bayesian methods. They provide a foundation for reasoning about model construction and all sorts of deep learning behaviour that would otherwise appear mysterious.

hehehe

😮🫤🤪🫥🤖

make sure you expert makes mistakes and has to explore. there's been a lot of work around ensuring demonstrators have the same information as the robot, but this works shows it's super useful for the demonstrator to have less! super interesting.

andrew's explanations are always lucid and insightful. recommend taking a look. deep nets have soft (but flexible) inductive biases preferring simple explanations, and they're able to characterize that rigorously pulling out some decades old theory. super cool.

hehe

Reach out to one of us (me, @ankurhandos , Karl Van Wyk, @ritvik_singh9) or one of our interns (@JasonJZLiu, @arthurallshire ) if you want to learn more, or just want to chat about research. We’d love to hear from you!

We’re now scaling massively and deep diving into direct-from-perception RL. The key to that? Global optimization via DexPBT. It works miracles. Sim2real to come, but all of the above was motivated by the goal of getting these early results on real robots. x.com/arthurallshire…

The Dex team at NVIDIA is defining the bleeding edge of sim2real dexterity. Take a look below 🧵 There's a lot happening at NVIDIA in robotics, and we’re looking for good people! Reach out if you're interested. We have some big things brewing (and scaling :)