ViktorM🇺🇦

The @ilyasut episode 0:00:00 – Explaining model jaggedness 0:09:39 - Emotions and value functions 0:18:49 – What are we scaling? 0:25:13 – Why humans generalize better than models 0:35:45 – Straight-shotting superintelligence 0:46:47 – SSI’s model will learn from deployment 0:55:07 – Alignment 1:18:13 – “We are squarely an age of research company” 1:29:23 – Self-play and multi-agent 1:32:42 – Research taste Look up Dwarkesh Podcast on YouTube, Apple Podcasts, or Spotify. Enjoy!

@tonyzzhao @sundayrobotics Congratulations Tony!

Today, we present a step-change in robotic AI @sundayrobotics. Introducing ACT-1: A frontier robot foundation model trained on zero robot data. - Ultra long-horizon tasks - Zero-shot generalization - Advanced dexterity 🧵->

@JasonMa2020 Congratulations Jason and Lindon!

We have raised $120M to accelerate our mission of building and delivering high-performance general-purpose robots to the physical world. Within one year, we have made research breakthroughs, showing that it is possible to achieve real-world reliability with large VLAs, and demonstrated commercial and deployment traction, with our DYNA-1 models running live in sites at SF, LA, and Sacramento. This is just the beginning, and I never felt more optimistic about a future where AI-powered robots can positively impact human productivity. From when I started PhD where robot policies barely worked even in highly controlled settings to now deploying DYNA robots with the confidence of out-of-box model performance, when I think about the trajectory of robotics, it’s astonishing how quickly we’ve gone from if it works in the lab to it works in the world. The next frontier isn’t about proving robots can move—it’s about proving they can reliably help in real-world environments, at scale, across industries. That’s what we’re building at @DynaRobotics The impact of this shift will be massive: - Unlocking productivity across logistics, manufacturing, and beyond. - Expanding what small teams and businesses can achieve. - Freeing humans to focus on higher-level creativity, problem-solving, and connection. The mission is bigger than any single deployment. It’s about ushering in an era where general-purpose robots are as ubiquitous and trusted as computers or smartphones. We’re just getting started, and I couldn’t be more excited for what’s ahead. Join us!🚀🤖

@TairanHe99 @GuanyaShi @ChangliuL Congratulations, Tairan!

Two weeks ago I passed my PhD thesis proposal 🎉 Huge thanks to my advisors @GuanyaShi & @ChangliuL, my committee, and everyone who has helped me along the way. Last week I also gave a talk at UPenn GRASP on our 2-year journey in humanoid sim2real—reflections, lessons, and outlook. Great Q&A, now online: 🔗 youtube.com/watch?v=sRpx2h… During the visit, I had inspiring 1:1s with 10 professors and even an hour with GRASP Lab founder Ruzena Bajcsy. I asked her: after nearly 60 years as a robotics professor, what defines a truly great researcher? She said: 1️⃣ Don’t publish many papers 2️⃣ Don’t pick trivial problems Two principles I’ll strive to follow for the rest of my PhD journey.

Today in 1969. Neil Armstrong took this - the first ever photograph on the surface of another world. Remastered here from a scan of the actual film from Armstrong's camera.

@bingjietang07 @Tesla_Optimus @gauravsukhatme @uscresl @NVIDIARobotics Congratulations Dr Tang!🎖️

Excited to share I’m joining @Tesla_Optimus! Huge thanks to my advisor, @gauravsukhatme, labmates @uscresl, collaborators @NVIDIARobotics, and all the friends and family who supported me throughout my PhD.

Are you a graduate student in #Ukraine interested in machine learning and neuroscience? My research lab at #UofT is now accepting applications for remote thesis supervision. (1/3) #neuroAI #compneuro @VectorInst @UofT @UofTCompSci @UHN

At #RSS25 next week. If you’re interested in chatting about humanoids, foundation models in robotics, RL, or sim2real over coffee, feel free to DM!

It’s time for Humanoid Robot Olympics now! We’ve had running competitions, established robot soccer leagues, and now boxing. Let’s bring them all together - tennis, wrestling, track and field, olympic weightlifting, soccer - all in one thrilling event!

Wow - it looks impressive! A technically sound Muay Thai bout. How about freestyle wrestling next? Robots can’t get KO’d like humans, but imagine the spectacular 4-point throws they could pull off!

Except it's far from been the cheapest - look at the statistics for the countries who heavily invested in solar + wind, vs those with nuclear + hydro, or just compare the electricity prices (and amount of CO2 generated) in Germany and France. Solar is also not the cleanest and CO2 free - all the materials needed for solar batteries have to be somehow mined first. And due to the energy density constraints the amount of mined materials (and land used) is tens times more per GW compared to the nuclear power. And it's even without taking into account a huge amount of batteries required. So it's rather the opposite - solar would be a nice addition where it's appropriate to a stable, clean and reliable nuclear energy (with hydro and fusion in the future) baseload.

Solar Power is now 10% of World's Power Generation I agree with Elon Musk on this point. Solar energy is currently the best available energy source. Solar energy is the best source of energy today because it is superior in almost all key areas - economically, environmentally, technologically and politically. It is now the cheapest form of electricity generation worldwide: The price per kilowatt hour is less than 3 cents in some cases, cheaper than coal, gas or nuclear. At the same time, the sun is an inexhaustible resource - every day it supplies many times the global energy requirement, free of charge and available everywhere. The sun radiates more energy onto the earth in one hour than the entire human race consumes in a year. -Cheapest electricity worldwide -Inexhaustible resource -CO₂-free and quiet -Decentralized & can be used independently -Quickly expandable and future-proof As long as we don't have fusion energy, solar energy is a herasurable source, but it must be supported by energy sources such as wind and, in some cases, nuclear energy to compensate for fluctuations.

@scaling01 o1-pro, GPT-4.5, DeepSeek R2

comment your: - favorite model of all time - most used model (presently) - model launch you are most excited about I'll start: Sonnet 3.5, o3, GPT-5 and DeepSeek R2 close 2nd

Impressive new work "VideoMimic" by @arthurallshire and team. The G1 humanoid robot masters complex contextual skills - like climbing stairs or sitting -directly from casual human cell-phone videos! Not only novel research-wise, but also exceptionally engineering-intensive, the project cleverly integrates computer vision, state-of-the-art physics-based animation methods, and RL in simulation (Isaac Gym) with distillation into a single real-to-sim-to-real pipeline.

Streamed by @breadli428

Exciting news—the first-ever human-robot half-marathon is happening right now in China! m.twitch.tv/breadli428 Robotics engineers and scientists really need to be in great shape these days—to lift heavy humanoid robots when they fall and run alongside them over long distances. I'm starting to wonder if this is becoming part of the standard interview process: 'Congrats, you've aced all our tests! Now, one final challenge—finish a half-marathon in under 2 hours, and you're hired!'😅

I noticed the same: o3/o4-mini started using my name, during reasoning: "Viktor asked me to do this..." or "Viktor is frustrated and angry on me for not solving the issue, let's go over the problem statement again.".😆 My account has memories, but o1 pro/o3-mini weren't using my name before.

o3 seems to know my name 'Aditya' while reasoning - even though I am using it on an account with no memories or custom instructions

I've experimented with many novel (at the time) optimizers for RL training on continuous control tasks. Yet, none outperformed Adam—even after extensive tuning. Either iteration times increased without improving rewards, or the best achieved rewards consistently fell short compared to Adam. Might be a different story for LLM training, though.

The fact that Adam, developed more than a decade ago, is typically the default optimizer is mind boggling. I've heard of newer optimizers (e.g., Muon) but not seen it used much for RL -- particularly interested in ones which are cheaper memory-wise (less optimizer parameters on GPU) but better performance.

Protoclone is the most anatomically accurate android in the world.

@ritvik_singh9 @Berkeley_EECS @pabbeel @JitendraMalikCV Congratulations Ritvik! Well deserved!

Excited to share that this Fall I'll be starting my PhD @Berkeley_EECS with @pabbeel and @JitendraMalikCV!

@chris_j_paxton Congratulations Chris!

So I recently joined Agility robotics to help lead AI efforts, and I wanted to share this as one of the first things I worked on: - whole body control, running sim-to-real RL, all day for like 4 days straight at GTC - manipulating previously unseen objects - we bought them monday and put them on a shelf