Conan (@conany08) header
Conan (@conany08) profile picture
San Francisco, CA
Joined July 2022
Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)Photo by Conan (@conany08)
From Data Dystopia to Data Utopia
3
1
21
3,490
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
58,173
118,748
24,069
1,132,823
52,981,725
Everyone's been showing Fable and GPT-6 Astra driving robot arms. We trained an open 27B Qwen model to do it, purely from its own attempts in simulation, with zero demonstrations. And it can run locally at 220+ tok/s on an RTX 5090. On a real robot it had never touched, it now succeeds in 18 of 30 tests, up from 0 before training. It also copies a painting and designs its own chemistry experiment. It generalizes and follows instructions better than π0.5, and on LIBERO it comes close to Astra max. All without losing its general capabilities, like math and coding. Built on Qwen3.8-27B. We've open-sourced the model weights, code, and full training infra! Check out our blog at mit-han-lab.github.io/pear-b….
52
168
50
1,224
110,569
This is the exact product i have been looking for
Backed by YC and Paul Graham, we're building Vexo, an always-listening AI bracelet that turns your conversations into action. Built by engineers from Tesla, AMD, and Amazon. One sentence to Vexo can: •Get Claude building the feature you wanted •Text the family •Block your calendar for dinner •Order the groceries •Find the founder you met at last week's event Every conversation adds context, so you explain less and get more done. Get yours now, shipping start January 30th 2027 vexoai.com/checkout
159
Atlas is now available to everyone!! I’m very excited to see what users can build with our newest world model! 🤩🌎
We’re opening up access to Atlas to everyone today. We’ve built apps and API endpoints to help you use Atlas any way you can imagine: bring your ideas for building a 3D world and come away with keyframes, video clips, or 3D Gaussian splats.
108
308
51
2,720
179,469
The really strange thing about the bitter lesson is that experts of many fields think the bitter lesson applies to all other fields except their special, chosen field. Therein lies the source of bitterness. Case in point: so many roboticists at CoRL at various points of the five stages of grief, and many in denial. If you accept the bitter lesson, and scaling laws, it could help you think more clearly about what approaches may stand the test of time as robotics gets solved. Robotic control, as we are learning, is not special and is adjacent to so many frontier modeling problems - multimodal reasoning, CAD, computer use and coding. If it can make slides, it is already a multimodal VLA. VLA experts for the longest time have emphasized over and over how generalization is the true test of intelligence, and how generalization is the core problem to solve. So it is strange that now when we see frontier models that can generalize way better than specialized robotics models - there’s so much pushback. It is human nature to cling on to a model of the world that feels familiar, to hold on to methods that defined your career and your research. The paradigm shift is here - and you can choose to be the bitter researcher who says LLMs will never be able to do xyz - and get proven wrong - or you can just accept that they will eventually get there and adapt your research.
23
27
16
328
26,907
It was a really good experience taking CyberCab in Austin Called three times, average waiting time was 2 min
Robotaxi works 168 hours a week
5
299
Lisa said robotics will drive compute to the next level today AMD
Great to be at #CoRL2026 in Austin today! Enjoyed the conversation with @drfeifei about the exciting advances in robotics, physical AI, and the opportunities ahead. Thanks to @PeterStone_TX, @yukez, and the @corl_conf team for the invite and the opportunity to connect with so many brilliant students and researchers!
1
133
Sim is important, but eval in simulation alone isn’t enough, models must be evaluated on real-world stations
Fable 5.2's impressive 64.9 on RoboDojo, the number @AnthropicAI reported, comes from simulation. We wondered: what would it score in the real world? So we rebuilt 10 RoboDojo tasks on a real two-arm robot, covering memory, generalization, precision, long-horizon tasks, and instruction-conditioned tasks. Zero-shot Fable averaged 38/100, nearly 2x a π0.5 fine-tuned on our own station (20.9). Here's where each one wins and breaks 🧵
1
4
214
Conan retweeted
Today, we’re making Praxis-1 available to everyone. Our state-of-the-art world action model turns video pretraining into control for real robots, and it ships with open weights under Apache 2.0, ready to run and fine-tune on your own hardware.
129
97
126
622
102,947
Fable 5.2 CaP’s coming
Replying to @claudeai
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
136
Conan retweeted
Real-world RL environments for robot models. Feels like this space could be much larger than software RL envs?
Today we're launching @se3labs, backed by @ycombinator. Robot models are getting better fast, but there's still a big gap between a great demo in sim or the lab and a robot that works reliably on a real task. We're building real-world RL environments for robots. We start with tasks from industry partners, run policies on real robots, and find exactly where they fail. Those failures decide what data we collect and curate for the next policy update. Next, we're bringing this to partner sites with physical sandboxes for real-world RL, so policies can train on the exact tasks and conditions they'll face in practice. If you’re training robot policies or have a real task you want robots to do, we’d love to talk: se3labs.ai/ or DM me directly! See you at CoRL 🤖👋
5
2
1
47
6,145
The gap between a great robot demo and a system that works reliably on a real task is almost always a data problem we haven’t measured yet. Real-world evaluation is how we uncover those failure modes, systematically address them, and ultimately move toward real-world reinforcement learning
Today we're launching @se3labs, backed by @ycombinator. Robot models are getting better fast, but there's still a big gap between a great demo in sim or the lab and a robot that works reliably on a real task. We're building real-world RL environments for robots. We start with tasks from industry partners, run policies on real robots, and find exactly where they fail. Those failures decide what data we collect and curate for the next policy update. Next, we're bringing this to partner sites with physical sandboxes for real-world RL, so policies can train on the exact tasks and conditions they'll face in practice. If you’re training robot policies or have a real task you want robots to do, we’d love to talk: se3labs.ai/ or DM me directly! See you at CoRL 🤖👋
117
We need more open-source robotics models that companies can run on their own infrastructure. For many enterprises, where sensitive data is processed and who can access it matter just as much as what the robot can do
Open-source is critical for robotics AI. It means a student, a researcher or a startup can build on years of progress and contribute something the whole community can use rather than concentrating everything in a few companies building a black-box robot. That’s why we’re building @huggingface into the platform for open-source robotics AI! Amazing to see the progress we've made in the six months and it's accelerating! - Almost 100,000 public robotics datasets on hf.co/datasets, up 48% - Over 30,000 public robotics models, up 41% - Microduck and Reachy mini, our open-source robots became the most popular robot of 2026 (probably ending the year with 50,000+ sold and 20,000+ shipped) - We’re a founding partner of @RoboColosseum, evaluating robots across 140+ real-world scenes. Every run is published publicly on HF, including failures, so everyone can learn from what actually happens. And @LeRobotHF got downloaded over 1M times and v0.7.0 just landed with new world action models, interactive robot control and remote GPU inference: huggingface.co/blog/lerobot/…! Huge thanks to everyone making this happen 🤗🦾
1
2
172
Watching a robot policy fail at the same task over and over again feels like watching a kid struggle with 1 + 2 and somehow never get to 3 You just sit there thinking, “Fuck, I wish I could just do it for you.” And then comes real-world DAgger You take over the leader arms, correct the policy, and show the robot exactly what it should’ve done I’ve never realized how therapeutic correcting a robot could be
1
120
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
2,366
7,256
5,753
55,941
10,068,081
Zero-shot is one of the hardest tests a robot model can face. Dream-1 from @dreamlabs_ai has real zero-shot skills. With no task-specific training, it grasps, carries, and places objects on a two-armed robot. Where it falls short is following instructions. We ran 100 real-robot episodes across 10 tasks, and all of them are public 🧵
7
6
4
69
8,071
Conan retweeted
We're doubling down on hardware and are looking for someone to manage our hardware shop ycombinator.com/companies/y-…
99
61
49
890
159,841
🤖 𝐒𝐒𝐈𝐎: 𝐒𝐞𝐥𝐟-𝐒𝐮𝐩𝐞𝐫𝐯𝐢𝐬𝐞𝐝 𝐈𝐧𝐞𝐫𝐭𝐢𝐚𝐥 𝐎𝐝𝐨𝐦𝐞𝐭𝐫𝐲 Since the first learned inertial odometry 10 years ago, every method has needed ground-truth labels. It doesn't have to: SSIO learns from raw IMU, fully self-supervised 🧵👇 ssio-imu.github.io
8
65
10
387
43,916
academia is experiencing unprecedented paper inflation
183
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×. Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours. Project: supermimic.github.io Demo: supermimic.github.io/demo/
98
317
124
1,913
290,157