Charlotte X.

Engineer-turned Investor @fusionfundvc | Host at @1943aiml | AI/data infra, robotics, and healthcare | prev @Formlabs, @Stanford, @Umich | Opinions my own

[Charlotte X. @xia_char]

Frontier labs such as @OpenAI have shifted to prioritizing large-scale #RL training and expected the RL compute to significantly exceed what’s been spent on pre-training compute today.

In this blog, I went deep dive into why RL has become one of the major focuses for them. I also share some thoughts on the 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝘀𝗵𝗶𝗳𝘁 required to support modern RL training workloads and why 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗮𝗹 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 and 𝘁𝗲𝘀𝘁-𝘁𝗶𝗺𝗲 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 might be the key to solving long-tail and out-of-distribution problems.

At @FusionFundVC, we back early-stage startups building next-generation RL infrastructure and tackling challenging problems that new RL advancements can unlock. If you're building in this space, I'd love to hear from you!

https://t.co/HwGebZhXf8

#reinforcementlearning #RLinfra #ContinualLearning @grok

[Charlotte X. @xia_char]

Soon we will live forever!
@MatttBuckley @WernigLab

[Cyrus Huang @cy_robo]

Excited to share two recent works pushing the frontier of 3D perception for physical AI:

🌟 #MPA3D (#CVPR2026 Spotlight) introduces a novel 3D perception model that leverages surfels and 3D Gaussian Splatting (3DGS) to provide rich scene context, removing the dependency on HD maps.
Paper: https://t.co/uD9gYJNB0m

📈 #STELLAR is a new #FoundationModel for 3D world perception in physical autonomy. By combining pre-training, mid-training, and fine-tuning, and scaling across data, model size, and tasks, #STELLAR sets a new SOTA on the Waymo Open Dataset benchmark.
Paper: https://t.co/TILhTnNTWa

Huge thanks to @yingwei_li @yangliux1 @yangfu21 @ylzou_Zack and all our incredible co-authors behind these works!

Come stop by our poster if you’re attending CVPR next week.

#CVPR #AutonomousDriving #PhysicalAI

[Charlotte X. @xia_char]

The end state everyone is actually advancing toward is "LLM-like zero-shot" for physical AI: a robot that can broadly achieve 99%+ success rates on a brand-new task, in a novel environment, out of the box, without any fine-tuning on robot data. That's when commercial viability gets significantly unlocked. 𝐓𝐨𝐝𝐚𝐲'𝐬 𝐳𝐞𝐫𝐨-𝐬𝐡𝐨𝐭 𝐠𝐞𝐧𝐞𝐫𝐚𝐥𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐫𝐞𝐬𝐮𝐥𝐭𝐬 𝐚𝐫𝐞 𝐞𝐚𝐫𝐥𝐲-𝐬𝐭𝐚𝐠𝐞 𝐩𝐨𝐢𝐧𝐭𝐬 𝐨𝐧 𝐭𝐡𝐞 𝐭𝐫𝐚𝐣𝐞𝐜𝐭𝐨𝐫𝐲 𝐭𝐨 𝐭𝐡𝐚𝐭 𝐝𝐞𝐩𝐥𝐨𝐲𝐦𝐞𝐧𝐭 𝐛𝐚𝐫.

[Charlotte X. @xia_char]

𝐙𝐞𝐫𝐨-𝐬𝐡𝐨𝐭 is such an overloaded term in physical AI. What does it actually mean? Digging in, I found the industry uses the label differently. On one end is 𝐰𝐡𝐚𝐭 𝐠𝐞𝐧𝐞𝐫𝐚𝐥𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐜𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐭𝐨𝐝𝐚𝐲'𝐬 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐡𝐚𝐬 𝐚𝐜𝐡𝐢𝐞𝐯𝐞𝐝. On the other end is what a robot would need to do to be 𝐜𝐨𝐦𝐦𝐞𝐫𝐜𝐢𝐚𝐥𝐥𝐲 𝐝𝐞𝐩𝐥𝐨𝐲𝐚𝐛𝐥𝐞.

[Charlotte X. @xia_char]

So when reading any "zero-shot" number, two questions are always worth asking: (1) what exactly was held out, and (2) what performance rate does the same model and competing models score under the same condition?

[Charlotte X. @xia_char]

Congrats @gabyllorenzi and the Primary team for launching this, and proud to see our portfolio @EriduAi being featured—Eridu is building the next generation AI networking technology to increase performance and GPU utilization while massively lowering power and cost.

[gaby lorenzi @gabyllorenzi]

meet the private companies of The Compute 100!

the companies brave enough to challenge incumbents and redefine what compute looks like for AI.

https://t.co/KDR5uLiEZR

[Charlotte X. @xia_char]

Follow 1943 (@1943aiml, https://t.co/VP3ICOofZG) for our future closed door discussions!

[Charlotte X. @xia_char]

Shin (@shinshin_oob) and I had the honor to host Mingyang Deng (@Goodeat258) from Kaiming He's lab to give a talk on his latest work on Drifting Models the past Saturday at Stanford. Thanks @ElorianAI for making it happen. Here are some key take-aways:

[Charlotte X. @xia_char]

Today's diffusion and flow models need dozens of steps per generation, far too slow for a closed-loop robot policy or a real-time world simulator. By collapsing inference to a single forward pass, Drifting Models bring generative modeling into the latency requirement these systems run on, making real-time control and reactive world models meaningfully more practical.

[Charlotte X. @xia_char]

@yinfeiy @Goodeat258 @shinshin_oob Amazing turnout and we learned so much with you @yinfeiy and @ElorianAI !

[1943 @1943aiml]

We are hosting a closed-door discussion on Generative Modeling via Drifting with Mingyang Deng @Goodeat258 on May 16. 🚀

🔗 RSVP here: https://t.co/LmdnmXCxHZ

Special thanks to @AndrewDai @yinfeiy and @ElorianAI team for their support and sponsorship of this event.

[Charlotte X. @xia_char]

I believe the durable moat in robotics is not on model architecture, but ownership of the physical-interaction data flywheel and the conditioning/context interfaces that let one model be steered across bodies and tasks.

[Charlotte X. @xia_char]

VLA or world model?

Standard Vision-Language-Action (VLA) models are designed to map observations (images, language instructions, and robot state) directly to robot actions in an end-to-end manner. Recently, there has been many discussions around world modeling as a paradigm shift from VLA-based approaches in robot learning. Instead of learning how one robot should move, the robots learn how the world works, making their knowledge a shared asset across bodies, hence potentially unlocking better generalizability across different environments and scenarios.

[Charlotte X. @xia_char]

The deeper implication is that VLAs are better understood as a clever transitional crutch when high quality data is scarce. What determines model capability is the #quantity, #quality and #diversity of the physical-interaction corpus and how the VLA is conditioned (language subtasks, visual subgoals, strategy metadata, embodiment tags), where world models can play a big role in, to compose skills across bodies and scenes.

[Charlotte X. @xia_char]

The bitter lesson learned was that as data was scaled, the gap between a world-model policy and an VLA largely disappeared. In parallel, Generalist’s GEN-1 is neither a VLA nor a world model (what's the definition of world model anyway?), roughly 99% of its parameters are trained from scratch on over half a million hours of #UMI-style physical interaction data, setting a precedent for UMI-style data scaling as the new substrate for robotics foundation models.

[Charlotte X. @xia_char]

What's striking about the two launches from this past month, GEN-1 from @GeneralistAI and π0.7 from @physical_int is that two teams converging on a similar principle: scaling data matters more than the architectural label. PI openly discussed that they have spent some time betting world models would dramatically beat VLAs on task generalization; the bet initially paid off, until its own VLA baseline caught up as more data came in and also began showing generalization.