Sabitlenmiş Tweet
Super proud of the team for live demoing at GTC.
- AI-driven whole-body control
- End-to-end autonomy
- Cost-effective hardware
Excited for what’s ahead.
English
Wenda Wang
111 posts
Wenda Wang
@wwd7086
MTS @ElorianAI | Co-founder @NobleMachines｜Autonomy @Apple SPG | MRSD @CMU_Robotics 🤖 | ❤️ in the work
Sunnyvale, CA Katılım Temmuz 2010
2.5K Takip Edilen368 Takipçiler
What interests me about ChatGPT Live isn’t just that voice interaction feels more natural. It points toward a very different kind of personal computer.
Imagine a wearable device that can hear what you hear, see what you see, and remain available throughout the day. You wouldn’t need to stop, open an app, and carefully formulate a prompt. You could simply speak whenever you wanted, while the assistant continuously understood the surrounding context.
ChatGPT Live doesn’t create that entire experience yet, but it makes the interaction layer feel surprisingly close. I would genuinely want a device like this, assuming the privacy, trust, battery life, and user-control problems were handled well.
There is also a broader convergence happening across AI research:
• Add continuous video input, and the model can understand what you are seeing.
• Add visual output, and it can have a generated physical appearance, present information spatially, or create interfaces on the fly.
• Add “action” as another output modality, and the model can begin affecting the physical world through a robot.
At that point, the boundary between a personal AI assistant and a robotics foundation model starts to blur. Both continuously perceive the world, maintain state over time, decide when to respond, and produce actions that influence what happens next.
Speech, video, generated interfaces, tools, and robot actions may ultimately become different output streams of the same kind of real-time interactive model.
Justin Uberti@juberti
Today, we're launching our third-gen voice model and architecture, GPT-Live. GPT-Live is a full-duplex model with built-in async delegation, which allows it to deliver incredibly natural conversation along with the intelligence you expect from ChatGPT. openai.com/index/introduc…
English
Wenda Wang retweetledi
🔥Vision as Unified Multimodal Generation🔥
🎯SenseNova-Vision🎯 unifies vision tasks (e.g., detection, keypoints, segmentation, depth, surface normals, point maps, and camera pose) as unified multimodal model (UMM) generation *with SOTA results*
- Code: github.com/OpenSenseNova/…
DailyPapers@HuggingPapers
SenseNova-Vision unifies computer vision as multimodal generation One model handles detection, OCR, depth, segmentation, keypoints, normals, and camera pose with no task-specific heads. Just follow natural-language instructions to generate text, images, or mixed outputs.
English
Wenda Wang retweetledi
A super long overdue (3+ years?) post on scaling laws.
Compute is expensive. Scaling laws are a way to help us reason about the optimal compute allocation between data and model size before committing to a large run.
The post covers what scaling laws predict, how compute-optimal allocation works, why Kaplan et al. and Chinchilla disagree, and how data limits + fitting details make extrapolation tricky.
lilianweng.github.io/posts/2026-06-…
English
Wenda Wang retweetledi
PPO had a second wave in the LLM era for reasons unanticipated by the original paper
- the importance-ratio objective fixes biases from numeric error, async training, and forward pass noise
- the clipping objective affects entropy through a mechanism that we didn't know about at the time of publication (DAPO, arxiv.org/abs/2509.26114)
John Schulman@johnschulman2
PPO: rejected from NIPS 2017
English
This is fascinating. In the future, a codebase won’t really be a codebase anymore—it will evolve into something closer to a knowledge and memory base for the agentic system, a pillar for continual learning.
Code is only one part of it. Other forms of knowledge will emerge alongside it. This base provides long-term context to agents, while the agents in turn produce and accumulate knowledge over time—sometimes as code, sometimes as natural language or other representations.
Separation of compute and storage, but for agentic systems.
Cursor@cursor_ai
We're launching code storage and git hosting. Origin gives teams and agents a place to host, review, and collaborate on code. Available this fall. Join the waitlist. cursor.com/origin-waitlist
English
Wenda Wang retweetledi
Proving value before scaling is a key pillar of Noble Machines’ execution, and stability and strength are at the core of that value.
For Noble Machines, real-world value means extending human capability, supporting teams in physical work, and improving operational flexibility.
This video demonstrates multiple Noble Machines robots handling payloads, including one robot managing a 50.3lb load while maintaining balance, control and mobility.
Watch below and see more videos showing Moby in action on our YouTube channel (link in thread).
#IndustrialAutomation #PhysicalAI @NVIDIARobotics
English
Wenda Wang retweetledi
💡What makes Noble Machines’ Moby different? How it learns.
Most #industrialautomation breaks when the situation changes. Moby adapts in real time — supporting industrial work, autonomously.
Meet our leaders at #Automate2026 to see how Moby learns in action and explore deployment use cases.
Book a 1:1 consultation: go.noblemachines.ai/book-a-meeting…
Booth #2488 | June 22–25 | Chicago
#PhysicalAI @NVIDIARobotics
English
Wenda Wang retweetledi
Industrial robots that adapt. Most automation breaks when conditions shift. #NobleMachines built Moby, a general-purpose robot that learns industrial tasks in hours, not months, and remains useful when loads shift and environments change.
Watch Moby work autonomously below, and book a 1:1 use case meeting at #Automate2026 👉go.noblemachines.ai/book-a-meeting…
See more of Moby in action on our YouTube channel: @noble-machines/" target="_blank" rel="nofollow noopener">youtube.com/@noble-machine…
#Robotics #IndustrialAutomation #AI @NVIDIARobotics
English
If everyone has access to the same model, the differentiator isn't intelligence — it's what you feed into the context.
And it compounds quickly: the output of one turn becomes the context of the next.
English
Wenda Wang retweetledi
As AI models become more capable, industrial autonomy is becoming more tractable. The bottleneck is no longer just high-level reasoning.
Legacy systems do not maintain stability under varying loads, do not adapt across tasks, and do not learn from failures. Our latest technical blog explains the learning stack behind Moby:
✅ AI-driven whole-body control
✅ Task-level intelligence
✅ Runtime correction loops
At #NVIDIAGTC, the same autonomy model powered multiple autonomous demos across distinct industrial workflows in live, unsanitized environments — trained on roughly 10 hours of demonstration data.
Read more: noblemachines.ai/blog/beyond-th… and meet us at #Automate2026 in Chicago to see Moby in action (link in thread below👇)
English
A few things worth calling out about our GTC demo:
The robot was running a single end-to-end autonomy model trained on only ~10 hours of demonstration data. In Q2, we expect to reduce that to 1 hour.
A lot of live autonomy / VLA demos are staged in controlled conditions: robot facing a wall, carefully managed background, extra lighting. We chose not to do that. Our robot faced the audience, with no special lighting, because we believe the model should generalize in the real world, not just in a sanitized setup.
We also made a point to be transparent about when the robot was in autonomy mode versus manual mode at all times. That should be the standard. Too often, people just hide the remote controller.
You’ll also notice there were no QR codes in the scene. When you see QR codes, it usually means part of the behavior was pre-scripted around the environment.
Most importantly, we did not build a demo for the demo’s sake. The software build we ran at GTC is the exact same weekly release build we ship to customers. What we showed is what customers get today.
Noble Machines@NobleMachines
#NVIDIAGTC was a strong show. We debuted Noble Machines Moby with 4 booths, 3 robots, and 2 live demonstrations doing real-world work at GTC – • Handling heavy payloads and navigating narrow spaces commonly found in factories and warehouses • Autonomously completing long-horizon material handling tasks that industrial customers expect • Learning and training for Physical AI skills. Hear more about our training process in our interview with Reuters here: reuters.com/video/watch/id… A glimpse of what our booth visitors think about Noble Machines Moby in the video below. Thanks to everyone for visiting the Noble Machines booths and engaging with the team. To @NVIDIARobotics and the NVIDIA team – thank you for the platform at GTC. What would you like to see Moby doing at the next event? Let us know here: noblemachines.ai/contact
English
Wenda Wang retweetledi
Short chat with @NobleMachines CEO and Co-founder Wei Ding at GTC 2026:
- why bipedals for industrial use
- challenges with high carry capacity
- AI architecture
- solving the reliability problem
English
Wenda Wang retweetledi
That's a wrap on #NVIDIAGTC. 4 booths, 3 robots, 2 live demos, hundreds of 1:1 conversations — and Moby is all in.
This is just the beginning.
DM us or sign up at qrto.org/s7GYG4 to schedule a VIP Assessment with CEO Wei Ding— a 90-minute working session to discuss your application requirements, walk through your operations, and assess the value of a general-purpose robot solution.
#NobleMachines #PhysicalAI #EmbodiedAI #Robotics #NVIDIAGTC #IndustrialAutomation @NVIDIARobotics
English
Wenda Wang retweetledi
Day 3 - 1:1 with Noble Machines CEO, Wei Ding. Last day left see Moby live at Booth #941 and #3303.
📅 DM us or sign up at qrto.org/s7GYG4 to book a 1:1 VIP Assessment with CEO Wei Ding before GTC wraps — a 90-minute working session to discuss your application requirements, walk through your operations, and assess the value of a general-purpose robot solution.
#NobleMachines #NVIDIAGTC #PhysicalAI #IndustrialAutomation
English