
PR Advisor @AntGroup @JohnsHopkins MA & @Mizzou J-school alumn Ex @Baidu_Inc @BitFuFuOfficial @SyncedTech Opinions are my own
Sunnyvale, CA
Joined January 2014
- Tweets5.2K
- Following2.7K
- Followers1.7K
- Likes5.7K
Pinned Tweet
Just redesigned the China AI Index site with a brand-new visual identity (and a DARK MODE). Hope you like the new version!
chinaidb.com/
Tony Peng retweeted
It’s really important to tell Shenzhen’s story to the world, because it is still so unknown. Though an unconventional roll for an investor, I helped with this video and other material for APEC. I hope this is an effective intro to Shenzhen.
Tony Peng retweeted
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
Tony Peng retweeted
Four months after we launched Manus, I was invited to a dinner with President Macron. I ended up seated directly across from Jensen Huang.
He was half an hour late, and Macron asked him to give a toast. Jensen picked up his wine glass and turned it into a fully improvised speech — in front of a table of European politicians and founders. It was electric. The charisma of a top founder, on full display.
I typed the whole thing into my Notes that night. Still one of the sharpest takes I’ve heard on software, AI and Europe.
What a pity. What a loss.
“There are decades where nothing happens. And then there are weeks where decades happen.” Love this opening.
1
120
Tony Peng retweeted
News: The US intel community is still working to understand the relationship between a Chinese businesswoman who had an affair with Nvidia CEO Jensen Huang and the Chinese government, with multiple sources telling CNN that the intelligence collected on her to date does not prove that she was acting as a spy for Beijing.
cnn.com/2026/11/12/politics/…
The photo was taken in 2019 at a Nvidia GTC press conference in San Jose.
As a then tech reporter, I asked Jensen Huang two questions: How would Nvidia deal with its GPU stockpile? And was Nvidia worried about other AI chip competitors?
Looking back from 2026, it’s almost surreal. Nvidia’s chips have become probably the most sought-after product in the world. And the company remains the leading player in AI chips.
Yet the biggest change is that the man behind all this just resigned.
I remember reading an article in 2024 (or 25) describing Jensen as Silicon Valley’s new poster child. He was different from the typical tech billionaire: humorous, humble, friendly. He loved street food, hanging out with people, and eating fried chicken with Korean business leaders. And he (seemingly) loved his wife.
I work in PR, so I know how a public image like this can be orchestrated. But still I believed Jensen was different.
For many Asians, Jensen was also a huge inspiration. His story made you believe that anything was possible.
That’s why all the saga, allegations in the past week have been so disappointing. Not just because a billionaire made mistakes, but because so many people held him in such high regard. Turns out he isn’t different after all.
He should resign. I imagine even Jensen wouldn’t want the Nvidia’s brand to be forever entangled with a personal scandal.
But Nvidia would unquestionably be losing an huge asset, an enormous advantage. It was Jensen’s prescient bet on accelerated computing that helped make today’s AI boom possible. I’m not sure any successor could replicate that success. Just look at what happened to Apple after Steve Jobs was ousted.
And imagine what would happen if Jensen joined another AI company, another chipmaker, or started a new venture of his own.
As someone who closely follows China’s AI industry, there’s also a lot for me to process.
Jensen has been one of the most prominent voices supporting open models and speaking highly of Chinese AI labs. Whatever business interests motivated those positions, he was willing to say things publicly. Losing that voice and support would matter. It could also mean a more challenging geopolitical environment for Chinese AI labs. There are a lot of unknowns ahead.
(Above are my own opinions)
1
203
Tony Peng retweeted
NVIDIA announces CEO transition. Read the release: nvidianews.nvidia.com/news/n…
Tony Peng retweeted
Stronger general-purpose models can still leave robots short of the capabilities they need most. Zhihu contributor 沈宇军 (YuJun Shen), whose team develops LingBot-Touch @AntLingAGI, argues that progress in Claude Fable 5.2 and GPT-6 Astra should sharpen a different question: what capabilities should a robot’s model be built around?
His answer starts with physical sensing, interaction feedback and continuous adaptation. A model may have more software engineering capability than a robot needs while lacking the physical understanding required for precise manipulation. The goal is the right mix of capabilities, not a particular parameter count.
Similar movements can produce different outcomes
When a robot twists a bottle cap, successfully opening it and letting its fingers slip can involve very similar hand trajectories. The contact forces differ. Visual motion and language descriptions alone can miss the distinction, so the robot must use physical feedback to adjust its actions.
Sensing is imperfect, and the world keeps moving
Distance and touch measurements contain errors. Models need to interpret those imperfect signals, particularly where sensors are still immature. The author presents LingBot-Depth and LingBot-Touch as work toward better physical perception.
Robots must also keep receiving information and revising actions while they operate. The author argues that simultaneous speaking and reasoning in language systems does not yet solve continuous, real-time processing across the physical modalities robots need.
Decide what to record before scaling the data
More training data cannot recover information that was never captured. Does the input preserve what is needed to interpret sensing errors? Can the output represent fine physical interactions? Do task recordings capture changes that require the robot to adapt?
Shen sees suitable data formats and standards as unresolved questions for both general-model developers and robotics companies.
His team therefore favors models designed around physical tasks from the outset, while drawing on general-purpose models where useful. He leaves open which route will ultimately work better: adapting increasingly capable general models or building specifically around robotics. Both must solve the same physical constraints.
The benefits of using JEPA, an approach @ylecun has championed for years, in LingBot-Touch:
1) Avoid wasting model capacity reconstructing mostly inactive tactile sensor readings.
2) Learn from thousands of hours of tactile data without manually labeling every contact event.
3) Help learn representations of physical contact that generalize across different sensors.
What changes when robots can feel—not just see? 🤖✋
Meet LingBot-Touch, our open-source tactile foundation model towards dexterous manipulation, pretrained on nearly 5,000 hours of real tactile data.
Through unified heterogeneous tactile modeling, LingBot-Touch supports transfer across sensors, dexterous hands, tasks, and action models.
Watch it in action. 👇 #LingBot #EmbodiedAI #Robotics
96
LingBot-Touch
A Foundation Model for Heterogeneous Tactile Representation Learning
huggingface.co/collections/r…
Tony Peng retweeted
i hate to be the one to say this, but as an asian woman myself, i feel really odd scrolling on x today...
seeing other chinese women play into the "honeypot" jokes for views. i know it's meant to be funny, but to me it just reinforces the racist and sexist stereotypes we already deal with, so let's NOT set us back even further by sexualizing ourselves...
Tony Peng retweeted
大火的“中国女间谍”,符号背后的隐喻:中国女性被物化成单纯的性吸引力和对物质的无度索取;中国技术的进步被单纯的认定为是窃取与模仿;中国男性的主体性是孱弱以及缺失的;而中国的国家形象无疑是神秘、隐蔽、充满诱惑,同时又具有危险性的,被想象的东方女性。
说到底,西方到现在,依旧不承认我方能够通过自身的劳动、知识、制度和社会组织,创造出属于自己的价值。
Tony Peng retweeted
LingBot Touch 🔥 A JEPA based tactile foundation model for robotic manipulation, from Ant Group Robbyant team.
- Tiny + Small
- Apache 2.0
- Unifies 3 tactile modalities: magnetic, piezoresistive & visuotactile
- Shared Transformer for heterogeneous tactile sensors
- Learns contact structure and temporal dynamics
LingBot-Touch is a JEPA-based tactile perception model for dexterous manipulation tasks (applying toothpaste, unlocking a lock).
It works across 3 sensor types and significantly improves accuracy by integrating with robotic control models, from diffusion policies to VLA.
What changes when robots can feel—not just see? 🤖✋
Meet LingBot-Touch, our open-source tactile foundation model towards dexterous manipulation, pretrained on nearly 5,000 hours of real tactile data.
Through unified heterogeneous tactile modeling, LingBot-Touch supports transfer across sensors, dexterous hands, tasks, and action models.
Watch it in action. 👇 #LingBot #EmbodiedAI #Robotics
1
154
Tony Peng retweeted
Introducing Ling-3.2-flash, Ant Ling's native multimodal model: 560B total / 25B active parameters.
Stronger complex visual understanding, long-horizon multimodal agents, 3D understanding & interaction, and medical intelligence.
Outscores Opus 5 (max) in release evaluations:
- OmniDocBench1.5: 92.21 vs 90.56
- ZEROBench(Sub): 43.71 vs 39.52
- Vision2Web: 85.86 vs 80.20
I left my dog Malatang (a shih tzu) with a carer for three months as I went back to Shanghai for work.
He’s always been introverted and rarely played with other dogs whenever I walked him.
And then yesterday, his carer sent me this…😂
As China’s AI industry races to build more powerful models and apps, its companies and users are increasingly giving AI softer, more familiar, and sometimes human-like identities.
This contrasts with Silicon Valley’s AI products. ChatGPT, Gemini, Claude have largely stuck to abstract, geometric logos and stay away from anthropomorphized characters.
I don’t have an answer yet.
open.substack.com/pub/recode…
96































