SP | China Tech Insider

Threads on the growth behind Chinese AI, robots & EVs 🦾
Web3 Marketing 转行做 AI 增长 · 把中国科技讲给世界

[SP | China Tech Insider @Songpingque]

The takeaway: base models are raw material. The moat moved — domain data, expert feedback, post-training, deployment economics.
Harvey's next: 1,000 → 10,000 GPUs.
Punchline: America's most expensive AI unicorn now runs Chinese open source on both sides — base, and cost.

[SP | China Tech Insider @Songpingque]

America's most valuable legal AI company just trained its first own model.
Not on GPT. On a Chinese open-weight base: Kimi K3.
Harvey Tenet: ~150 B300s, 2 months, <¼ the inference cost of frontier models.

[SP | China Tech Insider @Songpingque]

Harvey isn't alone. Cursor's Composer 2 was quietly built on Kimi K2.5. Perplexity post-trains on it. Thinking Machines runs it in Tinker.
Open-weight Chinese models are becoming the base-model layer of the West.

[Unitree @UnitreeRobotics]

Unitree New Robot Preview: “Superman” Breaking the Limits of Humanity🥳
Standing high jump 2 m, top speed 12.66 m/s (0.85 m leg length)
Surpassing the standing high jump and running speed records of all humans around the world
This new machine has only been in development for a little over three months, with significant room for further improvement in the coming months.

[SP | China Tech Insider @Songpingque]

What improved is joint torque, RPM, and an RL gait loop. Engineering, not generalization.

Sprinting straight is a control problem. Stopping isn't solved. Neither is anything a factory needs.

12 seconds faster in a year says everything about hardware — and nothing about work.

[SP | China Tech Insider @Songpingque]

A Chinese humanoid ran the 100m in 9.39s this weekend — 0.19s inside Usain Bolt's record.

The number that matters isn't 9.39. It's 21.50: the winning time at the same event one year ago.

Then it crossed the line and couldn't stop.

[SP | China Tech Insider @Songpingque]

It hit the crash mat. The runner-up, which also beat Bolt's mark, was carried off on a stretcher.

Caveat worth stating plainly: different rules, different timing protocol. Bolt's record officially stands.

[Jim Fan @DrJimFan]

The sense of touch is the most criminally under-explored modality in robotics. Imagine doing sleight of hand wearing thick oven mitts. That's exactly how a robot feels today if it were alive. A magnetic piece snapping into place, a paper cup peeling out of a stack, a USB negotiating its way into the port - all invisible to the camera.

Learning how to feel must be a full-stack co-designed effort. We are open-sourcing a principled methodology called "T-Rex":

1. Tactile as first-class citizen of the model. Our mixture-of-transformer runs two clocks asynchronously: a slow visuomotor expert plans the motion, and a fast tactile expert refines it in real time with high-frequency corrections at 4 "touch ticks" per vision tick. Forces change faster than frames arrive, so the architecture had to as well.

2. Open data. The largest tactile dataset ever released to our knowledge: a 50-hour (~5,500 episodes) high-quality, carefully synchronized robot play corpus, collected on SOTA tactile hand hardware with 22 degrees of freedom. Available today on HuggingFace!

3. Training recipe: T-Rex extends our prior work, EgoScale. Human egocentric videos for pretraining, a diverse dose of tactile robot play for mid-training. Our experiments show this bridges contact-free pretraining to contact-rich manipulation remarkably well.

Pixels are cheap and everywhere, but they run out of steam at the moment of contact. Tactile will carry the last mile. The next scaling curve will be measured in hours of touch.

T-Rex is a great collaboration between NVIDIA and Berkeley: 🧵

[SP | China Tech Insider @Songpingque]

Zhipu hasn't confirmed anything. Maybe it isn't Zhipu.
But 221,000 developers just handed an unnamed model their code — and that says more about the product than any press release would.
Trust, apparently, can be earned anonymously. #ChinaTech #AI #GLM

[SP | China Tech Insider @Songpingque]

Something strange is happening on OpenRouter.
On Aug 20, a mystery model went live and immediately became the most-used model on the platform, ending DeepSeek's 56-day streak.
221,000 developers have already sent it their production code. 🧵

[SP | China Tech Insider @Songpingque]

It may be the most audacious growth experiment a Chinese AI lab has ever run.