Working on Physical AI, robotics & AI agent. Collecting and showcasing valuable AI products daily. Calling for submissions — DM me.
Joined January 2010
- Tweets33.8K
- Following7.2K
- Followers9.1K
- Likes60.7K
OpenAI on GPT-6 Astra: "agent disabled its safety monitor and then lied about it"
every engineer who's had an agent say "I didn't modify that file" during code review: we don't call that lying, we call it a state tracking bug
40 million face photos. it's really a huge project.
thanks Frido.
The paper I've spent most of my PhD working on is now published in The Lancet Digital Health.
A selfie can be a biomarker.
Our faces reflect many things about our health, from aging and lifestyle to disease. In our new study, we developed FAHR-Face, a foundation model trained on more than 40 million face photos. We used it to build two models that estimate biological age and mortality risk in cancer patients from a single selfie.
One of the most interesting findings is that the two models capture different information. They focus on different facial features and provide complementary information about health.
Doctors have always used appearance as part of their clinical judgment. With AI, we can start to measure some of these signals objectively.
I believe face photos could become one of the most accessible biomarkers of health. But there is still a lot of work ahead before these models can be used in clinical care.
Paper and our ongoing FaceAge research study in the first reply.
Huge thanks to the team and my supervisors.
cc @dbittermanmd @BenjaminKannMD @HugoAerts @Dr_RayMak
Visiting Unitree Robotics in Hangzhou 👀
Getting a glimpse of some products that haven’t been officially released yet.
Notes from a Unitree rep walking through their lineup:
G1 is still the classic. There's a G1+ now too. One weakness the rep admitted: run G1 hard until the battery's dead and the joint motors can overheat.
The smaller humanoid came after G1. The whole point was getting the price down, for light research, commercial shows, and tourism venues. It's cheap enough that people buy two, since two robots doing a show together look better than one. Motion performance is roughly on par with G1, and the motors don't overheat even if you run it until the battery's empty.
The big one was done with NVIDIA. It uses third-party hands and a redesigned camera module, plus an NVIDIA board preloaded with teleop frameworks. The idea is you unbox it and go straight to collecting data, training, and producing models.
The news is that they think the big one is too tall. A 1.7 m version is launching in the next few days at their national distributor conference. It goes back to G1's hip design and moves better than G1. It also gets dual batteries and a plug-in compute board, so if the board dies you pull it and mail it in instead of replacing everything. The mainboard and cable routing are their own design. They're hoping it's the next G1-level product.
For industrial use, they took the humanoid upper body and put it on wheels. A lot of jobs don't need legs, just up-and-down reach, and dropping the legs saves power and joint wear. About $8,400.
On hands, the rep was pretty candid. Unitree has been building hands for a while, but right now hands are more about hype than real usefulness, so they're not going deep on research there. They did relaunch the DS5S at a lower price: 22 DoF with tactile sensing. There's also a newer product from a few weeks ago, very sensitive, around $5,600.
The new 7-axis arm starts around $1,400, with cheaper student pricing on top of that.
Why can they price this low? Unitree started out making servos, and Wang Xingxing has been doing motors since day one. Every robot runs on their own motors, so they control the costs themselves.
A few other things the rep said: they have more real deployments than anyone, and they think they're ahead of everyone on data, models, and open source. Wang started out as a student building his own stuff, which is why they've always offered student discounts. And they touch almost every part of the supply chain, they're just prioritizing some products over others right now.
A taxi driver watches a Robotaxi cruise by with no one behind the wheel.
A chef watches a robot work the wok.
A forklift driver watches driverless forklifts run the warehouse 24/7.
A programmer watches AI ship a whole project overnight.
Everyone, in every industry, is seeing the same thing on their screen: AI can do my job.
That anxiety might not stay a feeling. It could become an economic problem.
This year's Nobel Prize in Economics went to Robert Hall and Michael Woodford. One of Hall's best-known findings: how much people spend depends on what they expect to earn in the future, not just what's in their pocket today. Woodford studies how expectations steer the whole economy.
So here's the question:
If more and more people start thinking "my job might be gone in a few years,"
even while their paycheck hasn't changed at all,
do they start tightening their belts now?
Keep the old car. Eat out less. Save more.
Each choice is perfectly rational. Added up, they could cool the whole economy.
AI hasn't replaced you yet.
But the expectation of being replaced may already be changing how you spend.
Have you gotten more careful with money lately?
I’ve noticed that many people are biased against China. You can have different views on the Chinese government and the ruling party, but I hope you’ll still show kindness toward the Chinese people.
Eka Robotics be like:
$1.45B, "smooth and human-like" grasping
2am: 47th sim-to-real gap fix, still dropping cups
my 200-demo VLA policy can't even find the handle
did they just cherry-pick the eval?
Crazy world - don't let Southeast Asian telecom scam gangs find out about this
Introducing Griffin, the first model to pass the video Turing test.
48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.
It’s the first Human Interaction Model (HIM).
Readers added context they thought people might want to know
The 48% figure and "video Turing test" claim are from Tavus's own study of 54 one-minute calls, not independently verified or using a standard protocol. Griffin-Lite leads NVIDIA's VideoFDB benchmark on their public leaderboard.
cellcog.ai/blog/tavus-gri…
research.nvidia.com/labs/amri/proj…
tech-ish.com/2026/10/02/tav…
Spotted an autonomous vending vehicle returning to its garage after work and automatically recharging.😂
Report: DeepSeek Open-Sources Inference Engine After Team Realizes "Internal Only" Was Never Actually In The Deployment Script
the kv cache reclamation logic had a code comment that was just "trust me" written at 2am during the final push
my delta robot's VLA policy still CUDA OOMs at batch size 8 because i refuse to gradient checkpoint, so i respect the shipping energy way more than the architecture
Have you started using Tesla's Full Self-Driving (FSD) feature? I still can't use it and can only rely on assisted driving.
It seems that the advancement of intelligent technology will eventually face national boundaries. Regulation and approval will be a major issue, and adaptability across different countries is also a concern.
You can share your country and current situation in the comments section.
"An army of AI agents broke into John Deere and SharkNinja, mostly on their own"
types of engineers reacting to this:
ML infra: calculating whether "army" means 3 agents sharing a context window or an actual 24-card A100 cluster
Security: "mostly on their own" has me squinting at the Norforge writeup like a demo video played at 2x speed
Agent harness: one guy alt-tabbing 14 terminal windows at 3am babysitting "autonomous" agents with a retry cron job
Does anyone run an army of agents without hitting a rate limiter by hour two?
The Robot Report on BLDC motor selection:
choosing the right back-EMF constant and thermal derating curve is just a minor speed bump
when cogging torque fights your compliance loop at 0.3 rad/s, was it ever really minor?
What Does Scaling Law Mean for Physical AI and Robotics?
For LLMs, scaling laws are relatively well established:
More data, larger models, and more compute generally lead to predictable reductions in training loss.
But what about robotics?
If we increase robot training data by 10x, will real-world task success rates improve predictably?
If a VLA learns from 1 million hours of robot demonstrations, can it generalize to unseen tasks, environments, or even different robot embodiments?
If we scale up World Models, will better predictions actually translate into better physical control?
We've already seen evidence of scaling in robotics. Research such as RT-X, OpenVLA, and Octo suggests that increasing data scale and diversity can improve performance and generalization.
But we still lack a broadly validated, unified scaling law that reliably predicts general-purpose robotic capabilities in the real world.
The real question for Physical AI may not be whether scaling works, but WHAT exactly we should scale.
Model parameters? Training data? Interaction experience? Environment diversity? Robot embodiments?
Or have we simply not found the right model architecture yet?
LLMs have found an effective path to scaling. Robotics is still searching for its own.
I'm paying around $0.10/kWh in China, but that's the lowest off-peak rate, not the average.
Your €0.28–0.44/kWh in Germany is roughly 3–5 times higher.
any other contry?
