Robots have a brain, but still no hands

Let me share something. OpenAI says they’re making robots; reportedly, GPT-6 Astra has a high success rate for simple pick-and-place, but fine insertion drops to only 10%. NVIDIA spent $12.93 billion acquiring Hugging Face, and terms like world models and embodied brains suddenly became hot.

My first reaction was still that phrase: This demo is far from mass production.

Models understanding causal relationships in the digital world doesn’t mean they understand screws, fabric, oil stains, and tolerances in the real world. Looking at those two success rates together is honest. Pick-and-place looks like a demo; insertion looks like actual work. The trouble with robots is continuous operation: today it grabs accurately, tomorrow the bin shifts two millimeters; morning goes smooth, afternoon sensors get dirty; no matter how high the single-point success rate, if it runs eight hours a day, one anomaly is enough for people on-site to curse.

I’ve only been looking at embodied intelligence for two months, but after talking to hardware makers, autonomous delivery vehicle observers, and solution providers, the industry’s old habits remain the same: PPTs talk about brains, sites lack hands and feet, and after-sales lacks manpower. OpenAI has money, models, and compute power—clear advantages. But what it needs to cross are industrial verification, BOM costs, spare parts, on-site maintenance, and customer acceptance. Especially for humanoid robots, don’t be fooled by the word "general-purpose." The more general-purpose, the harder it is to prove where the first paid scenario is.

So, advice for startups is just one sentence: Don’t rush to bet on who owns the embodied brain; first dominate a narrow scenario. Core metrics shouldn’t just look at demo pass rates; look at stable runtime on real workstations, number of exception handovers, failure recovery time, unit task cost, whether data can flow back, and whether customers are willing to pay for labor savings. Once you can run through these, then talk about world models.

As for big companies, in the short term, they look more like selling shovels. Models, compute power, and data platforms will all compete fiercely. What might be valuable in the end are teams that can integrate themselves into production lines, warehouses, stores, and hospital corridors, and have someone fix faults in the middle of the night.

The physical world doesn’t care about launch events; it cares if it’s still running the next day.