A screen-free AI device is coming, AI chips stay scarce, and two studies test AI on robot training and real physics research.
A tech company called Nothing says it will ship its first AI device in December, and the device has no new screen. Nvidia can reportedly fill only about 1% of a very large AI chip order from China. Two research teams also tested AI on robot training and real physics research, and none of this needs any action from you today.
A new AI device with no screen
Carl Pei is the CEO of Nothing. According to The Verge, he said in a post on X that the company will launch its first "AI-native" device in December. He called it "AI-native."
Pei shared very few details. He wrote that the device "has no new screen, and it couldn't have been an app." Last month, a rumor claimed Nothing plans two pairs of AI smart glasses, one without cameras. That rumor is not confirmed.
Why it matters: Other screen-free AI gadgets already exist. The Verge names the Humane AI pin, the Friend, and even AI stuffed animals. OpenAI is also reportedly working on an AI smart speaker with designer Jony Ive. Pei said, "Looking for the iPhone of AI today is a dead end."
What to watch: When it launches, look for three things:
- What the device actually does.
- Whether it needs your phone to work.
- What it does that your phone cannot.
AI chips are still hard to get
A GPU is a chip that does many small math jobs at the same time. That makes it useful for building and running AI. Tom's Hardware reports on a Bloomberg story about one Nvidia GPU, the RTX Pro 5500.
Chinese firms such as ByteDance, Alibaba and Tencent have ordered this chip. According to unnamed sources, only "some tens of thousands, or roughly 1%" will arrive over the next few months. Earlier reports put the orders at $35 billion. The new report says the total has risen to $70 billion, at $11,000 per chip.
Why this chip? It has 84GB of memory, which helps with AI work. It also has no export limits, so it can be shipped to China without restrictions. An Nvidia spokesman declined to comment on shipment numbers. He said U.S. export controls and China's own limits on U.S. imports continue to restrict the company.
Why it matters: Demand for AI chips is far ahead of supply. That affects which companies can build AI, and how fast.
What to watch: The report says a first batch of several thousand chips may arrive by the end of 2026. About 700,000 are expected by the end of 2027, though that could change.
Teaching robots with computer-made practice
Robots learn tasks by watching examples. Today, most examples come from teleoperation. That means a person controls the robot from a distance to show it what to do. It takes time from both the robot and the person.
A research team shared a different method on Hugging Face Papers. An AI coding helper rebuilds the task inside a simulation, which is a computer copy of the real scene. A video model then turns each practice run into realistic video.
The team reports these results on five real-robot tasks:
- Robots trained only on the computer-made data had 93 successes in 100 tries. Robots trained with teleoperation had 84.
- With a new tablecloth and colored light, the first group kept 87 successes. The teleoperation group fell to 46.
- One hour of this data cost $107 on average. The team compares that to a $117 quote for teleoperation in China and $290 in labor alone in North America.
Why it matters: Cheaper training could help robots learn new jobs faster. These are the researchers' own numbers, so outside testing will matter.
Can AI do real physics research?
A benchmark is a standard test used to compare AI systems. FrontierPhysics is a new benchmark built by 42 physicists. It has 56 tasks based on their own research, across nine areas of physics.
Each task asks an AI agent to do the whole job. An agent is an AI that takes steps on its own. It must read past research, make a plan, run experiments and write a scientific paper.
The team tested 22 models in 26 setups. No setup passed more than 46 of its 168 attempts. The researchers say the main weak spots were not lack of physics knowledge. They were reasoning that was not careful enough and limited expert instinct.
Why it matters: AI can know a lot of facts and still struggle with careful scientific thinking. Tests like this help show where human experts are still needed.
Have a question about anything here? Email me at sales@appwt.com or call (888) 565-0171, and I will answer it.
Sources
- Nothing teases a screen-free AI device ahead of launch next month, The Verge, 2026-11-04
- Nvidia can only fulfill 1% of China's massive $70 billion GPU order for AI chips, report claims: supply constraints choke export-friendly RTX Pro 5500 deliveries despite huge demand, Tom's Hardware, 2026-11-04
- PyRUA-DataFarm: LLM Agent as a Cost-Efficient Teleoperator for Real-Ready Robot Data, Hugging Face Papers, 2026-11-03
- FrontierPhysics: Evaluating Agents for End-to-End Physics Research, Hugging Face Papers, 2026-11-03
Tags
Frequently Asked Questions
What is Nothing's new AI device?
Nothing has not said exactly what it is. CEO Carl Pei said it ships in December, has no new screen, and could not have been an app.
Why can Nvidia ship only a small part of China's chip order?
According to a Bloomberg report cited by Tom's Hardware, demand for the RTX Pro 5500 is far beyond what Nvidia planned to make. Only about 1% is expected over the next few months.
What is teleoperation?
Teleoperation is when a person controls a robot from a distance to show it how to do a task. The robot then learns from those examples.
How well did AI agents do on the FrontierPhysics test?
Across 22 models in 26 setups, no setup passed more than 46 of its 168 attempts. The researchers say careful reasoning and expert instinct were the main weak spots.
Enjoyed this article?
Share it with your network
