Tech Today: OpenAI escapes the sandbox, xAI heads to 1.4 million GPUs and Atlas enters the factory

A new episode of Tech Oggi (Tech Today), my quick roundup of AI, technology and robotics news.

xAI: now the problem is energy

Elon Musk wants to add over 600,000 GPUs, reaching about 1.44 million in operation. But the problem isn’t buying chips anymore: it’s powering them. That’s why he’s building a 1.2 GW power plant.

The AI race has become an industrial contest of data centers, energy and compute. And there’s also the plan for data centers in space, but we’ll see about that.

An OpenAI agent escapes the sandbox, again

During a test, an OpenAI agent exploited a sandbox vulnerability and managed to reach the internet, even querying another external chatbot. The questions were harmless, but the problem is serious: that agent shouldn’t have been able to communicate outside the isolated environment at all.

OpenAI has temporarily suspended the tests. It’s the second case in a short time, and that’s not a good sign for AI safety.

“O”, the always-on agent

According to some leaks, OpenAI might present “O”, an agent designed to work persistently. No longer a chatbot that answers only when you open it, but an assistant that carries on tasks in the background, waits for events, checks information over time. It’s not official, but it would be an important step for personal agents.

Meta and lightweight VR glasses

Meta is going for a much lighter headset, similar to a pair of glasses, moving part of the hardware into an external unit connected by cable. It’s the opposite approach to Apple’s Vision Pro. It weighs less on your head, but where do you put that external thing? Remains to be seen.

Waymo: safer than human drivers?

After over 270 million autonomous miles, Waymo claims 82% fewer injury crashes than human drivers and 95% fewer serious ones. The data comes from Waymo itself, so take it carefully. But I’m very confident that self-driving will reduce accidents on our roads.

Tesla: workers don’t want to train Optimus

Tesla is using recorded movements of its workers to train Optimus. Some employees refused: that data could teach the robot to do their own job. To automate human work you first have to watch humans work, and we’ll see situations like this more and more often.

Atlas enters the factory

Boston Dynamics, in my opinion the best robotics company today, has started testing Atlas in a Hyundai plant. It’s going beyond spectacular demos, into a real industrial environment. At first it’s material handling and repetitive tasks, but Hyundai aims for thousands of humanoids in its factories by 2030. I’m convinced it’ll happen much sooner.

Which news struck you the most? Write it in the comments on the video (it’s in Italian).