There is an episode of The IT Crowd where Moss and Roy convince Jen that a small black box with a blinking light contains the entire internet. She carefully carries it around like some impossibly important piece of technology.
Intern 2 gave me a little of that feeling.
I was taken aback by it in the way I imagine someone encountering ice for the first time might be. There was wonder, but also a little fear. I knew broadly what was happening inside this thing, yet having an AI agent physically sitting there on my desk still made it feel strangely unfamiliar.
This isn’t a review, not yet anyway. I’ve only had it running it a few hours as of writing.
As a first encounter, though, it immediately gave me something else to think about.
AI usually lives behind a browser tab for me. Suddenly, I had AI in a pyramid.
Intern 2 is basically a small, always-on computer built to give an AI agent its own dedicated hardware, microphones, speaker, and memory. You can talk to it directly, message it through Telegram, Slack, or Discord, or set tasks to run on their own in the background.
What actually counts as physical AI
Autonomous defines physical AI as AI that can sense and act in the physical world. The key difference is consequence. A chatbot can misunderstand a coffee cup. A physical system can actually knock it over.
Embodied AI is close enough to overlap: it’s the research idea that intelligence is shaped by having a body and interacting with an environment, while physical AI is the broader, more product-facing term.
Autonomous actually calls it a plug-and-play agent server, not a robot. It can hear me, speak back, and use lights to communicate what it is doing, but the tasks I gave it were still overwhelmingly digital.
Putting AI into a physical object gives it presence. Giving it agency over the physical world is another step entirely.
Having AI in the room feels different
The first thing I did was talk to it, which immediately made me aware of my own voice.
I do not have a particularly thick accent, but voice assistants still give me a tiny bit of anxiety. There is always that suspicion that I will have to repeat myself because the machine heard something else.
Intern 2 responded quickly enough that the worry stayed mostly theoretical. Naturally, I went into its voice settings next. A few options were unexpectedly sultry, which made me raise an eyebrow for reasons that probably say more about me than the device.
Still, talking to an object beside me felt different from typing into a browser. Not necessarily smarter or more useful. Just more present.
Physical AI probably arrives in stages
Eventually, I tried making Intern 2 useful. I set up a crawler to gather consumer-tech news from the previous 24 hours and ran it several times while troubleshooting the workflow.
By roughly the fourth attempt, I hit a usage limit.
I shrugged and unplugged.
It did make for a strangely appropriate first impression, though: we can now give an AI agent dedicated hardware and somewhere permanent to live, while the intelligence behind it is still metered separately, on a clock you don’t control. Intern 2 can even be configured around different model endpoints, including self-hosted ones.
Maybe that is how the road toward physical AI begins. Not with a humanoid robot suddenly walking through the front door, but with increasingly capable software gaining a presence in the world before it gains meaningful control over it.
I am nowhere near ready to say whether this particular AI pyramid is any good.
For now, I am more interested in the strange fact that there is an AI pyramid sitting on my desk at all.
