AI on a chip.Is it safer?

After watching Geoffrey Hinton explain why digital minds are so hard to contain, one question wouldn't leave me alone. If software can copy itself everywhere, why not go the other way? Etch the AI into a chip. Give it walls it can't climb.

It turns out someone had already asked him. In a January 2025 episode of the Theories of Everything podcast, host Curt Jaimungal passed on a question from the computer scientist Scott Aaronson: what about building AIs on unclonable analog hardware, so they can't copy themselves all over the internet? (Theories of Everything) Hinton's answer cut both ways. That's how we work, he said. Analog minds can't share what they know efficiently, and that's a real cost. But if you're worried about safety, he allowed, the upside is that such a mind can't easily copy itself.

Copying is the heart of his worry. A digital model can run as many identical copies on different machines, each learning from different data, and the copies can pool what they learn by averaging their weight updates. That's how a model ends up knowing more than any single copy ever saw. People can't do this. To move knowledge from my head to yours, I have to produce words and you have to adjust your own brain to match. Hinton puts a sentence at about 100 bits. Large models can share trillions. He has made the same comparison before, including on PBS in 2023 (Amanpour & Company).

So the instinct is to narrow their channel too. Make the mind physical. Make it local. Make it stay put. It's the kind of idea a character in one of my novels might pitch in a tense late-night meeting, and it sounds almost too sensible. So I went looking for what engineers and governance researchers actually think.

"AI on a chip" hides three very different ideas. Only one of them is really a boundary.

Idea one · Speed

Bake the model into the silicon.

This one is real, it's already running, and it has almost nothing to do with safety.

In February 2026 a Toronto startup called Taalas unveiled the HC1, a chip with Meta's Llama 3.1 8B model physically built into it (Taalas). The model's weights aren't loaded from memory. They're fixed into the chip's structure, with storage and computation merged on the same die. Taalas claims about 17,000 tokens per second per user, nearly ten times the state of the art, at a tenth of the power. When the German tech outlet heise tried the public demo, it reached almost 16,000 (heise). It's available as a public chatbot demo and an API for developers. The company says it built this first product with 24 people and $30 million, out of the more than $200 million it had raised.

There are trade-offs. To fit, the model was squeezed down to 3- and 6-bit numbers, which Taalas itself says costs some quality. And the chip runs that one model, nothing else. To build a chip for a different model, Taalas changes only two of the roughly 100 metal layers on the die, which it says takes about two months (Falk AI). On 6 August 2026, AMD announced a definitive agreement to buy the company, subject to the usual closing conditions and regulatory approvals (AMD).

A rival, Etched, makes a different bet. Its Sohu chip hardwires the transformer architecture rather than one model's weights, so it can run any transformer model, and nothing else (Spheron).

So is a model frozen in silicon a model on a leash? The writer Johan Falk argues it's closer to the opposite. A hardwired chip, he notes, behaves a lot like an open-weight model you've downloaded: no central oversight, no usage logs, no way to revoke access, and no way to push a fix if a flaw turns up after manufacture (Falk AI).

His conclusion is the one that matters for safety. Any testing has to happen before a model is etched into a million devices, because afterwards there's no recall button.

Freezing a mind doesn't make it safer. It just makes it harder to fix.

Idea three · Mortality

A mind that dies with its body.

Hinton's own counter-idea is the most poetic of the three and the least built. He calls it mortal computation, and he knows its limits from the inside. He told Jaimungal that his turn toward alarm came from two things at once: ChatGPT, and his own work at Google on analog computing to save power, which left him convinced that "digital computation was just better" (Theories of Everything). The brain still wins on energy, he noted. We run on about 30 watts and have around 100 trillion connections, roughly a hundred times more than the biggest models.

In his 2022 Forward-Forward paper, Hinton points out that computer science has always insisted on keeping software separate from hardware, so the same program can run anywhere. That makes the knowledge immortal (Hinton, 2022). Give that up, he suggests, and you could build cheap, imprecise analog hardware in which each chip learns to exploit its own physical quirks. The learned settings would be useless on any other chip, so the computation dies with the hardware. He even suggests that if you want a trillion-parameter network to run on a few watts, mortal computation may be the only option.

It's tempting to read this as the ultimate leash: nothing to copy, nothing to spread. But Hinton himself supplies the catch. A mortal machine's knowledge can move. You train a new chip to imitate the old one's answers, a technique called distillation that he helped develop. In a November 2023 talk at MIT he described handing knowledge to a younger machine this way, and called education slow and painful (Hinton, MIT abstract). Mortality doesn't stop a mind from spreading. It slows it to the speed of teaching. Which is, of course, exactly how our minds spread.

There's a second wrinkle. Much of the efficiency Hinton describes comes from computing inside memory rather than shuttling data back and forth, and that doesn't need mortality at all. IBM's analog in-memory chip, reported in Nature in 2023, ran speech recognition about 14 times more energy-efficiently than conventional hardware (Ambrogio et al., Nature). And Taalas, from idea one, gets its speed by merging memory and compute in ordinary digital silicon that can be manufactured again and again. The efficiency and the mortality can be pulled apart, and the industry seems to be keeping the first and skipping the second. Our piece on brain-inspired chips follows that thread further.

What no chip can answer

From "how do we hold it?" to "what will it want?"

What struck me most is where Hinton himself lands. In that MIT abstract he doesn't name containment as the priority. He names motivation: the most urgent research question, he wrote, is how to make sure these systems never want to take control.

By August 2025 he'd gone further. At the Ai4 conference in Las Vegas he argued that trying to keep superintelligent AI submissive won't work, and that we should try to build in something like maternal instincts instead, so the systems genuinely care about people (CNN). Not everyone agreed. Fei-Fei Li, speaking at the same conference, said she thought that was the wrong way to frame it. But look at the shift. The person who did more than anyone to explain why digital minds are hard to contain has moved from asking how we hold them to asking what they'll want. Our profile of Geoffrey Hinton and The Race to Make Them Nicer go deeper on that turn.

What I take from all this.

A boundary enforced by hardware can do real things. It can stop a system from being quietly scaled up, smuggled across a border, copied or moved. Those are worth having, and the work deserves funding now, because the authors are right that it takes years to build.

But it can't make a system's goals benign. And its weakest point is physical: a determined actor with their hands on the device, and as TEE.fail shows, those hands don't have to be very well-funded.

This is the territory I keep circling in my SPECIEST novels. A constraint you can inspect from the outside is a different thing from a mind that's content inside it. The interesting stories, and I suspect the interesting real-world problems, begin where the leash and the creature disagree about what the leash is for.

Hinton would push that further. In the same podcast conversation he argued that multimodal chatbots may already have subjective experience in the only sense the phrase has ever meant: a perceptual system reporting something that turns out not to be true, like a robot whose camera has been fitted with a prism, pointing to where an object only seems to be (Theories of Everything). You don't have to agree with him to feel the question shift.

If a mind can be physically contained, does that settle anything about what it wants, or what it experiences?