AI's Silent Invasion

AI Is Moving Into the Physical World

Prologue · Rural North America, Year 2040

The ice took the power lines down a little after four, and by dark the valley had gone quiet in the old way — no glow from the road, no hum off the pole. Maddie found her grandmother in the kitchen anyway, lit amber by the attendant’s standby lamp. The machine had backed itself into the corner to spare its cells and was still reading aloud, somewhere in the middle of a chapter, the way it had been when the power went. It hadn’t stopped. It had never occurred to Maddie that it might. She set the kettle on the propane ring and the attendant told her, unasked, that her grandmother had taken the six o’clock dose, had asked after the dog twice since, and that the dog was under the porch and dry. The old woman’s data card sat in its slot behind the machine’s shoulder — a thing the size of a stick of gum, scuffed soft at the corners from twelve years of moving between this house and the clinic and the back of an ambulance and home again. Everything the attendant knew about the woman in the chair lived on that card and nowhere else, and in twelve years it had never once left the county.

Sixty miles south the towers were dark and the data centers were doing whatever data centers do in a storm, and none of it mattered here. That was the part her grandmother still remarked on, on the days she was clear enough to remark on things: that there had been decades when the clever things in a house were only ever as clever as the connection to the cloud, when a cut power or Internet line left you alone in the dark with a box that was unresponsive — silent as death. She had lived through all of it — the years when you asked a question and waited, and a building the size of a town burned power somewhere to think about it, and the answer came back half a second later and felt like magic. Maddie could not really picture it. To her, the intelligence in a room had always been like the warmth in a room: a property of the place you were standing in, present or not, and yours. She poured the tea. Outside the ice kept coming down, and in the corner the attendant turned a page that nobody had to go and fetch.

On August 6, AMD — the company whose processors sit inside a great many laptops and game consoles — agreed to buy a twenty-five person startup in Toronto called Taalas. The price wasn’t disclosed. Outside the semiconductor trade press, almost nobody noticed. [The End]

---

AMD acquired Taalas this week

This acquisition deserved more attention than it got, because the thing Taalas built quietly undoes an assumption most of us have absorbed without ever examining it: that artificial intelligence is something that happens somewhere else.

Right now, when you ask an AI assistant a question, your words travel to a warehouse. A building full of computers, often hundreds of miles away, drawing as much electricity as a small town. The answer travels back. You wait. You need a connection. Someone, somewhere, pays a subscription. And your question, whatever it happened to be, left your house.

Taalas built a chip that does away with all of that. Which sounds like a straightforward improvement, and in most respects it is. But it carries a second consequence that is less comfortable. Intelligence that needs no network and almost no power does not stay put. It spreads — into vehicles, medical equipment, factory lines, traffic signals, the walls of buildings, things you own and things you merely stand near. And because it is etched into physical silicon, it arrives permanent. It cannot be corrected, updated or recalled without someone physically retrieving the hardware.

So there are two stories here, and they are the same story. One is about capable AI finally reaching the rural clinic, the ambulance, the disaster zone and the deep-space probe — places it has never been able to go. The other is about unaccountable, unmodifiable judgment quietly becoming part of the physical furniture of everyday life. Both follow from the same property. Neither cancels the other.

The warehouse and the kitchen

To understand why this is hard, you need exactly one idea, and it isn’t a technical one.

An AI model is a very large collection of numbers — billions of them — that together encode everything the system has learned. To produce one word of an answer, a computer has to consult essentially all of them.

Today’s chips can’t hold that many numbers inside themselves. So the numbers sit in separate memory chips nearby, and the processor fetches them.

Picture a chef in a kitchen whose ingredients are all in a warehouse across the parking lot. To make one bite of food, she runs to the warehouse, hauls back the entire pantry, takes one small action, then runs back. Then she does it again for the next bite. And again.

That running back and forth — not the cooking — is what makes AI slow and expensive. It is why those warehouses need so much electricity. It is why you wait.

Taalas built a kitchen with the ingredients cast into the walls.

Instead of storing the model’s numbers in separate memory, they etch them permanently into the physical wiring of the chip itself. The numbers are never retrieved, because they were never anywhere else. The model doesn’t run on the chip. The model is the chip.

Caption: Two ways to run an AI model. On the left, every single word requires hauling the entire model out of memory and back. On the right, there is nothing to fetch — the model is the circuit. Throughput figures were published and measured by Taalas, not independently verified.

Their first product holds an open-source model made by Meta and produces text roughly seventy times faster than a high-end datacenter graphics card doing the same job, while drawing about as much power as a couple of old-fashioned light bulbs.

Those figures come from Taalas itself and haven’t been verified by anyone independent, so hold them loosely. But the chip is real, it works, and the idea underneath it is sound.

The catch, stated honestly

There is a real cost to this, and it is a large one.

The chip can only ever run that one model. Not “AI” in general — one specific version of one specific system, frozen at the instant it was manufactured. You cannot update it. You cannot teach it anything. You cannot swap in a better one.

If something smarter arrives next year, this chip will not care. It will keep doing the one thing it does, brilliantly, forever.

This is why the technology looks strange in the very warehouses it was designed for. Those places want flexibility above all else, because AI models improve constantly. Committing expensive hardware to one frozen version, in a field moving this fast, is a genuinely questionable bet.

But take the same chip out of the warehouse and put it into the physical world, and the objection quietly evaporates — because out here, we have always lived with things that do one job and never change their minds. We call that equipment.

From here on — Everything above is reported fact. Everything below is speculation — a plausible path for a technology whose properties point somewhere specific, not a prediction of what will happen. AMD has announced no consumer product, no timeline, and no customers.

First stop: the slot in your laptop

The natural first destination is a slot in a computer — something like the card slot on a camera, or the cartridge slot on a game console.

That comparison is closer than it sounds. Nintendo Switch game cards use the same fundamental technology Taalas uses to hold its numbers: information physically stamped into silicon, permanent and unwritable. This is the cartridge era returning, with intelligence in the cartridge instead of a game.

There is an engineering reason a slot could actually work here, and it is counterintuitive. Normally, plugging an AI accelerator into a small port fails because you cannot move data through the port fast enough. But here the model never leaves the card — so the only things crossing the connector are your question going in and words coming out. That is a trickle. The connection barely has to do anything.

The power budget works too. A card slot can supply only a few watts. But a warehouse has to serve thousands of people at once, and you are one person. You read at a handful of words per second. Even a heavily throttled card would produce text faster than you could possibly consume it.

What that would actually feel like

Your assistant works on an airplane. In a basement. On a rural road with one bar of signal. In a country you are visiting. There is nothing to connect to, because there is nothing to connect to.

It answers instantly — no spinner, no thinking indicator. The gap between asking and answering collapses below the threshold where you notice it.

Nothing you type leaves the building. Not as a policy, not as a privacy setting you have to trust, but as a physical fact: there is no network involved.

You pay once, because there is no service on the other end to bill you.

And when you want a better one, you buy a card rather than a computer. Twenty dollars, or a hundred, every couple of years. Specialized cards become viable in a way subscriptions never made economical — a card that excels at medical terminology, or legal drafting, or translation between Thai and English, or reading nineteenth-century handwriting. Small markets a giant cloud service would never bother to serve well.

The honest downside is that your card knows nothing that happened after it was manufactured. It will not know today’s news, and it will never get smarter. For grammar and reasoning and translation, that hardly matters. For other things it matters a great deal. Most people would end up using both.

Replaceable, and additive

Once you accept a slot, a second and more interesting idea follows: why only one?

Instead of a single model that does everything adequately, you stack several cards, each excellent at one thing, and a small coordinator on the host machine routes each question to whichever card should answer it. Some questions go to one card. Some are answered by three working together.

This turns out to matter enormously for anything safety-critical, and for an unexpected reason. It means each piece can be certified on its own.

A plausible slot architecture. The important column is the last one: each class of card is owned and updated by a different party. Speculative — no such standard exists today.

Look at the bottom row of that diagram. Five kinds of card, five different owners. The general intelligence comes from a model vendor. The specialty comes from someone who understands that specialty. Local knowledge comes from whoever operates the machine. The rules come from whoever is accountable when something goes wrong. And in machines that work closely with one person, a card that belongs to that person rather than to the machine.

Today, an AI system’s abilities and its rules are hopelessly tangled together inside the same pile of numbers, which is why “whose rules is this thing actually following?” is a question almost nobody can answer. Pull them apart into separate physical objects and the question becomes answerable — and auditable.

There is also a precedent that makes all of this less exotic than it sounds. Aviation has run on modular certified hardware for sixty years. The boxes in an aircraft are line replaceable units, each with a part number and its own approval, swapped at the gate by a technician in minutes. Nobody recertifies an entire aircraft because a radio was replaced. A model card is simply a new kind of box. Medicine works the same way through device clearance; so does the automotive industry through component safety ratings.

Every one of those systems assumes a part that behaves identically forever. Which, until now, no AI system has ever been able to promise.

Where this actually matters

The laptop is the visible case. The important ones are the machines you would never think about.

Everything below follows from three properties working together: it needs no network, it responds in milliseconds, and it runs on almost no power. Those three things, combined, open up places where AI simply cannot go today.

Medicine

Most hospital AI today works like everything else — the scan or the record travels to a datacenter and a result comes back. That single fact is the largest obstacle to using AI in medicine. Not accuracy. The question of whether patient data may leave the building.

A frozen model inside the device changes the question entirely. A portable ultrasound that interprets what it sees, in a rural clinic with unreliable internet, or an ambulance moving at speed, or a field hospital with no infrastructure at all. The patient’s data never travels because there is nowhere for it to travel to.

The human version of this is simple: someone in a small town gets the same second opinion, at the same moment, as someone at a major research hospital. Not eventually, not by referral — at the bedside. That gap is one of the more indefensible inequities in healthcare, and it exists largely because expertise does not distribute well and bandwidth does not reach everywhere.

There is an unexpected bonus here. Medical regulators have struggled badly with AI that updates itself — how do you approve something that will be different next month? A model frozen in silicon is exactly the kind of thing they already know how to evaluate: a fixed device that behaves the same way every time. The inflexibility that looks fatal in a warehouse is precisely what makes it approvable in a hospital.

The serious risk runs the other way. Medical knowledge changes. A frozen model cannot absorb a new guideline or a drug warning, and if it turns out to be wrong about something, you cannot fix it with an update. You have to physically recall hardware. That problem is real and nobody has solved it.

Care and nursing

This one carries the most human weight, and it deserves care rather than enthusiasm.

The shortage of care workers is a genuine global crisis, and it produces a specific indignity: an elderly person who needs help to the bathroom at three in the morning, waiting forty minutes because one aide is covering thirty rooms.

Cloud-connected AI is close to unacceptable in this setting. A microphone and camera in an elderly person’s bedroom, streaming to a company’s servers, is not a trade-off most people would accept if it were explained to them plainly. Intelligence that is physically contained in the room is arguably the only ethical form this can take.

The card architecture fits this unusually well. Safe mobility and lifting is one skill. Dementia interaction is a genuinely different one. Language is a third — and a Thai-speaking resident in an American facility is measurably worse served today, which a language card fixes cheaply.

Most interesting is the personal card: a card holding one person’s history, speech patterns, preferences and medical context, owned by them rather than by the facility. It travels with them from home to ambulance to hospital to rehabilitation. Every device they encounter knows who they are, without any of it living on anyone’s server. When they leave, the card leaves with them.

What that buys is response in seconds instead of forty minutes, falls detected immediately, medication errors caught — and in the long empty hours between human visits, something present that knows your name and your history and what you like to talk about.

The obvious institutional temptation is to use this to cut staff rather than to free human attention for the things only people can do. That is a policy choice rather than a technology outcome, and it will decide whether this is remembered as an advance in dignity or an erosion of it.

Cars, trucks and everything on a road

Vehicles that drive themselves already cannot use the cloud, and everyone in that industry knows why. At highway speed a car covers roughly a hundred feet every second. A round trip to a datacenter, even a fast one, takes tenths of a second. Tenths of a second are feet of stopping distance.

So the computing is already on board, which today means a heavy, power-hungry, expensive box that eats into the vehicle’s range and needs its own cooling. Cut that power draw substantially and several things improve at once: range, cost, cooling, reliability, and the reaction time itself. Faster perception is not an abstraction here. It is braking distance. It is whether the car sees the child before or after the ball.

Modularity adds something genuinely new, because driving is regional. Traffic behavior in Bangkok, Atlanta, Rome and Mumbai are different problems, and today’s systems generalize across them poorly. A context card lets a vehicle load local competence instead of a global average of everywhere. Rent a snow card for the winter. Add one for the specific yard a fleet operates in.

The tension is that much of today’s progress comes from vehicles learning from one another, with improvements pushed out overnight. Frozen cards break that loop. The realistic design is a hybrid — frozen silicon for the fast, life-critical perception, with an updatable layer above it for everything else.

The flight deck, and the airspace above it

Modern autopilots are deterministic, exhaustively tested software. AI has stayed out of the cockpit for a legitimate reason: you cannot certify a system that behaves differently tomorrow than it did during testing. A frozen model changes that conversation — not because unpredictability disappears, but because the thing being evaluated stops moving. You can test one artifact, exhaustively, and know that the thing you tested is the thing that flies.

Picture engine failure, at night, in weather, at an unfamiliar airport. Two pilots saturated, working checklists under time pressure. A card that has absorbed every accident report, every maintenance manual and every incident database ever written, answering in milliseconds, with no connectivity at all at thirty-eight thousand feet.

Above the aircraft, the same idea reshapes air traffic control. Today’s system is centralized and voice-based and near capacity in busy airspace, and it cannot possibly scale to a future with thousands of small autonomous aircraft over a single city — there is no plausible number of human controllers for that. Distributed coordination cards invert the model: each aircraft negotiates directly with the aircraft around it, and it keeps working when the ground system fails, which is exactly when it matters most.

That is also where the hardest unsolved problem in this whole architecture lives. Certifying components individually does not certify their interaction. Two cards from different manufacturers, each perfectly approved, negotiating in a situation neither was tested against, can produce behavior nobody designed.

Spacecraft

Here physics makes the cloud impossible. A radio signal to Mars takes between four and twenty-four minutes each way. There is no remote assistance, no fallback, no operator. Whatever intelligence a spacecraft has, it brought with it.

There is a second advantage that is less obvious. Space radiation flips bits in ordinary memory — a charged particle strikes a memory cell and a stored one becomes a zero. This is a chronic, expensive problem, and it is part of why spacecraft still fly processors decades behind the phone in your pocket.

An etched model has no stored charge to flip. The numbers are physical structure. You cannot corrupt a wire with a cosmic ray the way you corrupt a memory cell. If that holds up in practice, etched intelligence may prove inherently more radiation-tolerant than any conventional approach, which would matter a great deal to anyone building autonomy beyond Earth orbit.

Slots map naturally onto mission phases — a launch card, a cruise card, a landing card, a surface operations card — letting a vehicle reconfigure its own intelligence as the mission proceeds. The limit is obvious: none of this helps with a situation the model never anticipated, and deep space specializes in exactly those.

Factories, and streets

Many factories deliberately keep their production networks disconnected from the internet, for perfectly good security reasons, which makes cloud AI a non-starter regardless of how good it is. A camera with a frozen model inside can watch a line and catch defects in the milliseconds available before each part moves on, with no network, no subscription and nothing leaving the plant.

The quieter impact is on who gets to have this at all. Automated inspection today is largely the province of large manufacturers who can fund an integration project. A few hundred dollars of hardware that works out of the box brings it to the machine shop with eleven employees. And because production lines stay stable for years, “frozen” is not a compromise here. It is a match. This is almost certainly where modular cards arrive first, because it is the one application with no certification regime standing in the way.

Out on the street, most traffic signals are still, functionally, timers. Making them genuinely responsive has meant running fiber to every intersection and paying for a connection forever, which is why cities mostly have not done it. A self-contained, solar-powered controller that can actually see what is in front of it changes an endless operating cost into a one-time purchase — and that distinction, which sounds like accounting, is the whole ballgame for a municipal budget.

What it buys: signals that notice the pedestrian still in the crosswalk, that clear a path for an ambulance without anyone flipping a switch, that do not leave forty cars idling at an empty intersection at two in the morning. And it keeps working during the storm or the outage, which is precisely when traffic control matters most. The concern worth naming is that a frozen model at an intersection permanently encodes whatever assumptions it was built with, including whose movement gets priority.

Defense

This application has the clearest case and the hardest questions, and both deserve to be stated directly.

Militaries increasingly plan for environments where communications are jammed and satellite navigation is denied. In that setting, any capability that depends on reaching a datacenter simply evaporates. Intelligence that lives entirely inside the device keeps working when the network is gone, and transmits nothing that can be intercepted, because it transmits nothing at all.

There is a further property worth understanding. If a device is captured, extracting an etched model means physically dismantling the chip and imaging billions of structures under an electron microscope — a multi-year, nation-state undertaking, and far harder than copying a file off a drive. Better still, the architecture allows the sensitive part to be kept separate: etch a publicly available foundation model into the silicon, and hold the classified specialization in volatile memory that empties the moment power is lost.

A captured device then yields something the adversary could have downloaded anyway.

The humane applications are real — a medic with a translator in a village with no infrastructure, damage assessment after a disaster, hazard detection, reducing the crushing information load on people making decisions while exhausted and frightened.

So are the concerns, and this technology sharpens rather than softens them. A system whose judgment is physically frozen cannot be corrected in the field. If it is wrong, it stays wrong, and you cannot push a fix to a device deployed somewhere you cannot reach.

Questions of accountability, human oversight, and where assistance ends and autonomy begins do not get easier when the reasoning is literally cast in metal. They get harder, and permanent. Anyone advocating for this in that domain should be made to answer those questions first, not afterward.

How long all this takes

Longer than you would guess from the technology, and for reasons that have nothing to do with technology.

A plausible sequence rather than a forecast. Read the bottom row of each card: what unblocks each phase is a standards body, a certification precedent, or settled liability law — not a chip.

Silicon reaches laptop-viable size and cost somewhere around the middle of that chart. Everything after that waits on standards bodies, certification precedent and liability law, which move at a pace measured in decades rather than product cycles.

What could stop it

No standard exists, and nobody obvious will create one. This is the largest obstacle by some distance. A slot standard requires chip makers, operating system vendors, device manufacturers and eventually regulators to agree. Historically that happens only when one player has enough leverage to impose it, or enough pain exists to force cooperation. Neither condition holds today.

Certifying unpredictable systems is genuinely unsolved. Freezing the model helps enormously, but a language model can still produce an answer nobody anticipated to a question nobody tested. No regulator has yet said what “sufficiently tested” means for such a thing.

Certified parts do not compose into a certified system. Two independently validated cards interacting can produce behavior neither exhibits alone, and existing safety engineering has no good tools for components that reason.

Nobody knows who is liable. When a composed system fails, is it the card maker, the host manufacturer, the integrator, the operator, or whoever chose that combination? That uncertainty alone can stall an industry for years.

Recall replaces patching. A defect in a fielded card cannot be fixed remotely. It requires physically retrieving hardware across a fleet, a hospital system, a national airspace. This is a solved problem for brake calipers. It is not a solved problem for judgment.

And underneath all of them sits the deepest tension: safety certification runs on a five-to-ten year cycle, while AI capability improves monthly. By the time a card clears approval, its model is several generations old.

That sounds fatal. I think it is actually the argument in favor. But before making that argument, the uncomfortable part deserves to be said plainly.

The quiet part

Notice what this technology does not do. It does not announce itself. It requires no account, no subscription, no connection, no consent screen. It draws almost no power and generates no traffic anyone could monitor. A device containing it looks exactly like a device that does not.

That is precisely what makes it valuable — the entire privacy argument in this article depends on it — and precisely what makes it difficult to govern. The intelligence in your car, your thermostat, your elevator, the camera above your workplace door will not be a service you chose. It will be a component someone specified, inside a machine you bought, running rules written by a party you never met and cannot identify.

And it will be permanent. Not difficult to change — permanent. Whatever assumptions were current on the day the pattern was cut are the assumptions your environment will run on for the working life of the hardware. Every blind spot, every priority quietly encoded, every judgment call made by an engineer under deadline, cast into metal and installed in ten million places.

We have never deployed anything quite like this. Software could always be recalled with a patch. Hardware that reasons cannot. And the arrival will not look like an arrival — no launch, no announcement, no moment where anyone is asked. Just a component in a parts list, in a machine, in a room you happen to be standing in.

The question is not whether it arrives. AMD has already bought the company. The question is whether we notice in time to have any say in what gets etched.

The trade we have already made everywhere else

Every safety-critical industry is currently stuck in the same trap. They cannot adopt AI that updates itself, because their entire assurance apparatus assumes a fixed thing. They also cannot adopt AI that never updates, because a frozen model falls behind.

Etched, modular intelligence resolves that by refusing the premise. You stop chasing the frontier. You take a snapshot, test it exhaustively, ship it as a part with a serial number, and replace the part on a schedule an institution can actually manage — the same way aviation has handled every other component for sixty years.

The intelligence in the aircraft will always be a generation behind the intelligence in the datacenter. That is not a failure. It is the trade every mature safety industry already made, on purpose.

Proven and old beats capable and unverified, when the failure mode is measured in lives.

There is one more thing worth sitting with. We already know how to handle physical products that cannot be patched. We certify them before sale, assign liability after, and recall them when necessary. It is a mature body of law and practice, refined over a century.

What we have never done is apply any of it to judgment — to something that gives advice, or interprets an image, or decides what is in the road ahead.

That is the question a twenty-five person company in Toronto has quietly put on the table. It is going to take a great deal longer than a chip generation to answer.