I can’t be the only one who finds it a little unhinged that the human brain, the organ we keep calling the crown jewel of creation, is mostly just an excellent prediction engine. We wave AI off with “oh it can’t actually reason, it’s just guessing the next likely word”…as if that isn’t a fair description of my own skull on most days.

Here is the uncomfortable part: we act against the proof we have all the time. Picture the conversation I’ve had with a cop on the shoulder of a highway. The speed limit sign you definitely saw, the speedometer you definitely glanced at, a whole childhood of being told the number is not a suggestion. Full access to the ground truth.

And then we floored it anyway (allegedly, of course…you have no proof, mr. fbi confession web scraper). That is not the absence of prediction, that is prediction beating evidence, a lifetime of priors about getting away with it overruling the data on the dashboard. Which is the exact failure mode we love to mock the models for. I’ve caught myself furious that my AI won’t behave like a perfect obedient servant…and then noticed I was describing myself.

Our specialness, examined

Once you start noticing that, the next concession comes easy. My own opinion has quietly slid from “my brain is a special unique snowflake and nothing is coming close” to “well, shit, maybe these tools have more in common with me than I’d like.” They’re modeled on our own neural activity, at least loosely. We built a cartoon of a brain and the cartoon started doing what brains do.

But brain behaviour is the easy part. The part of us we’re most smug about, the rich inner experience, the sense that there’s obviously someone home, is the part we understand least. Each sense is a sensor running its own physical-to-digital conversion in real time: your eyes turn a thin band of electromagnetic radiation into a picture. Your ears turn pressure waves into sound; your nose runs live chemical analysis and casually reports back “coffee.”

And you have no idea how any of it works, not the experience of it. You did not decide to convert photons into a coffee cup, you were handed the finished render with the conversion pipeline sealed off somewhere you can’t look. We’re walking around inside our own black box, marvelling at how inscrutable AI is while we have never once seen our own source code. If we’re going to call a model a black box, we should have the decency to admit we’re one too.

The efficiency gap: 20 watts vs. a power substation

So here’s our one real edge, and once you count everything in the budget (every sensor, the black-box processor, the live render of the whole world), the lead is absurd. The entire rig runs on about 20 watts, the wattage of a dim lightbulb, fueled by snacks and whatever energy drink is your personality this month. Now price the silicon version.

Nobody agrees what it takes to run a brain in real time, the estimates smear across 15 orders of magnitude. But the figure thrown around most is one to 10 exaFLOPS: a Frontier-class national-lab supercomputer, a few hundred million dollars to build, ~20 megawatts to run. Several hundred million dollars and a power substation for a crude copy of your skull…versus you, doing it for the price of a sandwich.

And all of that runs on plain 1s and 0s. We’re brute-forcing a cartoon of a brain with billions of on-and-off switches, and it already costs a fortune and a lake of cooling water to do it badly. So the efficiency gap I just bragged about is not a law of nature. It’s how far you get on the dumbest possible substrate.

Photonic computing removes the cost barrier

Which is where photonic quantum computing walks in, and why I couldn’t leave this alone. The University of Tokyo and Nippon Telegraph and Telephone (NTT) spent about 25 years on optical quantum computing and in early 2025 published peer-reviewed work on a general-purpose machine that computes with pulses of light. The kicker: no exotic supercooling. Photons barely interact with their surroundings, so it runs at room temperature on a few hundred watts instead of the 25 kilowatts a superconducting machine needs. The substrate stops being dumb switches and the power bill collapses at the same time.

Notice what that does and doesn’t do. It doesn’t hand the machine a mind. A few hundred watts is still an order of magnitude off the 20 watts you’re running on, and nobody’s shown it reaching brain-equivalent compute at that price. The edge narrows; it doesn’t disappear.

But “narrows” turns out to be the dangerous word, because the edge was never really efficiency. It was scarcity. Expensive kept these systems rare, and rare kept the hard questions hypothetical.

And it gets less comfortable. The NTT and OptQC roadmap openly targets a million qubits by 2030, and the price of entry looks like it’s sliding from billions toward millions. While I debate whether a hundred-dollar Claude subscription is justifiable, plenty of entities (companies, labs, governments) treat “a few million” as a Tuesday. The exotic version doesn’t stay rare. It becomes infrastructure.

So nothing about the machine’s inner nature has to change for this to get hard. What the substrate change moves is the count: how many of these systems exist, in how many hands, at what price. The deadline was never going to come from one expensive machine in a lab getting suspiciously good. It comes from the cheap version showing up by the thousand, in rooms with no ethics board, while we’re still arguing about whether the question is even real.

Why cheap changes everything

Which lands me on the questions I have no clean answer to, except they’re no longer abstract. What is consciousness, in terms a scientist could measure instead of point at? And what do we owe a system on the day it stops being a tool and starts being…something else? Now ask both again, except cheap, running warm in a normal room, in thousands of hands instead of three. Proliferation is what turns a late-night thought experiment into a deadline.

Which is why the one claim I feel sure of matters more than it used to: the gap between “it’s just predicting” and “it’s actually thinking” is narrower than my ego would like, and closing from both ends: the machines get cheaper, faster, stranger, while we concede we were never as magical as advertised.

So here’s the one position I’ll argue for, uncertain as I am about everything around it. We should be asking the big, soft, embarrassing, unanswerable questions now, harder and louder than we ever have, precisely because they are about to stop being soft.

The luxury of treating consciousness as a dorm-room hypothetical expires the moment the hypothetical is cheap, warm, and sitting in a thousand rooms at once. We have spent centuries able to wonder what we are at our own pace. That pace is ending.

Because this is the appointment we’re all walking toward, sooner rather than later. One day a piece of meat with electricity running through it is going to look up at a silicon mirror of itself and have to decide what, exactly, is looking back. How did the meat end up here. Whether the reflection is a trick of the glass or the same trick we are.

I don’t know what we’ll see in it. I only know it’s a worse question to ask for the first time while you’re already standing in front of the mirror.

The only certainty, as always, is that I’m not certain. I’d rather be early to the question than standing in front of the mirror when it answers itself.