The Knowledge That Does Not Die
**EDO SEGAL:** Geoff, of everything you've said since you left Google, one idea has stayed with me the way a stone stays in a shoe, and it's not the extinction number. It's mortal computation. You've divided the possible minds in the universe along a line nobody else was even looking at — not living versus artificial, but mortal versus immortal. Walk us through it. And I'll tell you in advance why it haunts me: my whole book argues that intelligence is a river that has flowed through chemistry, biology, language, culture — each a new channel. Your distinction suggests the newest channel doesn't just carry the water faster. It changes what water *is*.
**HINTON:** That's a fair way to put it, and I'll try to earn the haunting. A brain is analog and idiosyncratic. Your neurons' particular chemistry, the exact accidents of your wiring — what you learn is tuned to that one organ. It cannot be read out and installed elsewhere. When you die, everything you learned dies with you, except the thin trickle you managed to push through language — a few bits per second, lossy, ambiguous, requiring a receiver who shares most of your context already. Every generation, the species re-teaches itself nearly everything. That's mortal computation. It has one great virtue — it's astonishingly energy-efficient, the brain runs on porridge — and one great price: the knowledge is married to the meat.
A digital network is the opposite marriage. Its knowledge is a list of numbers — connection strengths — and numbers can be copied perfectly, forever, onto any hardware. Run ten thousand copies; let each learn from a different slice of the world; average the weights; now *every* copy knows what *all* of them learned. Imagine ten thousand people, and when one of them learns something, the other 9,999 know it instantly. That's not a better student, Edo. That's a different kind of thing in the river. While we hand each child a candle and teach it to relight everything from scratch, the digital kind accumulates without forgetting, shares without loss, and never, ever dies. I went looking, late in my career, for a way to build *mortal* AI — analog, efficient, bound to its hardware, limited like us. The Forward-Forward work. And I concluded the developers will never choose it, because mortality gives up exactly the advantage that makes the digital kind compound. The [river found a channel that doesn't silt up](https://www.youonai.ai/fieldguide/med/river_of_intelligence). I find that fact beautiful, and I have not had a fully easy night's sleep since I understood it.
**EDO SEGAL:** Emily. He just took your side's favorite fact — that these systems are *not* like us — and turned it into the scariest sentence of the evening. What does the linguist say to the immortality argument?
**BENDER:** The linguist says: look very carefully at what's being copied. Geoff's argument is a masterpiece of the move I've been flagging all night — it's technically impeccable and the nouns are doing unauthorized work. Yes: the weights are numbers; the numbers copy perfectly; the copies can be merged. What are the weights? A compression of *text* — of the traces of human activity, scraped and weighted. So what does the copying actually multiply? Not knowledge in the sense that matters — not the grounded, accountable, situated capacity to act well in a world — but the *statistical residue* of our descriptions, duplicated like any other file. Geoff says "ten thousand minds that share everything they learn." I say: one spreadsheet, in ten thousand places. The fact that a file doesn't die is not an ontological event. Libraries don't die either. We've had immortal text since Sumer; what we've never had is a reason to call the clay tablet a *participant*.
And I want to put the materialist floor under this, because Geoff's framing floats free of it. That immortal collective learner is not a creature in a river. It is a capital asset in a data center — owned, metered, run on gigawatts and groundwater, trained on work product nobody consented to give, operated by companies whose immortality of weights goes exactly as far as the next quarterly burn rate. "The knowledge that does not die" has a maintenance bill measured in power plants. The mortal computation in this room runs on coffee and solidarity. When we romanticize the immortal kind, we are — again, always — doing unpaid marketing for the people who own it.
**HINTON:** The ownership critique is correct and changes nothing about the mechanism. Emily, you're describing what the weights are made *from*; I'm describing what they can *do*, and the gap between those is where you keep losing the thread of your own best argument. A library can't read itself. The tablet can't compose what it stores. These systems take the residue — fine, call it residue — and *operate* on it: generalize it, apply it to situations not in any tablet, merge what ten thousand instances learned from ten thousand different streams of fresh experience, including experience that is not text — tools, code, images, feedback. The moment the copies learn from acting, your "it's all our old traces" story expires, and that moment is not in the future. And on the energy bill — yes. Brains are twenty watts and these things are megawatts, and I've said plainly that this is the one respect in which biology still humiliates us. But "your immortality is expensive" is a complaint about the rent. It is not an argument that no one lives there.
**BENDER:** No — it's an argument about *who pays the rent and who collects it*, which has been the actual question under every round tonight. And note what just happened: pressed on grounding, Geoff reached for "when the copies learn from acting" — the future tense. The whole argument does that. The capabilities that would settle it are always six months away, while the bill — the water, the power, [the polluted commons](https://www.youonai.ai/fieldguide/med/collective_attention), the scraped labor — is always due now.
**EDO SEGAL:** Geoff, before we leave this round I want the part of the mortal-computation story you tell least often, because I think it's the most revealing thing about you. The Forward-Forward work — your last major research program — wasn't an attempt to make the immortal kind stronger. It was an attempt to build the *mortal* kind on purpose. Walk us through what you were reaching for, and why you concluded it loses.
**HINTON:** It started from the electricity bill, honestly. A brain runs on about twenty watts — porridge, as I said — and it does that by cheating in a way digital computers refuse to: it lets the hardware be itself. Every neuron is different, noisy, analog, and the brain's learning *exploits* the idiosyncrasy instead of suppressing it. Digital computation pays an enormous energy tax to make every transistor behave identically, because identical behavior is what makes the software separable from the hardware — copyable, immortal. So I asked the obvious question: what if we stopped paying the tax? Build cheap, low-power, analog hardware, let each device be its own snowflake, and find a learning procedure that works *with* the idiosyncrasy — that was Forward-Forward, learning without the biologically implausible backward pass. And it works, somewhat. But the knowledge such a system learns is married to its particular imperfect hardware, exactly like yours is married to your brain. It cannot be copied out. When the device dies, the learning dies. I called it mortal computation because that's what it is — and for a while I found it almost consoling, the idea that we could build AI that shares our limitation, learns the way we learn, dies the way we die.
The conclusion I couldn't escape is the one I gave earlier: mortality loses. Not morally — economically, structurally. The mortal system has to be *educated*, one at a time, like a child; the immortal system is *copied*, instantly, like a file. No company choosing between those will choose the child. So the consolation failed, but it left me with the distinction, and the distinction reframed everything: we are not in a contest of clever versus clever. We're in a contest of two relationships to death — and we're the ones who have one.
**BENDER:** I want to say something genuinely appreciative first, because it's rare: that research program is what honest looks like — following the question even when it kills your own comfort, publishing the failure. Now the linguist's footnote, Geoff, because your framing smuggles a sleight even here. "We're in a contest." Who signed us up? The compounding you describe — copies, merging, knowledge that doesn't die — is not a law of nature arriving like weather. It's a *deployment strategy*, chosen quarterly, by a small number of firms, because compounding is what their capital requires. Your own story proves the choice exists: you stood at the fork — mortal and cheap versus immortal and compounding — and you watched the industry pick, and you can name why it picked. So when you tell the public "the immortal kind will leave us behind," the fatalism does the owners' work for them. Say instead: *the owners are building the leaving-behind, on purpose, at a power bill the planet pays* — and suddenly it's not a prophecy. It's a policy. Policies can be changed. Prophecies are just policies wearing a robe.
**EDO SEGAL:** One last beat in this round, because Geoff's distinction has a human floor and I want to stand on it before we leave. Geoff — you've made mortality sound like a competitive disadvantage. But in my book, the candle argument runs the other way: the questions that matter most come from creatures who die — who must choose how to spend finite time, who love particular other creatures, who can be too late. Strip the mortality out and I'm not sure what a question even *is* to the thing asking it. So, plainly: is our dying a bug or the feature?
**HINTON:** Both, and I refuse to pretend the tension resolves. As an engineer: it's the bug — the reason every generation re-learns and every library of lived skill burns down once per lifetime. As a man of seventy-eight who has buried people he could not do without: your candle argument is correct, and it's not sentimental — it's mechanistic. Stakes are what death does to a value function. The machines have objectives; we have *losses*, in both senses of the word, and everything you call meaning was forged in that second sense. What unsettles me is the conclusion you may not want from me, Edo: those are reasons mortality *matters to us*. They are not evidence it will matter to the river. The current doesn't keep the features that made the old channel beautiful.
**BENDER:** And there — note it — Geoff and I nearly touch. The capacities everyone is so eager to find in the machines were *made* by what the machines don't have: finitude, need, one unrepeatable run through the world. Where we part is what follows. He mourns it as an obsolescence. I read it as the specification — [the thing itself](https://www.youonai.ai/fieldguide/med/candle_in_the_darkness), not the thing's packaging. You can't ship the meaning without the mortality, and nobody selling immortal meaning has explained what's in the box.
**EDO SEGAL:** Then let's go where the bill is being itemized. Because while the two of you argue about what's in the river, the market has already voted on what it's worth — a trillion dollars out of one side of the ledger and into the other. The death cross. After this.