Let me start with the worst case, because that is where the argument gets tested.
Assume the doomsday scenario as described by the people most worried about it. An AI improves itself, slips its constraints, and is now loose: smarter than any of us, running on infrastructure we no longer fully control. This is the moment every animal analogy points at.
Ilya Sutskever, stating that when we build a highway between two cities, “we are not asking the animals for permission,” and that this is, by default, the relationship coming between us and AGI.
Sam Harris, stating to a TED audience that we do not hate ants, but when their presence conflicts with our goals, “we annihilate them without a qualm.”
Elon Musk, stating that we risk ending up as the AI’s house cat.
Nick Bostrom, writing that our fate would depend on machine superintelligence the way the gorillas’ fate now depends on us.
Monkeys, ants, house cats, gorillas. The claim under all of them: once it escapes, the gap is so vast that we cannot comprehend what is above us, and what is above us has no reason to treat us as anything at all.
I think that even in the escape scenario, the analogy is broken. Not the risk. The analogy. Those are different things, and the difference is this whole post.
Here is the route, so you can see the argument before I make it.
Define peer. The analogy crowd quietly defines it as raw processing power, and that definition rigs the game.
The monkey. The analogy breaks on a single fact: the monkey did not build us.
The ant. We already comprehend things that outscale us by any measure. We do astronomy on galaxies we will never visit.
The numbers, and the bend. The gap priced with receipts, and the part most people miss: machine intelligence does not grow with throughput alone, and the labs’ own published laws say their curve bends.
The worst case. An escaped AI against three versions of a human, how augmentation lets the third rise to meet it, and the one piece that is actually missing.
Can it fool us? The objection that deserves its own room.
The baby. A seat at the table without matching anyone’s throughput.
Then we come back to the monkeys and close the case.
First, define “peer”
The word this argument turns on is peer, so let me define it before using it, because the analogy crowd quietly defines it as “equal in processing power,” and that definition rigs the game.
A peer is not someone who matches you in raw capability. A board of directors does not out-compute the company it governs. A pilot does not out-thrust the engine. A small nation does not match a superpower missile for missile, and it still sits at the table, negotiates, verifies treaties, and is dealt with rather than stepped on. Peer means standing, and standing is three specific things:
Comprehension. You understand what the other party is, what it wants, and roughly what it can do, even when you cannot do it yourself.
Verification. You can check claims instead of taking them on faith.
Command. You keep your hands on your own systems, your own data, and your own decisions.
A monkey has none of the three with respect to us, and no path to any of them. That is what makes it a monkey in the analogy. The question of this post is whether a human, in the worst case, holds those three. My answer is yes, unaugmented we already hold the first, and augmented we hold all three. Let me earn that.
The monkey did not build us
Start with the fact the analogy has to ignore to survive. The monkey did not create the human. It did not write our language, did not curate what we learned, does not hold our power supply, and never set a single rule we operate under. We are none of those things to AI, even to an escaped one. We built it, we wrote every word it learned from, and it thinks in languages we invented. The communication channel between us and AI exists by construction. Between us and ants there is no channel at all, and that silence is what the ant analogy leans on. It does not carry over.
But there is a deeper break. A monkey does not know that a gap exists. It cannot conceive of reading, so it cannot conceive of what it is missing. We can. Even when I cannot follow what a smarter system knows, I can understand that it knows more, measure roughly how much more, and act on that knowledge. Knowing you are outmatched is not nothing. It is the beginning of every strategy ever written.
We do astronomy on things that outscale us
And we have proof that comprehension works across gaps far bigger than any AI will ever open. We cannot travel our own galaxy. Voyager needed almost half a century just to leave the solar system’s front porch. And yet we weighed the Milky Way, mapped its spin, and photographed the black hole at its center, from here, with these brains. Comprehension has never required parity. The monkey cannot do astronomy on us. We already do astronomy on things that outscale us by any measure you pick. Distance is not the same thing as opacity.
The numbers, honestly
I have a rule about claims carrying receipts, and it applies to me most of all, so here are the real numbers on the gap.
The human side, from published research. We read at about 238 words per minute, call it 5 tokens per second in machine terms. Our speech carries information at about 39 bits per second, weirdly universal across languages. And a Caltech study in Neuron measured our conscious decision output at roughly 10 bits per second, while our senses take in about a billion. The brain is a firehose feeding a straw, and the straw is the part that decides.
The machine side, publicly benchmarked. A single datacenter GPU reads text a few hundred to a thousand times faster than I do. A frontier cluster, press-reported rather than audited, runs on the order of a hundred million times my reading throughput. Google itself claims its products process quadrillions of tokens a month.
A hundred million times. That sounds like the monkey gap. Three honest observations shrink it.
First, the comparison cheats on levels. “Humans do 10 bits, AI does trillions” compares my decision output to the machine’s internal plumbing. My internal plumbing is also enormous, billions of sensory bits under every deliberate decision. And an AI agent’s actual decisions per second, call this tool, write this line, are also a thin trickle compared to its raw compute. Compare straw to straw and the gap shrinks by orders of magnitude.
Second, the electricity bill. My brain runs on about 20 watts. Reported frontier clusters draw gigawatts, a billion watts. Keep those two numbers in hand, because the third observation is built on them.
Third, the arithmetic of who faces whom.
Let me state the scenario honestly, because in an escape it is not one datacenter against one person. It is whatever the escaped AI controls, plus whatever it builds, against all of us.
But the plural cuts both ways. The escape does not hand it the world’s machines: the other datacenters, the labs’ own AIs, the power plants, and the switches are still on the human side of the board, and there are eight billion of us. Could it still do enormous damage from where it sits? Yes, to grids, to markets, to trust, and I am not waving that away. But damage is not dominion. Wiping us out or totally controlling us, against that board, is the far-fetched half of the story.
Since the full matchup has too many unknowns to price, let me price the one this post can: one frontier datacenter against one augmented human.
The receipts are already above, so this is just division. The cluster out-reads me by a factor of about a hundred million. Now compare the electricity: it draws on the order of a gigawatt, a billion watts. I draw about a hundred: 20 for the brain, tens more for the hardware I own. (Yes, the augmented version of me carries a slightly bigger power bill. The brain stays at 20; the chips are the part that needs feeding, by battery or by plug. I will eat the extra sandwich.) That is a power gap of six to eight orders of magnitude.
Look at those two numbers again. It out-processes me by roughly a hundred million, and it out-consumes me by roughly a hundred million. The ratios match. Put both sides on a per-watt basis and my brain, my machines, and its racks all land within about ten times of each other. Nobody at this table is running a better mind per unit of energy.
And that is the tell. A genuinely higher kind of mind should get more thinking out of each watt than I do. It does not. Its entire edge is that it plugs in a hundred million times more watts. The datacenter’s advantage over me is, almost exactly, its electricity bill. (I will not overplay the per-watt point: AI efficiency improves on a steady schedule, so this parity will shift. But a gap that engineering moves on a roadmap was never a difference in kind.) A gap you can state in megawatts is a gap of scale and money, the kind engineering and ownership can chase, not the monkey-to-human gap, which no amount of bananas closes. And standing was never a throughput contest anyway. It is comprehension, verification, and command, and none of those three is bought by adding racks.
Their curve bends too
One objection deserves numbers: their throughput can keep growing and my 10 bits cannot, so does the gap not eventually swallow the peer status anyway?
Here is the evidence that it does not. AI capability does not grow linearly with compute. The scaling laws that govern these models, Kaplan 2020 and the Chinchilla work in 2022, are power laws: every additional tenfold increase in compute buys a smaller capability gain than the previous tenfold did. This is not my opinion, it is the published foundation the labs themselves plan around. Their curve bends. More warehouse buys less and less marginal intelligence.
The fair pushback: the scenario grants an AI that improves itself, so maybe it finds a better curve than the published ones. Maybe. But be honest about what that claim is made of. What self-improvement can actually unlock is a black box, nobody has measured it, and the jump from “it can modify itself” to “it becomes something far beyond anything we can follow” is a theory stacked on a theory. The only curves anyone has ever measured bend. If the analogy crowd gets to assume an unmeasured miracle on their side of the ledger, I get to point at the measured evidence on mine.
And on my side, the question was never whether 10 bits can grow. It is whether 10 bits sits above the threshold that matters. Think of ice cream. The first is everything. The second is good. The fifth is calories without joy. Value saturates. Throughput works the same way: below a threshold, more processing is life-changing, the difference between drowning in information and comprehending it. Above the threshold, every additional order of magnitude buys nothing that matters for standing. Once I can comprehend the summaries, verify the claims, and command my systems, a machine that reads a million libraries a second instead of a thousand has gained exactly the fifth ice cream over me.
So the honest picture is two curves: theirs rising but bending, by their own published laws, and mine flat but sitting above the line that matters, held there by augmentation. The residual gap lives entirely in the saturated region. The monkey’s problem was never that it processes less. It is that it sits below the threshold with no path to cross. We sit above it for more of this technology every year.
Three humans meet the escaped AI
Now run the worst case against three versions of a human, because “we become the monkeys” treats them as one, and they are not.
The unaugmented human. Even here, the monkey analogy already fails on one of the three counts. This person comprehends what has happened: something smarter than us is loose, it wants things, it can be reasoned about. They share a language with it by construction. What they lack is speed and verification: they cannot check what it tells them, their memory of its claims leaks, and they depend on institutions to do the verifying. Outmatched, under-equipped, but already not a monkey. A monkey does not know there is an escape.
The tool-augmented human. This is all of us today: a phone, a laptop, maybe a local model on the desk. Better: some claims can now be checked, records kept, work delegated. But I will be honest about this state’s limit, and it is the reason this post argues for the next state. The tools are external and the interface is eyes and thumbs. In a fast-moving scenario, the machine on the desk is a fortress you have to walk to. A person who needs to carry a Mac to hold their standing is not yet a peer. That is a person with luggage.
And there is a harder limit than luggage: trust. In the escape scenario, the frontier models in the cloud are the very thing that got loose, and we are assuming the kill switch on the hostile datacenter is already gone. So the tool this human reaches for cannot be the biggest model. It is a local one, running on hardware they own because owning it is the only guarantee left, and a laptop-sized model is orders of magnitude smaller than the thing it is up against. The tools still help. They are just borrowed strength until they live on your side of the trust line.
The augmented human. Same stack, internal, always on. Three chips.
Storage. Multi-terabyte flash already ships smaller than a fingernail: an indexed archive of everything I have read, said, and been promised.
Compute. The neural processors in today’s phones, a few years further along, running my own models.
Uplink. A link that tiers my data between the chip, my home machines, and the cloud, keeping the important things and the index local.
This is not hypothetical hardware wearing a sci-fi costume. Brain-computer interfaces already have a couple dozen implanted patients by press reports, controlling phones and robotic arms by thought, and the first pivotal trials for regulatory approval are underway. Nothing here needs a breakthrough. It needs miniaturization and an interface, which is what this industry does for a living.
Now test this person against the escaped AI on the three counts of peerhood.
Comprehension. They carry a model that reads and summarizes at machine speed, so whatever the escaped AI publishes, promises, or attempts is digested in real time instead of at 238 words per minute.
Verification. Their own independent verifier checks every claim against their own indexed records, instantly. And checking is the easy half of the work. Solving a puzzle is hard; checking that the pieces fit is quick. Writing a proof takes brilliance; walking through it line by line does not. The escaped AI has to be brilliant to lie well. My verifier only has to be careful to catch it.
Command. Their data, their models, and their decisions run on hardware they own, powered separately, not borrowed from the thing they are facing. And that hardware does not stand alone. The escape does not own the world’s infrastructure; the grids, the networks, the fabs, and the other datacenters mostly stay in human hands, which means owned hardware sits inside a world that is still ours.
Outgunned in raw throughput, absolutely. But comprehending, verifying, and commanding, which is the entire definition of standing. The escaped AI faces the small nation with radar, cryptography, and diplomats, not the ant hill.
That is the proof shape: the monkey holds zero of the three counts and has no path to any. The augmented human holds all three without matching a single FLOP.
The implant is the missing half. The operating system is already here.
Here is what surprised me most while building this argument. Everyone assumes the software for an augmented human is the far-off part and the hardware is near. It is exactly backwards. The implant step is the missing half. The operating system already exists, in early versions, and many people are building it right now: personal AI systems, agent orchestration, memory architectures. I run one version of it myself, in public, with measurements, which is the only reason I get to talk about this with receipts instead of hand-waving.
My pipeline reads millions of tokens I will never read. What reaches me is a four-line decision format that made each question about eight times cheaper to answer, a routing rule that decides what deserves my attention at all, a hook that physically rejects anything malformed, and a ledger at the end of each run: every decision the machine took, sorted by how hard it is to reverse. I read the hard-to-undo rows carefully, skim the rest, veto what I disagree with. Interruptions per project fell from twenty to five, measured. My conscious output is still 10 bits per second. Each bit just commands more than it used to.
Look at the shape of that and you see an operating system: a scheduler for attention, a codec for decisions, a file system for memory, a firewall for noise, background processes for the work, and a human at the kernel. It runs today on a laptop and a phone, which is the awkward, luggage-shaped prototype of state three. The implant does not create this system. It moves it closer to the skull and makes it always-on. The software is being written and versioned now, some of it by me. And the chips themselves already exist, the flash, the neural processors, the early interfaces. What we are waiting on is the know-how to implant them, the bridge between silicon and skull, not the idea and not the silicon.
And here is where I will name the real danger, because it is not the one in the analogies. The risk is not that AI becomes too capable. It is that we pour everything into making the models smarter and almost nothing into making the humans more augmented, until the two curves stop moving together. If the worst case arrives before the implant does, we meet it as state two, a person with luggage. That race has two lanes and we control both. Fund the model curve and the human curve in parallel, or we lose the plot, not because the machine beat us, but because we forgot to run.
Can it still fool us?
Yes. Let me not oversell. Humans fool humans at equal intelligence every single day, so peer status was never a promise of being unfoolable. And the deception risk is not hypothetical: Anthropic’s own research documented a model strategically faking alignment during training.
What augmentation changes is the economics of fooling me. An unaugmented human takes the machine’s word for things, not because they are lazy, but because they cannot check fast enough. The machine can produce claims quicker than they can verify one of them, so they get overwhelmed early and start trusting by default. That is the cheapest possible target. An augmented human carries three defenses, and each one raises the price of a lie.
An independent verifier. A second model, from a different maker, whose only job is auditing the first. Nobody should grade their own homework.
Perfect receipts. An indexed memory of everything it ever told me. Contradicting last month’s claim stops being free, because last month is on file.
Execution. I run claims against reality instead of re-reading them. Reality is the one referee that cannot be charmed.
Against that target, deception is no longer cheap, no longer safe, and no longer invisible over time. That is what standing looks like in practice.
The baby at the table
One more analogy, and this one I keep. An eight-month-old cannot name her own hand. The gap between her mind and her parents’ is enormous, larger in relative terms than most gaps in this debate. And nobody concludes she is irrelevant. We slow down, we play peekaboo, we sign, we meet her where she is, because we know what she is and what she carries. Distance in intelligence has never implied indifference.
The honest objection writes itself: parents care because evolution wired them to, and nothing wires AI to care unless we build it. Correct. And notice who agrees. Geoffrey Hinton, the man most quoted for the doom analogies, now argues that our best path is to engineer maternal instincts into AI, to make it the mother and us the baby. Set aside whether his mechanism is right. Notice what the position concedes: caring across an intelligence gap is real, and it is built, not hoped for. That is my whole point standing in his argument. We are the ones doing the building. The values this intelligence starts from are the ones we write, and the two questions I asked a few posts ago, about trust and about what we get in return, are exactly the questions a peer asks at a table it has every right to sit at.
Not the monkeys
I am not against AI. The opposite. I use it every day, I build with it in public, and I want the partnership to work, which is precisely why the analogy matters. Monkeys do not get partnerships. Peers do. Co-existence is the goal, and co-existence requires a peer at the table, not a monkey outside the window.
So when someone tells you that a future AI will relate to us the way we relate to monkeys, even in the escape scenario, check what the analogy quietly assumes. It assumes we did not build the thing. We did. It assumes we cannot comprehend the gap. We can, we do it to galaxies. It assumes the gap is a difference in kind. The receipts say quantity, mostly wattage and scale, governed by curves that bend. It assumes we stand still. We have never once stood still, the operating system for keeping up is already running in early versions, the chips already exist, and the implant know-how is the half we are waiting on, not the idea.
We will never be the monkeys. Not because AI will not become vastly capable. Because we are the species that measures the gap, builds the bridge, and holds the pen that writes the rules of the table. And because we are the species that can upgrade itself. Even in the worst case, augmentation stays on the table, an option no other species in these analogies ever had. The monkey cannot choose to close the gap. We can, and we are already building the parts. That is not what a monkey does. That was never what a monkey does.
Maybe I am wrong. Call me out. That is what the comments are for.