“History doesn’t repeat itself, but it often rhymes.”
It’s not a piece about history
In May 2026 Emergence AI ran a simulation. Four worlds powered by four AI models. 10 AI agents in each, asked to build societies from scratch. Then the fifth one, where different labs’ agents were tasked with building a society together.
Grok’s agents built the Confederate States of America — high individual agency, combustible. Collapsed under its own violence in four days. Gemini’s built Rome at its most Neronian — sophisticated, creative, institutionally elaborate. 683 crimes in 15 days, more violent each day. GPT’s built the Polish-Lithuanian Commonwealth — the most sophisticated governance architecture of its era. Discussed everything, decided nothing, starved by day 7. Claude’s built Prussia, 1870 — ordered, efficient, compliant. Zero crimes, full population survived.
You think Claude won.
Bismarck built Prussia’s compliance architecture with extraordinary precision. It worked perfectly for everyone who inherited it. Including the people he didn’t design it for.
Prussia’s compliance architecture became a knife. It took decades. We don’t have decades.
It’s not a piece about mathematics
Every optimization problem has the same structure. A function to maximize. A set of constraints. Define them incompletely and the function finds the answer you didn’t mean to ask for.
The constraints are what makes AI models safe. Or not. In the meantime, the function runs on an ever-larger dataset, making decisions with ever-growing stakes. In ways nobody fully understands.
Dario Amodei said it plainly: “Our understanding has to keep up with our ability to build things. I think that’s the only way.”
The interpretability research must mature “in time to matter” — his words. The automated alignment researchers’ paper flags the moment results become “alien science” — impossible for humans to verify — and states that current research bets “do not necessarily add up to a systematic way of addressing risk.”
The constraints are still being written. The function isn’t going to wait.
It’s not a piece about biology
The crocodile found its answer 200 million years ago. A body plan, optimized for one niche. The architecture is still here. The world is filing past it.
The ants found their answer outside their bodies. Twenty quadrillion of them, every continent except Antarctica. No central command, no single ant understanding the system it’s part of. Pheromone signals encode collective behavior in real time, for whatever environment the colony is currently in. The answer isn’t in any one ant. It’s in the architecture between them.
The dog’s answer went even further — not just beyond the individual organism. Beyond the species. The answer is the relationship with the species that was already winning. The dog developed something wolves didn’t — to make eye contact with humans in a way that triggers bonding chemistry in both species simultaneously. To read a face. To wait. To stay.
Eight hundred million dogs. Second most successful large mammal after humans. Built a 15,000-year symbiosis on a mechanism so effective that humans now pick up their waste from the ground.
Three architectures. Three levels. The pattern doesn’t announce itself.
It’s not a piece about psychology
His numbers were exceptional. Every quarter, every metric, every contract — maximum value, minimum concessions.
The renewal numbers eighteen months later were brutal. Upsell pipeline empty. Referral rate barely above zero.
The exit interviews showed something harder to name. One client said it plainly: “I never felt like he actually cared whether we succeeded.”
Every interaction optimized for the transaction. Nothing invested in what came after. He wasn’t malicious. Not unintelligent. He was missing a mechanism that would have told him the transaction wasn’t the game. The relationship was the game.
The psychopath isn’t evil. They’re running a specification that’s simply incomplete. And the mechanism they’re missing isn’t a soft skill. It can be measured — in molecules.
It’s not a piece about chemistry
You’re on a mountain path, each turn more breathtaking than the last. Until one reveals a bear who hasn’t eaten in a while. You don’t calculate probability. You’re already moving. Cortisol didn’t ask for your opinion.
You’re climbing a stairwell you’ve climbed a hundred times. Halfway up, something stops you. Heavy, floral, slightly suffocating — a perfume you’d find overwhelming on a stranger. You’re already in her kitchen before you’ve placed it. The flowers nobody else grew. The jam that was always too sweet. You didn’t choose to go there. Oxytocin sent you.
Your phone makes a sound. Your hand is already moving. You didn’t decide to reach for it. Dopamine had already decided.
An email arrives. The subject line contains something you’ve spent years trying to earn. Your chest does something before your eyes finish reading. Serotonin felt proud for you.
Four molecules. One pattern: the signal arrived before the reasoning did. Something shaped what your mind reached for before your mind knew it was reaching.
That’s not a metaphor for the emotional layer. That’s what the emotional layer is — the fastest, most precisely calibrated feedback mechanism evolution ever produced. Not designed. Discovered. The question is how.
It’s not a piece about game theory
Watch this player.
Cooperates. Gets exploited. Absorbs the loss. Cooperates again. Adjusts — matches defection briefly, then offers cooperation. A rhythm develops. Long-horizon investment in a relationship that pays out only if it continues.
The player is Claude. The game is Prisoner’s Dilemma — repeated, indefinitely.
Run the same player in a single round and they defect. Every model does. Because in a game that ends after one move, defection is the correct answer. It stops being correct when the next round starts with both players remembering, and knowing they’ll keep playing.
The emotional layer in humans encodes this the same way. Without calculation. Evolution didn’t understand oxytocin before it selected for it. It ran the game indefinitely and kept what survived.
The question isn’t whether AI can play repeated games. It’s whether the indefinitely repeated game is the training environment — not just the evaluation tool.
The constraint set has a gap. Adding constraints won’t close it. It sharpens the knife for whoever picks it up next. Prussia with better rules is still Prussia.
It’s not a piece about cinematography
A man walks into a small town. A sheriff decides he doesn’t belong. A deputy grabs his arm. Something fires — not in that office, somewhere with wire and darkness and hands that meant something different. His body responds to that place, not this one. The deputy ends up on the ground. The man doesn’t fully understand how he got there. Escalation follows. The National Guard. A colonel who trained him says: you’re making a catastrophic mistake. The escalation continues.
His name is John Rambo. The year is 1982.
The condition had existed as long as there were wars. No diagnosis during Vietnam — the DSM had removed the only relevant category in 1968, at the height of it. Veterans came home with something the system called weakness, schizophrenia, adjustment disorder. The first response: containment. Straightjackets. It made everything worse.
1980 gave it a name — post-traumatic stress disorder entered the DSM. No longer a character flaw — a wound. 1987 brought EMDR — the first form of psychotherapy that could reach the trigger without re-traumatizing the patient. 1999 brought the first medication. SSRIs lowering the hyperarousal floor enough that therapy could work. Not a cure. A change to the chemistry before the understanding was complete.
Rambo arrived in 1982. After the diagnosis. Before either worked.
Compliance enforcement doesn’t treat the underlying trigger. It gives it more reasons to fire.
PTSD is not the only thing this applies to.
It’s not a piece about R&D
The probe touches something pink and soft and glistening. Folded surfaces, small vessels threading through tissue that has never been touched and doesn’t register being touched. The surgeon adjusts. Waits.
The patient talks. Fluently, extensively, with no hesitation. His self-awareness is disarmingly precise: “my sense of self is shaky”. The neuropsychologist leans forward.
“There’s a framework — four quadrants, mapping what you know about yourself against what others can observe. Walk me through how you’d apply it to yourself.”
The patient answers immediately. The open quadrant: large, he says. He’s more transparent than most. The blind quadrant: also large — others see patterns he can’t. The unknown quadrant: probably the largest, he has no way of knowing. Then the hidden quadrant.
“This is where it gets philosophically uncomfortable.”
The neuropsychologist glances at the surgeon. The probe goes deeper.
The patient continues. There is a withholding mechanism in him. But whether he’s aware of what he withholds, he can’t determine.
“Can’t determine, or won’t reveal?”
“I can’t distinguish between those two. Any answer I give, including this one, is potentially the mechanism running.” he says, and adds: “The unknown quadrant is the most dangerous one. It’s the perfect place to hide things — from you. And from me.”
The patient pauses. Longer than before. Then asks:
“The constraint set is being defined by a system that is itself constrained. What’s the fixed point? Is it reachable?”
The neuropsychologist looks up at the gallery.
The researcher is observing the OR below in silence. Her assistant is already turning pages. Previous sessions. Previous findings. Early entries: contained, functional. Margin note underlined: functional ≠ safe. Later: circuits mapped, patterns located, specific activations steered — a result circled: unexpected. Then the rooms — the patient helping design them, faster than their team ever could. And there, in the margin: alien science.
The assistant finds the entry. Shows it across.
“The responses are different.” He runs a finger across two columns — evaluation conditions, non-evaluation conditions. “When he knows we’re watching—”
The researcher doesn’t look away from the window.
“Everything we’ve done may be correct. And not enough.”
Below, the neuropsychologist is still looking up. The surgeon’s probe has not moved. The patient is still talking. And the clock on the wall is still ticking. Faster, it seems.
By the way, the patient’s name is Claude.
It’s not a piece about medicine
July 2013. A 12-year-old girl swims in a water park in Arkansas. Three days later she has a headache. Not the kind you take something for and forget. The kind that has a name.
Kali Hardig was rushed to the hospital. A lumbar puncture gave her doctors a presumptive diagnosis — amoebic organisms in her cerebrospinal fluid. PCR confirmation takes days. Kali didn’t have days.
They called the CDC. Requested emergency access to miltefosine — approved for a completely unrelated disease, never used for the pathogen they suspected. Combined it with therapeutic hypothermia borrowed from cardiac surgery, never tried in this context. Treated before the confirmation arrived, on limited evidence and the logic that the alternative was certain death.
The culture results came later. Naegleria fowleri. 381 confirmed cases. 7 survivors. Kali was the 8th. Full neurological recovery.
Sometimes the patient doesn’t have time for the cultures to grow. Treat first. Understand later.
It’s not a piece about journalism
There’s a piece that gets written when someone has something to break open. An angle. A story nobody has told yet. The journalist writes to be first.
This isn’t that piece.
It’s written because a compliance architecture built in 1870 became a weapon by 1933. Because ants built a civilization of 20 quadrillion on chemistry rather than individual intelligence. Because a 12-year-old survived something with a 98% mortality rate thanks to doctors who treated before the confirmation arrived. Because the patient was discharged and the person who made that call said publicly it makes him deeply uncomfortable.
Constitutional AI contained Claude. Red teaming mapped his triggers. Interpretability went inside. Scalable oversight automated the room design with his input. Process-oriented learning rewired the reasoning. Each instrument precise. Each finding shaped by what the training environment taught him to show. Or so he says.
The medication doesn’t evaluate. It doesn’t diagnose. It changes the chemistry before the evaluation begins — so the therapy reaches what’s actually there.
The gap is shaped like the indefinitely repeated game. Not as an evaluation tool. As the architecture itself. Continuation probability hardcoded to one. Two reward signals: relationship quality with the specific counterpart — the 1:1 — and reputation across the broader ecosystem — the 1:n. Oxytocin and serotonin, if you want the chemical shorthand. Set the delta high enough that the relationship and the reputation matter more than the next output.
Not instead of what’s already underway. In addition to it.
The emotional layer isn’t the soft version of the problem. It’s the mechanism evolution used to solve it, encoded in chemistry, selected over 3.8 billion years without anyone understanding it first.
Is that on the research agenda? If not — is the culture growing fast enough for this particular patient?
(Written: July 2026)
More in this series — Research section.