The View from Above

How a smarter mind might see our predictable mistakes

This is a bonus essay, something I’ve been thinking about for a while. Feel free to provide feedback.

Epistemic status: Conceptual synthesis. I am not claiming that a superintelligence could perfectly predict every human action indefinitely. The claim is narrower: many important human failure modes are simple enough that a much smarter system could model and exploit them systematically.

Relevance: If human behavior is cheaply legible and steerable from “above”, that changes how we should think about alignment, manipulation, and who we trust around powerful AI systems. This post is my attempt to spell out that frame.

Note on process: I used ChatGPT, Claude, and Gemini as editing and brainstorming assistants while drafting this. The core ideas, structure, and final wording are mine, and I have spent significant time revising. I personally vouch for every claim in the post.

---

I. Chicken Dog

A four-year-old girl named Sylvia hides behind a couch during hide and seek. Her feet are clearly visible. She doesn’t know this. From her perspective, she’s behind something, and being behind something means being hidden. She has a theory of hiding. The theory is coherent. It’s just incomplete in a way she can’t see from inside it.

A dog finds a rotisserie chicken in a specific spot during a walk. Big reward, unexpected, salient. Two years later, the dog still checks that spot every single day. The chicken is never there. The dog keeps checking. From the dog’s perspective, this is rational - the spot produced chicken once, and the dog has no way to model “one-time human accident” versus “reliable food source.” The heuristic is reasonable given its constraints. It’s also, from our vantage, obviously wrong in a way the dog will never understand.

A cat watches ants crawl across a TV screen. The ants move to the edge and disappear. The cat walks behind the TV to find them. The cat isn’t stupid. It has a model of object permanence - things that move out of sight still exist somewhere. The model is correct for physical objects. It just doesn’t account for abstract representations of objects. The cat can’t make that distinction.

These examples share a structure: an agent with a coherent model, operating reasonably within that model, arriving at conclusions that are obviously wrong from a higher vantage point. The error is systematic - the kind of mistake you can predict once you see the gap between the agent’s model and objective reality.

The claim I want to explore:

To a sufficiently advanced intelligence, much of human behavior would look like these examples.

Not because humans are stupid, but because we’re running on models that are coherent, locally reasonable, and systematically incomplete in ways we can’t see from inside them.

---

II. Legible, Patterned, and Deeply Predictable

What would it mean for human behavior to be “legible, patterned, and cheaply predictable” in the way the dog’s behavior is to us?

It would mean that a superintelligence could look at a human - or a human institution, or a human civilization - and see something like: “Ah, they found ‘chicken’ here once, and they’re still checking this spot. They’ll keep checking. I can tell you right now they’ll check this spot for the next decade, even though its highly unlikely the chicken will ever be there again.”

Four patterns that would likely be obvious from above:

Reward hijack. Humans run on dopamine systems designed for scarcity. Sugar was rare, so we evolved to find it intensely rewarding. Now we live in abundance, and we’re still “checking the sugar spot.” The same structure applies to social media, pornography, outrage content, gambling. The human thinks they’re making choices. From outside, it’s a simple circuit being exploited, and the behavior is predictable precisely because the mechanism is so simple.

Mimetic desire. A superintelligence wouldn’t need to understand why everyone suddenly wanted to move to Austin, or why being a Product Manager in tech became high-status in 2014 while others didn’t, or why the same baby names spike and crash across a generation. It would just see the mechanism: humans watching each other, copying what the high-status ones want, bidding things up - then moving on. Content changes, shape doesn’t.

Epistemic stubbornness. Once a belief gets tied to identity, evidence becomes almost irrelevant. Present someone with evidence against a belief they hold as part of their self-concept, and they often believe more strongly, not less. The belief is a badge, a membership card. This makes large swaths of human disagreement predictable: not in what people believe, but in knowing they won’t update, that they’ll reinterpret contradictory information, and that they will cluster with others who share the belief.

Institutional sclerosis. When internal incentives diverge from stated mission, internal incentives almost always win. You can treat that as the default outcome unless someone has gone out of their way to design around it. A company that says it values innovation but promotes people who don’t rock the boat will get less innovation. An LP says they want outsized returns, but invest in venture managers who invest in backwards looking tech. An agency meant to regulate an industry whose staff rotate into that industry will get captured. This pattern is really just mimetic desire, reward hijack, and identity protection poured into an org chart and left to run for a decade.

---

III. Checking the Chicken Spot

Here’s where it gets uncomfortable.

Everything I just described is already visible to humans. We know about reward hijack. We know about mimetic desire. We’ve studied institutional failure and motivated reasoning. We have names for all of it.

And yet;

The person who reads about dopamine loops still picks up their phone and scrolls for an hour. The person who understands mimetic desire still wants what their peers want. The person who studies institutional failure still builds failing institutions.

Knowing the pattern isn’t guaranteed to break it. At best, it provides a running commentary. You can watch yourself check the chicken spot while narrating “I’m checking the chicken spot again.” The narration doesn’t stop the checking.

This is different from the dog. The dog can’t see its pattern. Many humans can see the pattern - sometimes - but seeing doesn’t translate into escape. We lack what you might call executive access: the ability to actually intervene on the process rather than merely observe it.

So from a superintelligence’s view, there might not be a huge difference between “human who’s never heard of cognitive biases” and “human who’s read Kahneman.” Both are checking the chicken spot. One just provides more sophisticated commentary while doing it.

---

IV. Tiers of Tears

But humans aren’t a uniform category. There are tiers of meta-awareness, even if they compress when viewed from far enough above.

Tier 1: Fully inside the pattern. No meta-awareness. Sylvia behind the couch. They’re not aware there’s a pattern. Their behavior is just “what they do.” This is a large fraction of people, a large fraction of the time.

Tier 2: Aware that patterns exist. Has read some Kahneman (or a listicle about Kahneman), can identify biases in others, occasionally catches themselves. Still stuck in the patterns, but with better post-hoc narration. Can sometimes pause between stimulus and response.

Tier 3: Actively building guardrails. Constructing external structures - Ulysses pacts, environment design, social contracts, blocking tools, checklists, public commitments. Still has all the same impulses, but has learned not to trust raw willpower and instead builds fences around the chicken spots.

Tier 4: Moments when the scaffolding actually works. The combination of self-awareness and pre-built structure changes the output. The person still has the impulse, still has the same wetware, but the fences, commitments, or environments catch the impulse before it hits action. It looks like willpower from inside. From outside, it’s a system with more layers.

The temptation, reading this, is to place yourself in Tier 3 or 4. That temptation is itself part of the pattern. Most of the time, even very reflective people are running Tier 1 software with Tier 3 narration. The gap between how often you think you’re operating with executive access and how often you actually are is itself something a superintelligence would see clearly - and could exploit.

From far above, these tiers might blur into each other. But from the ground, when we’re deciding who writes policy or runs labs or shapes the trajectory of powerful systems, the differences matter enormously. A Tier 1 person will reliably confuse their identity with the job, their tribe with the world, their fear with risk assessment. A Tier 3 person will at least try to build external guardrails against their own blind spots - not because they’re wiser, but because they’ve stopped trusting themselves to be.

---

V. Elderly Children

It’s important to be precise about the claim here.

“Elderly children” is a phrase that captures something real, but it can mislead. Adults are not just kids with bigger toys. The same brain that doomscrolls also runs multi-decade projects, builds supply chains, designs chips, and writes down plans so that future selves and strangers can coordinate around them. Dogs don’t build institutions. Four-year-olds don’t create scientific communities that outlast them. Those are genuinely new capabilities.

The claim is narrower: in the domains that most often break our world - emotional regulation, tribal politics, magical thinking, identity-bound belief - a huge fraction of adults are still running strategies that wouldn’t look out of place in a middle school cafeteria.

As a species, we systematically struggle with some specific things. We struggle to delay gratification when the payoff is beyond a few years. We struggle to model other minds as genuinely different from our own. We struggle to separate identity from belief, to sit with uncertainty without rushing to fill it with story, to build institutions that scale beyond a few hundred people without being quietly captured by internal incentives. We struggle to resist superstimuli and to update on evidence that threatens our self-concept.

None of these are absolute limits. Some people, some of the time, push against them. But if you were modeling humanity from above, these are the biases you would bake into your simulator.

---

VI. Legible - to a Superintelligence

What would a superintelligence actually see, looking at humanity through this lens?

Not a world of rational agents occasionally making mistakes. It would see a world of beings who evolved for small-group politics and immediate-return foraging, who developed language, writing, agriculture, cities, industry, nuclear weapons, global financial systems, and climate-altering technology - all while running on the same basic cognitive architecture. They’re running hundred-thousand-year-old firmware on nuclear-era hardware. The technology scaled, the cognition is barely holding on.

From the superintelligence’s view, this wouldn’t be mysterious. It would be legible. It would be: of course this is happening. The errors are precisely what you’d predict from the mismatch.

And this is where prediction becomes something more concerning.

---

VII. Steerable - to a Superintelligence

If human behavior is this legible, it’s not just predictable. It’s steerable.

The same map that lets you forecast someone’s next move also tells you where to put the rocks in the river.

This isn’t speculation about some future superintelligence. We’re already running the experiment. Political campaigns A/B test messages until they find the phrases that move specific demographics. Social platforms learn which notifications pull you back into the app. Casinos and loot boxes are built around reinforcement schedules that keep you playing long after you want to stop. Recommendation algorithms know which videos will keep you scrolling better than you know yourself.

These are narrow optimizers exploiting legible human patterns. They work imperfectly because the systems doing the exploiting are themselves limited - they can only see certain data, optimize on certain axes, update at certain speeds. But the basic mechanism is already operating at scale.

A superintelligence wouldn’t invent a new kind of manipulation. It would run the same play with more resolution, more speed, and less internal noise. It would see, in real time, which status cues, which fears, which rewards move which clusters of people, and it could shape the choice architecture around billions of individuals in parallel.

From inside, it would still feel like choosing. From outside, it would look like routing electricity through a relatively simple circuit whose layout is already known.

---

VIII. We Are Not Just Chicken Dog

Is there any hope in this picture?

Maybe a few threads.

First: we’re not only the dog. Humans have, occasionally, figured out some of their own patterns and built structures to work around them. The institutions that deserve any hope are the ones built on the assumption that we are short-sighted, status-driven, and incentive-responsive - and that explicitly design around those traits. Independent central banks, some open scientific communities, some legal structures aren’t miracles. They’re what you get when people stop pretending we’re noble and start engineering for the primate that’s actually there.

Second: the tiers are real, even if they compress from above. The difference between someone who’s never reflected on their own patterns and someone who’s spent decades building self-binding structures matters - for the individual’s life and for what roles they should play in high-stakes decisions. You can’t escape the human condition, but you can build better scaffolding within it.

Third: if we know the patterns that make us manipulable, we can at least try to construct environments, habits, and institutions that don’t exploit them - or that exploit them in directions we’d endorse on reflection. This is hard, because the people building the environments are also running the buggy software. But it’s not impossible. It’s just adversarial engineering against our own nature.

---

IX. Who Places the Chicken

The deepest lesson, if this picture is roughly right, is not that “Humanity is doomed.” It’s that we’re early to seeing ourselves clearly - and badly mismatched to the environment we built.

A species that learned to build powerful tools before it learned to see itself clearly. Still hiding behind the couch with our feet out, still checking the chicken spot, still trying to find ants behind the TV.

We already share the world with narrow systems that see our patterns well enough to exploit them. Those systems are getting broader, faster, and better at modeling us. If we do end up sharing the planet with minds that see us the way we see Sylvia and the Chicken Dog, they will be able to place the rotisserie chicken wherever they want us to go.

The only real question is whether we will use what clarity we have to bind ourselves - to build the fences, the commitments, the structures our future selves would thank us for - or whether we wait, and let something smarter do the binding for us.