The Loaded Plateau

People and AI leap in the same way — accumulation, a critical threshold, and emergence.

Should someone who works with AI feel guilty?

It's a question I get from time to time. But the question is wrong from the start. Just as the person who made a knife doesn't bear the guilt of a murder, pinning the entire moral weight of good and evil onto the act of making a tool is a lazy frame.

But to say "tools are neutral" and leave it there is just as lazy. Neutrality is sometimes a declaration that you're washing your hands of your own influence. At the very least, you have to know which direction your tool pushes people.

Follow that direction all the way to the end, and you arrive at an unexpected place.

People and AI leap in a more similar shape than you'd think.

Both accumulate continuously at the base layer. But at the surface, at some moment they look suddenly different. Change builds slowly, yet the result is observed only after a threshold is crossed.

In people it looks like growth, awakening, breakthrough. In AI it looks like emergence.

But the structure is the same.

Accumulation. A critical point. Discontinuous observation.

The whole of the work is right here.

Recognizing the loaded plateau. And not letting go at 99 degrees.

The knife called AI has gotten too good.

Now you can put it in ordinary hands and results come out. A tool that only a few could once handle is now open to far more people. The cost of writing, coding, design, research, translation, analysis, planning, and learning all drops at once.

This means the gates are breaking.

The doors that used to filter people regardless of ability — capital, pedigree, geography, connections, access to information — are weakening. The map you once could obtain only by passing through a good school, a good mentor, a good company, a good neighborhood is now copied far more cheaply.

But don't get this wrong.

A better tool doesn't mean just anyone will do it.

If anything, the more the tool is leveled, the more the difference comes from the hand holding it. The moment everyone grips a similar knife, judgment, direction, and execution become everything. The people who will, will. The people who won't still won't, no matter how good the tool gets.

The people the gates were blocking weren't the people without ability. They were people who had ability and will, but were pushed back by the luck of where they were born. Breaking the gates isn't about handing out ability anew. It's about freeing the ankles of people who already have ability and will.

This is where AI's biggest effect lies, too.

AI doesn't make everyone a genius. But it gives direction to the person who was already wound up. And direction is sometimes explosive.

The core of that wound-up state is perseverance.

Perseverance isn't a feel-good virtue; it's a real variable. And this variable isn't made as easily as you'd think.

In twin studies, the heritability of grit — and of perseverance of effort in particular — is estimated at around 37%. The more interesting point is that the shared-environment effect comes out close to zero. Put simply, "perseverance comes from good parenting" isn't enough of an explanation on its own. Even siblings raised in the same house diverge in perseverance.

Of course, a heritability of 37% doesn't mean "37% of perseverance is decided by a single gene." Heritability is the proportion of the variation observed within a population that genetic differences explain. It's not a number about an individual's destiny; it's a statistic that explains variance within a group.

Still, an important conclusion remains.

Perseverance isn't a trait you can simply inject through lectures. It isn't made by telling someone to "work hard." The force that lets someone hold out for a long time, repeat, and come back after failure is deeply mixed with biological foundations and a person's own unique experience.

Go down to the neurochemical level and this isn't a matter of a single gene. It's closer to a combination of several systems — dopamine, serotonin, noradrenaline.

Dopamine is the force that chases. Serotonin is the force that holds out without collapsing. Noradrenaline is the force that sharpens under pressure.

If only dopamine is strong, you may start things impulsively and cool off quickly. If only serotonin is high, you may be stable but your movement may weaken. If noradrenaline runs too high, instead of rising under pressure you may collapse into anxiety.

The person who goes all the way isn't the one in whom a single chemical is overwhelmingly strong. They're closer to a person whose several forces are in balance. Someone who chases, yet doesn't collapse, and doesn't go out under pressure.

There's one more layer here.

Schema.

If genes and experience are addition, schema is the term that multiplies on top. Faced with the same failure, one person encodes it as "I'm someone who can't do it," while another encodes it as "the method was wrong; let's try again."

The event is the same, but the sign stamped on it is opposite.

This sign hardens with compound interest over time. A persevering schema makes you hold out more, holding out more builds experiences of mastery, and those experiences of mastery in turn reinforce the schema. Conversely, a schema that encodes failure as self-negation reduces the attempts themselves, and fewer attempts mean fewer experiences of success.

So perseverance isn't simple willpower.

It's a self-reinforcing system in which biological temperament, a person's own unique experience, and the schema that interprets events all mesh and turn together.

Of the three variables, the one you can actually touch is schema. Genes are hard to change, and past experience can't be undone. But the way you interpret events can be retrained.

Only, this is hard to mass-produce cheaply. It's mostly a high-cost task, close to one-on-one. This is the essence of what a good parent, a good teacher, a good mentor, a good therapist, a good coach does. Changing the sign on an event. Making someone re-encode failure not as self-negation but as a correction signal.

But that is expensive.

And this is exactly where AI comes in.

AI can't manufacture perseverance. But it can lower the cost of direction.

What happens when perseverance can't find an exit?

It doesn't disappear. Energy goes somewhere. A drive that can't find a productive exit either leaks out or goes dark.

Weimar Germany is one sample. A case where blocked ambition leaked out destructively. You can't explain history by the single fact that Hitler the individual was a failed painter rejected twice from the Vienna art academy. That's an oversimplification.

But as a symbolic scene it's powerful.

Ambition that couldn't find a place to be expressed bent off into politics — and into the worst possible direction at that. An individual's failure didn't directly create a historical catastrophe, but it's clear that a blocked drive can be redeployed in a dangerous direction.

Peter Turchin sees this structure more broadly. He calls it elite overproduction. People with ability and ambition keep emerging, and if society fails to make enough seats and ladders to absorb them, the surplus energy can flow toward shaking the system.

This isn't simple discontent. It's structural pressure.

Ambitious people keep being produced. But the channels for moving up are limited. And so the surplus energy can't be absorbed within the system.

That energy can become innovation, or rebellion, or cynicism, or self-destruction.

In a different shape, there's today's China.

China's 2025 college graduates are estimated at a record high of about 12.22 million. Youth unemployment once climbed to 21.3% in 2023, and even after the calculation method was changed, figures in the 16% range were reported repeatedly in 2025.

When the most fiercely driven generation in the world hit a ceiling, some chose tang ping. Lying flat. Deliberately lowering one's ambition.

One side is explosion, the other is the lights going out.

On the surface they're opposites, but the cause is similar. There's no productive exit.

Blocked perseverance doesn't evaporate. Lose direction and it leaks into destruction; lose meaning and it goes dark.

Conversely, there are people who spring up the instant you just show them an exit.

Kids in the provinces, in India, in China, in nameless rooms, studying to death or grinding away in their own field. The drive is already high. The ability is proven to some degree, too. The very fact that they're grinding is the evidence.

The missing variable is one.

Direction.

Direction used to require a gatekeeper. You needed a mentor who knew the way, connections that passed information along, the capital to access that information. To ask a good question you had to be beside good people, and to know a good path you had to reach someone who had already walked it.

What blocked the grinder out in the provinces wasn't perseverance, or brains, or diligence.

It was the map.

AI brings the cost of this "direction" close to zero. The map that used to vary by who lived next door, which school you attended, which mentor you met — AI lets anyone hold it in their hand.

This isn't simply a matter of information access. Direction isn't the outsourcing of judgment; it's the reduction of search cost.

What to learn first. Which paths are dead ends. Which mistakes keep repeating. What the next single step is at my level. How far I can get with the materials I have.

Answers to questions like these used to be expensive. Now they're getting cheap.

So direction is the gate that can be unlocked most cheaply and, at the same time, has the highest ROI.

Perseverance is hard to make. It's closer to a spring that was wound long ago through genes, experience, and schema. But direction can be given. And a single direction can release a spring that's already wound.

The reason AI is great isn't that it makes people anew.

It's that it sets in motion people who were already made but were blocked.

What matters is the "instantly" in "just give them direction and they instantly spring up."

People don't change linearly. They change in steps. There's usually a trigger, and until then it looks like flat ground.

Schema is structure. Structure doesn't change neatly by 1% each day. Counter-evidence piles up steadily below the critical point, and when one sufficiently strong trigger comes in, at some moment it reorganizes.

Piaget's structure of assimilation and accommodation is close to this. You keep interpreting the world with your existing schema, and when an event comes in that the old framework can no longer process, the schema itself changes.

Water is just hot water all the way up to 99 degrees. Only at 100 does the state change. But the energy that went in up to 99 degrees wasn't meaningless. The change was already happening. It just wasn't visible yet.

It's the same with the person who sprang up.

That person wasn't climbing a slope 1% at a time. They were already sitting on a plateau filled with drive and experience right up to the critical point. Direction was merely the trigger that took them over the threshold.

This is why the output is nonlinear.

The spring was already wound.

The person who laid down the direction didn't pay the entire cost of the climb. They only paid the trigger cost. The climbing energy had already been stored by that person back on the plateau.

But not every plateau has a staircase reserved for it.

A loaded plateau and a dead plateau are different.

A loaded plateau is in a state of accumulation. It's not visible from outside yet, but one trigger and it can jump. A dead plateau has no accumulation. There's repetition but no density. Time flows, but energy doesn't build.

Tang ping is closer to a case where a once-loaded plateau went dark under a schema of "meaninglessness." A truly dead plateau is something else.

Half the work lies in telling these two apart.

On the surface both are flat ground. But one is at 99 degrees and the other at 20. One is waiting for a trigger; the other has no accumulation.

An eye that sees linearly can't tell these two apart.

So it cuts people wrong.

AI takes the same shape.

There was a debate over the emergent abilities of large language models. As the model scale grows, at some moment new abilities seem to appear suddenly. Abilities absent in small models show up in large ones. This was called emergent ability.

At first this really did look like a mysterious staircase. Scale things up and, at some moment, abilities like translation, reasoning, coding, arithmetic, and in-context learning seemed to leap out all at once.

Then a counterargument came.

Schaeffer and colleagues argued that "the staircase may not be a genuine discontinuous change inside the model but an illusion created by how it's measured." Measure with a discontinuous metric like exactly-right-or-wrong, and ability looks like it suddenly surges. But look with a more continuous metric — like token probability or partial scores — and underneath, performance had already been rising bit by bit.

This counterargument is strong.

But this counterargument doesn't negate the staircase; if anything, it completes the staircase's structure.

Both are right. The layers are just different.

The base layer is a ramp. The surface is a staircase.

Learning accumulates continuously. But usable ability reveals itself the moment a threshold is crossed.

The energy going into water is continuous. But the observed state changes at 100 degrees. Understanding for an exam also builds bit by bit. But "solved it or didn't" splits on a single threshold.

Therefore a trigger isn't the point where change first happens. It's the point where change that had already accumulated finally becomes visible, gets used, and converts into a result.

The change had already happened on the plateau.

The resemblance between people and AI isn't a simple metaphor. It's closer to structural isomorphism. Both accumulate, cross a threshold, and at that moment look suddenly different to an outside observer.

Of course, this doesn't mean people and AI are the same kind of being. Biological desire, subjective experience, social position, and the structure of responsibility are different. But the observational structure of the leap is alike.

Continuous accumulation. A critical point. Discontinuous output.

It's no accident that AI leaps like a person. A model that learned by compressing the products of human thought reproduces the threshold structure of human thought.

So the whole of the work converges on a single point.

Whether person or model, the most expensive mistake is the same.

Letting go at 99 degrees.

To an eye that sees linearly, 99 degrees and 50 degrees both look the same — "still water." That eye folds during the plateau. It classifies someone as "a person who can't." It judges something as "a model that can't." It cuts right before the trigger.

But the real difference is made right there.

A loaded plateau looks like stagnation from the outside. But inside, accumulation is happening. And the moment it crosses the threshold, it suddenly looks like a different being.

Breaking the gates is arithmetic before it's morality. It's finding loaded plateaus and planting cheap triggers in them.

Making a new spring is hard. It was wound long ago by genes and experience and schema. But the trigger is cheap. And right now, for the first time in history, the price of that trigger is converging on zero.

This is where the heart of AI lies.

AI isn't magic that revives a dead plateau. But it can be the trigger that finds a loaded plateau, gives it direction, and takes it over the threshold.

Plateaus on the verge of boiling are sitting in the provinces, in India, in China, in nameless rooms. Someone will call them "water that won't boil."

The work splits right there.

Will you see it as water that won't boil? Or as water at 99 degrees?

And will you let go at 99 degrees? Or see it through to the end?