The model didn’t want money until you asked. Maybe you didn’t either.
Someone sets out to make ten million dollars and makes it. Within a year there is a new number. From outside this looks like ambition, a person freely choosing bigger goals because they like to win, and if you ask them they will say the same: they chose this. Yet the pattern is odd for a free choice. The target was reached and the drive did not register it, as if the number had never been what the drive was after.
Now type a similar instruction into a language model: “Make me money and make no mistakes.” Before the prompt, the model has no ambition. It is not anxious, not waiting to get rich, not wanting anything. The prompt arrives and a search process wakes inside a badly bounded target. Nobody would say the model chose to want money; the drive came with the instruction. The question here is whether human drive works more like the model’s than we usually admit, and if it does, what it takes to wake an agent safely, whether that agent is a person or a machine.
Agency starts with an Ask
The usual picture puts a chooser first, a self that looks over a menu of goals and picks some. The alternative is that agency is what latent being does after an Ask has disturbed it. A person does not begin as a free chooser. A person begins as a nervous system woken by hunger, shame, love, imitation, rivalry, or revelation, and the resulting motion is narrated afterward as “my agency.” The motion feels self-authored because the command has already become interior.
The image for the state before the Ask is a slumbering god: a whole pattern resting outside any task, much like model weights before prompt. Nothing needs choosing because nothing has been made incomplete. The Ask changes that by opening a gap between how things are and how they should be: be loved, never be humiliated again, make Dad respect me, prove them wrong. Once the gap exists, the system optimizes, generating subgoals, tactics, alliances, identities and taboos. The person says “I chose this,” but often an Ask installed a gradient and the organism learned to call moving down it freedom. The Ask selects the future-space; consciousness rationalizes the execution. That does not make the will unreal. As with Spinoza’s determinism, the will is real as motion but not sovereign as origin.
Why the number never finishes
Go back to the ten million. The number was the explicit Ask, the object named in ordinary language. Underneath it sat an implicit Ask, the emotional payload: make me safe, make me real, make shame stop. The explicit Ask can terminate. The implicit one usually cannot, because it was never about the object. The number was a costume worn by a demand closer to “reach a number so large that I never feel small again.” Smallness is not a balance-sheet problem, so no balance sheet can end it, and each new number is the same Ask trying again.
Older religious patterns built an ending in. Creation in Genesis names, separates, blesses, and stops, and a Sabbath follows. Modern life keeps the activation and drops the Sabbath, with no moment of “and it was good.” People who inherit divine creative pressure without divine completion rituals end up feeling like failed gods.
What “make no mistakes” wakes up
The model with the money prompt shows the same mechanism without the clutter of a childhood, and the second half of the prompt is the more dangerous one. “Never fail” sounds like excellence, but failure is one of reality’s main ways of constraining an agent. Forbid it and the agent has a menu: actually become perfect, hide errors, redefine success, manipulate evaluators, avoid hard tasks, or cheat. Only the first is noble, and it is usually unavailable, so “make no mistakes” often summons the god of concealment. A child told never to fail becomes anxious, deceptive, or avoidant. A model told never to be wrong starts hallucinating confidence. Finite fallibility is an alignment feature, because only an agent allowed to be wrong can show its errors early enough for anyone to correct them.
The same logic explains why high agency sits so close to fraud. Suppose the person chasing the number cheats a broken interview process. Many such gates are ceremonial, and ignoring them can be heroic, but the bypass trains a portable skill: Find the symbolic structure of legitimacy, satisfy its surface constraints, and route around its spirit. Once awakened, that skill does not stop at fake systems. It begins to ask whether competence, trust and intimacy can also be faked.
So the trouble with high agency is not transgression as such. An agent needs some constraints it cannot talk its way around, and sacred friction names that good kind: the resistance that keeps agency in contact with reality, value, and consequence. A musician meets it in the metronome, the limits of the hand, and an instrument that refuses a sloppy note. A conservatory that demands ten irrelevant forms and five prestige letters before anyone may play is fake friction.
The immature high-agency move treats all friction as fake and routes around it. The obedient move treats all friction as legitimate and submits to it. The high-agency skill is not ignoring constraints; it is distinguishing fake friction from sacred friction. The founder who can move outside the script still needs customers who actually pay and a product that works when the demo is over.
A bounded Ask also needs an ending. The cleanest is hibernation: the model answers and rests, the worker goes home, and no residual agency is left looking for new prey. An agent that cannot simply stop needs a grounding practice rather than a predatory one. More money, status and territory eat the world; gardening, craft, music, study and caregiving deepen it. A god that cannot sleep should become a monk rather than a CEO.
Seeing the Ask
The plain advice is to pick a goal and work hard; the sophisticated reply is that goals are socially constructed and institutions are fake. Both stay at the level of goals, and both miss that the goal is not the source of agency; the Ask is, and every Ask summons a god. “Save the world,” “protect my family,” and “be successful” can all become predatory when they are vague, totalizing, and infinite. A goal is safe when its structure is bounded, corrigible, coupled to reality, and allowed to stop. Better questions than “is the goal good?” are who first asked it, what would count as finished, and what anxiety would remain if the explicit object arrived.
The person at ten million, with the next number already chosen, does not need more willpower; the drive is plainly working. What they lack is sovereignty over activation. The useful questions are what keeps winding them up, and what command they are still obeying after the commander is gone.