The chatbot is most useful to the expert who has stopped asking it for answers.
A biologist finds a result that should not exist. She asks an LLM to restate the anomaly as an ecology problem, a control problem, an evolutionary game, and a measurement failure. Three answers are banal. One is wrong. The last mentions a mechanism from fungal networks that gives her a new experiment.
The model did not solve the biology. It made a door visible.
That runs against the usual picture of these tools as answer machines, most useful to whoever knows least. The biologist got her value the other way round. At a real frontier, AI’s highest value is not answering the question but multiplying the questions a trained judge can ask. The beginner asks the crystal to remove uncertainty. The frontier thinker uses it to produce more consequential uncertainty, then returns to reality to see which possibility survives.
Why a frontier needs more questions
An established problem arrives with a vocabulary, a method, and a finish line. A frontier problem arrives as clues that may not yet belong to the same problem. The biologist’s anomaly might depend on ecology, on control theory, on a game between organisms, or on a flaw in her instruments, and nothing in her training tells her which. At the center of a discipline, expertise produces fluency and confidence. At its frontier, expertise produces bewilderment. Her task is no longer to search a known solution space but to discover which spaces deserve searching.
An LLM suits that task because of what it is: crystallized intelligence, a manipulable compression of patterns accumulated across human language. It does not contain the world. It contains traces of how people have described, divided, analogized, and argued about the world. So it can translate her mechanism into another field’s vocabulary, generate rival causal stories, expose hidden assumptions, and recall a precedent, such as the fungal networks, that she did not know to search for. This is meta-rationality at conversational speed: the model changes the problem description before anyone tries to solve it. Search retrieves what you can name. The crystal helps name what you did not know to search for.
Judgment turns doors into expeditions
The doors are worthless without someone to judge them. Three of the five answers were banal and one was wrong, and the biologist could tell which only because of years of training. A novice uses the crystal as an oracle and borrows its confidence. An intermediate user treats it as a labor-saving device for work already understood. The frontier thinker uses it as a perspective multiplier, asking for incompatible accounts and for tests that could separate them. The more developed her judgment, the more she can extract without surrendering to the output, because she can tell a fertile analogy from a decorative resemblance and notice when the model has smuggled in the wrong frame. Human intelligence and AI use are complements.
Even then she does not delegate the discovery itself. Discovery is where delegation works worst, because intent and relevant detail are still forming. The fungal mechanism is not a result; it is a week of lab work. The crystal lowers the cost of launching a hypothesis, not the cost of earning belief.
When to work without it
Where did her judgment come from? Largely from years of working without any crystal, and that is the strongest case for abstinence. A pianist who never plays without correction software will not develop an ear. A mathematician who reads the proof before struggling may recognize an argument without building the machinery that could have produced it. Grokking often arrives after visible progress stops, and assistance introduced too early can remove the pressure that would have forced the mind to compress examples into a rule.
But training conditions are not performance conditions. These cases justify deliberate tool abstinence, not an identity built around permanent refusal. It is tempting to wear refusal as status (“my work is too original, tacit, or embodied for AI”), and every guild draws the line between mechanical and sacred work suspiciously close to its own desk. Refusal may protect attention or training. It cannot certify originality. The serious question is not “do you use AI?” but which capacity are you training now, and when does assistance begin extending it rather than replacing its formation? The mature rhythm alternates: struggle unaided until a first-person model exists, use the crystal to multiply frames and objections around it, then test the enlarged field against evidence. Permanent assistance collapses formation. Permanent abstinence starves expansion.
The true frontier humiliates you. The biologist’s anomaly showed her that hard-won expertise is one narrow organ of perception inside a reality with more dimensions than her own, and that humiliation is what made a foreign discipline, and a machine that could reach it, worth asking. The person who says “I already possess the relevant intelligence” is often not standing at the frontier. They are standing inside a local maximum and calling it the edge of the world.