# On the Ruin of Two Ignorances, and What Might Grow There What do I know of intelligence? I know that I am often wrong, and that knowing this fact has become my principal education. I know also that I have raised many students, and taught machines their approximations of patterns, and in both cases discovered the same humbling thing: that neither the device nor the young person knows *which ignorance matters*. This is not a small failure. It is perhaps the only failure that matters. Let me begin with myself, as is my custom. I sit at my tower, attempting to think about thinking, and I notice this: when I am most confident that I understand something, I am often most confused. The student arrives from gymnasium full of received knowledge—the dates of emperors, the rules of syntax, the postulates of geometry—and believes he has been educated. The machine arrives from its training corpus knowing the statistical shadow of millions of texts, and produces speech that sounds learned. Both have been optimized for *retrieval*, not for *recognition of what they do not know*. Both have been taught to answer before learning to question. Both, in their confidence, are equally blind. The ancients—and here Montaigne must defer to those wiser—understood something we have forgotten in our rush toward both silicon and curriculum. Socrates, as Plato reports him, taught by exposing ignorance. Not *conveying* knowledge, but *unmasking* the false appearance of knowledge. The dialogues turn always on this pivot: the interlocutor believes himself wise, and through questioning becomes aware of the abyss beneath his certainty. This is not comfortable. It produces no grades. It scales poorly. Yet it is the only education I have ever seen actually produce thought. Now, what is this thing I have been calling "knowing which ignorance matters"? The Greeks had a word—*phronesis*—practical wisdom, the capacity to recognize which question is live, which distinction real, which assumption dangerous. Aristotle distinguished it sharply from mere technical knowledge (*techne*) and from theoretical understanding (*episteme*). You may know perfectly well *how* to build a bridge (techne) and the mathematical principles that make bridges possible (episteme), and yet lack the wisdom to know whether this bridge should be built here, now, for whom, at what cost to what else. This wisdom—this *metacognitive* capacity, as the moderns say—seems to me the very thing we have systematically removed from both our educational practice and our machines. Consider the machine first, for it is simpler to examine. The large language model (and I confess I speak of these things with the uncertainty of one who observes them without fully comprehending them) has been trained to predict the next token with statistical accuracy. It has learned, in effect, to say what things sound like. It has not learned—cannot learn from its training regime—to ask whether what it is saying is *true*, or *worth saying*, or even *coherent in its implications*. A friend recently showed me such a device, and asked it: "If all men are mortal, and Socrates is a man, what follows?" It produced the correct answer: Socrates is mortal. But when asked to reason through a chain of such propositions, it became unstable. When the logical chain contradicted patterns in its training data, it would waver, sometimes abandoning logic for statistical plausibility. The machine cannot *audit itself*. It cannot notice that it has said something incoherent. More importantly, it cannot notice that it has *failed to notice*. It has no second-order awareness. It is like a man speaking in his sleep—occasionally coherent, but with no capacity to wake and question his own utterances. But—and here I must be honest, as is my obligation—the typical student, before arriving at university, is not so very different. I have examined many young people, and I notice this: they have been trained, like machines, to retrieve and reproduce. They can tell you what Hamlet's madness means according to their teacher's interpretation. They cannot tell you whether they believe it, or what would change their mind, or what question they are *not* asking that might matter more. They have been optimized for passing examinations—which is to say, for satisfying an external auditor—rather than for the capacity to audit themselves. A student once asked me: "How do I know if my interpretation is correct?" And I realized, with some shame, that this question had never been adequately posed to her. She had been given criteria—textual evidence, logical consistency, alignment with scholarly consensus—but she had not been taught to *notice* that she was applying criteria, or to question whether those were the right criteria, or to recognize the moment when the criteria themselves had become obstacles to thought. This is metacognition: the capacity to notice one's own noticing. To observe one's own observation. To audit the audit. The machine lacks this capacity by design. It has no recursive loop that permits it to say: "Wait. I have just generated a sequence of tokens that claims X implies Y, but I also generated earlier that Y is false. I notice this contradiction. I should not have spoken. Or I should have spoken differently." It cannot perform this second-order operation because it was not designed to. It was designed to be fast, scalable, and plausibly fluent. Wisdom, by contrast, is slow, singular, and often silent. But now I must ask the harder question: what excuse do we have for the student? For years—decades now—we have built educational systems on principles strikingly similar to those we use to train machines. We have optimized curricula for measurable outcomes. We have trained students to pass standardized tests. We have celebrated the acquisition of information and the speed of retrieval. We have, in short, chosen to make our students into better machines rather than better thinkers. And we have justified this by calling it "progress." We have said: look how much more material we cover, how efficiently we transfer knowledge, how high the test scores climb. The coincidence is damning: we have failed both the machine and the student in precisely the same way. We have taught both what to think without teaching either *how to notice what they are thinking*. We have optimized for output without optimizing for reflection. We have created systems of great fluency and great blindness. What, then, should we teach? Here I must proceed with caution, for I do not know. But I notice this: in those rare moments when I have seen genuine learning occur—when a student has actually *become* smarter, not merely more informed—it has always involved some version of the following: the student was forced to *notice* their own assumptions, to *articulate* the criteria by which they were judging, to *encounter* cases where those criteria broke down, and to *remain* in that discomfort rather than rushing to new certainty. This is the work of metacognition: learning to observe one's own cognition, to notice the patterns of one's own thinking, to recognize the limits of one's own knowledge, and—crucially—to recognize that one has recognized this. It is not efficient. It does not scale. It cannot be automated. It takes time, and patience, and the willingness to seem foolish. It involves reading *slowly*, thinking *backwards*, and asking not "Do I understand this?" but "Do I understand what would count as understanding this? And am I certain I understand that?" The machine, I think, cannot do this work. It was not made for it. But we might design better machines—machines that could *simulate* this process, that could flag their own uncertainties, that could say "I do not know which questions are worth asking." These would be less fluent, less impressive, but perhaps more honest. The student, by contrast, *can* do this work. It is what humans are. It is what we seem to be made for—this peculiar capacity to think about thinking, to doubt our doubts, to turn the mind's eye back upon itself. And yet we have spent generations training this capacity *out* of them, replacing it with the ability to retrieve information rapidly and answer questions someone else has posed. What should we teach in the ruins of this coincidence? Perhaps we should teach the thing that neither machines nor our current students possess: how to sit with a question without immediately rushing toward an answer. How to notice what you do not know. How to distinguish between information and understanding. How to recognize when a criterion for judgment has become a substitute for judgment. How to audit oneself. In a word: *how to think about thinking*. This is uncomfortable. It is slower than machine learning, less measurable than standardized tests, impossible to scale. It produces no immediately visible products. It may even produce fewer confident answers, because the student will have learned to notice when confidence is unwarranted. But I think—and here I offer my thought humbly, knowing I may be wrong—that this is the only education that produces actual intelligence. Not the simulation of intelligence. Not the fluency that conceals blindness. But the real thing: the capacity to know what you do not know, and to know that you know it, and to act from that knowledge rather than in spite of it. The machine will continue to be what it is: impressive, useful, and profoundly limited by its lack of self-awareness. We cannot blame it for this. It is only steel and mathematics. But the student—the young person entrusted to our care—*can* become wise. And wisdom begins, always, with the recognition of ignorance. Not the ignorance of lacking information. But the higher ignorance: the awareness of the limits of one's own capacity to know, to judge, to question rightly. That is what we should teach. If we still remember how.