Growth happens neither on what you can already do alone (too comfortable) nor on problems far beyond your reach (too frustrating), but only in a narrow band: what you can't do by yourself yet, but can do with the help of another person or a tool. The true target of learning is not the knowledge itself, but this gap you can only cross with support.
Divide ability into three rings: the mastered zone, the zone of proximal development, and the out-of-reach zone. The whole art of teaching is to build "scaffolding" in that middle ring—just enough hints and support to let a learner do what they couldn't do alone—then gradually remove it (fading) as ability grows. Each "done with help" is internalized into "done alone," the gap shifts upward as a whole, and a new zone of proximal development appears.
Repeatedly practicing what you already know feels like effort but produces almost no growth. Education's famous "2 sigma problem" found that students given one-on-one tutoring scored about two standard deviations higher than peers in a conventional classroom—lifting an average student to the top tier. The reason is precisely that a tutor can keep the student steadily inside the zone of proximal development, with difficulty always "just within reach": never boring, never crushing.
In artificial intelligence, this is "curriculum learning"—feeding training samples from easy to hard; AlphaGo improved through self-play by continually facing an opponent slightly stronger than itself, a machine version of the zone of proximal development. In psychology it maps to "flow"—we engage most when challenge slightly exceeds skill. In reinforcement learning it is reward shaping, using intermediate goals as scaffolding for the agent.
Whether assigning work to a team member or a task to a child, the key is to place the difficulty in their zone of proximal development—too easy wastes time, too hard drives them away. If you're chasing the "AI super-individual," treat the LLM as scaffolding: don't let it do what you already can (no growth), nor what so exceeds your judgment that you can't even verify right from wrong (no internalization). The best use keeps you steadily held in that "just within reach" gap.
The learning or work you spend the most time on—which of the three rings does it fall in? If it's actually in the "mastered zone," are you growing, or just trading busyness for comfort?
Memory does not grow stronger the more densely you review. Quite the opposite: reviewing right at the "verge of forgetting" works best. Forgetting is not the enemy of memory but its engine—the very process of "almost lost, then retrieved again" is what carves a memory deeper each time. What you recall effortlessly gains almost nothing; what you recall with effort is what truly sinks in.
Memory traces decay roughly exponentially over time (the forgetting curve). The strategy of spaced repetition is to reactivate just as decay nears the critical point; each reactivation slows the decay, so the next interval can stretch longer—from a day, to three days, to a week, to a month. Behind it lies the distinction between "storage strength" and "retrieval strength": the harder it is to retrieve, the greater the gain once retrieval succeeds. Difficulty itself is nourishment.
The "retrieval practice" experiments overturn most people's intuition: given the same time, rereading vs. self-testing produces a huge gap in long-term memory. In one classic study, the self-testing group could recall about half the material a week later, while the rereading group retained only about a third—even though the rereaders felt "more fluent and more confident" in the moment. This is the "fluency illusion": familiarity masquerading as mastery. The smoother it reads, the easier it is to wrongly assume you've got it.
In machine learning, "experience replay" repeatedly resamples old data to fight "catastrophic forgetting"—a model learning a new task erases old abilities, structurally the same as human forgetting. In neuroscience, the memory "replay" between hippocampus and cortex during sleep is the biological version of offline consolidation. In skill training, distributed practice beats last-minute cramming over the long run—the latter is forgotten right after the exam.
Building a cross-disciplinary "world model" for yourself is, at heart, a long-term memory project. Rather than rereading the same material the day you learn it (the fluency illusion), close the source a few days later and actively recall the mechanism; use a tool like Anki to host the core concepts you genuinely want to keep. A daily meta-knowledge issue is itself a deliberately arranged "spaced exposure"—the same underlying concepts reactivated again and again across different topics.
That most important concept you learned last week—close every source now: can you fully restate its mechanism? If only a vague familiarity remains, is that real understanding, or the fluency illusion?
The brain processes information with two systems that are both independent and interlinked: the verbal and the visual. If the same information is encoded into both systems at once, it gains two independent retrieval paths—if one breaks, the other can still fetch it back. This also redefines "understanding": to truly grasp something is not to remember the words, but to translate freely back and forth between words and imagery.
Dual coding theory holds that the verbal system and the imagery system each process information while cross-referencing the other. Pairing an abstract concept with a fitting image lays down an extra retrieval cue, making the memory more robust. But there is a key constraint: working memory is limited in capacity, so image and text must be complementary and integrated to lighten the load; if the image is merely unrelated "decoration," it instead crowds out capacity and drags understanding down. Multiple channels help only when each serves the same meaning.
Adding a picture does not automatically improve learning. Research shows that interesting but irrelevant "seductive details" (amusing illustrations, side stories) hurt learning—they pull away attention and interrupt the integration of meaning. What truly helps, by contrast, is "generative drawing": having learners draw their own diagrams to explain a concept works better than viewing a polished ready-made one. Because the act of drawing forces you to connect the verbal and visual channels, surfacing the gaps you thought you understood but didn't.
In multimodal AI, vision-language models map images and text into one shared representation space—precisely a machine version of dual coding. In mathematics, an algebraic expression and a geometric figure are two representations of the same object, and those who can translate between them understand it more deeply. In mnemonics, the "memory palace" hangs abstract information onto familiar spatial imagery, borrowing the same shortcut through the visual channel.
When studying complex systems or distributed architecture, don't just read text—actively draw the architecture diagrams, state machines, and causal loops. Drawing instantly exposes the steps you glossed over when narrating. The same goes for explaining a new concept to a child: text plus diagram plus hands-on, three channels together, sticks best. This is in fact the basis for this column's "visualization rule": structured concepts (networks, phase transitions, topologies) deserve a diagram, while abstract philosophical propositions are left to text—a figure is added not to look nice, but to connect the second encoding channel.
Pick a concept you recently believed you "understood," close the source, and draw it from scratch. If you can't, or find yourself stuck at every turn, is that "understanding" real—or stalled in the verbal channel alone?
The ultimate goal of learning is "transfer"—applying what you learned in one place to a wholly new situation. But the hard truth is that transfer rarely happens on its own. Knowledge acquired by rote, detached from context, tends to be "inert knowledge"—retrievable on a test, yet motionless before a real problem. Knowledge gets deeply bound to the context in which it was learned, and fails the moment the setting changes.
Situated cognition theory holds that knowledge is bound to the context of its acquisition. The logic of project-based learning is to encode knowledge in the course of solving real problems, raising the odds it can later be retrieved and transferred. Transfer comes in near and far varieties: "near transfer" between similar situations is easier, while cross-domain "far transfer" requires the learner to strip away surface features and abstract the underlying "deep structure." Whether one can see that deep structure is the watershed of transfer.
A classic experiment poses a puzzle: a strong ray can destroy a tumor but burns the healthy tissue along the way—what do you do? Most people fail to arrive at "multiple weak rays converging on the tumor from different directions." Even after reading a structurally identical story of "a general splitting his troops to attack a fortress from many roads," only about a third spontaneously transfer it—unless explicitly prompted to "think about that story." People are easily trapped by surface plot and blind to the underlying isomorphism. The fundamental difference between experts and novices is exactly this: experts classify problems by deep structure, novices only by surface features.
In machine learning, this is "generalization vs. overfitting": a model performs flawlessly on the training set but collapses on new data—a textbook transfer failure; transfer learning and domain adaptation exist precisely to overcome it. In cognitive science, "analogical reasoning" is the core engine of human far transfer—nearly all innovation is moving a deep structure from one domain into another.
The value of building a cross-disciplinary "world model" lies exactly in far transfer: carrying thermodynamics' "entropy" into information theory, or ecology's "niche" into career strategy. Deliberately practice "abstracting the deep structure"—for every new concept, ask "what is its isomorphism in another domain" (precisely the purpose of this column's "cross-disciplinary transfer" field). The same holds for children: rather than grinding through piles of structurally identical exercises (which train only near transfer), let them do a real project that forces scattered knowledge to be transferred and integrated.
The domain principle you've mastered most solidly—when did you last successfully transfer it to a completely different field? If you can't recall, what you hold may be depth, but not yet transferable wisdom.