"Whether it makes tools" was long treated as the uncrossable line between human and beast. Then chimpanzees were filmed stripping a twig, poking it into a termite mound, and fishing for termites — and the line collapsed on the spot. The deeper shift: tool use isn't an on/off switch but a bundle of cognitive prerequisites — seeing an object as what it "could become" rather than what it "currently is." What makes humans unique was never the tools themselves; it lies elsewhere.
Making a tool demands a means–end mental representation: for a goal not yet in hand, you first reshape something currently useless. A chimp stripping leaves before fishing for termites is "pre-processing" — its mind already holds the future action. More striking are New Caledonian crows: they bend a straight piece of wire into a hook to reach food, though wire does not exist in the wild. That reveals flexible causal reasoning (grasping "hookness" as a functional property), not instinct hard-coded into genes. The depth of tool behavior is measured by how far it sits from instinct and how close to "understanding why."
A crow with a brain the size of a walnut can go toe-to-toe with apes on multi-step problems. A New Caledonian crow will first use a short stick to extract an out-of-reach long stick, then use the long stick to get food — "using a tool to get a tool," a metatool inference that requires laying out the whole causal chain in advance. Crows and primates split on the evolutionary tree some 300 million years ago and have utterly different brains, yet each independently evolved higher intelligence — living proof of "convergent evolution": intelligence is no primate monopoly, but the same blade carved again and again by particular ecological pressures.
In AI, this is the cognitive core of large-language-model "function calling" — the leap from "compute it yourself" to "know which external tool to call" is exactly that means–end threshold; whether it can chain and recursively compose tools mirrors the crow's metatool inference. In cognitive science it is Gibson's "affordance" — we perceive not objects but the actions they permit. In software engineering, using tools to build tools, bootstrapping whole toolchains layer by layer, is the crow's short-stick-for-long-stick writ large.
When designing AI agents, the gulf between two kinds of system is precisely "instinctive" versus "insightful" tool use: one merely calls an API by a fixed script; the other, in a scene it has never met, assembles a few tools on the fly into a solution. The former is leaf-stripping instinct; the latter is wire-bending causal understanding. To gauge an agent's ceiling, don't count how many tools it wields — watch whether it can invent, for a brand-new goal, a toolchain it was never taught.
Think of the last automation or AI workflow you built: is it more like a chimp fishing for termites (highly skilled at a fixed scene), or a crow's metatool (composing on the fly for unfamiliar goals)? To cross that threshold, what it lacks — is it more tools, or a layer of causal reasoning?
Humanity's real superpower is not how smart any single brain is, but cumulative culture — knowledge that ratchets across generations, only forward, never back. No one person could invent a smartphone from scratch; we merely stand on a knowledge ratchet that has been cranked for millennia. Animals have "traditions" too, yet almost none have this ratchet: each generation must start over nearly from zero. The divide between human and beast hides in that one word — "cumulative."
Cumulative culture needs three things at once: social learning, high-fidelity transmission, and a non-slipping ratchet effect (proposed by Michael Tomasello). The key is imitation rather than mere emulation — imitation copies the whole process, emulation copies only the result. Copy only the result and each generation must rediscover the method, so improvements never accumulate; copy the process faithfully and you inherit predecessors' know-how intact even without understanding it, then add a little more. The ratchet doesn't slip precisely because of this "copy without asking why."
A classic controlled study places children and chimpanzees before the same transparent puzzle box, with the demonstrator deliberately inserting several actions irrelevant to retrieving the reward. The chimp is "shrewd" — it sees at a glance which steps are useless and skips them; the human child copies even the useless actions — this is "over-imitation." The child looks dumber, yet this is the very engine of cumulative culture: it lets humans faithfully reproduce operations they don't understand, thereby preserving knowledge whose "reasons are unclear but which really works." The chimp's "rationality" is exactly why its culture cannot ratchet upward.
In economic growth, this is the compounding of the knowledge stock — only non-retreating accumulation yields long-run exponential growth. In machine learning it maps to knowledge distillation and transfer learning: a new model inherits the previous one's "know-how" without training from zero. In evolutionary biology it is gene–culture coevolution, culture as a second channel of inheritance. In organizational management it is institutional memory — why documentation and retrospectives matter so much: without high-fidelity transmission, every staff turnover is a slip of the ratchet.
"Over-imitation" is a double-edged sword for a tech team. On one side, copying a best practice you haven't fully digested (an architectural paradigm, a review process) can bank the know-how first and understand it later — that is how a team's capability ratchets up. On the other, mindlessly reproducing rules that stopped working long ago becomes the org's appendix. Real judgment is telling apart which "incomprehensible extra steps" are wisdom waiting to be unlocked, and which are merely historical baggage no one dares delete.
Is your team more like a ratchet that accumulates, or one that slips back to zero at every handover? To pass a core member's "hard-to-articulate but very effective" know-how to the next person — are you relying on high-fidelity process replication, or have you left behind only the results?
Do animals know that "others also have a mind" — beliefs, knowledge, intentions? This debate, running from the 1978 paper "Does the chimpanzee have a theory of mind?" to today, exposes a brutally hard threshold: how to separate "mind-reading" from "behavior-reading." Much that looks like mind-reading may be exquisitely sophisticated behavior prediction. And that ambiguity is itself profound — it hangs equally over humans and today's AI: how do you know the other party truly understands, versus merely fitting a pattern?
Theory of mind means attributing unseen mental states to others. The gold-standard test is the "false-belief task": can you predict what a person who holds mistaken information will do? Chimpanzees generally fail explicit false-belief tests; but switch to a competitive food scene and they turn shrewd — a subordinate chimp clearly knows which piece of food a dominant rival "can" and "cannot" see, and reaches for the rival's blind spot. So the debate narrows to a subtle question: chimps can track "what others see and know," yet may not be able to represent "another holds a belief contrary to the facts."
The most counterintuitive finding: chimps look dumb on "cooperation" tasks but sharp on "competition" tasks. Ask two chimps to cooperate and share, and they often look blank; but the moment it means stealing food from a dominant's mouth and having to reckon with what the rival "knows," they turn instantly calculating. This hints that their social cognition was honed for competition, not sharing — the opposite of the human case. Scrub jays echo this: a jay that has pilfered others' caches, on noticing it was watched while caching, will secretly re-hide its food elsewhere — projecting its own thievish mind onto the other. It takes a thief to know a thief.
In AI, the argument over "does the large model have a theory of mind" is almost a carbon copy of this epistemic trap: when the model passes a false-belief test, has it acquired a model of other minds, or fitted the behavioral statistics of vast dialogue? In philosophy it is the ancient "problem of other minds." In game theory, every strategic deduction presupposes "I model you modeling me." In developmental and clinical psychology, the false-belief task is a yardstick for understanding differences in autistic social cognition.
When you say "this model understood my intent," pause a second: are you reading its mind, or reading its behavior? The chimpanzee debate's warning for the AI age is to resist attributing "mind-reading" to a system that may only be doing super-behavior-prediction — over-attribution makes you overestimate its reliability in unfamiliar situations. The same holds for judging collaborators: distinguish whether they truly grasp your goal, or are merely skillfully following your reactions.
The last time you judged that someone (or some AI) "got it," what was your evidence? Drop it into a scene it has never seen, where the cues run contrary to the usual — would it still act as if it truly understood, or reveal the tell of "only fitting behavior"?
A "sense of fairness" — refusing an unequal split even at a cost to oneself — was thought to be a uniquely human moral invention. Then capuchin monkeys threw a fit over "equal work, unequal pay," and that assumption was shaken. It hints that fairness has deep evolutionary roots: not a morality taught by culture, but an ancient mechanism for sustaining cooperation and preventing chronic exploitation. The bedrock of morality may be more "biological" than we assumed.
The core is inequity aversion. In a famous experiment, two capuchins do the same task for food: as long as both get cucumber, the monkey munches contentedly; but the instant it sees a partner get a tastier grape for the same work, it refuses to continue — even hurling the cucumber back at the experimenter. Note: the cucumber itself is fine; the problem is the comparison. This is "first-order" inequity aversion — I protest when I get less. Rarer, and closer to humans, is "second-order" — I feel uneasy even when I get more. The latter is almost never seen in animals.
A monkey rejecting a perfectly good cucumber is "irrational" in the economics textbook — throwing away a positive payoff for nothing. But seen differently, it is precisely rational: what it rejects is not the cucumber but "being systematically undervalued." In a relationship built on long, repeated cooperation, tolerating one unfairness is like signaling "you may keep taking advantage of me"; flaring up on the spot is a costly but effective anti-exploitation alarm. Even dogs show a version: ask two dogs to shake, and if only the partner is rewarded, the other soon quits — the arithmetic of fairness runs deep in the instincts of cooperative species.
In behavioral economics, this is the root of why people in the "ultimatum game" reject an unfair offer even at mutual loss — single-handedly overturning the "rational economic man." In organizational design, whether staff stay is often decided by perceived fairness rather than absolute pay — one opaque raise can wreck a whole team's morale. In AI alignment it presses the question of how to design rewards and resource splits that "feel fair" to many parties. In evolution it is the emotional enforcer of reciprocal altruism and "partner choice."
That capuchin flinging back the cucumber is the top engineer on your team watching a same-level peer land the "grape." Human responses to reward are relative, not absolute: one opaque promotion or bonus can ignite the "unfairness alarm" even when the person's own package hasn't worsened. What a manager truly has to govern is often not the absolute numbers, but that ever-present, exquisitely sensitive "comparison" — quietly driving retention, engagement, and trust.
In your team or family, which behaviors are quietly driven by "relative comparison" rather than "absolute gain or loss"? The last time you felt unfairness and wanted to "throw the cucumber back," what was the trigger — a real loss to yourself, or merely the sight of someone else's "grape"?