Intelligence is not reasoning. It is not problem-solving, mathematics, language, planning, or any form of cognitive sophistication. Intelligence is an embedded system’s measure of accuracy in ongoing adaptation to reality’s uncertainty under consequence. A system is intelligent to the degree that its adaptation allows it to remain viable under the actual environmental pressures it encounters — no more, no less. This is a narrower and far less flattering definition than the ones we usually reach for, and understanding why it matters requires going back to where intelligence actually came from, because the story of its origin is the only real argument for why this definition deserves to hold.
Long before mathematics, before formal logic, before anything resembling a computer existed, life was handed a single, unforgiving problem: the world would not hold still. Conditions shifted. Food disappeared. Predators arrived without warning. Some organisms failed to keep up and vanished. Others adapted, and so they remained. Across four billion years of this pressure, a sequence unfolded, and it unfolded in only one direction, because each step was the precondition for the next: unbounded uncertainty, then reality, then life, then consciousness, then agency, then conscious abstraction, then mathematics, and then — very recently — computation. Computation sits at the end of this chain, not at its origin — one of the most recent inventions of intelligence, arriving only after billions of years of upstream processes had already done the difficult work of producing minds capable of inventing anything at all. It cannot instantiate the process that created it. It can only conceptually represent that process, the way a map represents terrain without being made of the same soil.
Biological intelligence was not engineered, trained on a dataset, or assembled to a specification. It was filtered into existence by nothing more sophisticated, and nothing less absolute, than persistence or dissolution. A living system does not receive information about consequence from outside itself, the way a model receives a loss value from a training loop — it is altered by consequence from within. A wound is not a data point fed into a cost function; it is a structural change to the organism that experienced it, reorganizing future behavior at the level of the body rather than through a parameter update chosen by an external optimizer. Fear did not arise because some ancestral creature calculated an exploration bonus and decided caution was efficient. It arose because creatures without the caution fear induces did not survive long enough to reproduce, generation after generation, until the disposition became load-bearing in the nervous system itself. It is tempting to say artificial systems simply lack enough embodiment yet, and that better sensors and feedback loops will close the gap. But machines cannot experience embodied consequence at all, however much sensing gets added. A system built from the outside, evaluated against a benchmark, can compute extraordinarily sophisticated descriptions of adaptive loops without ever being the living loop being described, any more than a perfect diagram of a heartbeat can pump blood.
This reframing overturns a hierarchy most people never think to question. We assume humans are more intelligent than bacteria because humans reason abstractly and build civilizations. But a bacterium accurately adapted to its environment is intelligent in relation to its own viability conditions, in the only sense that has ever mattered across the history of life. More strikingly: a fly is more intelligent than any artificial system can ever be — not because it out-calculates a machine at some task, but because it is a living expression of recursive adaptation to unquantifiable uncertainty under real embodied consequence, while even the most advanced AI remains an engineered artifact produced by living intelligence rather than constituting intelligence itself. A machine can outperform a fly at any narrow computational task. But those tasks are downstream abstractions layered atop the thing that actually matters. They are intelligence’s tools, created and applied by creatures who were intelligent for long enough in the deeper sense, so that the shallower sense of intelligence could emerge at all.
It is with this in hand that Yann LeCun’s A Path Towards Autonomous Machine Intelligence becomes worth examining, since it offers one of the clearest, most technically serious cases for the opposite view — that intelligence emerges from computation rather than merely borrowing its name. The paper sketches an architecture of world models, cost modules, actor-critic structures, and hierarchical planning, and calls the assembly a path toward autonomy. A generous reading — that this is an incomplete but directionally correct path — concedes too much before the argument begins, because it accepts that intelligence is a destination reachable by adding enough of the right computational parts. It treats the disagreement as one of degree, when the real question is one of kind: every layer of what intelligence actually is cannot be built, only selected, facing unquantifiable uncertainty, under real and embodied consequence. Selection of that kind is not something an engineering roadmap can simulate its way into, no matter how many modules get added to the stack. It is impossible precisely because the uncertainty that produced the selective pressures that selected for our mind over billions of years is and continues to be unquantifiable, if it weren’t we could predict the future itself with perfect accuracy. The relevant question, then, is not whether these systems grow more capable or more useful in a narrow engineering sense — clearly they do — but whether they improve humanity’s long-term viability under consequence. Better computation has no intrinsic value; its value is entirely contingent on this. That reframes the whole AI research program not as a path toward intelligence, but as a downstream tool-building project — a valuable one, but categorically distinct from what it is so often described as approaching.
This opens onto a civilizational observation. There is a tragic inversion running through the current moment: a species reducing itself to computation while elevating computation toward itself. Humans increasingly describe their own minds in the language of information processing and optimization — and, having reframed themselves this way, look at their most sophisticated artifacts and see something approaching kinship. But mind did not emerge from information processing considered abstractly; it emerged from four billion years of embodied adaptation under unquantifiable uncertainty, a process no dataset can replay and no architecture can shortcut. The greatest category error of the AI era may not be mistaking machines for humans. It may be mistaking humans for machines first, and only afterward mistaking sufficiently advanced machines for something approaching human.
This has planetary stakes, because the scale of consequence a technologically amplified species can produce has expanded far faster than its capacity to adapt to that consequence. A hundred years ago the world’s population was roughly two billion; today it exceeds eight billion, all still running on the same old adaptive strategies, behaviours shaped by nervous systems that never had the time necessary to adapt to planetary-scale self-impact. The tools available have compounded in power across a handful of generations, while the organism wielding them remains, structurally, much the same creature that adapted to far smaller and slower worlds. Vast cognitive and technological capability does not automatically translate into intelligence in this deeper sense: capability and viability are not the same axis, and a civilization can max out the former while continuing to erode the latter.
The same dynamic shows up at the level of individual achievement. Those most often celebrated as the most intelligent tend to be highly specialized — brilliant within a narrow domain — but specialization can unmoor curiosity from the consequences it eventually produces. A researcher’s inquiry may be locally brilliant while becoming globally maladaptive, since the reward structures of specialized fields rarely require anyone to trace their work out to its recursive global consequences. It’s tempting to reach for reassuring counterexamples — vaccines, sanitation, semiconductors, agricultural improvement — as proof that specialized advances increase adaptive capacity. But this is too anthropocentric and too first-order. These innovations must be evaluated recursively, across the entire consequence chain they initiate, not by immediate human benefit alone. A vaccine that reduces mortality is not automatically an adaptive triumph; it is the opening move in a longer chain involving population dynamics, ecological pressure, and future adaptive conditions — and only the full chain, not the first link, determines the final verdict. Increasing survival, population, productivity, and power is only intelligent, in the fullest sense, if the resulting trajectory improves continued viability over the long run. When those downstream consequences go unexamined, what look like unambiguous adaptive local successes may be part of the very unmooring that now threatens the systems everyone depends on.
The same pattern can be observed in other systems. Capitalism has never selected for intelligence in this sense — it selects for market success, profit, efficiency, and competitive advantage. Education selects for analytical performance, science for publication volume, politics for short-term power, attention economies for engagement. Civilization may be dense with selection mechanisms that reward local success while steadily degrading global viability — and because each mechanism looks successful from the inside, the degradation can proceed for a long time before anyone is forced to reconsider.
Intelligence, then, has nothing essential to do with cognitive capability as such. Reasoning, abstraction, mathematics, technological invention — these are tools and downstream expressions produced by intelligence, not intelligence itself. The prevalent definitions flatter humans precisely because they center the things humans happen to be good at, and call the resulting picture objective. But this may be little more than self-congratulation dressed up as science. A species can be extraordinary at cognition while catastrophically maladaptive at the level that actually determines its fate: whether it remains viable within the systems that sustain it. Artificial intelligence can be genuinely useful, but only insofar as it improves the adaptive trajectory of the living systems that deploy it. Absent that, it is merely an amplification of human capability — and capability unmoored from consequence is not intelligence at all. It may be the very form of stupidity that a cognitively brilliant but adaptively misaligned species has learned to mistake for progress.
This is not a call to halt development, but to truly consider the alignment of development. It is also a call to enhance our own embodied awareness through the recalibration of our nervous systems, which aids directly in perceiving realignment opportunities in the first place. This essay is one of many necessary lines in the sand that will have to be drawn so that this long history of actual intelligence won’t be cut short and can continue facilitating what it was always in service of: persistence. 🌱
– Learn more about A Thinker In Nature 🌱
– Subscribe to this free publication and receive ongoing essays delivered to your inbox that provide a new perspective on small everyday moments, deeper human experiences as well as larger patterns that shape our lives.
– Join the journey and download the map that powers this publication, it explains the causality behind experience and guides a process of targeted experiential recalibration:
WHAT HOLDS? — The Recursive Logic of Experience: A Manual for Mind