A Worm With 302 Neurons Already Makes Trade-offs. AI Is Still on the Same Ladder.

Nathan Labenz bought Max Bennett's book on the title alone. He had been asking how many conceptual breakthroughs still sit between today's models and functional parity with humans. He would be surprised, though not shocked, if attention really were all we needed and hyperscaling closed the rest. He would be more surprised if the number were larger than five. His best guess is two or three — which leaves the harder questions of what humans do that models still cannot, and whether embodiment, memory, and theory of mind are load-bearing or evolutionary accidents.

Bennett's day job was commercializing what he now calls pre-transformer AI. As co-founder of Bluecore he spent years shipping deep-learning recommenders and customer-segmentation models to large e-commerce companies. The work paid off, but the mismatch never stopped bothering him: models could be astronomically good at some tasks and hopeless at things a person does without thinking, like putting an outfit together. That curiosity turned into emails to neuroscientists, textbooks, a few papers, and then a year off to write *A Brief History of Intelligence*. Along the way he collected mentors including Dileep George, Karl Friston, Joseph LeDoux, and David Redish, who, he says, were especially good at telling him which of his ideas were stupid.

The book's method is to force three fields into one story. Evolutionary neuroscience reconstructs how brains were assembled over 600 million years. Comparative psychology infers what ancestral animals could actually do from what living species can do. AI is the litmus test: if a concept is understood, a stripped-down version should be implementable, and many cherished intuitions fail that test. Reinforcement learning is his favorite exhibit. Trial-and-error looked obvious until people tried to code Thorndike's intuition; the version that finally worked in machines turned out to look a lot like what vertebrate brains were doing all along.

Labenz treats the five breakthroughs as a progress bar he can hold new model releases against. The labels are steering, reinforcing, simulating, mentalizing, and speaking. This conversation walks the first four; language is left for the book. He also flags a limit on the analogy. Birds inspired the Wright brothers, but jet engines left feathers behind. Brains are messy, homeostatic, and stuck with incremental fitness. Engineers can throw a design away. Inspiration is not a ceiling.

Readers keep asking why a book about intelligence never defines the word. Bennett catalogued the leading definitions and noticed that they do not change anyone's research agenda. Fei-Fei Li's version is broad: an agent that meets challenges and solves problems in a changing environment. Another researcher he spoke with simply equates intelligence with learning. An entity could, in principle, satisfy the first without doing the second. Because neither framing would have rewritten the evolutionary story, he skipped the fight.

Labenz is happy to skip it too. AI, he says, has taught him that definitional debates — especially "does it really understand?" — tend to go nowhere, and that weird systems are easy to miss if you start from a slogan. The deeper split is not semantic. It is whether intelligence will someday fall out of a single principle, or whether, a decade or two from now, people will admit there was never such a thing as AGI, only degrees of generalization across the particular suite of tasks an entity faces.

He puts himself with Yann LeCun on the second side. Bennett's hunch matches: there will be general principles, but not a panacea AGI, and the intelligences that matter will be specialized to domains. When people talk about intelligence, he estimates, about 80 percent of the time they mean something that *feels human* — not one capacity, but a bundle. One reason to build that bundle is usefulness. The other is that humans are trying to understand themselves.

Six hundred million years ago there was no life on land. Most organisms were single-celled; the earliest animals most resemble today's cnidarians — sea anemones, jellyfish, coral polyps. They have nerve nets, not brains: largely independent reflexes. A coral polyp does not hunt. It waits with tentacles out and grabs whatever bumps into it. Researchers have looked for even the simplest associative learning in coral polyps and have not found it. Pair a flash of light with a shock, and the animal still does not retract to the light alone.

Around 550 million years ago the first animals with brains appear in the fossil record: grain-of-rice worm-like creatures with bilateral symmetry. Nematodes are the living proxy, and *C. elegans* has been taken apart neuron by neuron. Its entire brain is 302 cells. It has no lens-shaped eyes, only a handful of neurons for smell, light, and touch, and yet it finds food extremely well. The algorithm is taxis navigation, which Bennett nicknames steering. You do not need a map. If an action raises the concentration of a food odor, keep going; if it falls, turn. The same rule, inverted, steers the worm away from predator smells.

A brain is required because the world presents trade-offs, and a web of separate reflexes cannot settle them. In a classic dish experiment, food sits on one side and a copper barrier — toxic to nematodes — sits in the middle. They will cross. How readily depends on food concentration and on how hungry they are. Sensory neurons for good and bad things have to land in one nucleus, inhibit one another, and emit a single movement. The interneurons that do that job already look, Labenz notes, like a toy version of a network's middle layer: inputs, a hidden choice, then motor output.

Steering also installed two templates that later brains kept. One is valence: a hardcoded split of the world into good and bad, which the AI field would call reward, though Bennett is reluctant to use that word yet. In a nematode, smell neurons map straight onto forward motion; in a human, the retina does not encode whether a sight is good. The other is affect. Real odor plumes are noisy, so a clean gradient-follower would fail. Dopamine neurons in the worm actually poke outside the body, detect food, and dump a persistent "dwelling" state — local search, on the bet that food clusters. Serotonin, fired when food hits the throat, produces rest and satiation. Seeking versus satisfaction is still the backbone of human motivation. Rodney Brooks's early Roomba did the same trick without a map: bump, turn, and, once dirt-detect landed, spin in place because a dirty patch usually has neighbors.

Fast-forward about fifty million years into the Cambrian explosion. Arthropods dominate the seas; some predators are human-sized. Vertebrate ancestors are small, fish-like, and mid-food-chain, and they evolve the basic vertebrate-brain template. Bennett's most useful piece of trivia: if you want to understand the human brain, a fish brain gets you about 70 percent of the way. Forebrain, midbrain, hindbrain, cortex, thalamus, basal ganglia, dopamine system — the parts are there, and they work in much the same way. What fish can do that nematodes cannot is learn from trial and error. People train fish to jump hoops and run mazes. You cannot train a nematode to do either.

Edward Thorndike went looking for insight and imitation in puzzle boxes and found neither. Cats did not suddenly "get" the latch; a second animal watching did not speed things up. Escape time fell on a smooth curve. He called it the law of effect: random attempts, successful acts strengthened, unsuccessful ones weakened. B.F. Skinner later claimed all learning was trial and error, which later breakthroughs would refute, but the reach of reinforcement is still enormous. Fish learn Thorndike's puzzles the same way. AI, copying the intuition, did not. In the 1950s Marvin Minsky built SNARC, a maze agent that reinforced recent actions after success, and it failed. The trap is temporal credit assignment. The move that wins a chess game is often in the middle, not the last one. Reinforce only the recent act and you get a dull agent; reinforce every act and you need too much data.

Richard Sutton's temporal-difference learning broke the stall. The training signal is not the reward itself but a change in predicted future reward. One process picks the next move; another estimates how good the current state is; a sudden jump in that estimate is the rush that trains the policy. TD-Gammon was the first practical win, then checkers and chess. Look back at the basal ganglia and the anatomy is uncomfortably close: a mosaic that biases brainstem actions up or down with dopamine, and another that turns dopamine on and off from predicted outcomes. Direct recordings in monkeys and mice show phasic dopamine bursting not at reward delivery but at a cue that raises expected reward — a neuroscience puzzle until Sutton's algorithm worked in silicon. Fish have not had the same dopamine-cell recordings, Bennett says, but surrounding structures already show prediction-error signals. Breakthrough two sits on breakthrough one: without a grounded notion of good and bad, there is nothing for expected future reward to be about.

Once arbitrary trial-and-error is online, other vertebrate talents become useful. Cortex does pattern recognition; some fish can recognize human faces. They also build a map of 3D space. In a grid of 25 containers, with only a blue dot on one wall as a cue and food under one dish, a fish returns to that location from a new starting point. It is not memorizing a left-right motor tape. It is learning a homing vector, a biological tweak on standard RL that is far more robust to changed starts, and a structure that will become the mammalian hippocampus. Labenz hears familiar AI rhymes: GANs bootstrapping a generator against a discriminator, RLHF's reward model scoring outputs, OpenAI's 2023 result that step-by-step math rewards beat a single right-or-wrong signal, and early web agents that never trained because they almost never got a positive. Sparse reward is 600 million years old.

About 150 million years ago our ancestors were four-inch mammals hiding from dinosaurs, coming out at night for insects. Inside that small head the old three-layer cortex of early vertebrates grew a six-layer sheet: the neocortex. Olfactory cortex and hippocampus kept the old three layers; everything folded on a human brain scan is the new one. In people the neocortex seems to do almost everything you consciously experience. Damage visual cortex and patients say they are blind, yet they blink at a looming object and their heart rate jumps at a scary image they cannot describe — blindsight, because a second optic path still hits the amygdala. Damage motor cortex in a primate and you get paralysis. Damage it in a rat and the animal still runs. It just cannot learn new motor skills, and fine, real-time paw placement falls apart. Bennett reads that as the original job: motor planning, simulating the act before the body moves. Watch a lizard run. Most non-mammalian vertebrates, birds excepted, look clumsy because they are not planning the next step.

The same sheet is a generative model: it recognizes the world and it can sample from its own representation. Helmholtz already had the idea as perception-as-inference — hold a prior about the world's state, predict, and update when the senses disagree. Visual illusions are the byproduct; once someone points out the Dalmatian in the speckles, you cannot unsee it. Fish without a neocortex are already superb at object recognition, including 3D rotations, so Bennett doubts recognition was the adaptive driver. What is new is the ability to cut off the senses and play out a state that is not the current one. David Redish recorded hippocampal place cells while rats ran mazes. At a choice point the animal pauses and looks both ways — the old "vicarious trial and error" hypothesis. The recordings made it literal: place cells stop coding the rat's location and sequentially play future paths until a decision is made. You can watch a rodent imagine.

Simulation unlocks counterfactuals that pure trial-and-error cannot touch. Teach a monkey rock-paper-scissors. If it plays rock and loses to paper, model-free RL would punish rock and split the leftover probability between paper and scissors. A simulator can notice that scissors would have won, and monkeys do shift toward scissors. (Humans can go further and see that the game is random.) Rat orbitofrontal cortex has been caught rendering the outcome of the path not taken. Episodic memory is the same machine pointed backward: remembering the past and imagining the future are both renderings of a world that is not here. None of it helps unless breakthrough two is in place. Imagined futures only change behavior if the animal can vicariously reinforce itself on the imagined result.

Labenz asks whether chain-of-thought and tree-of-thought are the AI version. Bennett's answer is that planning itself is 1950s-old; the unsolved problem is making it work in the real world. He splits mammal simulation into three subproblems: when to stop and think, how to prune the search, and how to score imagined endings. His hypothesis on the first is an uncertainty threshold. The second is a robotics nightmare — even a robot arm's degrees of freedom explode. The third is evaluation. You could chain LLMs to do all three; how well that works is an empirical question arriving soon. The deeper gap, he thinks, is continual learning. Fine-tuning on a small set overfits, forgets old tasks, and loses generalization. "I don't think this is a nuanced edge case," he says. "I think this is a foundational problem with the way these models work." Mammal brains may update weights only in surprise, when neuromodulators bath synapses after a failed prediction. People also have a eureka gate: tell a person that one plus one equals four and they will not swallow it as fine-tuning data; they will try to fit it, and reject it if they cannot. Transformers accept the batch anyway. Labenz hears a rhyme in Mamba-style selective state-space models — a long-lived state that can refuse an update — and both of them expect hybridization with transformers, not a killing.

Comparative psychology gets more argumentative once you reach primates, Bennett warns, but three abilities have decent evidence in non-human primates and are consistently missing in other mammals. The first is theory of mind. Chimpanzees allowed to play with two pairs of goggles — one see-through, one opaque — later approach the experimenter who is wearing the transparent pair when both are holding food. In another setup, animals learn that a red mark on a box means food; an experimenter then marks one box on purpose and "accidentally" drops the marker on the other. The stimulus is identical. They go to the intentional mark. The second ability is imitation. Teach one group member a special rake trick for food, return it to the troop, and the technique spreads. There is more controversial evidence that chimpanzees actively teach, correcting youngsters at termite fishing. The third, with the thinnest evidence, is anticipating future needs. Squirrel monkeys will take a lesser treat if water comes with it, skipping a richer treat that is followed by a long thirsty wait, even though they are not thirsty now. Rats in a matched setup will not. Humans doing the weekly grocery shop are running the same trick.

The brain change is not a wholesale redesign. Primates got a bigger brain, new patches of posterior sensory cortex, and — most of all — granular prefrontal cortex stacked on older neocortex. A reasonable first approximation, Bennett says, is that it builds a model of the older model. Rat prefrontal areas still take sensory input from hippocampus and amygdala; granular prefrontal cortex takes almost none of that, only input from agranular prefrontal cortex. Damage studies split neatly. People with hippocampal damage can put themselves in an imagined future, but the world they describe is thin. People with granular prefrontal damage describe the future world fluently and leave themselves out of it. That is a clue that this tissue is where you model *you* — the home of metacognition.

The three new abilities then look like three applications of one algorithm: simulate your own inner simulation. Put yourself in someone else's situation and you get theory of mind. Imagine yourself performing someone else's action, infer the intent, and vicariously reinforce that intent — not the literal joint angles — and imitation learning starts working, which is also what the AI imitation literature found. Imagine a future self that will be thirsty or will regret a choice, and you can act for a need you do not currently feel. Bennett is careful: this is nowhere near an implementable blueprint. It is a principle. It also has a warning for human-AI relations. People understand each other in part because brains are similar enough that projecting yourself is, on average, a decent proxy. If a machine's internals work very differently, that projection will fail, and both sides will need other ways to read intent. Labenz adds that GPT-4 is clearly off zero on theory-of-mind tests, but a human theory of mind *for* AI may be harder, because our instincts keep pointing at ourselves.

The story Bennett likes most is closer to an anecdote than a controlled paper, and he flags it as such. Emil Menzel was testing map-based memory in chimpanzees in a one-acre forest, showing one animal — Belle — where food was hidden. A second chimpanzee, Rock, began taking the food from her. She waited until he looked away. He learned to look away on purpose, then spin and chase. She started leading him the wrong way until he was far enough, then doubling back. Deception and counter-deception of that kind, Bennett argues, is extremely hard to explain unless each animal is trying to change what is in the other's head.

Biological brains are messy, cyclic, and not differentiable. Today's leading AI is linear and requires differentiability for backpropagation. Bennett would be surprised if backprop were discarded wholesale in ten years, and surprised if the field left deep learning. What brains may add is taste about *when* weights update, which is how they pull off continual learning without wiping prior knowledge. Backprop is almost too precise: every weight gets an accurate partial derivative. A less accurate rule plus a gate — pause, actively interrogate an uncertain object, allow updates only then, perhaps replay during something like sleep — could sit on top of the same optimizer. If the idea is found first in AI, he expects another Sutton moment: neuroscientists will look back and recognize it. If it is found first in brains, someone will have to read that messy literature with enough taste to port the right piece.

Macro-architecture is already drifting away from one soup of weights. Brains are specialized subsystems in conversation, not a single scaled model. Bennett's prediction for a household robot is not one foundation model. It is a convnet for objects, a language model, a separate world model, a continual-learning episodic memory, wired into a whole-brain layout. Labenz offers OpenWorm and the fruit-fly connectome as a tempting shortcut: take an evolved wiring diagram and soup up the parts. Bennett is unconvinced that connectivity is enough. Neurons differ in baseline rates, thresholds, phasic versus tonic firing, dendritic-spine learning rates, excitatory versus inhibitory versus neuromodulatory identity, and receptor expression. Knowing who talks to whom is a small percentage of the computation. Either we eventually measure all of that, or a modeler with good taste guesses the modules and the messages and tests whether the simulation behaves like a fly.

That is the Wright brothers point again. They watched birds and did not perform avian surgery. The useful move is to steal a principle, not to photocopy tissue. Evolution also refuses lookahead: every generation has to survive, which is why bolting a convolutional predator-detector onto a nematode is a thought experiment, not a lineage. Associative learning in simple bilaterians is locked to about a two-second window. An expensive pattern recognizer that can only bind events inside two seconds may not be worth the energy, and it may not have been an available mutation. Lens-shaped eyes remain the canonical puzzle of that constraint — every increment had to be neutral or helpful. Silicon does not have to care. That is why Bennett thinks near-term AI will keep diverging from brains, and why borrowing from neuroscience is a taste problem, not a mandate.

Labenz closes with a question Bennett answers only as intuition, not as moral authority: how much should one care about a worm, a fish, an ape, and then an AI? The jump that matters to him is from model-free reinforcement learning to model-based — the animal that can pause, think, and hold a model, which mammals have and which birds seem to have evolved independently. Qualia, he speculates, are rendered in the neocortex, which would put rats in the experiencing club. Identity and thinking about one's place in the world look more primate. He would put more moral weight on experience than on identity. Meditation, in that frame, is dropping the primate rumination and sitting in perception-as-inference: simulating only the present.

If suffering attaches to experience rather than to a self-story, a rat should have moral weight close to a human's, because it does not need an identity to feel pain or fear. Fish are harder; their brains look different enough that the correlation is uncertain, and he admits the risk that we are wrong. Non-human primates, he thinks, are almost impossible to rank clearly below humans. Then comes what he calls the trillion-dollar ethical question of the century, and one he thinks the field under-discusses: not only how AIs will treat us, but how we will treat them. We already lack a philosophical grounding for anyone else's consciousness besides our own. The usual argument is mechanistic similarity — your brain looks like mine. That test will not travel to silicon, and most people in AI and philosophy are materialists, so biological goo is not supposed to be required. He does not have an answer. He does think it will be an emerging problem of the twenty-first century, and, Labenz suggests, maybe the next book.