Burden Neurological Institute · Bristol · winter 1949
Two radio valves. One eye that never stops turning. No memory, no map, no plan — and a machine that explores a dark room, avoids furniture, and dances with its own reflection.
Act one · You are the eye
It is dark. You are a photocell the size of a thumbnail, bolted to the top of a spinning steering column inside a plastic shell. You cannot see shapes. You cannot see colour. You cannot see anything at all except one number, which changes as you swing around: how bright is it, straight ahead?
There is a candle somewhere in this room. Find it — using the only instrument you have.
Drag the dial (or use the slider) to swing your eye around the room. The meter is the only thing you can perceive. When you think you are pointing at the candle, press Fix on it.
This is the whole sensory world of Grey Walter's tortoise: one moving number. Everything it ever did came out of that.
Act two · Four rules, and that is all
In the spring of 1948, in a hospital laboratory on a hill outside Bristol, a neurophysiologist named W. Grey Walter began building an animal out of war-surplus junk. He had spent the 1930s and 40s wiring people's heads to electroencephalographs, watching the brain's alpha rhythm crawl across a paper chart. He had a suspicion he could not test on a human being: that the astonishing thing about brains is not how many cells they have, but how richly the few are connected.
So he built the smallest possible nervous system and gave it a body. Two miniature radio valves — thermionic tubes, the glowing glass things inside your grandparents' wireless set — wired one into the other. Two relays. Two motors: one to drive, one to steer. One photocell for an eye. One contact on the shell for a sense of touch. He called the species Machina speculatrix, the machine that explores, and the first one ELMER. He showed it to the press in December 1949.
Here is the entire brain. Four situations; four settings of two motors. Nothing else. No map of the room, no memory of where it has been, no list of instructions. Read the four boxes below — then go and drive it.
The eye is bolted to the steering column, so the eye always looks where the driving wheel points. Turning the wheel is looking around. That one piece of carpentry is why two valves are enough.
Act three · The floor
Walter used to photograph his tortoises in a darkened room with the camera shutter held open, so the little lamp on the machine's nose drew its own path across the plate in a curl of light. The paths look like the tracks of a beetle. They were made by four if-statements.
Below is that room. Drag the candle. Drag the stool. Break the rules one at a time and watch the animal stop being an animal.
Drag the candle, the stool or the mirror. The glowing trail is the path of the machine's own headlamp — exactly what Walter's open-shutter photographs recorded.
As the battery falls the amplifier's gain falls with it, the bright-light aversion weakens, and the machine can finally get into its own brightly-lit hutch. Walter's real hutch, he later admitted, hardly ever worked — and eventually the 20-watt lamp set it on fire.
With no fear of bright light, the machine drives at the candle and parks there for ever. Walter's tortoise had to dislike what it liked, or it would starve at the first bright thing it found.
Now it grinds against the stool and stays there. Touch is not an extra sense here — it hijacks the very same two valves and turns the amplifier into a one-beat-per-second oscillator.
It never drives straight at the candle. It arrives in a series of arcs, losing the light, finding it, losing it — what Punch called “a series of slow waltzing movements”. That waltz is not a bug. It is the scan.
Act four · The dance
Now for the strangest thing in the machine, and it is an accident of wiring. The little headlamp on the tortoise's nose is not decoration — it is part of the circuit, sitting in the same branch as the first relay. Which means: the moment the machine locks on to a light, its own lamp goes dark.
Put a mirror in front of it and something uncanny happens. It sees its own lamp. It locks on. Its lamp goes out. Now there is nothing to see, so it goes back to searching — and the lamp comes back on. And it sees itself again. The machine sits in front of the mirror flickering and jerking at a couple of beats a second, in a way it will never do in front of anything else in the world.
Walter's joke, which he only half meant, was that if you saw an animal behave this way in front of a mirror and in no other situation, you might reasonably call it self-recognition. Turn the switch below off and on, and decide for yourself what you are looking at: a machine that knows itself, or a loop that eats its own tail.
The green trace is the headlamp: high means lit. Watch what the mirror does to it.
Unwire the lamp from the relay and the dance stops dead: the machine simply drives at its reflection and dazzles itself. The dance needs the loop, not a soul.
Each tortoise carries a lamp, and each is drawn to lights. So two of them find each other, rush together, and — as they lock on, both lamps go out — instantly lose each other. Then the lamps come back, and they find each other again. Walter called it a mating dance. Go back up to the dark room, tick Release a second tortoise, and watch two machines with no concept of one another perform a duet.
Act five · The machine that learns
In 1950 Walter added a third thing to the tortoise: a microphone and a small circuit he called CORA — the COnditioned Reflex Analogue. The new species was Machina docilis, the teachable machine. Its brain still had almost nothing in it. What CORA added was one honest piece of memory: a capacitor that charged a little every time two things happened close together in time.
That is Pavlov, in a component you can hold. Blow a police whistle. Then, immediately, kick the machine — or push a stool into its path so its shell bumps. The whistle means nothing; the bump means swerve. Do the two together enough times and the capacitor holds enough charge that the whistle alone makes it swerve. It has learned. And if you stop, the charge leaks away, and it forgets.
Learning rule, written out honestly: a whistle sets a trace that decays over about a second and a half; whenever the unconditioned event arrives while the trace is still up, the association grows by Δw = η · trace · (1 − w) with η = 0.06. Nothing is stored between visits; nothing is pre-trained. Everything on the meters is what you did.
Association strength after each pairing you performed. The dashed rule at 0.60 is the level at which the whistle alone is enough to fire the response.
A British police whistle has two notes. Walter tried to teach one note to mean danger and the other to mean food, and then blew both at once. The machine had two strong reflexes pulling it in opposite directions and no way to choose. Walter reported that it “went right off into the darkness on the right-hand side of the room and hovered round there for five minutes in a sort of sulk.” He called it an experimental neurosis, and listed three cures: let it rest, switch it off and on, or perform surgery — cut CORA out and turn it back into an ordinary tortoise. Try all three above.
Six things this page is being careful about
A thermionic valve is an amplifier with a hot wire in it. A neuron is a living cell with thousands of synapses, its own chemistry and its own timing. Walter never claimed they were the same thing. His claim was narrower and more interesting: that a great deal of what looks like animal behaviour needs astonishingly few elements, provided they are wired back into each other.
Popular retellings say the tortoises returned to their hutch to recharge, like pets. Owen Holland, who studied the surviving machines and Walter's papers closely, reports that the recharging arrangement was never reliable — and that the original hutch burned down, set alight by its own 20-watt lamp. Our simulation lets it work. Reality was scruffier.
Walter gave both numbers for how many pairings CORA needed, in different tellings. Our bench uses a rule that crosses the firing threshold at around twenty — close to the larger figure — and shows you the count, so you can watch the disagreement rather than be told an answer.
Every path you see is computed from the four rules with floating-point arithmetic. The real machines had sticky relays, drifting valves, a floor that was not level, and batteries that sagged. A great deal of the tortoises' famous unpredictability was hardware noise — which is not a flaw in the argument, but it is not in this page either.
Walter wrote that the mirror dance “may be regarded formally as being diagnostic of self-recognition”. Formally. The tortoise has no model of itself, no concept of a mirror and no concept of anything else. The lesson runs the other way: a behaviour that looks like self-awareness can be produced by a loop with two valves in it — so seeing the behaviour is not, by itself, evidence of the mind behind it. That caution is worth carrying into every conversation you have about machines today.
ELMER and ELSIE are usually expanded as ELectro-MEchanical Robot and Electro-mechanical robot, Light-Sensitive, with Internal and External stability, but the wording shifts from account to account, and so do the dates for individual machines. What is solid: Walter began in spring 1948, showed a working tortoise to the press in December 1949, and six improved machines were built at the Burden in early 1951 by the engineer W. J. “Bunny” Warren and colleagues, three of which went on show at the Festival of Britain.
The line from there to the machine on your floor
The examination
No penalty for a wrong answer except a small amount of teasing.
For the next forty years, artificial intelligence went the other way. It built worlds out of symbols and rules and long chains of reasoning, and left the body behind. Then in 1986 a young MIT professor called Rodney Brooks published a paper arguing that a robot does not need a model of the world, because — his phrase — the world is its own best model. He built insect-like machines with layered reflexes and no central plan, and freely credited the ancestors: Grey Walter's tortoises, wandering a Bristol laboratory in the dark.
Today's systems are enormous. A large language model has hundreds of billions of numbers in it where Walter had two valves. But the tortoise still has one thing to say to them, and it is not a comfortable thing: behaviour that looks intelligent, curious, sociable, even self-aware, can come out of a mechanism far simpler and far stupider than it appears from the outside. Walter proved it in 1949 with a plastic shell and a torch bulb, and everybody in the room agreed the machine was exploring.
So here is the question to take away: if two valves are enough to make something look alive, what exactly would you need to see before you decided that something wasn't just looking?
Sources: W. Grey Walter, “An Imitation of Life”, Scientific American 182(5), May 1950, 42–45; “A Machine That Learns”, Scientific American 185(2), August 1951, 60–63; The Living Brain (Duckworth/Norton, 1953). Owen Holland, “The first biologically inspired robots”, Robotica 21 (2003) and “Exploration and high adventure: the legacy of Grey Walter”, Phil. Trans. R. Soc. A 361 (2003). Phil Husbands & Owen Holland on the Ratio Club. cyberneticzoo.com on CORA, the hutches and the Festival of Britain; Punch, June 1951. Science Museum Group collection, “Conditioned Reflex Analogue, ‘CORA’, England, 1950–1953”. Rodney Brooks, “A Robust Layered Control System for a Mobile Robot”, IEEE J. Robotics and Automation 2(1), 1986, 14–23.