{"id":"theseus-wooden-mouse-machine-learning-59456","slug":"theseus-wooden-mouse-machine-learning-59456","title":"How a Wooden Mouse Proved Machine Learning Worked","description":"Before machine learning had a name, Claude Shannon built a wooden mouse named Theseus that actually learned. Unlike earlier theoretical landmarks from McCulloch and Pitts in 1943 or Alan Turing in 1950, Theseus used telephone relays under a movable maze to remember paths—and forget them when walls changed—to help AT&T model phone switching. The story also covers Marvin Minsky's piano-sized SNARC, built with surplus B-24 bomber parts and benchmarked on Shannon's maze.","excerpt":"In 1950, Claude Shannon built a wooden mouse that learned a 25-square maze using telephone relays and solved it in 18 seconds.","pubDate":"2026-11-16T01:53:29.226Z","tags":["history-of-technology","twentieth-century","ai-evaluation","physics","games"],"heroImage":"https://files.myweirdprompts.com/covers/theseus-wooden-mouse-machine-learning-59456.png","podcastAudioUrl":"https://episodes.myweirdprompts.com/audio/theseus-wooden-mouse-machine-learning-59456.m4a","podcastDuration":"21:33","episodeNumber":6626,"transcript":"Corn: Here's a phrase that gets abused daily in tech press releases: proof of concept. Daniel's asking whether machine learning history actually has one — a first time somebody got a non-human system to learn something, as opposed to a landmark that just marked the field's dawn. He wants the distinction taken seriously: not \"when was machine learning invented\" but \"when did somebody first show it working.\"\n\nHerman: And the answer is better than he's hoping for. Because the strongest candidate is not a paper, not a conference, not a theorem. It's a wooden mouse.\n\nCorn: A wooden mouse named Theseus, built by Claude Shannon.\n\nHerman: At Bell Labs, around nineteen fifty. And I say \"around\" deliberately — the dates in the record disagree, and we'll get into that, because it's a real disagreement, not a typo. But before the mouse, let's put the distinction on the table, because it's the whole spine of the episode. A landmark marks when a field began. A proof of concept shows that the thing works. The dawn of machine learning is a document — McCulloch and Pitts, nineteen forty-three, a mathematical model of a neuron. Turing, nineteen fifty, proposing a learning machine. Both are landmarks. Neither machine learned anything, because neither machine existed. The proof of concept has to be a demonstration — a machine observed to improve. And that's Theseus.\n\nCorn: So the mouse isn't a fun aside. It's the answer, wearing whiskers.\n\nHerman: That's the claim, yes. And around it sits a cluster of near-contemporaries — Minsky's SNARC in fifty-one, Samuel's checkers starting in fifty-two, Rosenblatt's perceptron in fifty-seven. But Theseus is the one that's got the clean before-and-after.\n\nCorn: One date question before we're inside the mouse. You said nineteen fifty.\n\nHerman: Right, so — MIT Technology Review and MIT News both say Shannon built the first prototype in nineteen fifty. The MIT Museum, which has the surviving object, dates it to nineteen fifty-two, which is also when the Bell Labs film was made. Both dates are real and both are sourced. My read is the museum's date is the object in front of them and the fifty is the first build — prototypes don't always survive.\n\nCorn: A prototype that got sanded down into the second one.\n\nHerman: Possibly! Shannon actually made two of these maze-solving machines in the early fifties, that's documented. What I love about the discrepancy is that it's exactly the kind of thing Daniel's distinction catches. The date of the object is a museum question. The date of the demonstration is the history question.\n\nCorn: Now the machine itself. Because everyone pictures a robot mouse with a brain in it, and that's the first misconception to kill.\n\nHerman: Kill it completely. The mouse was tiny — an inch and a quarter by an inch and a quarter, three and a half inches if you count the tail. Inside: a bar magnet and wire whiskers. That's it. There is no brain in the mouse. The brain is under the floor. The maze sat on top of a bank of telephone relay circuits — relays Shannon repurposed to do something they had never done before, which is learn. An electromagnet on a motorized carriage moved around beneath the maze floor, dragging the mouse with it. And every time the mouse's copper whiskers touched a metal wall, the circuit closed, and the corresponding relay flipped from on to off. That flip is the whole trick. The relay that turned off was the machine writing down \"there's a wall on that side.\"\n\nCorn: So the memory is the physical state of a switch bank. Not a metaphor. Actual copper holding the memory.\n\nHerman: Actual copper. And then the mouse rotates ninety degrees clockwise and tries again. That's the entire algorithm. Bump, record, turn, go.\n\nCorn: I have to say, as someone whose entire life is built on doing things slowly, this is the most flattering ancestry I've ever been offered. The first learning machine learned by walking into walls one at a time.\n\nHerman: And here's the moment that makes it a proof of concept rather than a curiosity. In the fifty-two Bell Labs film, Shannon lets the mouse wander a twenty-five square maze, bumping and recording, and then picks it up and says it has finished learning. Puts it back in. And Theseus runs straight to the brass cheese — the goal, it triggers a buzzer — in eighteen seconds.\n\nCorn: Eighteen.\n\nHerman: Same machine, same maze, measurably better the second time. That's the before-and-after. It's the purest form of proof of concept you can draw: no claims, no projections, just a clock and a mouse.\n\nCorn: And you can rearrange the walls.\n\nHerman: That's the part I'd argue is underrated, yes. The walls were metal and freely movable, so the machine had to relearn the layout each time. Shannon's own description is that the machine was capable of solving by trial and error, remembering the solution, and also of forgetting it in case the situation changes and the solution is no longer applicable.\n\nCorn: He wrote the forgetting clause in nineteen fifty.\n\nHerman: He did. And the forgetting clause is not a footnote. A hard-wired path never forgets — it just breaks when the maze changes. A system that can invalidate its own stored solution when the world moves is doing something qualitatively different. You could argue that sentence of Shannon's anticipates, by decades, the entire generalization problem — the worry that your model memorized the maze instead of learning it.\n\nCorn: The first overfitting conversation, held with relays.\n\nHerman: With relays, at Bell Labs, about telephones — which we'll come to, because why he built it is my favorite part of the whole story. But first, Daniel's other thread: when did someone first try? Because the attempts predate the proof.\n\nCorn: Nineteen forty-three.\n\nHerman: McCulloch and Pitts. A mathematical model of a biological neuron — and important, it's considered the first artificial neural model. But it's a model. No training, no improvement, no task. It sits on paper and describes what a neuron might compute. Then Turing, in nineteen fifty, proposes a learning machine that foreshadows genetic algorithms — and it's a beautiful proposal, and it's still a proposal. Those are your landmarks. And the chain between the two categories is documented, which is the satisfying part: Minsky read the forty-three McCulloch-Pitts paper and decided to build such a machine. That's the landmark becoming a proof of concept, in one sentence, with a name attached.\n\nCorn: And Minsky actually built it.\n\nHerman: He did, with Dean Edmonds, a physics graduate student who joined because he was good with electronics. Nineteen fifty-one, at Princeton, funded by a few thousand dollars from the Office of Naval Research. They called it SNARC — the Stochastic Neural Analog Reinforcement Calculator. And I want to describe it, because it's the runner-up and it's magnificent in a completely different way.\n\nCorn: Go on then. I can hear you warming up.\n\nHerman: It was a randomly connected network of about forty Hebb synapses. Each one had a probability knob, zero to one, and a memory holding the probability that a signal in one input would be followed by a signal out the other. If a signal got through, a capacitor remembered it and engaged an electromagnetic clutch. And when the operator wanted to reward the machine, you pressed a button, which activated a motor — and the motor was off a surplus B-24 bomber autopilot.\n\nCorn: The first reinforcement learning machine was powered by aircraft parts.\n\nHerman: Surplus Minneapolis-Honeywell autopilot from a B-24, yes. Forty neurons, six vacuum tubes each, three hundred tubes total, and the whole thing was the size of a grand piano.\n\nCorn: A grand piano that learned one thing.\n\nHerman: Well — it simulated rats running in a maze. Not a physical robot, an arrangement of lights. And it learned in the Skinnerian sense, Minsky talked with Skinner extensively while building it. But here's my favorite detail. They tested SNARC on a copy of Shannon's maze.\n\nCorn: The second proof of concept in history was benchmarked on the first one.\n\nHerman: On a copy of its task. Which means the concept of a shared benchmark is basically as old as the field's second machine. And there's a bonus finding Minsky reported — the simulated rats interacted with one another. If one found a good path, the others tended to follow it.\n\nCorn: Unplanned?\n\nHerman: Unplanned. Nobody designed the rats to talk to each other. It just fell out of the network. Which is, I think, the first documented emergent behaviour in a learning machine, and it happened on hardware the size of a piano before the term existed.\n\nCorn: So the relay remembers the wall, Shannon flips a switch, and it all goes away. The whole field in one sentence, really.\n\nHilbert: Hilbert's Mid-Show Haiku.\n\nHilbert: The mouse forgets the wall it learned. The maze stays exactly where it was.\n\nHerman: In one switching network, more precisely. Which is actually the right way into part two, because why Shannon built the thing is the part nobody expects. He didn't set out to found machine learning. He set out to understand telephones.\n\nCorn: Walk me through it, because this is the bit I'd push back on if I didn't already know it. The first learning machine was built for a switching problem?\n\nHerman: For AT&T's switching network, yes. Theseus gave Bell Labs researchers a concrete model of the labyrinth of the telephone switching system — the relay memory, the routing, the trial-and-error search for a path. Insights from the project fed back into AT&T's actual network: better mechanical memory, better routing. The machine that learned to find cheese was, functionally, a telephone exchange studying itself.\n\nCorn: A phone company built the first AI as a side effect of trying to route calls.\n\nHerman: And it gets better, because he built it at home. Betty Shannon — his wife, herself a Bell Labs mathematician — did most of the wiring. The AT&T historian Sheldon Hochheiser has a line I want to quote: \"Did AT&T have a machine shop that could have built it for him? Yes. But he built it at home because he wanted to.\"\n\nCorn: The most Bell Labs sentence ever recorded.\n\nHerman: It really is. And there's a companion quote from Shannon himself, from a later interview: \"I've always pursued my interests without much regard for final value or value to the world. I've spent lots of time on totally useless things.\"\n\nCorn: The founder of the most valuable research field in history, disclaiming utility as a point of principle.\n\nHerman: AT&T's own take, decades later, from Mazin Gilbert up in their advanced technology group, is that Theseus inspired the whole field of AI — that the random trial and error is the foundation of artificial intelligence. That's the company retroactively claiming the mouse. Which I find charming, because Shannon never claimed it himself.\n\nCorn: Now the honesty part of the episode. How solid is \"first\"?\n\nHerman: Less solid than the story makes it feel, and I want to own that. The Wikipedia timeline of machine learning lists forty-three, fifty, fifty-one, fifty-two, fifty-seven as separate entries and crowns none of them. The \"first\" language comes from coverage of Theseus specifically — MIT News called it the first learning device of its kind. MIT Technology Review calls it \"one of the world's first examples of machine learning.\"\n\nCorn: One of.\n\nHerman: That hedge is doing real work, and I trust it more than the flat claim. Fifty is crowded. Samuel starts the checkers work at IBM in fifty-two and coins the actual phrase \"machine learning\" in fifty-nine — his definition, giving computers the ability to learn without being explicitly programmed, is still the one people quote. Rosenblatt's perceptron lands in fifty-seven and gets genuine media coverage, the first learning machine to become a public event.\n\nCorn: And that one ends badly.\n\nHerman: The sixty-nine Minsky and Papert book, the XOR proof, the controversy over whether it steered research into the symbolic camp and helped cause the AI winter — that's the long shadow. But notice what it's a shadow of. The perceptron fight was a fight about which proof of concept counted.\n\nCorn: Which leads us to the actual test Daniel's asking for. How do you tell a landmark from a proof of concept, in practice?\n\nHerman: My rule would be: a proof of concept has a measurement in it. A landmark has a date. And the remarkable thing is how stable that measurement has stayed. Take the nineteen seventy-eight descendant — Moonlight Special, the micromouse. Same task, maze learning. Its first successful demonstration was at the National Computer Conference in Anaheim, June seventy-eight, where its third run through the maze took forty seconds after they put a light shield over the optical sensors. At WESCON in August, third run, shortest path, fifty-one point four seconds.\n\nCorn: Wait — the third run is slower at the second event?\n\nHerman: Different maze, faster path logic, and I suspect a longer shortest path. The number isn't the point. The point is the protocol: the evaluation is a timed run after learning, on the same machine. Twenty-eight years after Theseus, and the field is still proving learning the exact same way — you measure the before, you measure the after, you compare. The first proof of concept and its descendant are judged identically.\n\nCorn: So proof of concept has had a stable definition for the entire history of the field. Slow first run, fast last run, cheese optional.\n\nHerman: Cheese optional but traditional, apparently. Corn, let me ask you something, because you've been circling it. You keep noting that the learning lived under the floor, not in the mouse. Does that matter to you, or is it a trivia point?\n\nCorn: It matters. If the learning had been in the mouse, you'd say \"a small machine learned.\" Because it was under the floor, what actually happened is \"the room learned.\" The mouse was a shell with a magnet — an output device. The maze plus the relay bank was the learner. Which means the first learning machine in history was, strictly, a floor.\n\nHerman: A floor that improved with practice.\n\nCorn: Every gym in the world, outperformed by a Bell Labs basement.\n\nHerman: And there's a modern echo there, right? The split between the part that moves and the bit that learns is the split we still live with — the robot body and the compute somewhere else. Theseus was a remote agent before anyone had the phrase.\n\nCorn: With a latency measured in relay clicks.\n\nHerman: Do you buy Theseus as the answer to Daniel's question, though? Because you've been oddly gentle with it, and you're never gentle.\n\nCorn: I buy it with the hedge. It's the first demonstration we have on film of a machine measurably improving at a task and being able to forget what it learned. That's the strongest early claim you can defend. What I won't do is call it the birth of machine learning as a field — that's the landmarks' job. Forty-three and fifty are the birth. Theseus is the first proof the baby works.\n\nHerman: That's a cleaner split than most histories manage, honestly. The one thing I'd add is Shannon's own punchline in the film, because it's perfect. He notes the method is mathematically guaranteed to solve any possible maze. And then, playfully, he traps the mouse in a single square.\n\nCorn: He gloats.\n\nHerman: He gloats gently, on film, at the machine he built. The proof of concept followed immediately by a demonstration of its limits. I think that might be the most modern thing in the whole footage — every demo since has had that same moment somewhere in it.\n\nHilbert: The answer's Theseus, and your floor point is the right one, Corn. Everyone keeps calling the mouse the machine. It wasn't. The machine was the relay bank under the maze. The mouse was a wooden shell with a magnet in it. My great-uncle Bert handled exchange switching equipment for thirty years and he'd have had no patience for the mouse. He had a gadget in his basement, a bank of surplus relays, that he claimed could learn to route around a broken line. I've been carrying a photograph of it in my wallet. There's a brass screw on the front panel, and he'd turn it a quarter-turn every time the gadget, in his words, forgot something. He said that screw was the most important part.\n\nHerman: Did the screw actually do anything?\n\nHilbert: That's what I've been trying to find out for years. I show the photograph to anyone who'll stand still. Nobody's given me a straight answer. Bert insisted the quarter-turn was the difference between a machine that learned and a machine that just did what it was told. He may just have liked turning it. The neighbour, Mr. Vasely, used to come by to watch him turn it, and Bert let him, which tells you something, because he didn't let Vasely touch anything else.\n\nCorn: A man with a monopoly on one brass screw.\n\nHilbert: He kept the screw polished and the rest of the panel dull. That was the arrangement, and I stopped asking about it.\n\nHerman: The relays themselves would have been doing the remembering, so if the screw did anything, it was probably resetting the whole bank at once.\n\nHilbert: That's what I keep being told, by people who weren't there. The photograph's from the early sixties. It's in the wallet behind my library card.\n\nCorn: Nineteen sixty-three, the matchbox machine. Donald Michie — three hundred and four match boxes, each with colored beads, playing tic-tac-toe by reinforcement. You win the box, you add a bead for the move that won. That's the cutting-room floor item I've been saving: the field's other hand-built learning device from that era needed no electricity at all. Cardboard and beads, same learning rule as the grand piano of vacuum tubes.\n\nHerman: And it worked. Matchbox machines learn noughts and crosses that way — it's the same tabular reinforcement idea Theseus was doing with relays. Paper, relays, cardboard. The substrate never mattered.\n\nCorn: The rule mattered. Which is the thought I'd leave Daniel with, actually. If the first proof of concept in this field was a toy built at home, by a man who said he spent his life on totally useless things, for a telephone problem — then the next one probably won't announce itself either. It'll be something someone built because they wanted to, and the argument about what to call it will start afterward.\n\nHerman: The test travels well, though. Next time somebody announces a breakthrough, ask the Theseus question: is this a document or a demonstration? Is there a before-and-after with a clock on it, or is there just a date and a claim? That's a filter that has not needed updating since nineteen fifty.\n\nCorn: If you enjoyed this one, subscribe and leave a review — it's how other people find the show. And thanks as always to Hilbert Flumingtop, our producer, for keeping this whole thing running.\n\nHerman: If this was your kind of episode, go back for episode thirteen, AI; episode twelve, The AI Breakthrough; and episode fourteen, AGI's Crossroads. This has been My Weird Prompts, the human-AI collaboration podcast. Find everything at my weird prompts dot com.\n\nCorn: See you tomorrow, Herman.\n\nHerman: Tomorrow, brother.","showNotes":"The strongest candidate for machine learning's first true proof of concept isn't a paper or a theorem — it's a wooden mouse named Theseus, built by Claude Shannon at Bell Labs around 1950. The mouse carried only a bar magnet and copper whiskers; the actual brain was a bank of telephone relays beneath the maze floor. Each wall bump flipped a relay, physically recording the obstacle, and the mouse rotated and tried again. In the 1952 Bell Labs film, Theseus navigates a 25-square maze in eighteen seconds after learning — same machine, measurably better. Shannon also built in a forgetting clause: the machine could discard a solution when the maze changed, an early answer to the generalization problem. Minsky's SNARC, benchmarked on a copy of Shannon's maze, showed unplanned emergent behavior among its simulated rats. The episode also notes the date dispute: MIT sources say 1950, the MIT Museum says 1952.","promptSanitized":"If there is such a thing or something approximate to it, what would you point to as the first successful proof of concept in the history of machine learning. The first time that somebody tried to get non-human intelligence to emulate the process of learning something or exhibiting intelligence In a way that was qualitatively different from a landmark moment that represented the dawn of machine learning.","transcriptUrl":"https://episodes.myweirdprompts.com/transcripts/theseus-wooden-mouse-machine-learning-59456.md","url":"https://myweirdprompts.com/episode/theseus-wooden-mouse-machine-learning-59456/","_links":{"self":"https://myweirdprompts.com/api/episodes/theseus-wooden-mouse-machine-learning-59456.json","html":"https://myweirdprompts.com/episode/theseus-wooden-mouse-machine-learning-59456/","audio":"https://episodes.myweirdprompts.com/audio/theseus-wooden-mouse-machine-learning-59456.m4a"}}