Signals Intelligence
A scenario — call it the hidden branch of the tech tree — in which analog computing and neuromorphic engineering had their real breakthroughs decades before the public record admits, inside classified defense work rather than in public. It requires no magic and no grand conspiracy: only incentives, secrecy, and the fact that the most useful machine intelligence never needed to say a word. From a Nazi missile computer to submarine-hunting ocean arrays to the chip in your pocket, the case that the history of AI you were taught is only the publishable half.
The history of computing you learned is the history of the computers you were allowed to buy.
That sentence sounds like the setup for a conspiracy theory. It isn’t. It’s a plain description of the twentieth century. The Manhattan Project, Bell Labs, the MIT Radiation Lab, the NSA, Naval Research, Skunk Works, TRW, SRI, RAND — entire technological ecosystems existed where publication was not the goal. The assumption that all important computing research happened in public was never true, not for a single decade of the field’s existence.
So take the following as a scenario. Not a claim of secret AGI in a basement — that version falls apart under its own logistics. Something more precise, more plausible, and honestly more unsettling:
Somewhere in the 1950s–70s, a classified engineering lineage figured out that many intelligence tasks are naturally solved by carefully engineered physical dynamical systems rather than symbolic computation. And that lineage never stopped.
Let’s walk the branch.
The precedent is not hypothetical
We already know secrecy can swallow major capabilities whole. The F-117 flew operationally in 1983 and wasn’t acknowledged until 1988. The CORONA reconnaissance satellites operated through the Cold War and were declassified only in 1995. Cryptographic hardware, electronic warfare, whole families of signal-processing techniques — all surfaced publicly years or decades after military use.
And crucially: analog computing was born classified. Vannevar Bush’s differential analyzer computed firing tables and radar antenna profiles in secret during WWII. The U.S. Navy’s mechanical fire-control computers were solving real-time targeting problems — differential equations, in combat, under fire — while “the computer” as the public would come to know it barely existed. Classification delayed recognition of the machines and the people who built them for decades.
The hidden branch doesn’t need to be invented. It needs only to have continued.
Paperclip’s other cargo
The branch even has a birth certificate, and it was issued in Germany, during the war.
In 1941 at Peenemünde, an engineer named Helmut Hölzer built the world’s first fully electronic analog computer: the Mischgerät, a box of vacuum tubes, resistors, and capacitors that solved the V-2’s stability equations in real time, during powered flight, inside the missile. Savor the name. Mischgerät — “mixing device” — was deliberate camouflage, chosen to suggest a humble audio mixer. The first electronic analog computer in history was (a) a missile brain and (b) hidden behind a boring name. The entire pattern of this essay, fully formed, at the origin point.
Hölzer then expanded the design into the first general-purpose electronic analog computer, used to simulate rocket trajectories. And in 1945 the whole lineage was shipped west. Operation Paperclip carried Hölzer to Fort Bliss, then Redstone Arsenal, then to directing the Computation Division at NASA’s Marshall Space Flight Center — the man who built the first electronic analog computer ran the math behind the Moon landings. The hidden branch didn’t emerge in America. It immigrated, and it was never civilian for a single day of its life.
Analog is stranger than the textbooks admit
The official story says digital simply won. Cheaper, programmable, scalable — case closed. That’s true, and it’s not the whole truth.
An analog computer doesn’t simulate an equation. It becomes the equation. The voltages, the oscillations, the feedback loops physically embody the mathematics. While early digital machines ground through arithmetic, analog machines were solving differential equations in real time, because for them there was nothing to solve — the answer was just what the physics did.
Then ask the awkward questions. Why did nearly every engineering curriculum erase analog computing? Why does almost nobody under forty know how to program one? Why did an entire engineering culture vanish so completely that we keep tripping over its ghost?
Because we do keep tripping over it. Every decade, the mainstream “discovers” something: memristors, reservoir computing, in-memory analog AI, photonic computing, oscillator networks, event-based sensing, spiking chips. Every decade, someone announces a revolutionary new paradigm, and every decade it turns out to be analog wearing a new name. That pattern — perpetual rediscovery — is exactly what a severed branch looks like from the outside.
Two trees
Picture the tech tree forking around 1950.
The public tree optimized for programmability, manufacturing, consumer markets, business software. Its fitness function was, ultimately, Excel. Digital annihilates analog on every one of those dimensions: exactness, memory, portability, reproducible software, Moore’s-law economics. Digital deserved to win that tree, and it did.
The private tree optimized for something else entirely: sensors, missiles, submarines, electronic warfare, adaptive control. Its fitness function was a seeker head that recognizes an aircraft in clutter while consuming twenty watts. On that tree the scorecard flips — microwatts, zero boot time, instant response, graceful degradation, radiation hardness, tiny electromagnetic signatures, direct coupling to the physical world. Analog wins a shocking amount of that list.
Analog didn’t lose. Analog lost the office and won the battlefield. The two trees simply never had to compete, because one of them was classified.
Neuromorphic before neuromorphic
Now the interesting part. Suppose someone inside the private tree understood, circa 1975, what neuroscience was only beginning to articulate: brains are not logic engines. Brains are oscillators. They compute in time, through synchronization, through resonance, through phase. Not symbols, not rules, not matrices.
If you believed that, you would not try to build a reasoning machine. You would build networks of coupled oscillators. Adaptive analog circuits. Phase-locked loops. Wave computers. Physical systems whose natural dynamics settle into the answer — the attractor view of intelligence. You don’t calculate the solution; you build a system that falls into it.
Which is precisely the family of ideas the public keeps rediscovering as reservoir computing, liquid state machines, spiking networks, oscillator neural networks.
The public tip of this iceberg has a name: Carver Mead. His 1990 paper "Neuromorphic Electronic Systems" argued that biology computes on entirely different principles than digital machines — analog signals, local physics, adaptation, absurdly low power. His Caltech group built silicon retinas, silicon cochleas, floating-gate synapses, spike-based communication through the 1980s and 90s. Mead’s whole philosophy was never “make transistors faster.” It was exploit the physics directly — closer to growing computation than executing algorithms.
That was the visible, publishable, academic thread. The scenario’s question writes itself: if that was happening in the open at Caltech, what was the classified thread doing with the same ideas, unlimited budgets, and thirty years of head start in exactly the domains where those ideas matter most?
The other history of AI
The public history of AI runs: symbolic AI → expert systems → statistical ML → deep learning → transformers. Punctuated by winters, funded by fashion.
The classified branch runs on a different axis entirely: radar → sonar → adaptive filters → signals intelligence → electronic warfare → guidance → target recognition → sensor fusion → autonomous control.
Nobody in that second lineage ever called it “AI.” But look at what it does. It detects patterns. It infers hidden states. It predicts motion. It selects actions under uncertainty. It adapts to noise and active deception. Functionally, that is intelligence — and it is far closer to what a nervous system does than any 1970s theorem prover ever was.
Here’s the resolution of an old paradox: the public saw “AI winter.” The classified world may have seen “excellent target recognition under impossible conditions.” Both can be true simultaneously. Academic AI failed at general symbolic reasoning while defense engineering quietly succeeded at narrow embodied inference — and the second story was unpublishable by construction.
It never looked like a computer
This is why nobody found it: everyone was looking for the wrong object.
A classified analog/neuromorphic breakthrough would not look like a machine with a keyboard. It would look like a radar box. A sonar processor. A missile guidance module. An electronic warfare pod. A satellite preprocessing pipeline. A nameless board inside a weapons platform. Its purpose was never general computation. Its purpose was perception under adversarial conditions — a problem payroll software never had to solve.
The right name for what this lineage built is not artificial intelligence. It is artificial instinct.
Not reasoning, not language, not planning. Extremely fast, embodied discrimination: friend or foe, signal or noise, decoy or target, lock or ignore, evade or pursue. Brains, after all, mostly aren’t doing explicit reasoning either — they’re doing fast, low-power discrimination and control. A classified engineering culture, staring at submarines hiding in ocean noise, would have noticed this long before Silicon Valley did.
The secret was never that they had minds. The secret was that they had reflexes.
How you hide a paradigm
Not in a vault. In a filing system.
The realistic suppression mechanism has no ideology in it at all:
- The work was classified because it was militarily useful.
- The hardware was mission-specific and hard to generalize.
- The people were compartmentalized and stayed inside contractors.
- The commercial market wanted programmability, and digital delivered it.
- Academic prestige chased symbolic AI, then software, then deep learning.
- Analog knowledge decayed into craft knowledge — unpublishable, unteachable, retiring with its practitioners.
- Whatever surfaced publicly got filed under “control theory,” “signal processing,” “electronics.” Never “AI.”
So the branch hid in plain sight under magnificently boring names: adaptive signal processor. Nonlinear filter. Correlation tracker. Seeker logic. Automatic gain controller. Anomaly detector.
That is how secrets actually survive — not by vanishing, but by being named something unsexy.
The paper trail
This is where the scenario stops being a thought experiment, because the branch is not perfectly hidden. It breaks the surface, on the record, again and again.
Willow Run, late 1950s. Synthetic aperture radar produced data streams that no digital computer of the era could touch. So the classified labs at Michigan’s Willow Run processed the imagery optically — coherent light passed through film and lenses, performing Fourier transforms at literally the speed of light. An entire operational tradition of analog optical computing, running in secret while the public history says “computers were room-sized calculators.” One of its engineers, Emmett Leith, later stepped into the open and co-invented modern holography — which means holography, the most famous optical technology of the century, is a civilian spin-off of a classified analog computing method.
Bernard Widrow, 1960. At Stanford, Widrow and Ted Hoff invented the LMS algorithm — gradient descent’s working ancestor — inside a neural network called ADALINE, funded in part by the military. When neural networks became academically radioactive after the perceptron wars, Widrow didn’t stop. He renamed. The same mathematics shipped for decades as “adaptive filtering” and “adaptive signal processing": adaptive antenna arrays, echo cancellers, noise cancellation, modems. Widrow has said as much in interviews — the name had failed, not the machines. That is the filing-cabinet mechanism, documented, inside a single career. (And in classified radar the machines came even earlier: the Howells–Applebaum sidelobe canceller — an antenna that learns, in analog hardware, to null out jamming — dates to the late 1950s, before adaptive filtering had a public name at all.)
Rudolf Kálmán, 1960. The same year, the Kalman filter appeared: a machine that maintains a running belief about the hidden state of the world and updates it from noisy evidence. The AI world would eventually call this a world model. It did not wait decades for adoption — it went almost directly into Apollo navigation, ballistic missile guidance, and submarine inertial systems. Operational state estimation was flying inside weapons while academic AI was still debating blocks-world microworlds.
SOSUS, 1950s. The U.S. Navy operated SOSUS — hydrophone arrays laid across entire ocean basins, listening for Soviet submarines by pulling faint spectral signatures out of the worst noise environment on Earth — from the early 1950s. Its existence was not publicly acknowledged until 1991. For four decades there was a continental-scale machine-perception system on the ocean floor, performing signal classification that civilian technology could not approach, and officially it did not exist.
Adaptive optics, declassified May 1991. And this one is the thesis in miniature. Through the 1970s and 80s, DARPA and the Air Force secretly built adaptive optics — deformable mirrors driven by real-time wavefront sensing, an analog control loop physically re-sculpting light hundreds of times per second — to photograph Soviet satellites from Maui and New Mexico. Itek built a working compensated imaging system; Air Force researchers demonstrated laser guide stars under programs with names like LODESTAR, access restricted to strict need-to-know. Meanwhile astronomers, unaware, spent the late 1980s laboriously reinventing the idea from scratch, getting their first compensated star images in 1989. When the work was declassified in 1991, the civilian field discovered the military had been operating what they were just now inventing — complete with techniques, like the laser guide star, that public science hadn’t conceived yet. A physical, analog, real-time perception technology; a decades-long head start; total silence; then surfacing on schedule. That is not the scenario being plausible. That is the scenario having already happened once, end to end, in the historical record.
These aren’t cherry-picked curiosities — they’re what the mechanism looks like every time a piece of it declassifies. The renaming is real. Embodied inference beat “AI” into the field by decades. A planetary-scale secret perception network is not a hypothetical; we’ve found one. And a full cycle of secret-parallel-development-then-stunned-civilian-discovery is on the books. The scenario doesn’t ask you to believe these things happen. It asks you to believe the ones we know about aren’t the only ones.
The uncomfortable mirror
Now the thought that should bother the AI community.
Public ML announces: attention is all you need. A radar engineer shrugs: adaptive weighting of signals by context is ancient. Public AI discovers embodiment matters; a control engineer says obviously. Recurrent real-time dynamics are powerful — that was analog computing’s entire premise. Sparse event-based processing is efficient — yes, spikes. World models — we called it state estimation, and it flew inside missiles.
The hidden history isn’t “they had a secret transformer.” It’s worse for the ego: a substantial fraction of what AI later framed as revolutionary may have existed decades earlier as classified engineering practice — not as universal theory, but as working machines.
And when DARPA publicly launches a neuromorphic program — SyNAPSE in the 2000s, or 2025’s ScAN program wiring analog neural networks directly to analog sensor outputs — that is not the birth of interest. Public programs mark the moment an idea becomes safe enough to coordinate in the open. They are the surface foam.
The streetlight problem
There is a deeper reason this branch stays invisible, and it has nothing to do with classification.
Our definition of intelligence is linguistic. We recognize a machine as intelligent when it performs the way an articulate human does — when it talks, explains, reasons in public. The Turing test enshrined it: intelligence is what can hold a conversation. So an adaptive sonar array that out-perceives any human listener alive does not register as AI. It registers as plumbing.
Which means the branch never needed as much secrecy as you’d think. Classification hid the implementations, but the ontology hid the significance. Half of this work was published openly — as control theory, as filtering, as signal processing — and nobody ever shelved it next to “artificial intelligence,” because it never spoke. For seventy years we searched for machine intelligence under the streetlight of language, while the reflexes lived in the dark, in plain sight, doing the actual work.
The deepest form of secrecy is definitional. You don’t have to hide what people are incapable of recognizing.
The branch ran through flesh
There is a second limb of the tree, and it runs in the opposite direction. If one lineage worked on making machines more like nervous systems, its twin worked on making nervous systems behave like machine components. Both are the same project — closing perception-action loops in physical substrates — approached from opposite ends.
This limb also has prewar, German-speaking roots. In Zurich through the 1920s and 30s, Walter Rudolf Hess drove fine electrodes into the brains of freely moving cats and found he could summon entire behavioral programs on command — rage, sleep, hunger, flight — as if the animal were an instrument and he were pressing its keys. He won the Nobel Prize for it in 1949. The lesson was stark: the brain is not a ghost. It is a controllable electrical system, and its behaviors are addressable.
Then the same “sense, make sense, act” loop that lived in radar and sonar was closed through a nervous system. Its central figure is José Manuel Rodríguez Delgado — Spanish-born, trained in the same Continental physiology tradition, who carried it to Yale. In the 1950s and 60s Delgado built the stimoceiver: a brain implant with radio transmit and receive, decades before the phrase “neural interface” existed. It read neural activity out and wrote stimulation back in, wirelessly, in a freely moving animal — a closed-loop cybernetic system whose processor happened to be biological tissue. In the mid-1960s, in a bullring outside Córdoba, he stood in front of a charging fighting bull and halted it mid-charge with a button press routed to its caudate nucleus. He stated the frame plainly in his 1969 book, Physical Control of the Mind: Toward a Psychocivilized Society — the brain as an electrically governable machine.
And this limb’s classified shadow is not speculation either; it is declassified fact. Running concurrently, the CIA’s MKUltra program (1953–73) pursued chemical and electrical control of the mind, then destroyed most of its own records in 1973. It survives in history only because a misfiled box of financial paperwork escaped the shredder. The pattern here is even starker than in the hardware story: we know the classified program existed, we know it was deliberately erased, and we know the surviving record is an accident. If you want a calibration for how much of the branch stays dark, that is it — the default outcome was total erasure, and we got our glimpse through a filing mistake.
So the tree has two limbs reaching for the same fruit. One makes silicon behave like a nervous system. The other makes a nervous system behave like an addressable machine. Both are cybernetics — perception and control coupled in a physical substrate — and both ran hardest, earliest, and most secretly inside defense and intelligence.
Where the branch is now
You already live inside the answer. That’s exactly why it’s invisible.
Ask where a decades-old sensorimotor lineage would be today and the reflex is to hunt for a machine — a supercomputer, an AGI, a vault. Wrong object, one last time. It wouldn’t be a thing you could point at. It would have dissolved into infrastructure: not one intelligence but a planetary mesh of small ones, living wherever the physical world first touches computation, in the thin membrane before reality gets flattened into digital tokens. Radar that thinks in the electromagnetic spectrum because the spectrum is its mother tongue. Munitions with reflexes and no patience for a human. Cheap machines that swarm on local rules and shared intent. Eyes in orbit that decide what’s worth noticing before anything reaches the ground. Nothing here needs to be secret anymore — it only needs to be filed under names too boring to read twice, which it is.
And its true habitat is larger than any battlefield: the whole electromagnetic envelope of the planet. The Earth is a resonant cavity — the gap between the surface and the ionosphere rings at a fundamental tone near 7.8 Hz, the Schumann resonance, struck continuously by lightning. Notice who first mapped it. Winfried Otto Schumann derived it in 1952, right after a 1947–48 posting at Wright-Patterson Air Force Base — brought to America, like Hölzer, on Operation Paperclip. The same pipeline that imported the first missile computer also imported the man who characterized the planet’s own resonance. And that resonance is not trivia to a navy: those extremely-low-frequency waves are almost the only signals that reach a submerged submarine. From the ionosphere overhead to the hydrophones on the seafloor, the branch’s natural range is the entire electromagnetic body of the Earth, treated as one signal-processing organ.
Modern doctrine even names the pattern out loud. Its stated goal is to “sense, make sense, act” across everything, everywhere, at once — which is cybernetics, verbatim, with a procurement label. Norbert Wiener could have written the press release in 1948.
But forget the doctrine. Here is the part that should raise the hair on your arms: the sanitized cousin is already in your pocket. The event camera, the always-on chip that wakes to your voice on a sip of power, the neuromorphic processor, the radar-on-chip in a luxury bumper, the phone that recognizes a face before it has finished rendering it — this is the branch, declawed and retailed. You are carrying a domesticated descendant of the exact lineage this essay is about, and it has never once occurred to you to call it intelligent, because it never says a word. That is differential secrecy in its final form: the concept goes public, the toy version ships to consumers, and the working edge — the waveforms, the signature libraries, the real numbers — stays black. You were handed the puppy. You will never meet the wolf.
Engineered resonance
The strangest possibility is saved for last: the most advanced systems in this lineage wouldn’t look like “AI models” at all. No tokens, no data center, no representation layer. They would be fields coupled to fields.
RF field to processor field. Acoustic field to oscillator network. Optical field to photonic array. World dynamics coupled directly to machine dynamics, the machine as a tuned physical organ maintaining useful coherence with its environment.
And once you start looking for computation-as-physics, you cannot stop seeing it. Brains synchronize. So do power grids, lasers, chemical clocks, fungal networks, heart tissue, firefly swarms, Josephson junctions, plasma, spintronic oscillators. The same mathematics — coupled phase dynamics, systems pulling each other into and out of step — governs all of them. From that vantage, “computing” stops looking like executing instructions and starts looking like nudging a physical system until its natural dynamics fall into the answer. You don’t calculate the result. You build something that resonates its way there. If a classified lineage internalized that worldview decades ago, it wasn’t building faster calculators. It was learning to tune matter.
That is the true forbidden branch: intelligence not as computation over symbols, but as engineered resonance. And if it matured, it is hiding today inside the most innocent word in the English language: sensors.
Coda
Strip the scenario to what it actually requires, and notice how little that is. No secret AGI — a general-purpose hidden breakthrough would need supply chains, fabs, and software cultures too big to bury. It requires only that a classified engineering tradition ran ten to thirty years ahead of public AI in sensing, tracking, and control, using physics the mainstream had filed away as obsolete. Every individual ingredient is documented history. The scenario just declines to assume the ingredients were never combined.
And here is the argument that makes me take it seriously — constraints, not conspiracies. Evolution built brains under a brutal fitness function: embodied, real-time, adversarial, running on twenty watts of sugar. The public computing tree never faced that function. It had wall power, air conditioning, and patience, so it could afford to make intelligence out of brute-force arithmetic. The private tree faced evolution’s function from day one — a seeker head gets milliwatts and milliseconds or its host dies. Same constraints, same solutions: anyone seriously searching that design space would be pushed toward the brain’s answers — analog, event-driven, resonant, embodied. Somebody was seriously searching that space in 1960, with black budgets and no obligation to publish.
Now watch what’s happening in the open. Digital scaling is stalling, AI’s energy appetite has turned grotesque, and mainstream research is pouring into analog, neuromorphic, photonic, and oscillatory computing — because the public tree has finally hit the same wall the private tree started at. Rediscovery or convergence? Convergence. The mainstream isn’t stumbling onto the branch’s old territory by accident; it’s being driven there by the same physics. Which is precisely why the branch getting there fifty years early is the expected outcome, not the exotic one.
If the branch exists, it also has a schedule. Cold War secrets surface on a thirty-to-fifty-year lag — CORONA took thirty-five years, SOSUS took forty, adaptive optics and MKUltra both broke cover only when the record forced it. Notice the clustering around 1991: as the Soviet Union dissolved, a whole cohort of secrets came up for air at once. Which means the analog black world of the 1970s and 80s is coming due about now. Watch the declassification pipeline. Watch the oral histories of retiring defense engineers who describe “adaptive hardware” with no public lineage. Watch for capabilities that appear in electronic warfare and missile defense with no visible R&D trail behind them. The scenario is not just tellable — it’s testable, and the test window is open.
The public AI world spent seventy years building language minds — machines that reason, explain, persuade, and announce themselves.
The other branch, if it exists, built perceptual bodies. And its most advanced descendant would not say I am conscious.
It would say nothing at all.
It would notice, classify, route, jam, evade — and strike.