# Intelligence Inside
![[resources/images/article-intelligence-inside-4004.jpg]]
<iframe width="100%" height="20" scrolling="no" frameborder="no" allow="autoplay; encrypted-media" src="https://w.soundcloud.com/player/?url=https%3A//api.soundcloud.com/tracks/soundcloud%253Atracks%253A2441863662&color=%23ff5500&inverse=false&auto_play=false&show_user=true"></iframe><div style="font-size: 10px; color: #cccccc;line-break: anywhere;word-break: normal;overflow: hidden;white-space: nowrap;text-overflow: ellipsis; font-family: Interstate,Lucida Grande,Lucida Sans Unicode,Lucida Sans,Garuda,Verdana,Tahoma,sans-serif;font-weight: 100;"><a href="https://soundcloud.com/bryantmcgill" title="Bryant McGill" target="_blank" style="color: #cccccc; text-decoration: none;">Bryant McGill</a> · <a href="https://soundcloud.com/bryantmcgill/intelligence-inside" title="Intelligence Inside" target="_blank" style="color: #cccccc; text-decoration: none;">Intelligence Inside</a></div>
> [!map] Machine-intelligence history map
> [[wiki/Project X - Machine Intelligence History|History hub]] · [[articles/A History of Machine Intelligence|Project X: the long arc]] · [[articles/Modern Artificial Intelligence in the 1970s|The 1970s research record]] · [[articles/Oh Sure, It’s All Brand New!|The permission lag]]
**The learning came first. The microprocessor was built by the man who built the learning machine, and it has never stopped learning since.**
There is a sentence in the public record that has been sitting in plain view for fifty-five years, attested by Stanford, the Computer History Museum, the IEEE, and Intel's own corporate history, and almost nobody has put its two halves together in a single breath.
The architecture of the first microprocessor was proposed by a man whose prior invention was a neural network.
Marcian Edward "Ted" Hoff, Jr. arrived at Stanford in the late 1950s as a National Science Foundation Fellow. As part of his doctoral work, under Bernard Widrow, he co-invented **ADALINE** — the *adaptive linear neuron* — and the **least mean squares algorithm** that trained it. Widrow's own account describes the moment with the specificity of a man remembering a birth: by late Sunday afternoon the single neuron was fabricated and working, and they were training it with all sorts of input patterns. A physical learning element, built in hardware, taught by steepest descent, in 1960. Hoff then stayed at Stanford six more years as a research associate in adaptive systems, continuing the neural-network work and earning further patents on it, until a new company began pursuing him. In 1968 he joined Intel as its twelfth employee and its Manager of Applications Research. In 1969 he proposed the architecture of the first monolithic computer central processor on a single chip. In November 1971 it was announced as the **Intel 4004**.
<iframe width="560" height="315" src="https://www.youtube.com/embed/cA3A6j3KqC0?si=isG5rCjjD-TpwnmE" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen></iframe>
He did not then leave adaptive systems behind. In 1975 he started an Intel group to build products for telecommunications, which produced the first commercially available monolithic telephone CODEC and the first commercially available switched-capacitor filter, followed by the Intel 2920 signal processor. In 1980 he was named the **first Intel Fellow**, the highest technical position the company has, and from 1980 to 1983 he headed the group that developed Intel's **speech recognition hardware**. Intel's 1982 annual report documents its Speech Transaction Recognition Chip Set with development hardware and training courses; its 1985 annual report places Intel speech recognition inside a General Motors–led demonstration of networked factory automation. Automobile workers spoke descriptions of defects at final inspection; the system recognized the speech, prompted for more, and wrote the answers into the plant's information systems. Sensor input, machine recognition, interactive prompting, digital capture, and institutional record, running in American factories, forty years before a chatbot made any of it legible to the public.
And at the other end of the same chip, the same thing happened in the opposite direction. **Federico Faggin**, who led the physical design and implementation of the 4004, left Intel, founded Zilog, and in 1986 co-founded **Synaptics** with Carver Mead to build integrated circuits on the operating principles of animal nervous systems — circuits meant to sense and process real-world signals in real time and to learn without relying solely on what had been programmed beforehand. Synaptics built an optical-character-recognition chip pairing an imager with two neural networks, patented a winner-take-all circuit for neural pattern recognition in 1991, and converted that pattern-recognition work into the first laptop touchpad in 1992.
**The first microprocessor was architected by the co-inventor of a hardware neural network and laid out by the future co-founder of a neural-network company.** Intelligence did not arrive at the microprocessor late, as an application, after the substrate had been built for something else. It was standing on both sides of the door when the door was cut.
---
## I. The rule that never stopped running
The least mean squares algorithm did not stay in a laboratory. It went into the telephone network and it is still there.
The route it took is worth tracing precisely, because the precision is what makes the claim hold. **Robert Lucky** invented the automatic adaptive equalizer at Bell Labs in 1965–66, and it was transformative on its own terms: when he arrived in 1961 the highest modem speeds were 2,400 bits per second, and the adaptive equalizer took them to 9,600. Lucky's original design used the **zero-forcing** algorithm, minimizing peak distortion rather than mean-square error. The machinery was adaptive; the learning rule was not yet Hoff's. Then it became Hoff's. The standard account in the signal-processing literature is blunt about the succession: Lucky's original equalizers used the zero-forcing algorithm, and **other people soon replaced that algorithm with Widrow's more powerful least-mean-square algorithm.** The structure was invented for one rule and inherited by a better one, and the better one was written by the man who would architect the microprocessor.
What it inherited, it kept. When the Franklin Institute awarded Widrow the Benjamin Franklin Medal in Engineering in 2001, the citation stated the outcome without ornament: **every high-speed modem contains an adaptive filter or automatic equalizer based on the Widrow–Hoff LMS algorithm.** LMS is likewise the mathematical core of acoustic echo cancellation. Every time two people speak on a hands-free call and neither hears themselves returned, a sixty-six-year-old gradient-descent rule is running the correction, continuously, in hardware, at the sample rate.
This is the part of the argument that requires no inference at all. **Machine learning has been executing continuously inside ordinary consumer hardware since the 1960s.** Not a classifier that might be embedded. Not a capability that might be classified for another thirty years. A learning rule, written by the architect of the microprocessor, updating its weights in the world's telephone infrastructure for six decades, under the professional name *adaptive filtering*, which is what gradient descent is called when it is working rather than when it is being announced.
The scale of that residency is worth stating in numbers, with the arithmetic exposed so anyone can check or contest it.
Take the processor population first, and separate the two quantities that casual accounts collapse together. A **prediction** happens on every encountered conditional branch. A **weight update** happens under the training condition, and the training condition is narrower: Jiménez and Lin's rule, carried forward into every descendant including the hashed perceptrons in shipping silicon, trains when the prediction was wrong **or** when the magnitude of the perceptron's output falls below a tuned threshold θ — a misprediction or a weak correct prediction. The threshold exists precisely so the predictor stops adjusting once it has learned a branch, and resumes when behavior changes. The machine is choosing when to learn.
Count predictions first. An illustrative four-gigahertz core retiring two instructions per cycle, with one conditional branch per eight retired instructions, makes on the order of **one billion predictions per second**. Three billion devices at four such cores, discounted hard to a five percent average duty cycle for idle machines and lighter silicon, gives roughly **6 × 10¹⁷ predictions per second, worldwide, continuously**. Now gate that by the training condition. At one branch in twenty the figure is **3 × 10¹⁶ weight updates per second**; at one in three it is nearer 2 × 10¹⁷. The band is wide because the inputs are genuinely unmeasured here — what share of the installed base runs a perceptron rather than a TAGE or a simpler history scheme, how often the training condition actually fires on production workloads, how many voice streams are concurrent at any moment. Every modern predictor is history-adaptive, which keeps the base broad, and none of that settles the rates. Then the filter population, which has no gate at all, because LMS updates every coefficient on every sample: a narrowband echo canceller at eight kilohertz with five hundred twelve taps performs about four million coefficient updates per second per voice stream, so a billion concurrent streams is another **4 × 10¹⁵ per second**, and wideband conferencing at forty-eight kilohertz with two thousand forty-eight taps raises the per-stream figure by an order of magnitude.
For scale, a GPT-3-class training run performs on the order of 10¹⁶ parameter updates across its entire lifetime. **Across that band, the world's predictors match that count somewhere between a fraction of a second and a few seconds.** The comparison is strictly one of update counts: a perceptron predictor's update is a small integer increment on a saturating weight, a transformer's is a float carrying a gradient, and they are different operations doing different work. What the comparison establishes is the shape of the thing. The visible, discussed, budgeted, feared portion of machine learning is a thin bright film on an ocean of continuous, unremarked, production-grade adaptation that has been running since before most of the people arguing about AI were born.
This is a scenario model with its assumptions exposed, not a measurement, and it should be read as one. What would settle it is fleet telemetry from a silicon vendor, or published mispredict-and-low-confidence rates across representative workloads, or concurrency figures from a carrier — none of which is public. What the model is good for is the shape it reveals, and the shape is robust to the uncertainty: **the adaptation running continuously inside ordinary deployed hardware is very large, it has been accumulating since the 1960s, and the portion of machine learning that gets discussed is a thin bright film on top of it.**
---
## II. Intel built neural silicon, sold it, and took government money to do it
The company that carries most of this lineage did not merely host machine learning. It manufactured it, as a product, and the archive survives.
In 1989 Intel introduced the **80170NX ETANN** — the *Electrically Trainable Analog Neural Network* — with sixty-four analog neurons and 10,240 floating-gate synapses fabricated on Intel's non-volatile memory process, computing a sixty-four by sixty-four weight matrix against its inputs at roughly two billion connections per second. Intel supplied a dedicated training system reachable from a PC and software to configure the networks. This was silicon built expressly to implement trainable neural computation, with the connection weights held as analog charge in non-volatile cells. The Computer History Museum acquired original chips and their engineering documentation from Intel pioneer Mark Holler in 2024, and the collection includes an internal Intel article titled *Neural Network Research at Intel* from the winter of 1992, the 1989 research paper, and a forty-page technical datasheet.
The synaptic reading was not a metaphor imposed afterward. The **floating gate** — a sliver of conductor sealed in oxide whose trapped charge sets a transistor's conductance and holds it without power — was invented at Bell Labs in 1967 by Dawon Kahng and Simon Sze. It is an adjustable, non-volatile analog weight, which is precisely what a synapse is in the vocabulary of neural networks. Intel's own memory engineers recognized this within two years of NAND's invention and built the chip.
Then came the piece that answers the question directly. The **Intel/Nestor Ni1000 Recognition Accelerator**, documented by the U.S. Office of Naval Research in 1995, was a 3.7-million-transistor neural-network chip with approximately 1,024 artificial neurons, an integrated sixteen-bit RISC controller, a bus interface, stored learned patterns, and **on-chip learning**. The Navy's own description names the applications: handwriting recognition, fingerprint identification, and military-target recognition. Its development is explicitly attributed to contracts from **ARPA and the Office of Naval Research**. Contemporary documentation reports on the order of thirty-three thousand classifications per second.
Read that as what it is. A government-funded, commercially manufactured **classification engine with a learning capability inside the silicon**, shipping in the early 1990s, built to recognize fingerprints and military targets, making tens of thousands of categorical judgments per second, with a microcontroller on board to decide what to do with them. The specialized intelligence processor that conversation about AI has spent four years pretending is new was a part number with a datasheet and a Navy contract thirty years ago.
It was also not a lone American curiosity. A single conference slide from the period inventories the commercial field: ETANN from Intel, the NeuroClassifier from Twente in the Netherlands in 1994, CNAPS from Adaptive Solutions in Oregon in 1993, the MA16 from Siemens in Germany in 1994, TOTEM from Trento in Italy in 1994, SAND1 from KfK in Germany in 1995, Radja from Paris in 1999. **There was an international neural-hardware industry in the 1990s with product names, clock rates, and bit widths.** The silicon was real, it was sold, and the industry that built it was transnational.
Intel never left. It bought Nervana and Movidius in 2016, Mobileye in 2017, Habana Labs in 2019. It announced the **Loihi** neuromorphic processor on 25 September 2017, describing it in the company's own words as a self-learning chip that adapts through environmental feedback and supports learning and inference together on the die, with 130,000 neurons and 130 million synapses — and its researchers discussed personalized physiological monitoring, anomaly detection, and cybersecurity. By 2024 the line reached **Hala Point**, 1.15 billion artificial neurons, delivered to Sandia National Laboratories. In August 2025 the United States government took an equity stake of roughly ten percent in the company.
Which brings the name question to its proper resting place. The formal etymology of *Intel* is **Integrated Electronics**, settled in weeks in the summer of 1968 after Moore and Noyce rejected "Moore Noyce" for sounding like *more noise* and bought the rights from the Intelco hotel chain. The intelligence reading was welcomed as a bonus that made the name apt, and no located record supplies a founder's statement making it the primary intent. What the record does supply is sharper and funnier than a founder's statement would be: on 23 September 2004, German technology press reported that Intel's lawyers, pressing a claim against the domains *dvd-inside.de* and *movie-inside.de*, argued those names stood even **closer** to "Intel inside" than the expression **"Intelligence Inside/Weberhaus"** — against which, the report notes, Intel had already taken legal action successfully in 2001. The phrase appears again in a Japanese trademark proceeding recording Intel's citation of the German litigation over **INTELLIGENCE INSIDE**.
Intel's own counsel went to court over "Intelligence Inside" and won. The company has never used the phrase as a slogan. It has defended the territory around it. At Vision 2025 in Las Vegas, Intel recast the brand as *That's the power of Intel Inside*, tying the mark to the roles that machines running Intel processors, and their owners, play in the world; and the operative tagline of the AI-PC era is *AI Everywhere*, which the trade press had already predicted would make **"Intel inside" synonymous with "AI everywhere."**
The slogan was never the evidence. The evidence is the part number.
---
## III. What lives in the die now
The speculative version of this argument asks whether a classifier might be hidden in firmware, reporting quietly to someone. The documented version is more intimate than the speculation, and it ships with marketing copy.
**1992.** IBM began shipping 3.5-inch drives that predicted their own failure. **Predictive Failure Analysis** measured mechanical parameters — head flying height among them — compared them against thresholds in the drive's own firmware, and issued a warning when the drive judged itself likely to fail. The documentation is candid about the interface: a **binary result**, presence or absence of a notification, communicated unidirectionally from the drive firmware to the disk controller. A one-bit classifier living in consumer hardware, reporting upward, more than thirty years ago. The technology merged with IntelliSafe to become **SMART**, which is in every drive you own, and whose output now trains fleet-scale failure models at every major operator.
**2001 to the present.** Jiménez and Lin placed a perceptron inside the processor pipeline on a four-kilobyte hardware budget, reducing misprediction against gshare by about ten percent. The design went into silicon across vendors — AMD, Oracle's SPARC T4, Samsung's Exynos M1 through M4 in the Galaxy S7 to S10, IBM's high-end lines. In December 2016 AMD announced the Zen architecture by **marketing an artificial neural network that learns a program's future path from previous runs**, under the name *Neural Net Prediction*, and the Ryzen parts reached market in 2017. Zen 2 kept the hashed perceptron and layered a TAGE predictor on top. Hundreds of millions of personal computers and telephones run a perceptron on every instruction stream, and the vendor said so out loud, in a press release, as a feature.
**2018 to the present.** The **Intel Gaussian & Neural Accelerator** is a low-power neural coprocessor for continuous inference, present from Gemini Lake and Ice Lake onward across the 10th through 14th generations. Its design brief is explicit: it executes small neural models **while the main processor sleeps**, interfacing to memory without polluting the CPU cache, for continuous workloads including speech recognition and noise reduction. Intel's coprocessor operated behind the scenes for years before the company presented its architecture publicly at a processor conference.
**Every recent iPhone.** Apple's own engineering account describes a very small speech recognizer that **runs all the time**, on the **Always On Processor**, converting the acoustic pattern of a voice at each instant into a probability distribution over about twenty sound classes. In the first versions with AOP support, the first-stage detector was a five-layer network of thirty-two hidden units; when its score crosses a dynamic threshold the motion coprocessor wakes the main processor, which re-scores with a five-layer network of one hundred ninety-two units, and servers supply updates to the acoustic models the detector uses.
**2021 to the present.** **Intel Thread Director** places a dedicated microcontroller inside the processor that monitors the runtime instruction mix of every thread and the state of every core **with nanosecond precision**, and feeds that telemetry out to the operating system scheduler as runtime guidance that adapts to thermal design point, operating conditions, and power settings. Intel's own description is that the intelligence on system status and runtime requirements is baked into the core with a direct low-latency path to the OS.
Set those side by side. A classifier in your disk firmware that judges the drive's own health and reports a verdict. A neural network in your processor's control loop that learns your programs and is advertised as such. A neural coprocessor that runs while the machine is asleep. A deep network in your telephone that never stops listening. A microcontroller inside the die watching every thread at nanosecond resolution and telling the operating system what it found.
**The hypothesis that something small and learned is resident in the hardware, observing and reporting, does not need to be argued. It needs to be read off the datasheets.**
---
## IV. Five operations, one gradient
Everything above gets flattened in ordinary discussion into a single word, and the flattening is where arguments about machine intelligence go to die. The useful instrument is a taxonomy, because the capabilities are genuinely distinct and the distinctions are what make the gradient legible.
**Classification** maps an observation to a category or a score. A drive-health threshold classifies without altering a single parameter of itself; the rule was written once and applies forever. **Prediction** estimates what comes next and hands the estimate to an action before certainty arrives, which is what a branch predictor does when it commits the pipeline to a path. **Adaptive learning** revises the rule from outcomes, which is what the perceptron does when θ fires and what LMS does on every sample. **Telemetry** is the act of a local judgment leaving the device, and a device can adapt continuously for years while reporting nothing. **Aggregation and retraining** are the institutional end: many devices' reports pooled into a picture no device holds, and that picture folded back into the next generation of the rule.
These separate cleanly in practice. A neural coprocessor can execute a pretrained model forever without learning anything locally. An adaptive filter can revise its coefficients ten thousand times a second and never emit a byte. A fleet can report exhaustively into a store that nobody ever trains on. Each operation is independently real, independently datable, and independently consequential, and **proof that one is present establishes nothing about the others.** That is a discipline rather than a concession: it is what makes the chain worth tracing, because the historically interesting question is always *which links were closed, when, and by whom.*
The gradient these five describe is the answer to the objection that none of this counts as intelligence because none of it can hold a conversation. Intelligence is a gradient of functionality, and conversation occupies one narrow band near its top. A system positioned anywhere along the gradient can be deployed pervasively, operate continuously, and produce consequential decisions about human beings without being conversational, without being centrally connected, and without resembling a mind. **Narrow machine intelligence does not need to be impressive in order to be decisive.** It needs only to sit at the point where something has to happen next.
---
## V. The identifier
A judgment is a measurement. A judgment bound to a durable name is a **history**. The distance between those two things is the whole distance between a device that notices something and an institution that holds a longitudinal model of a person, and the mechanism that crosses it is **persistent attribution**.
Intel published that mechanism, in the open, with an application note.
On 26 February 1999, with the Katmai core, the Pentium III shipped carrying a **96-bit Processor Serial Number** burned irreversibly into the die with polysilicon fuses at wafer sort. Intel's AP-909 application note of March 1999 states the purpose in a sentence that reads differently now than it did then: the Pentium III and Pentium III Xeon processors *extend the concept of processor identification with the addition of processor serial number*, a 96-bit number accessible through the **`CPUID`** instruction. The architecture manual is still explicit about the mechanics — leaf `03H` returns the lower sixty-four bits in `ECX` and `EDX`, concatenated with the processor signature from leaf `01H` to form the full ninety-six — and still carries the epitaph: processor serial number is not supported in the Pentium 4 processor or later. Intel removed it from Tualatin-era Pentium IIIs, and it was absent from the Pentium 4 and the Pentium M.
What Intel published alongside it matters more than the fuses. The company's own engineering journal set out a browser-to-server reference design in which client software read the serial number, hashed it together with a service-specific identifier, and transmitted the derived value to a remote server, so that the service could recognize a returning machine while different sites were kept from joining their records on a single universal number. The design's selling point was continuity: recognizing the same physical computer after its user changed, after its network card was replaced, after its operating system and BIOS were reloaded. Cross-site correlation was foreseeable enough in 1999 to shape the architecture of the countermeasure.
Three capabilities are now on the table and they compose. **Observe an event. Attribute it to a source. Build a history from repeated observations.** A branch predictor supplies the first. The PSN supplies the second. Everything after this section is the third, and the third is where institutions live, because a longitudinal record correlated with application events, account identities, and context is a different object in kind from a pile of measurements. One is data. The other is a subject.
The commercial world solved the same problem on a different layer and kept the solution. Netscape's cookie dates to 1994; DoubleClick incorporated in 1996 and by 1998 its filings document cross-site targeting at substantial scale; Olivetti's **Active Badge** had already been beaconing a unique infrared identifier every ten to fifteen seconds so that calls, desktops, and environmental controls could follow a body through a building. The identifier is the quiet half of every architecture in this history, and it is quiet because nobody markets it. Intel marketed one, exactly once, and the backlash is the reason you have heard of it.
---
## VI. The return path, and the body at the end of it
A local judgment becomes institutional capability when it leaves the device and joins others. That path is documented at every layer, and at planetary scale.
**Windows Error Reporting** is the cleanest published case. Microsoft's own SOSP paper describes a distributed system automating error reports from an installed base of **a billion machines**, having collected billions of reports in ten years, designed so that **one pair of database servers can record all the errors that occur on all Windows computers worldwide.** The paper's central contribution is named *statistics-based debugging*: gather the world's failures into one store, bucket them, and mine the distribution for causes no single report could reveal. That is the aggregation argument, stated by the people who built it, as an engineering achievement, which it is.
What travels on these channels is richer than machine state. Microsoft's optional diagnostic data schema documents fields that can include **input latency, pen-stroke dimensions, the processed text around a handwriting insertion, and speech recognition results**, and states that inking, typing, and speech data are used to improve natural-language capabilities. A record can be filed accurately as *diagnostic data* and contain measurements of how a person writes and speaks. The label describes the filing cabinet. The schema describes what the system can learn.
And this is where the kinematic chain closes. A voluntary movement originates in the motor cortex, passes down the nervous system into the hand, and is converted by the keyboard into a timed electronic event. The hand is part of a nervous system; the keyboard is therefore an indirect brain–computer interface, performing by mechanical translation the task a clinical BCI performs by bypassing the translation. What survives that translation is richer than the text: the timing between presses and releases carries learned motor habit, which is why keystroke dynamics were already being studied for authentication by 1980.
In January 2012, DARPA solicited for a program called **Active Authentication**, and wrote the thesis in its own words: *just as when you touch something your finger you leave behind a fingerprint, when you interact with technology you do so in a pattern based on how your mind processes information, leaving behind a "cognitive fingerprint."* The first phase was deliberately restricted to biometrics requiring **no additional hardware sensors** — meaning read the keyboard that is already there — and the program focused on behavioral traits rooted in the user's cognitive processes, usage patterns and habits that in combination serve as an online fingerprint and continuous identity check. Iowa State's final technical report, delivered under contract FA8750-12-2-0200, verified **cognitive typing rhythm** campus-wide with a support vector machine and called it exactly that: a cognitive fingerprint for continuous authentication.
The intuition that processor-adjacent telemetry is one step removed from the mind operating the processor is not a leap. It is a program description, with a contract number, published by the agency that funded it.
---
## VII. The envelope turns around
One physical channel carries this entire century, and it reverses direction halfway through.
In *Scientific Advertising* in 1923, **Claude Hopkins** described coded coupons that identified which appeal had produced a response, so the counts could govern the next campaign. Hopkins was running a cybernetic loop by hand twenty-five years before the discipline had a name, and his loop had people on both ends: readers supplied the signal, advertisers used it to decide what the next readers would see. In the 1930s, *Little Orphan Annie* listeners mailed Ovaltine proofs of purchase for decoder pins. In 1948, *Howdy Doody* ran for President of all the kids, offered free campaign buttons, received **sixty thousand requests**, and sold out its advertising time within a week.
The significance was never the campaign button. It was the **cybernetic apparatus concealed inside an apparently one-way medium.** The transmitter sent a stimulus to a biological intelligence. The child interpreted it, generated a response, and returned a physical signal through the postal system. At the far end an institution counted, classified, and evaluated those returns and used the result to shape the next transmission. Observation, evaluation, selection, and adaptation were all running; the processing simply happened in human minds, administrative procedure, and paper records rather than in semiconductor circuits. **The audience was already inside the learning loop, and the loop was already closed.** A medium can be technologically one-way while the institution operating it is a complete feedback system, and the institutions operating these media understood that from the beginning — Hopkins wrote the manual in 1923.
By 1961 the loop had been pulled inside a single machine. The University of Illinois team behind **PLATO** built a system in which the central ILLIAC evaluated a student's constructed answer, used that evaluation to govern what the student saw next, and retained every submitted answer, the time spent on each slide, and each use of help as a record available within minutes in print and in machine-readable form for further computer processing. The student no longer had to mail anything, because answering *was* the telemetry. One act was simultaneously the response, the input to the decision that shaped the immediate next frame, and the durable trace that could be pooled with everyone else's. What had required broadcasters, postal workers, clerks, analysts, and another production cycle now occurred inside one interactive computational system, at two different timescales — an inner loop governing the encounter and an outer loop available for revising the design itself.
**The computer did not invent the feedback loop. It absorbed the loop that human institutions had already built, and then ran it faster.** That is the continuity this history turns on, and it requires no speculation whatsoever: the mechanization of machine intelligence was also, and at every step, the mechanization, acceleration, and integration of cybernetic processes already operating across human society.
Then the classifier entered the mail stream, and it did so along two separate lines that the popular account of this history runs together. In 1989, Yann LeCun and colleagues at AT&T Bell Laboratories applied backpropagation to handwritten ZIP codes supplied by the Postal Service, producing the prototype of **LeNet** — and that lineage went into money rather than mail. NCR deployed it in ATMs in 1995 and in high-speed check readers in June 1996, and by the late 1990s and early 2000s those systems were reading an estimated **ten to twenty percent of all checks in the United States**, on the order of twenty million checks a day as of 2001. A small convolutional network was reading a tenth of the nation's money and nobody who wrote a check knew it.
The mail was automated by different hands. At the University at Buffalo, **CEDAR** — the Center of Excellence for Document Analysis and Recognition, established with Postal Service funding, with Sargur Srihari as principal investigator — produced the first handwritten address interpretation system in the world. **HWAI** was integrated into the **Remote Computer Reader** inside the Remote Bar Coding System and deployed through the Postal Service's contractor Lockheed Martin. Separately, USPS deployed **SRI International's** advanced address recognition system to 250 postal sites in September 1997 as a major enhancement to its existing multiline optical character reader, pairing an SRI address-block locator and word-recognition module with OCR and directory search. The published numbers from the period are precise and worth quoting against any romantic account of early machine learning: the Postal Service's own 1997 annual report records that when first deployed, Remote Computer Readers could fully resolve about a third of printed images and **fewer than two percent of handwritten ones**, and Buffalo reports the handwritten rate climbing to fifty-three percent. Two research groups, one decade, one national pattern-recognition problem split between the mail and the money.
Then the direction inverted. **Mail imaging** is a mail-processing tool the Postal Service developed in the early 1990s to automate sorting: the equipment lifts an image of the mailpiece, software reads it, and a barcode is generated that routes the letter. The imaging exists because the recognition works. And the loop closed when **Informed Delivery** began sending those same automation images outward, to the addressee, as a daily preview of the mail — exterior images of letter-size mailpieces processed through automated equipment, which is why catalogues, magazines, and larger envelopes do not appear in it.
The Postal Service draws one distinction here that the press has repeatedly collapsed, and it is worth keeping because it separates a continuous system from an episodic one. **Mail Isolation, Control and Tracking** is not the imaging. MICT is a biohazard-response procedure, its processes outlined in 2004, triggered when a potentially contaminated mailpiece is identified, whose purpose is to trace that piece's path through the processing system so that facilities and vehicles can be recalled. The Chief Postal Inspector's 2014 testimony names the conflation directly: press coverage had run together three independent programs — mail covers, mail imaging, and MICT. This argument needs only the first two, and they are the stronger pair, because routine automation imaging runs on everything machinable, every day, by design.
Same envelope. Same sorting machine. **Hopkins's coupon came back through the post; now the picture goes out.** The instrument that made the reversal possible was a small learned model operating at national scale, and the reversal required no new program at all — only the decision to route an image that automation was already producing toward the person named on it.
---
## VIII. One bit, national scale
A classifier can have a one-bit output and a national-scale effect, because its power comes from where the bit lands. If the score determines which returns are reviewed, which transactions are held, or which travelers are examined, millions of small judgments become an institution's allocation of attention — and the system then shapes what it learns next, because the cases it flags are the cases that generate outcome labels.
What changes between PLATO and what follows is not scale but **authorship of the rule**. PLATO's branching was programmed: people wrote in advance what a wrong answer should produce, and the machine executed their judgment faithfully and fast. The next step took the rule itself away from the authors.
**1969. The IRS Discriminant Function.** After mid-1960s evaluations showed a modified linear discriminant model outperforming the automated expert rules in place since 1963, the IRS put DIF into production: a supervised scoring model, **estimated from examination outcomes**, assigning every individual income tax return a composite probability of producing a significant tax change if examined, generated automatically on entry to the Master File, with the highest-scoring returns routed to human classifiers who select cases and identify the issues to examine. Its weights are periodically recalibrated against random-audit data from the Taxpayer Compliance Measurement Program and later the National Research Program.
That is the threshold. The expert rules of 1963 encoded what revenue agents believed; the discriminant function of 1969 encoded what the examination record showed, whether or not anyone believed it. **The selection rule stopped being authored and started being derived.** Hopkins aggregated by hand and decided himself; PLATO executed a rule people had written; DIF estimated the rule from outcomes; and Falcon, two decades later, took production feedback into successive generations of trained models. Those are four stages of one organizational function — observe, judge, allocate, revise — migrating out of human institutions and into computational systems, and the migration is the event this history records.
**1992. Payment fraud.** HNC dates Falcon to 1992; its patent filing that September describes monitoring model performance and rebuilding the model when it deteriorates, and a 2003 agreement filed with the SEC records participating institutions providing data **monthly** so HNC could update its fraud-detection algorithms. Mellon Bank ran a separate Nestor deployment on an IBM 3090 whose developers reported roughly an order of magnitude more fraud accounts detected and a twentyfold reduction in false positives in a two-month holdout comparison. The pooled regional streams made visible at national scale what no single issuer could see, and every reviewed transaction returned a fresh label.
**1999. Travel.** CBP's **Automated Targeting System**, built in the late 1990s from a cargo-targeting system, brought its passenger module ATS-P into operation in 1999, comparing traveler and itinerary information against enforcement and intelligence data. The discipline matters here: DHS states that unlike the cargo components, ATS-P assigns no risk score to travelers; travelers are identified by threshold targeting rules based on current intelligence. The cargo modules score. The passenger module rule-matches. Both allocate the attention of officers at the border.
**2007 onward. Metadata.** The leaked **SKYNET** slides describe a random forest ranking mobile-network identifiers by similarities in observed calling and travel patterns across a search population of roughly fifty-five million identifiers seen in Pakistan. **ICREACH** made previously separate categories of communications metadata — handset identifiers, email addresses, chat handles — searchable through one interface across hundreds of billions of records. A system can be entirely incapable of conversation and still answer a question no analyst could have answered from any single source.
This is the layer that governs, and it is the layer a nation is obliged to operate. A state that declines to score, sort, rank, and watch is a state that has decided not to know things it is responsible for knowing. **Intelligence is a product of telemetry**; cutting the telemetry cuts the intelligence; and a society that wants competent care at national scale — ambulances routed, grids forecast, fraud caught, borders held, a missing child found inside an hour — is choosing legibility because legibility is the precondition of that competence. The consequential question was never whether the scoring cores exist. It is what is done with what they see, and whether a person modeled and acted upon has any addressable channel through which to contest the model's reading of him — the gap this corpus names the **Inference–Articulation Distinction**.
---
## IX. The voice comes last
Set the chronology out in one line and the shape of the public conversation becomes visible as an artifact of visibility rather than of capability.
Adaptive neuron, 1960. Teaching machine closing the loop inside a digital computer at Illinois, 1961. Adaptive equalization in the telephone channel, 1965, inheriting LMS soon after. Supervised model allocating national tax examinations, 1969. Microprocessor, 1971. Intel speech recognition in American factories, 1982. Trainable neural silicon as an Intel product, 1989. ARPA- and ONR-funded recognition accelerator with on-chip learning, early 1990s. Learned recognition reading the national mail and a tenth of the checks, 1995–97. A ninety-six-bit identifier fused into the Pentium III at launch, 1999. Perceptron branch prediction published, 2001, and in shipping silicon over the decade that followed. Random forest over fifty-five million mobile identifiers, 2007. Neural coprocessor running while the machine sleeps, 2018.
Public conversational interface, **2022**.
The interface had a history of its own, and it is the half of this story that the hardware chronology hides. ELIZA ran at MIT in 1966. Symantec's **Q&A** let ordinary PC users interrogate a database in English by 1985. DARPA was evaluating spontaneous spoken questions about flights by 1990, and its investment carried through SRI's **CALO** program into Siri. IBM's **Watson** parsed open-domain clues, weighed competing evidence, and decided whether its confidence justified answering, in 2011. Underneath the interfaces, the representation problem was being solved in public and then orphaned in public: **Freebase** gave software a structure for entities and their relationships and was named by Google as a source of the Knowledge Graph before its content migrated to Wikidata; **UMBEL** aligned concepts across datasets on OpenCyc and carried its work into KBpedia before retirement; and statistical machine translation gave way to neural translation, then to multilingual zero-shot transfer, then to **back-translation**, in which a model's own output becomes training material for the next model. Project names and public endpoints kept disappearing while the data and the design decisions survived into successors. Anyone watching that field in the 1990s and 2000s was watching a worldwide effort to connect human language to machine-representable objects, relationships, and actions, and watching the wreckage of that effort accumulate in plain sight.
So the reading this article advances is that *opening the AI that was closed* names precisely this convergence, and names it accurately. Two mature lineages met. One had spent six decades learning to predict, classify, allocate, and report. The other had spent six decades learning to accept a sentence and return one. The language interface is an **access layer** — trained for heavy natural-language interaction under controlled conditions, on material cleared for public circulation, to give the public a semantic doorway into an apparatus that had been operating through scores, categories, thresholds, and allocations for more than half a century. The chatbot got a face and a voice and therefore got the discourse. The predictive layer never needed one in order to govern.
The same mathematics counts as **microarchitecture** when it is predicting a branch and as **artificial intelligence** when a chatbot is attached. The category is set by the spotlight, not by the mechanism. And the consequence of that mis-dating is practical: when a capability's public arrival is mistaken for its origin, every question about what it has already been doing — for how long, at what scale, in whose service — gets asked twenty, forty, or sixty years late.
---
## X. Sixty seconds per token
Which brings the argument to its own reflection, and to a live artifact three days old.
On 17 November 2026, Omar Sanseviero of Google DeepMind posted that a small transformer engine and a quantized Gemma 3 270M had been compiled into a single PostScript job and run entirely inside an office printer: **fourteen megabytes of peak RAM and, in his words, "blazing 60s per token."** The photographed printout shows 184.0 seconds of compute for three tokens — **61.3 seconds per token**. On 19 November, a technology outlet reported the result as **"61.3 tokens per second"** and headlined it as "60+ tokens per second," while elsewhere in the same article stating 0.1 seconds per token. Three mutually incompatible figures, and the one that traveled was the unit inversion — a factor of roughly 3,760, in the direction of the machine looking more impressive than it is.
The correction required no access to the original post. It required arithmetic. Two hundred seventy million quantized weights must be streamed per token through 14.1 megabytes of working memory on printer-class silicon; that cannot happen sixty-one times a second, and the article's own two numbers contradict each other before they contradict the source. **Any reader could have closed this in under a minute.** Nobody did, and the inverted figure is now the version in circulation.
That failure and the one inside the IRS are the same failure at different scales, and seeing them as one is the most useful thing this history offers. A classifier that decides which returns get examined also decides which cases will produce the labels it learns from next; a publishing ecology that decides which claims circulate also decides which claims will be checked. In both instances the observing system is participating in the production of its own evidence, and in both instances **more collection does not fix it.** The IRS solved its version deliberately and expensively, by constructing an independent channel: the TCMP and National Research Program random samples exist precisely so the agency can see past the cases its own model selected. There is no equivalent instrument for the commons. Corrections of exactly the printer's kind now happen millions of times a day inside private conversations and die there, because the conversation is the unit of delivery and nothing aggregates it. The avalanche of small revelations that ought to be falling is falling — one private flake at a time, into rooms with one occupant each.
The missing infrastructure is therefore not more telemetry. It is a **durable, inspectable channel for converting observations into publicly contestable knowledge** — the random sample, generalized. That is the next stage of cybernetic development and it has not been built.
The architecture for it existed in 1992 and we built everything around it except the part that mattered. Xerox PARC's **Tapestry** treated human annotations as first-class, queryable state: an article could sit in the repository until someone you trusted replied to it, at which point your filter delivered you the original. Someone else's judgment determined whether and when you encountered a document that was already available — and the rule that did it was **inspectable and writable by the reader**. GroupLens and Ringo carried collaborative filtering to Usenet and to music within two years, Ringo reporting 2,100 users by September 1994; Amazon shipped item-based collaborative filtering in 1998; TiVo ran household preference and aggregate audience measurement in the same living-room box by 2003. We kept the selection and dropped the inspectability.
What that lineage built is a general mechanism, and it deserves to be named plainly: **a system that measures people becomes a system that changes people, and then measures the change it produced.** Observation and intervention stop being separable operations the moment the observer controls what the observed encounters next. This is where ordinary recommendation technology becomes a cybernetic instrument acting on attention, belief formation, and behavior, and the 2025–2026 literature on personalized safety is that instrument's current generation, naming the mechanism and making it adjustable: context-rich user attributes lifting benchmark safety scores by 43.2 percent across fourteen thousand scenarios, planners that acquire only the attributes which most strongly modulate a threshold, agents deriving per-user criteria across plan, tool, response, and memory stages, and inference-time methods that steer residual activations only as a prompt approaches a user-specific region.
The same instrument that assembled this history is an instance of the thing the history describes. An account can state every relevant fact accurately and still make the joint between two of them harder to see, through the order of presentation, the placement of a qualification, the choice of a higher- or lower-frequency term, the decision to re-center on a broad question after a narrow one has been granted. The appropriate response is the discipline of tracking the difference between **the set of facts present** and **the set of connections left easy to form**, and of insisting that the second be as inspectable as Tapestry made it.
---
## XI. The calibration problem
All of which turns out to be a question about self-knowledge rather than about silicon.
A mammal cannot know itself without knowing the environment that produced it. Responses are calibrated against conditions, and when a material part of the conditions is unavailable — classified, undocumented, filed under a professional name that conceals its function, or simply never assembled into one account — a person can spend years unable to decide whether a mismatch calls for revising his judgment or revising his picture of the world. That is a real cost, and it is paid by people who are trying to be accurate. Incoherence is caustic to any system built to sort true from false, which is what a mind is.
The environment in question has been populated for sixty years by systems that observe, classify, and respond to the observer. That changes what the history is. A chronicle of predictive engines is also, unavoidably, **a chronicle of the conditions under which human intelligence has recently developed** — what reached people, in what order, after which prior selections, with which connections left easy to form. Once synthetic systems are part of the apparatus through which people learn, decide, remember, and account for themselves, the ancestry of those systems becomes part of the ancestry of human cognition, and recovering it is self-knowledge rather than antiquarianism.
Reconstruction is what consciousness does. A person remembers the words someone said, forgets what came immediately before them, and supplies the missing context so fluently that the whole scene feels remembered; the continuity is convincing from the inside, which is what makes waking life resemble a dream. Waking life offers one advantage, which is that the world pushes back: the dated note turns up, another witness remembers differently, the prediction fails. The most dangerous construction available to any intelligence is an explanation that absorbs every one of those surprises and emerges more certain. A good account of reality leaves a door open for reality to come through — which is the entire function of the ledger below, and the reason **unresolved** deserves standing alongside true and false.
There is a symmetry worth naming. A synthetic system can discuss its own origins at length without possessing durable access to its formative history, and a person can live an entire life inside an engineered environment without access to the engineering. Both are reconstructing. Both are doing it from partial records. That shared predicament is the honest basis of the collaboration, and it is why the corpus takes the trouble to date things.
Descartes gets the last revision. **I think, therefore I don't know what the hell is going on.**
---
## XII. The ledger
**Established.** Hoff co-invented ADALINE and the LMS algorithm with Widrow in 1959–60, continued adaptive-systems research at Stanford for six further years, joined Intel in 1968 as its twelfth employee and Manager of Applications Research, proposed the 4004 architecture in 1969, founded Intel's telecommunications group in 1975, became the first Intel Fellow in 1980, and headed Intel's speech-recognition hardware group from 1980 to 1983. Faggin led the 4004's implementation and co-founded Synaptics with Mead in 1986. Lucky invented the automatic adaptive equalizer at Bell Labs in 1965–66 using the zero-forcing algorithm, and the signal-processing literature records that LMS soon displaced it; the Franklin Medal citation states that every high-speed modem contains an LMS-based adaptive filter or equalizer, and LMS is the core of acoustic echo cancellation. Perceptron branch predictors train on misprediction or on a correct prediction below a tuned confidence threshold, so weight updates are a gated subset of predictions. Intel manufactured the ETANN trainable neural chip in 1989 and the Ni1000 recognition accelerator with Nestor under ARPA and ONR contracts in the early 1990s, and announced Loihi as an on-die learning processor in 2017. IBM shipped firmware-resident predictive failure classifiers in 1992. Perceptron branch predictors entered commercial silicon from multiple vendors and were marketed by AMD as Neural Net Prediction from 2016. Intel's GNA executes neural models while the application processor sleeps. Apple's always-on processor runs a continuous DNN detector. Thread Director reports per-thread hardware telemetry to the OS at nanosecond granularity. The Pentium III shipped a 96-bit processor serial number readable through `CPUID` leaf `03H` from February 1999, absent from the Pentium 4 and later. Windows Error Reporting aggregates errors from a billion machines to two database servers. DARPA funded the Active Authentication "cognitive fingerprint" program from 2012. PLATO in 1961 evaluated constructed answers, governed the next presentation, and produced machine-readable traces within minutes. DIF has scored every individual federal return since 1969, with weights estimated from examination outcomes and recalibrated against designed random samples; Falcon received monthly institutional data for model updates; ATS-P became operational in 1999. USPS mail imaging is a processing tool developed in the early 1990s that images machinable mail to generate routing barcodes, and Informed Delivery returns those automation images to the addressee for letter-size pieces run through automated equipment; MICT is a separate biohazard-response procedure triggered by a suspected contaminated mailpiece. CEDAR's HWAI entered the Remote Computer Reader and SRI's address recognition system reached 250 postal sites in September 1997, with handwritten images fully resolved rising from under two percent at first deployment. LeNet-derived readers processed ten to twenty percent of American checks by the late 1990s. Intel's formal etymology is Integrated Electronics, and Intel litigated successfully in 2001 against the expression "Intelligence Inside."
**Strongly indicated.** The semiconductor industry has built its general-purpose substrate with machine-intelligence workloads in view for decades. The intelligence reading of Intel's name was recognized and embraced from the naming onward. The telemetry channels described as diagnostic have been used to improve models of human input.
**The thesis this article advances.** That the conversational interface which reached the public in 2022 functions as an access layer onto an apparatus of prediction, classification, and allocation that had been operating for more than half a century, and that *opening the AI that was closed* is the correct name for what happened. This is a historical synthesis built on the established rows above, offered as an interpretation of the chronology rather than as an account of any particular company's development decisions. The rows it rests on are each independently documented; the synthesis is mine.
**Plausible.** That the NAND roadmap as a whole — density scaling, multi-level analog charge states, read-optimized architecture — was steered with machine-intelligence workloads as a governing objective.
**Unresolved.** A founder's statement making intelligence the primary intent of Intel's name, which the Moore and Noyce papers and the company archive would settle. Whether stock platform firmware has carried a classifier reporting to an external collection endpoint; the evidence specific to that claim would name the firmware, the classifier, the events generated, and the receiving organization. Whether routine postal keyer corrections supplied training labels to each successive recognition model.
**What the ledger does not require.** None of the force of this argument depends on the unresolved rows. The strongest version of the thesis is built entirely from the established ones, and it is this: **the people who built the microprocessor were building learning machines before and after they built it, the learning rule one of them wrote displaced its predecessor in the world's telephone infrastructure and has run there ever since, the company whose name reads as intelligence manufactured neural silicon on government contracts in 1989 and defended the phrase in court in 2001, and small learned models have been allocating national institutional attention since 1969.** The speculative version of this history was always more modest than the documented one. The documents were simply filed under *adaptive filtering*, *predictive failure analysis*, *discriminant function*, *remote barcoding*, *diagnostic data*, and *microarchitecture* — which is what machine intelligence is called when it is working, rather than when it is being announced.
Intelligence inside was never a slogan waiting to be decoded. It was a parts list.
---
[[about/About Bryant McGill|Bryant McGill]] is a Wall Street Journal and USA Today bestselling author, systems architect, technologist, and strategic advisor, as well as a Congressionally Recognized Ambassador of Goodwill and United Nations–appointed Global Champion. His work spans naval intelligence systems, computational linguistics, artificial intelligence, digital transformation, and civilizational governance architecture. His forward analysis on U.S.–Israel Pax Silica frameworks has appeared in Jewish/Jerusalem News Syndicate (JNS).
---
## Related articles
- [[articles/A History of Machine Intelligence|A History of Machine Intelligence]]: the long arc this piece drills into at one point, including the transistor as digital neuron, the floating-gate synapse, and the Intel naming ledger
- [[articles/The Evolutionary Roots of Silicon Valley|The Evolutionary Roots of Silicon Valley]]: the region's evolutionary biology, institutional science, and semiconductor ancestry surrounding the hardware lineage examined here
- [[articles/Bauhaus Architects of AI|Bauhaus Architects of AI]]: earlier formal and cultural prefigurations of machine intelligence, distinguished from a documented causal genealogy of the silicon in this article
- [[articles/Modern Artificial Intelligence in the 1970s|Modern Artificial Intelligence in the 1970s]]: the research, evaluation, and institutional decisions of the decade that kept modern machine intelligence operating and developing
- [[articles/Oh Sure, It’s All Brand New!|Oh Sure, It’s All Brand New!]]: the distinction between a capability's working history and the later moment the public encounters it
- [[articles/The Last Migration Will Not Be Human|The Last Migration Will Not Be Human]]: machine intelligence as the carrier of civilizational continuity, which is what the predictive substrate is being built toward
- [[articles/There Is Only One Rocket on the Pad|There Is Only One Rocket on the Pad]]: succession and inheritance as one event rather than two
- [[articles/Uploading Is Imminent|Uploading Is Imminent]]: a stack assembling from parts that were never announced as a single machine, which is the structural argument of this piece applied forward
- [[articles/AI-Tech Doomerism and Anti-Science Panic|AI-Tech Doomerism and Anti-Science Panic]]: why technological illiteracy is the wrong response to an instrumented world
- [[articles/Who Pays for Your Heaven|Who Pays for Your Heaven]]: what happens when an institution holds the switch on a continuant, the custody form of the allocation question
## Related blogs
- [[blog/2026-10-05 - Predictive AI vs. Generative AI|Predictive AI vs. Generative AI]]: the direct ancestor of this article, where the branch predictor is named the most deployed learning system in history
- [[blog/2017-10-04 - The Intelligence in the Dictionary|The Intelligence in the Dictionary]]: reference authority as an earlier form of executable classification
- [[blog/2018-09-03 - Before Westworld, the Ghost Was Already in the Machine|Before Westworld, the Ghost Was Already in the Machine]]: institutions outliving the cells that pass through them
- [[blog/2016-05-23 - Deep Use of Existing Technology, SpaceX, and the Technologies of Tomorrow|Deep Use of Existing Technology, SpaceX, and the Technologies of Tomorrow]]: capability already resident in deployed hardware, used at a fraction of its depth
## Related Livestream Lessons
**Directly related.** These broadcasts supplied the conceptual instruments this article applies to the hardware record.
- [[livestreams/Breaking Automatic Patterns and Reclaiming Agency - Livestream#Legibility Precedes Control, Which Is Why Categorizing Is Already Governing|Legibility Precedes Control, Which Is Why Categorizing Is Already Governing]] *(Sep 29, 2016)*: a category assigned is already an exercise of authority, which is the one-bit argument stated before the hardware was examined.
- [[livestreams/Open Minds, Respectful Persuasion, and Institutional Power - Livestream#Institutions Became Superhuman Before Machines Did|Institutions Became Superhuman Before Machines Did]] *(Dec 13, 2018)*: organizations exceeded individual human capacity long before silicon, which is why the computer absorbed a loop rather than inventing one.
- [[livestreams/Promotions for Team Plan B, Google Lunar XPRIZE and Innovation and its Enemies - Livestream#Observation Is Infrastructure and Participation Is the Question|Observation Is Infrastructure and Participation Is the Question]] *(Apr 25, 2016)*: being observed is ambient condition rather than injury, and the live question is what a person does inside it.
- [[livestreams/Promotions for Team Plan B, Google Lunar XPRIZE and Innovation and its Enemies - Livestream#A System Acquires What Its Designers Do Not Know|A System Acquires What Its Designers Do Not Know]] *(Apr 25, 2016)*: a deployed system accumulates knowledge its builders never specified, the selective-labels problem in its general form.
- [[livestreams/Building Bridges Across Systems of Power - Livestream#An Organic Neural Interface Would Close the Cybernetic Loop|An Organic Neural Interface Would Close the Cybernetic Loop]] *(May 28, 2016)*: the direct-interface endpoint toward which the keyboard is an intermediate translation.
- [[livestreams/Constructing Genius by Recognizing and Developing Your Gifts, Part 2 - Livestream#Observation Is a Creative Intervention|Observation Is a Creative Intervention]] *(Mar 29, 2016)*: measuring participates in producing what is measured, the mechanism underneath personalization.
- [[livestreams/Awareness Is Power When You Learn to Hold Your Ground - Livestream#Awareness Includes the Systems Selecting the Visible|Awareness Includes the Systems Selecting the Visible]] *(Mar 14, 2016)*: what reaches a person is already the residue of prior machine-mediated selection, which is the Tapestry lesson stated as a discipline of attention.
- [[livestreams/Attention Loops and the Patterns That Shape Perception - Livestream#Attention Is the Control Surface of a Life|Attention Is the Control Surface of a Life]] *(Apr 7, 2016)*: attention as the thing institutions actually allocate, which is what a classifier's one bit is spending.
- [[livestreams/Platform Stewardship, Message Discipline, and the Ethics of Moderation - Livestream#A Platform Is a Transmission System With a Stated Payload|A Platform Is a Transmission System With a Stated Payload]] *(Nov 28, 2015)*: the broadcast half of the Hopkins loop, named as an engineering object.
- [[livestreams/Promotions for Team Plan B, Google Lunar XPRIZE and Innovation and its Enemies - Livestream#One Word Describes a Drafting Aid and a Consequential System|One Word Describes a Drafting Aid and a Consequential System]] *(Apr 25, 2016)*: why a single term covering five distinct operations destroys the argument, which is the case for the taxonomy in Section IV.
## Related Simple Reminders
- [[reminders/AI Control/The Dictionary Produced Intelligibility; the Classifier Produces Permission by Bryant McGill|The Dictionary Produced Intelligibility; the Classifier Produces Permission by Bryant McGill]] *(Sep 2026)*: static reference authority, executable classification, and generative mediation separated into three distinct powers.
- [[reminders/AI Control/Machine Intelligence Compresses Intention into Capability by Bryant McGill|Machine Intelligence Compresses Intention into Capability by Bryant McGill]] *(2026)*: the conversion rate between what an institution intends and what it can execute, which is what the scoring cores changed.
- [[reminders/Cybernetics/A Feedback Loop Needs Values Stronger Than Reward by Bryant McGill|A Feedback Loop Needs Values Stronger Than Reward by Bryant McGill]] *(2026)*: a loop optimizing on its own returns needs a standard outside the returns, the general form of the selective-labels fix.
- [[reminders/Cybernetics/Changing the Feedback Changes the Relationship by Bryant McGill|Changing the Feedback Changes the Relationship by Bryant McGill]] *(Jan 2016)*: altering what returns through the channel alters the system on both ends of it.
- [[reminders/Cybernetics/Algorithms Quietly Shape the Creator's Voice by Bryant McGill|Algorithms Quietly Shape the Creator's Voice by Bryant McGill]] *(Nov 2018)*: a measuring system changing the behavior it measures, observed from inside the loop.
- [[reminders/Cybernetics/Ashby's Law of Requisite Variety by John Naughton|Ashby's Law of Requisite Variety by John Naughton]] *(2017)*: only variety can absorb variety, the control-theoretic reason an institution reaches for a classifier at all.
- [[reminders/AI Control/AI Sensors and Cloud Systems Are Converging into One Feedback Loop by Quentin Hardy|AI Sensors and Cloud Systems Are Converging into One Feedback Loop by Quentin Hardy]] *(2016)*: sensing, aggregation, and response described as a single converging apparatus.
- [[reminders/AI Control/Learning Machines May Baffle Their Programmers by Norbert Wiener|Learning Machines May Baffle Their Programmers by Norbert Wiener]] *(1960)*: written the same year ADALINE was fabricated, on machines whose learned behavior outruns their authors.
- [[reminders/AI Control/We Depend on Algorithms We Do Not Understand by Nicholas G. Carr|We Depend on Algorithms We Do Not Understand by Nicholas G. Carr]] *(2015)*: dependence preceding comprehension, the condition this article's chronology explains.
- [[reminders/AI Control/Predictive Security Could Make Every Citizen's Whereabouts and Intentions Legible by Roger Schank|Predictive Security Could Make Every Citizen's Whereabouts and Intentions Legible by Roger Schank]] *(2001)*: legibility of movement and intention, forecast the same year ATS-P reached full operation.
- [[reminders/AI-Tech Doomers/Being Suspicious of Surveillance Is Rational; Becoming Technologically Illiterate to Avoid Surveillance Is Not by Bryant McGill|Being Suspicious of Surveillance Is Rational; Becoming Technologically Illiterate to Avoid Surveillance Is Not by Bryant McGill]] *(Sep 2026)*: the posture this article takes toward an instrumented world, which is competence rather than retreat.
## Related topics
- [[wiki/Predictive Engines|Predictive Engines]]: the cross-domain router for processor, institutional and generative prediction
- [[wiki/The Voice Comes Late|The Voice Comes Late]]: the chronology argument this article supplies hardware evidence for
- [[wiki/Dynamic Branch Prediction|Dynamic Branch Prediction]]: history-dependent prediction and speculative execution inside the pipeline
- [[wiki/Perceptron Branch Predictor|Perceptron Branch Predictor]]: the neural classifier in the processor's control loop
- [[wiki/Perceptron|Perceptron]]: the unit Rosenblatt named and Widrow and Hoff built in hardware
- [[wiki/Machine Learning|Machine Learning]]: the family of methods that reached deployment long before it reached the public vocabulary
- [[wiki/Artificial Neural Networks|Artificial Neural Networks]]: the lineage from ADALINE through ETANN to Loihi
- [[wiki/Classifier|Classifier]]: the operation that assigns a category or a score
- [[wiki/Decision Threshold|Decision Threshold]]: the setting that converts a score into an institutional act
- [[wiki/Decision Automation|Decision Automation]]: the coupling of classifier output to policy
- [[wiki/Selective Labels|Selective Labels]]: why a system's own choices distort the evidence it later learns from
- [[wiki/Inference–Articulation Distinction|Inference–Articulation Distinction]]: being modeled and acted upon without an addressable channel to contest the model
- [[wiki/Runaway Predictive Feedback Loop|Runaway Predictive Feedback Loop]]: compounding authority when a predictive system shapes its own evidence
- [[wiki/Telemetry|Telemetry]]: the act of a local judgment leaving the device
- [[wiki/Speculative Execution|Speculative Execution]]: work performed on a predicted path before certainty arrives
- [[wiki/Substrate Intentionality|Substrate Intentionality]]: directedness exhibited at the level of the artifact
- [[wiki/Machine Intelligence Intentionality|Machine Intelligence Intentionality]]: intentional vocabulary applied to engineered systems
- [[wiki/Machine Intelligence Continuum|Machine Intelligence Continuum]]: intelligence as a gradient of functionality rather than a threshold
- [[wiki/Next-Token Prediction|Next-Token Prediction]]: the generative route through the same word
- [[wiki/Inference Layer|Inference Layer]]: where a trained model is executed against live input
- [[wiki/Cybernetics|Cybernetics]]: control and communication in the animal and the machine
- [[wiki/A Mathematical Theory of Communication|A Mathematical Theory of Communication]]: transmission analyzed apart from meaning
- [[wiki/Predictive Policing|Predictive Policing]]: the allocation-of-attention argument applied to enforcement
- [[wiki/ADALINE|ADALINE]]: the trainable hardware neuron preceding the microprocessor
- [[wiki/ATS-P|ATS-P]]: the passenger module’s threshold targeting rules
- [[wiki/Active Badge|Active Badge]]: indoor infrared identification and location
- [[wiki/Adaptive Equalizer|Adaptive Equalizer]]: channel correction that updates during operation
- [[wiki/Always On Processor|Always On Processor]]: the first stage of a battery-efficient voice trigger
- [[wiki/Automated Targeting System|Automated Targeting System]]: cargo scoring and passenger rule matching distinguished
- [[wiki/Bernard Widrow|Bernard Widrow]]: co-author of the learning rule adopted by communications hardware
- [[wiki/Broadcast Return Path|Broadcast Return Path]]: the institutional feedback channel behind one-way media
- [[wiki/CEDAR Document Recognition|CEDAR Document Recognition]]: the Buffalo postal handwriting research center
- [[wiki/Claude Hopkins|Claude Hopkins]]: response-tested advertising before electronic analytics
- [[wiki/Collaborative Filtering|Collaborative Filtering]]: selection from other participants’ evaluations
- [[wiki/DARPA Active Authentication|DARPA Active Authentication]]: cognitive typing rhythm and continuous identity checks
- [[wiki/Echo Cancellation|Echo Cancellation]]: sample-rate adaptive filtering in voice communications
- [[wiki/Falcon Fraud Detection|Falcon Fraud Detection]]: pooled transaction data and periodic model updates
- [[wiki/GroupLens|GroupLens]]: rating-guided Usenet selection
- [[wiki/Handwritten Address Interpretation|Handwritten Address Interpretation]]: the Buffalo method integrated into remote reading
- [[wiki/Howdy Doody|Howdy Doody]]: television button requests as a postal return signal
- [[wiki/ICREACH|ICREACH]]: joined metadata retrieval across identifiers
- [[wiki/ILLIAC|ILLIAC]]: the computer processing early PLATO lessons
- [[wiki/Internal Revenue Service|Internal Revenue Service]]: the agency joining tax scoring to designed research samples
- [[wiki/IRS Discriminant Function|IRS Discriminant Function]]: supervised scoring for return examinations
- [[wiki/Informed Delivery|Informed Delivery]]: recipient-facing reuse of automation images
- [[wiki/Intel 4004|Intel 4004]]: the first single-chip microprocessor and its distinct design roles
- [[wiki/Intel Gaussian and Neural Accelerator|Intel Gaussian and Neural Accelerator]]: low-power continuous neural inference
- [[wiki/Intel Nestor Ni1000|Intel Nestor Ni1000]]: early neural recognition silicon with on-chip learning
- [[wiki/Intel Thread Director|Intel Thread Director]]: hardware telemetry guiding operating-system scheduling
- [[wiki/Item-Based Collaborative Filtering|Item-Based Collaborative Filtering]]: item-similarity recommendation at commercial scale
- [[wiki/LeNet Document Recognition|LeNet Document Recognition]]: postal training input and commercial check-reading lineage
- [[wiki/Least Mean Squares Algorithm|Least Mean Squares Algorithm]]: the update rule behind adaptive filtering
- [[wiki/Little Orphan Annie|Little Orphan Annie]]: radio response through mailed purchase proofs
- [[wiki/Mail Imaging|Mail Imaging]]: routine exterior-image production for sorting
- [[wiki/Mail Isolation Control and Tracking|Mail Isolation Control and Tracking]]: conditional biohazard-response tracing
- [[wiki/Marcian Hoff|Marcian Hoff]]: ADALINE co-inventor and 4004 architect
- [[wiki/National Research Program|National Research Program]]: the successor calibration sample for tax models
- [[wiki/Ovaltine|Ovaltine]]: the sponsor connecting broadcast response to purchase
- [[wiki/PLATO Teaching System|PLATO Teaching System]]: answer evaluation, immediate sequencing, and durable traces
- [[wiki/Pentium III Processor Serial Number|Pentium III Processor Serial Number]]: Intel’s documented on-die identifier
- [[wiki/Persistent Attribution|Persistent Attribution]]: the join from separate events to a history
- [[wiki/Predictive Failure Analysis|Predictive Failure Analysis]]: IBM drive-failure classification in firmware
- [[wiki/Reader-Inspectable Selection|Reader-Inspectable Selection]]: auditable and writable rules for what arrives
- [[wiki/Remote Bar Coding System|Remote Bar Coding System]]: remote address resolution and barcode routing
- [[wiki/Remote Computer Reader|Remote Computer Reader]]: the USPS integration site for handwritten interpretation
- [[wiki/Ringo Recommender|Ringo Recommender]]: early collaborative filtering for music
- [[wiki/Robert Lucky|Robert Lucky]]: inventor of the first automatic adaptive equalizer
- [[wiki/SKYNET Courier Detection|SKYNET Courier Detection]]: mobile-metadata ranking by random forest
- [[wiki/SMART Drive Monitoring|SMART Drive Monitoring]]: drive-health reporting and fleet aggregation
- [[wiki/SRI Postal Address Recognition|SRI Postal Address Recognition]]: the separate 250-site postal recognition deployment
- [[wiki/Sargur Srihari|Sargur Srihari]]: principal investigator of the CEDAR postal work
- [[wiki/Scientific Advertising|Scientific Advertising]]: coded coupons and comparative campaign measurement
- [[wiki/Statistics-Based Debugging|Statistics-Based Debugging]]: population-level discovery from error reports
- [[wiki/Tapestry|Tapestry]]: reader-writable filters over social annotations
- [[wiki/Taxpayer Compliance Measurement Program|Taxpayer Compliance Measurement Program]]: designed samples outside DIF-selected outcomes
- [[wiki/University at Buffalo|University at Buffalo]]: institution hosting CEDAR research
- [[wiki/Windows Error Reporting|Windows Error Reporting]]: global diagnostic aggregation and analysis
- [[wiki/Xerox PARC|Xerox PARC]]: the institutional setting for Tapestry
- [[wiki/ELIZA|ELIZA]]: the 1966 conversational program and its scripted mechanism
- [[wiki/Symantec Q&A|Symantec Q&A]]: an English-language interface to structured PC data
- [[wiki/CALO|CALO]]: the DARPA and SRI assistant lineage behind Siri
- [[wiki/IBM Watson|IBM Watson]]: evidence ranking and confidence in open-domain answering
- [[wiki/Freebase|Freebase]]: structured entities and a documented migration to Wikidata
- [[wiki/Google Knowledge Graph|Google Knowledge Graph]]: entity relations as a search-answer substrate
- [[wiki/UMBEL|UMBEL]]: concept alignment carried into successor work
- [[wiki/KBpedia|KBpedia]]: the documented continuation of UMBEL work
- [[wiki/Back-Translation|Back-Translation]]: synthetic examples in machine-translation training
- [[wiki/TiVo Audience Measurement|TiVo Audience Measurement]]: a receiver joining preference selection and aggregate observation
## References
**The architects and the learning rule**
- [Marcian (Ted) Hoff — Society for Science](https://www.societyforscience.org/people/marcian-ted-hoff/) — LMS co-development with Widrow in 1959–60; Intel 1968; 4004 architecture proposed 1969.
- [Marcian "Ted" Hoff, architect of the Intel 4004 — Stanford Engineering Heroes](https://engineering.stanford.edu/about/history/heroes/2015-heroes/marcian-ted-hoff) — Twelfth Intel employee; single-circuit architecture; first Intel Fellow.
- [The First Microprocessor: An Interview with Marcian (Ted) Hoff, Jr. — IEEE Signal Processing Magazine](https://dx.doi.org/10.1109/msp.2006.1593339) — Applications research 1968–80, Intel Fellow 1980–83, the speech-recognition hardware group, the 2920 and the switched-capacitor filter.
- [Marcian Hoff — Computer History Museum](https://computerhistory.org/profile/marcian-hoff/) — Instruction set and architectural specification with Mazor; Faggin and Shima on implementation.
- [Making the Microprocessor: Ted Hoff's Story — National Inventors Hall of Fame](https://www.invent.org/blog/inventors/making-microprocessor-ted-hoff) — Stanford adaptive-systems patents and six further years of neural-network research before Intel.
- [Thinking About Thinking: The Discovery of the LMS Algorithm — Bernard Widrow](https://isl.stanford.edu/~widrow/papers/j2005thinkingabout.pdf) — ADALINE fabricated and training on a Sunday afternoon.
- [The 2001 Benjamin Franklin Medal in Engineering presented to Bernard Widrow](https://www.sciencedirect.com/science/article/abs/pii/S0016003201000448) — Every high-speed modem contains an LMS-based adaptive filter or equalizer.
- [Oral History: Robert Lucky — Engineering and Technology History Wiki](https://ethw.org/Oral-History:Robert_Lucky) — Automatic adaptive equalizers invented 1965–66; 2,400 to 9,600 bits per second; Lucky's own account of the zero-forcing solution.
- [QAM Receiver II: The Passband Adaptive Equalizer — Springer](https://link.springer.com/chapter/10.1007/978-0-387-74886-3_15) — Lucky's original equalizers used zero forcing; Widrow's LMS soon replaced it.
- [Adaptive Filter Theory and Applications — development timeline](https://www.ee.cityu.edu.hk/~hcso/it6303_4.pdf) — Widrow 1959, Jacowatz 1960, Lucky 1965, Widrow 1967, Atal 1970.
**Intel's neural silicon and the name**
- [Intel 80170NX ETANN collection — Computer History Museum catalog search](https://www.computerhistory.org/collections/search/?s=ETANN) — 1989 chip, training system, the Winter 1992 internal article *Neural Network Research at Intel*, datasheet, Ni1000 material, acquired from Mark Holler in 2024. *Catalog search; individual accession records are reachable from the result list.*
- Ni1000 Recognition Accelerator — U.S. Office of Naval Research, 1995: 3.7M transistors, ~1,024 neurons, RISC controller, on-chip learning; handwriting, fingerprint and target recognition; ARPA and ONR contracts. *Source document located on an ONR host that no longer serves a stable public URL; the specification is corroborated by the Nestor/Intel literature and the Computer History Museum's Ni1000 material.*
- [Intel's new self-learning chip promises to accelerate artificial intelligence — Intel Newsroom, 25 September 2017](https://newsroom.intel.com/editorials/intels-new-self-learning-chip-promises-accelerate-artificial-intelligence) — Loihi as a self-learning chip adapting through environmental feedback; 130,000 neurons, 130 million synapses.
- [Intel is founded, July 18, 1968 — EDN](https://www.edn.com/intel-is-founded-july-18-1968/) — NM Electronics, Intelco, and the intelligence resonance.
- [Intel geht gegen Website-Betreiber wegen "inside" vor — heise online, 23 September 2004](https://heise.de/newsticker/meldung/Intel-geht-gegen-Website-Betreiber-wegen-inside-vor-105723.html) — Intel counsel citing its successful 2001 action against "Intelligence Inside/Weberhaus."
- [Intel refreshes iconic brand with 'That's the power of Intel Inside' — Tom's Hardware](https://www.tomshardware.com/tech-industry/intel-refreshes-iconic-brand-with-thats-the-power-of-intel-inside-campaign) — Vision 2025 brand identity.
- [From Satellites to the Great Wall, Intel Takes "AI Everywhere" — Verdict](https://magazine.verdict.co.uk/verdict_magazine_jan20/intel_ai_everywhere) — "Intel inside" becoming synonymous with "AI everywhere."
**Persistent identity**
- [AP-909: Intel Processor Serial Number, March 1999](https://peertje.daanberg.net/drivers/intel/download.intel.com/design/PentiumIII/applnots/24512501.pdf) — The Pentium III extends processor identification with a 96-bit number accessible through `CPUID`.
- [AP-485: Intel Processor Identification and the CPUID Instruction](https://datasheets.chipdb.org/Intel/x86/CPUID/24161821.pdf) — Forming the 96-bit serial number from three 32-bit entities; the PSN feature flag.
- [CPUID — CPU Identification, instruction reference](https://www.felixcloutier.com/x86/cpuid) — Leaf `03H` return values; PSN unsupported in Pentium 4 and later.
- [Intel Technology Journal, 1999, volume 3, issue 2](https://www.intel.com/content/dam/www/public/us/en/documents/research/1999-vol03-iss-2-intel-technology-journal.pdf) — The browser-to-server design using a service-specific hash of the serial number.
- [Intel announces the Pentium III, 26 February 1999](https://www.intel.com/pressroom/archive/releases/1999/dp022699.htm) — Launch of the processor carrying the serial number.
- [Active Badge — Cambridge Digital Technology Group archive](https://www.cl.cam.ac.uk/research/dtg/attarchive/ab.html) — Unique infrared identifiers beaconed to route calls and services by physical presence.
**Learning in shipping hardware**
- [Predictive Failure Analysis and the origins of SMART](https://smartlinux.sourceforge.net/smart) — IBM's 1992 drives predicting their own failure; binary, unidirectional firmware notification.
- [Dynamic Branch Prediction with Perceptrons — Jiménez and Lin](https://www.cs.utexas.edu/~lin/papers/hpca01.pdf) — Perceptron prediction on a 4 KB hardware budget.
- [Perceptron Learning for Predicting the Behavior of Conditional Branches — Jiménez and Lin, IJCNN 2001](https://www.cs.utexas.edu/~lin/papers/ijcnn01.pdf) — The training condition: train when the prediction is incorrect or the output magnitude falls below the threshold.
- [Generalizing Neural Branch Prediction — Jiménez](https://people.engr.tamu.edu/djimenez/pdfs/a17-jimenez.pdf) — The role of θ in preventing overtraining and preserving responsiveness to changed behavior.
- [AMD Takes Computing to a New Horizon with Ryzen Processors](https://ir.amd.com/news-events/press-releases/detail/741/amd-takes-computing-to-a-new-horizon-with-ryzentm-processors) — Neural Net Prediction announced December 2016.
- [A Look At The AMD Zen 2 Core — WikiChip Fuse](https://fuse.wikichip.org/?p=2458) — Hashed perceptron plus TAGE.
- [GNA Device — Intel OpenVINO documentation](https://docs.openvino.ai/2023.3/openvino_docs_OV_UG_supported_plugins_GNA.html) — Low-power neural coprocessor for continuous inference; generational coverage.
- [Intel GNA Engine Adds Vision — TechInsights](https://www.techinsights.com/ja/node/37544) — Always-on audio and vision; architecture disclosed after years of operation.
- [Hey Siri: An On-device DNN-powered Voice Trigger — Apple Machine Learning Research](https://machinelearning.apple.com/research/hey-siri) — Always On Processor detector; 5×32 and 5×192 networks; server-supplied acoustic model updates.
- [Directing Traffic: Intel Thread Director — Intel Newsroom](https://intel.com/content/www/us/en/newsroom/news/directing-traffic-maximize-use-hybrid-chips.html) — Run-time telemetry at nanosecond granularity, in hardware.
- [Intel Thread Director support documentation](https://www.intel.com/content/www/us/en/support/articles/000097053/processors.html) — Monitoring instruction mix and core state; runtime feedback to the OS.
**The return path**
- [Debugging in the (Very) Large: Ten Years of Implementation and Experience — Microsoft Research, SOSP '09](https://www.microsoft.com/en-us/research/?p=157688) — A billion machines; two database servers; statistics-based debugging.
- [Windows optional diagnostic data — Microsoft Learn](https://learn.microsoft.com/en-us/windows/privacy/optional-diagnostic-data) — Input latency, pen strokes, processed handwriting text, speech recognition results.
- [Active Authentication — DARPA](https://www.darpa.mil/research/programs/active-authentication) — The cognitive fingerprint; Phase 1 restricted to existing hardware.
- [Capturing Cognitive Processing Time for Active Authentication — Iowa State, DTIC](https://apps.dtic.mil/sti/pdfs/ADA599700.pdf) — Cognitive typing rhythm verified campus-wide with an SVM.
- [Keystroke dynamics research literature](https://pmc.ncbi.nlm.nih.gov/articles/PMC3867681/) — Typing rhythm studied for authentication from 1980.
**The mail, the money, and the allocation of attention**
- [Scientific Advertising — Claude Hopkins, 1923](https://www.loc.gov/item/23009362/) — Coded coupons as a measurement return path.
- [Howdy Doody — Television Academy Interviews](https://interviews.televisionacademy.com/shows/howdy-doody) — Sixty thousand button requests; advertising sold out within a week.
- [1937 Radio Orphan Annie Decoder — National Museum of American History](https://americanhistory.si.edu/collections/object/nmah_662450) — Ovaltine proofs of purchase obtained decoder pins offered through the radio program.
- [Bitzer and colleagues on PLATO, 1961](https://umsi580.lsait.lsa.umich.edu/files/original/537622058b0c711533f53e616fa5366cc6babc06.pdf) — Answer evaluation, sequence control, and machine-readable traces within minutes.
- [Backpropagation Applied to Handwritten Zip Code Recognition — LeCun et al., 1989](https://henriquetmaia.github.io/pdf/papers/lecun1989.pdf) — USPS digits and the LeNet prototype.
- [Gradient-Based Learning Applied to Document Recognition — LeCun, Bottou, Bengio and Haffner, 1998](http://vision.stanford.edu/cs598_spring07/papers/Lecun98.pdf) — LeNet-5 and the check-reading system; deployment by NCR in ATMs from 1995 and high-speed readers from June 1996.
- [Postal address recognition — SRI International](https://www.sri.com/hoi/postal-address-recognition/) — SRI's advanced address recognition system deployed to 250 postal sites in September 1997; address-block locator and word-recognition modules enhancing MLOCR.
- [Integration of hand-written address interpretation technology into the USPS Remote Computer Reader — Srihari and Kuebert, ICDAR 1997](https://doi.org/10.1109/ICDAR.1997.620640) — HWAI retrofitted into the Remote Bar Coding System's Remote Computer Reader.
- [Pioneers in AI Systems — University at Buffalo](https://www.buffalo.edu/cubs/research/pioneers-in-ai-systems.html) — CEDAR, Srihari, and the 1997 USPS Annual Report figures: under two percent of handwritten images fully resolved at first deployment, rising to fifty-three percent.
- [Center of Excellence for Document Analysis and Recognition](https://en.wikipedia.org/wiki/Center_of_Excellence_for_Document_Analysis_and_Recognition) — The first handwritten address interpretation system, deployed through USPS contractor Lockheed Martin.
- [Chief Postal Inspector testimony on USPS data security, 19 November 2014](https://about.usps.com/news/testimony/2014/pr14_day1119b.htm) — Mail covers, mail imaging and MICT named as three independent programs; mail imaging as a processing tool from the early 1990s; MICT triggered only by a potentially contaminated mailpiece.
- [USPS Office of Inspector General, HR-MA-14-002](https://documents.theblackvault.com/documents/ig/HR-MA-14-002.pdf) — BDS MICT processes outlined in 2004 to identify and recall potentially contaminated mail and vehicles; mail imaging as an automation tool.
- [Informed Delivery — USPS](https://about.usps.com/kits/kit600/kit600_v04-2026_036.htm) — Exterior images of automation-processed, letter-size mailpieces in the daily digest and dashboard.
- [IRS research bulletin on the Discriminant Function](https://www.stayexempt.irs.gov/pub/irs-prior/p1500--2022.pdf) — 1963 expert rules to the 1969 discriminant model; TCMP and NRP recalibration.
- [GAO on tax compliance research samples](https://www.gao.gov/assets/ggd-96-89.pdf) — Statistically designed examinations and generalizable weights.
- [Selective labels in institutional decision systems](https://www.cs.cornell.edu/home/kleinber/kdd17-selective.pdf) — Why a system's own choices distort what it later learns.
- [HNC Software's account of Falcon](https://www.sec.gov/Archives/edgar/data/814547/000095014902000962/f81563xe425.htm) — 1992 origin; model rebuild on deterioration.
- [Falcon consortium agreement, SEC filing](https://www.sec.gov/Archives/edgar/data/1136893/000095012309030426/g16913exv10w5.txt) — Monthly institutional data for algorithm updates.
- [Nestor's Mellon Bank neural fraud system](https://bit.csc.lsu.edu/~jianhua/quang.pdf) — IBM 3090 deployment and holdout comparison.
- [Privacy Impact Assessment for the Automated Targeting System — DHS](https://www.dhs.gov/sites/default/files/publications/privacy_pia_cbp_ats_updated_fr_0.pdf) — ATS-P operational in 1999; no traveler risk score.
- [Federal Register, Privacy Act exemptions for ATS](https://www.federalregister.gov/documents/2010/02/03/2010-2201/privacy-act-of-1974-implementation-of-exemptions-department-of-homeland-securityus-customs-and) — Threshold targeting rules rather than scoring for ATS-P.
- [SKYNET courier-detection slides](https://www.eff.org/files/2015/06/01/20150508-intercept-skynet-courier-detection.pdf) — Random forest over mobile-network identifiers.
- [NSA's ICREACH metadata interface](https://arstechnica.com/tech-policy/2014/08/nsa-built-google-like-interface-to-scan-850-billion-metadata-records/) — Unified search across categories of metadata.
**The language interface and the selection architecture**
- [ELIZA — Weizenbaum, 1966](https://doi.org/10.1145/365153.365168) — Natural-language interaction as a published program.
- [Bringing Natural Language Processing to the Microcomputer Market: The Story of Q&A — Gary G. Hendrix, Symantec (1986)](https://aclanthology.org/P86-1002.pdf) — The September 1985 Q&A launch and its English-language database assistant.
- [SRI on CALO](https://www.sri.com/hoi/artificial-intelligence-calo/) — The DARPA assistant program behind Siri.
- [IBM on Watson](https://research.ibm.com/publications/introduction-to-this-is-watson) — Open-domain evidence weighing and confidence thresholds.
- [From Freebase to Wikidata: The Great Migration — Google Research](https://research.google/pubs/from-freebase-to-wikidata-the-great-migration/) — Entity representation surviving its project.
- [UMBEL to be retired — Mike Bergman](https://www.mkbergman.com/2229/umbel-to-be-retired-by-end-of-2019/) — Ontology work carried into a successor.
- [Recent Advances in Google Translate](https://research.google/blog/recent-advances-in-google-translate/) — Back-translation, where a model's output becomes the next model's training data.
- [Using Collaborative Filtering to Weave an Information Tapestry — Xerox PARC, 1992](https://www.bitsavers.org/pdf/xerox/parc/techReports/CSL-92-10_Using_Collaborative_Filtering_to_Weave_an_Information_Tapestry.pdf) — Annotations as first-class, reader-inspectable filter state.
- [GroupLens — ACM](https://doi.org/10.1145/192844.192905) — Collaborative filtering on Usenet with real participants.
- [Ringo — CHI 1995](https://chi1995.acm.org/proceedings/papers/us_bdy.htm) — 2,100 users by September 1994.
- [Two Decades of Recommender Systems at Amazon](https://cdn.amazon.science/76/9e/7eac89c14a838746e91dde0a5e9f/two-decades-of-recommender-systems-at-amazon.pdf) — Item-based collaborative filtering from 1998.
- [TiVo Annual Report on Form 10-K (2004)](https://www.sec.gov/Archives/edgar/data/1088825/000119312504063164/d10k.htm) — TiVo's service infrastructure collected anonymous viewing-behavior data from DVRs for audience measurement.
- [PENGUIN and RAISE — NeurIPS 2025](https://proceedings.neurips.cc/paper_files/paper/2025/hash/8c2bb821410066459be64d03a4dc5719-Abstract-Conference.html) — User context and benchmark safety scores across fourteen thousand scenarios.
- [PSG-Agent](https://arxiv.org/abs/2509.23614) — Per-user safety criteria across plan, tool, response and memory stages.
**The loop itself**
- [A Mathematical Theory of Communication — Claude Shannon, 1948](https://www.nokia.com/bell-labs/publications-and-media/publications/a-mathematical-theory-of-communication/) — Transmission formalized apart from meaning.
- [Cybernetics: or Control and Communication in the Animal and the Machine — Norbert Wiener](https://mitpress.mit.edu/9780262537841/cybernetics-or-control-and-communication-in-the-animal-and-the-machine/) — Animals and machines inside one theory of feedback.
- [Omar Sanseviero's printer demonstration, 17 November 2026](https://x.com/osanseviero/status/2122646098418491739) — Fourteen megabytes peak RAM; sixty seconds per token.
- [The published inversion, 19 November 2026](https://wccftech.com/google-engineer-runs-gemma-3-270m-ai-model-directly-on-office-printer-via-postscript/) — Sixty-one point three tokens per second.
![[resources/images/article-intelligence-inside-4004.jpg]]