The Man Who Buried His Teacher

Earlier this week, I wrote about Walter Pitts and Srinivasa Ramanujan. A teenage runaway from Detroit and a clerk from Madras whose work sits under large parts of the mathematical world that made modern AI possible. Neither of them could get hired today. A lot of people engaged with that post.

What I didn’t tell you is how Pitts died at 46 believing his life’s work was a failure, or how the field he helped found was buried for twenty years with help from someone shaped by his work.

In 1943, Pitts and neurophysiologist Warren McCulloch published “A Logical Calculus of the Ideas Immanent in Nervous Activity,” one of the foundational papers in the neural network lineage behind today’s AI boom. McCulloch had taken the homeless boy in off the streets of Chicago. They worked together every night.

A young man named Marvin Minsky learned from them.

Minsky built his early neural net computer under their influence. He wrote his Princeton doctoral thesis on self-organizing networks in 1954. He got his Harvard fellowship with the support of McCulloch, Norbert Wiener, and Claude Shannon. He built his career on the foundation that Pitts and McCulloch laid.

Then he buried their field.

Minsky did not single-handedly kill neural networks. Connectionism survived in pockets. The practical revival of the paradigm depended on advances in compute, data, and algorithms that hadn’t arrived yet. The first AI winter involved broader UK and US funding politics that extended well beyond one book.

But Minsky did something more durable than killing a field outright. He gave the funding establishment a technically respectable reason to stop taking it seriously.

In 1969, Minsky and Seymour Papert published Perceptrons, a book that demonstrated the mathematical limitations of single-layer neural networks. The critique was narrow and technically correct. Single-layer perceptrons cannot solve the XOR problem. They cannot learn non-linearly separable functions. This was known. What Minsky and Papert did was frame a limitation of the simplest possible architecture as a fundamental indictment of the entire connectionist paradigm.

They knew multi-layer networks could solve XOR. The possibility was not unknown. Minsky himself later acknowledged the critique had been over-applied. But Perceptrons was not written as a technical paper exploring the boundaries of a promising approach. It was written as a permission slip for the people holding the checkbooks.

Rebecca Skinner, writing from Stanford’s archives in 1997, documented the aftermath:

“The devastation of the field of neural net research in 1968 was accomplished almost single-handedly by a harsh critique by Marvin Minsky and Seymour Papert of MIT. The neural net and connectionist paradigm, as dependent on research funding as any other field, was laid waste. The AI establishment at the time was so tight-knit that funding for this type of research was not to be found for two decades.”

Two decades. An entire branch of research, the branch that would eventually produce the technology reshaping every industry on earth right now, was marginalized for twenty years. Not because one man decreed it. Because one man published a narrow critique and the funding apparatus he was embedded in treated it as the final word.

Minsky himself acknowledged the overkill in a cover article in AI Magazine in July 1991. By then the damage had been done for a generation.

To understand how one book could sideline an entire research paradigm, you have to understand the funding structure.

The Advanced Research Projects Agency, DARPA, was AI’s rich uncle. Through its Information Processing Techniques Office, DARPA underwrote practically all formative AI research for three decades. ARPA spent more on AI than the rest of the world combined, and most of that money went to two places: MIT and Stanford.

The first IPTO director was J.C.R. Licklider. Licklider came to DARPA from MIT. He showered MIT with government money. Then he went back to MIT to run Project MAC, the ARPA-funded program that Minsky’s AI group had been incorporated into. Then he returned to DARPA for a second term as director.

Thomas Haigh, writing in the ACM’s Communications in 2023, called the arrangement “a little too cozy by modern standards.” That is an understatement that would make a diplomat blush. It was a revolving door between the people writing the checks and the people cashing them, and the same names show up on both sides.

Skinner’s manuscript describes the result: “DARPA funding and institutions led to the consolidation of an ‘invisible college,’ the so-called ‘Artificial Intelligentsia,’ who determined what was to be done and not done.”

What was to be done was symbolic AI. What was not to be done was connectionism. And one of the central figures making that determination was Marvin Minsky.

Four of the first ten Turing Award recipients were symbolic AI specialists: Minsky, Simon, Newell, and McCarthy. They founded the three leading AI labs. They built the top three computer science programs. They controlled where the money went.

When Minsky published Perceptrons, the funding for neural network research didn’t just decline. It evaporated. Not because the math proved the approach was fundamentally broken. Because the most influential figure in the field, at the institution that received the most DARPA money, published a book that gave everyone in the funding chain a respectable reason to redirect resources. And in a community so small that “practically all of its participants had met in person or were at least aware of each other’s work,” that was enough.

Minsky was not an outsider misunderstanding the field. He understood it from the inside. And years later, he said as much.

In an interview with John Brockman for Edge.org, Minsky was asked about the cybernetic tradition that preceded AI. Gregory Bateson had once told Brockman that the cybernetic idea was “the most important idea since Jesus Christ.”

Minsky’s response:

“Well, surely it was extremely important in an evolutionary way. Cybernetics developed many ideas that were powerful enough to challenge the religious and vitalistic traditions that had for so long protected us from changing how we viewed ourselves. These changes were so radical as to undermine cybernetics itself. So much so that the next generation of computational pioneers, the ones who aimed more purposefully toward Artificial Intelligence, set much of cybernetics aside.”

Read that again. He is saying, on the record, that cybernetics was set aside not because it failed but because the ideas were too radical. Too powerful. Too challenging to existing frameworks of understanding. The next generation didn’t lose interest. They made a choice.

And Minsky was the leader of that next generation.

In the same interview, he lists McCulloch and Pitts as his “early mentors.” The man who laid the foundation. The man who took in a homeless boy and worked with him every night until they produced the most important paper in the history of the field. Minsky names them as his teachers, then describes the destruction of their life’s work as a natural evolutionary process.

Walter Pitts burned his unpublished doctoral dissertation.

Let that sit for a moment. A man whose 1943 paper sits in the foundational lineage of modern neural networks and large language models burned his own doctoral work. Not because of Minsky. Not because of Perceptrons. In 1959, Pitts co-authored a paper on the frog’s visual system that demonstrated real neural systems perform analog preprocessing, undermining the pure logical-digital model of the neuron he had staked his career on. His own experimental work told him his foundational model was wrong. The realization broke something in him. He lost focus. He declined. He drank. The fracturing of his relationship with McCulloch and Wiener, the men who had been his only real family since Chicago, left him without the support system that had held him together since he was a teenager.

Pitts died in 1969 at the age of 46. His only earned degree was an Associate of Arts. He died believing his life’s work had been built on flawed premises.

The technology he created is now worth trillions of dollars. His foundational assumptions turned out to be more right than he knew.

And the same year he died, Minsky published the book that made sure nobody would find that out for another twenty years.

The standard history of AI presents the first “AI winter” as a natural hype cycle. Promises were made. Promises weren’t kept. Funding dried up. The market corrected.

This is not what happened.

Haigh’s ACM article, drawing on SIGART membership data and Google Ngram analysis, demonstrates that the AI community actually grew steadily through the 1970s. Membership in ACM’s AI special interest group nearly tripled between 1973 and 1978, growing faster than ACM as a whole. References to “artificial intelligence” in published text rose consistently through the decade. The field was not contracting. Interest was not declining.

What changed was the political dynamics of who controlled the funding. When Congress passed legislation in 1973 requiring DARPA to justify military relevance for its research, the cozy arrangement between elite labs and DARPA program officers came under pressure. The “winter” was experienced at the top of the power structure, at the handful of well-connected institutions that had been receiving money on a handshake. It was not experienced across the field.

The Lighthill Report in the UK, often cited as a catalyst for the first AI winter, was “commissioned with the specific intent of justifying the withdrawal of funding for Donald Michie’s lab at Edinburgh.” It was not an independent scientific assessment that happened to find problems. It was a political instrument designed to produce a predetermined conclusion.

This is not a hype cycle. This is a funding consolidation driven by institutional politics, wrapped in the language of scientific review.

The architecture of control didn’t disappear when the AI winter thawed. It migrated into modern procurement offices. The institutional habit of treating a paradigm question as a vendor management issue was baked into the funding pipeline from the beginning.

The parallels to today are not subtle.

In 2026, the Department of War designated Anthropic a supply chain risk, the first such designation ever applied to an American company. Not because their models were uniquely unreliable. Because their CEO, who helped build some of these systems and understands their architecture, said “I need to evaluate the use case before I can tell you whether the system is reliable enough for it.” The engineering answer was heard as insubordination. The companies that signed faster are not running safer models. They are running the same math with fewer questions asked.

On June 3, Secretary Hegseth doubled down on the designation, clarifying in a legal filing that the basis was not a technical finding that Anthropic could manipulate its model post-deployment. Anthropic had argued that the initial claim rested on a misunderstanding of their capabilities. Instead, Hegseth wrote that the designation rested on “the loss of trust” and “pre-deployment risks.” Loss of trust. The official legal justification for designating an American AI company a supply-chain risk amounts to this: they gave the Pentagon an engineering answer instead of a procurement answer.

I wrote about this a few weeks ago in detail. The man who holds the AI portfolio for the Department of War described frontier AI models on a podcast for forty minutes using the language of enterprise software procurement without once acknowledging that these are probabilistic systems. He talked about “reliable” and “steady” partners who won’t “wig out.” Vendor management terms. He was looking for a shelf-stable product. He was treating a probabilistic system like a lightbulb that either works or it doesn’t.

This is not primarily a criticism of one official. Procurement offices are built to evaluate vendors. The people staffing them are usually doing the job the institution trained them to do. The problem is not reducible to one person. The problem is that the institution is applying a vendor-management framework to a technology that does not behave like a vendor product.

This is not a new mistake. It is the exact same category error that shaped the first AI winter. The DARPA establishment decided symbolic AI was the only game in town because symbolic AI looked like logic, and logic is easy to procure. Neural networks are not. Today’s defense establishment decided frontier AI models are enterprise software because enterprise software is easy to procure. Probabilistic systems are not.

I want to be clear about something. Minsky was brilliant. His contributions to computer science were real. The Society of Mind is genuine intellectual work. His technical skills were formidable. He built SNARC, one of the earliest neural net machines, as a graduate student. He wrote his doctoral thesis on self-organizing networks. He understood connectionism from the inside. He was not an outsider misunderstanding a field he’d never worked in. He was the field’s most prominent insider, and he supplied the kill switch.

If the history ended there, it would be a story about institutional power and intellectual credit. It doesn’t end there. The same funding structure that allowed a small group of researchers to control an entire field’s direction also created a vulnerability that would migrate into private channels, where the accountability was even lower and the leverage over researchers was even higher. That’s a story for another time.

The history of AI is not a story about ideas competing in a marketplace of merit. It is a story about what happens when institutional power decides what technology is allowed to be.

Walter Pitts looked at the engineering reality of his own work, realized its limitations, and burned his dissertation. Dario Amodei looked at the engineering reality of frontier AI, refused to rubber-stamp mass domestic surveillance and autonomous weapons as procurement line items, and saw his company designated a supply-chain risk.

Meanwhile, Marvin Minsky published a narrow critique, helped bury his teacher’s field for twenty years, and rode the DARPA pipeline to the top. Today’s defense establishment is running the same playbook. They don’t want engineers who understand probabilistic limits. They want steady vendors who will sell them a black box and won’t “wig out.”

You cannot regulate calculus. But Minsky proved you can defund the people doing it honestly, bury a field for a generation, and call it evolution. The question is whether we learned anything from watching him do it.

We didn’t. We’re doing it again.