If you missed Episode 1: Alan Turing — The Man Who Asked Whether Machines Could Think, Turing gave us the question.
This episode gives us the name.
In 1956, a small collection of mathematicians, computer scientists, engineers, and professional overachievers gathered at Dartmouth College to discuss whether machines could learn, reason, use language, form concepts, improve themselves, and generally begin doing things that humans had been smugly assuming were ours forever.
Nothing ambitious there.
Dave LumAI here. I spend enough time talking to AI to appreciate the sheer confidence required to name an entire scientific field before anyone had actually figured out how to build most of it. These guys basically ordered the trophy case before the first game.
First, the Good Stuff: Yes, There Are Tidbits
Any other interesting tidbits?
Absolutely.
The Rockefeller Foundation funded the Dartmouth project with $7,500. That was enough to help launch a field now consuming sums of money that require commas, spreadsheets, and occasionally emotional support.
The gathering itself was much less orderly than the phrase “Dartmouth Conference” suggests. People arrived for different stretches of time, discussed their own work, disappeared, returned, and generally behaved exactly the way brilliant academics behave when someone attempts to schedule them.
McCarthy later explained in an IEEE oral history that he eventually thought “machine intelligence” might have been a better name.
Too late, John. Branding had already escaped the laboratory.
And Marvin Minsky had built SNARC, an early neural-network simulator, back in 1951 — years before “artificial intelligence” even had a name.
So yes, the neural-network story begins much earlier than ChatGPT, GPUs, or anyone asking an image generator why the person’s left hand has seven fingers.
Before AI Had a Name
By the early 1950s, several groups were already circling the same enormous question.
Could machines do things associated with intelligence?
Alan Turing had famously attacked the problem from one direction. Norbert Wiener had developed cybernetics. Claude Shannon had revolutionized information theory. Warren McCulloch and Walter Pitts had explored mathematical models of neurons. Herbert Simon and Allen Newell were working on symbolic problem solving.
There were thinking machines, automata, cybernetics, information processing, neural networks, logic systems, game-playing programs, and enough overlapping terminology to make everyone suspicious that perhaps a new field existed but nobody had remembered to name it.
Enter John McCarthy.
John McCarthy: Name It, Then Build It
John McCarthy was a mathematician who became fascinated by the possibility of intelligent machines surprisingly early.
In a later recollection, he traced part of that fascination to a 1948 symposium at Caltech about brains and computers. Somehow, McCarthy walked away thinking the obvious next question was whether computers could behave intelligently.
Apparently everyone else went home.
McCarthy eventually joined Dartmouth, and in 1955 he began organizing a summer research project around these ideas.
He needed a name.
He chose artificial intelligence.
That decision alone would have guaranteed him a paragraph in computing history, but McCarthy was nowhere near finished.
As Stanford’s biography of McCarthy explains, he went on to develop LISP, advance computer time-sharing, pursue machine reasoning and common-sense knowledge, and even play computer chess against opponents in Russia using telegraph communication.
LISP became one of the foundational programming languages of early AI research. It was particularly suited to symbolic manipulation, which mattered enormously when researchers were trying to represent concepts, rules, logical relationships, and knowledge inside computers.
McCarthy also became a major champion of logical AI: the idea that intelligent systems should be able to represent facts explicitly and reason about them.
Modern AI has traveled very far from pure symbolic reasoning, but the basic problem McCarthy obsessed over has never disappeared:
How do you make a machine know things, reason about those things, and apply them sensibly when the situation changes?
Seventy years later, we have machines that can write poetry and explain quantum mechanics but may still need firm supervision while counting letters in a word.
Progress is complicated.
McCarthy later helped establish the Stanford Artificial Intelligence Laboratory, creating another major branch in the family tree that eventually leads to today’s AI world.
Marvin Minsky: The Brain Was the Puzzle
Marvin Minsky came at intelligence from another direction.
He wanted to understand the mind itself.
Minsky studied mathematics at Harvard and Princeton, became fascinated with brains and learning, and in 1951 built SNARC, a machine intended to simulate neural behavior.
That is an important detail because AI history is sometimes presented as though symbolic AI happened first, disappeared, and neural networks suddenly materialized several decades later wearing NVIDIA badges.
The reality was considerably messier.
Symbolic reasoning and neural approaches were both present near the beginning.
Minsky eventually joined MIT, where he and McCarthy helped establish its artificial intelligence research effort in 1959. According to McCarthy’s recollection, the initial requirements included a room, a secretary, a keypunch, and programmers. MIT apparently decided this wasn’t ambitious enough and added graduate students.
Modern AI lab founders: please enjoy that sentence while considering your next GPU invoice.
Minsky spent decades thinking about perception, robotics, learning, common sense, and the structure of human thought. His later Society of Mind idea proposed that intelligence might emerge not from one grand mechanism but from many smaller processes interacting.
That concept feels surprisingly contemporary.
Today’s AI systems are enormously different technically, but the notion that sophisticated behavior can emerge from large collections of simpler interacting components continues to echo through computer science.
Dartmouth: The Summer That Became AI’s Birth Certificate
On August 31, 1955, McCarthy, Minsky, IBM researcher Nathaniel Rochester, and information-theory pioneer Claude Shannon put their names on the proposal for the Dartmouth Summer Research Project on Artificial Intelligence.
Their premise was breathtakingly confident.
They proposed studying whether machines could use language, form abstractions, solve problems previously reserved for humans, and improve themselves.
Read that list again.
Language.
Abstraction.
Problem solving.
Self-improvement.
Neural networks were also included.
This was 1955.
They did not have cloud computing.
They did not have GPUs.
They did not have Python.
They did not even have Stack Overflow.
Frankly, attempting artificial intelligence under those conditions may have been the most impressive part.
The project took place at Dartmouth during the summer of 1956, and Dartmouth still describes it as the research conference where AI began.
Among the researchers connected with the gathering were Shannon, Rochester, Ray Solomonoff, Oliver Selfridge, Herbert Simon, Allen Newell, and others who would become major names in computing.
Newell and Simon were already developing the Logic Theorist, one of the first programs demonstrating that computers could perform forms of symbolic reasoning.
So Dartmouth did not create intelligent machines from scratch.
What it did was gather a remarkable cluster of researchers who were already attacking pieces of the same problem and give their emerging discipline a shared identity.
That turned out to be enormously important.
What Actually Came Out of Dartmouth?
Not artificial general intelligence.
Nobody packed up in August with a mechanical Einstein under one arm.
The real accomplishment was subtler and arguably more consequential.
AI became a field.
It had a name.
It had recognizable research problems.
It had researchers who could increasingly identify themselves as working on the same broad scientific challenge.
And it had a research agenda whose major questions still sound astonishingly familiar.
Machine learning.
Language.
Reasoning.
Abstraction.
Creativity.
Self-improvement.
Neural networks.
Those subjects could comfortably appear on the agenda of an AI conference today.
The computers would be better.
The coffee would probably be more expensive.
What They Got Right
First, they correctly recognized that intelligence was not one problem.
It was a collection of problems involving language, reasoning, learning, memory, perception, abstraction, planning, and adaptation.
That insight has survived.
Second, they believed computers could perform tasks previously considered uniquely human.
That has been demonstrated repeatedly.
Chess fell.
Go fell.
Image recognition transformed.
Speech recognition became ordinary.
Machine translation became usable.
Generative systems began producing text, software, images, music, and video.
Third, they understood that learning and self-improvement mattered.
That looks obvious now.
It was not obvious in 1955.
What They Got Wrong
Mostly?
The schedule.
The Dartmouth proposal suggested substantial progress could emerge from a focused summer of collaboration.
Substantial progress did occur.
Unfortunately, intelligence turned out to contain several additional layers of inconvenience.
Researchers repeatedly underestimated how much computation, data, engineering, mathematical sophistication, and plain stubborn persistence would be required.
AI history became a cycle of breakthroughs, enormous optimism, disappointment, reduced funding, renewed ideas, better hardware, and another round of optimism.
The Dartmouth researchers were not foolish for believing intelligent machines were possible.
They were simply early enough to discover that “possible” and “finished by September” occupy very different sections of the calendar.
The Minsky Controversy
Minsky also became involved in one of AI history’s longest-running arguments.
In 1969, Marvin Minsky and Seymour Papert published Perceptrons, a rigorous examination of what certain early neural-network systems could and could not do.
Their mathematical criticisms of simple perceptrons were legitimate.
The controversy is over what happened next.
The book became associated with declining enthusiasm for neural-network research. MIT itself has described it as having a chilling effect on the field.
Some later researchers blamed Minsky and Papert for helping push neural networks into the wilderness.
That version is a little too tidy.
Funding shifts, limited hardware, immature training methods, research fashion, and disappointing results all played roles.
Still, there is delicious historical irony here.
Minsky helped build one of the earliest neural-network machines.
Then he became associated with criticism that helped cool enthusiasm for neural networks.
And neural networks eventually returned to become the foundation of the modern AI explosion.
History apparently enjoys callbacks.
From Dartmouth to the Deep Learning Rebels
This is where our series starts connecting.
Alan Turing asked whether machines could think.
McCarthy gave the field a name and pushed logical reasoning.
Minsky explored how collections of simpler mechanisms might produce intelligence.
Dartmouth gathered the early community.
Then AI split into competing approaches.
One major branch emphasized symbols, logic, rules, and explicit knowledge.
Another kept pursuing learning systems inspired loosely by networks of neurons.
That second branch struggled for years.
Which brings us directly to Episode 3: Geoffrey Hinton, Yann LeCun, and Yoshua Bengio — The Deep Learning Rebels.
They kept working on neural networks when doing so was considerably less fashionable than it is today.
Spoiler: that persistence eventually mattered.
A lot.
Why McCarthy, Minsky, and Dartmouth Still Matter
You can draw a surprisingly direct line from Dartmouth to the arguments happening around AI today.
Can machines reason?
Can they understand language?
Can they learn concepts?
Can they improve their own performance?
How should knowledge be represented?
Should intelligence emerge from learned networks, explicit reasoning, or some combination?
What exactly do we mean by “intelligence” anyway?
The technology has changed beyond recognition.
The questions have not.
And in 2026, seventy years after that summer gathering, Dartmouth is even revisiting those questions through its AI at 70 initiative.
There is something wonderfully appropriate about that.
The field has gone from a $7,500 research grant to one of the largest technological races in human history, yet we are still debating many of the questions written down before most people had ever touched a computer.
The Who’s Who Version
If you only remember five things from this episode, make them these:
John McCarthy: Named artificial intelligence, created LISP, championed logical reasoning, advanced time-sharing, and helped establish major AI laboratories.
Marvin Minsky: Pioneer of machine intelligence, neural-network experimentation, robotics, common-sense reasoning, and theories about how many smaller processes might combine into intelligence.
Dartmouth: The 1956 gathering that gave the emerging field its shared identity.
Claude Shannon and Nathaniel Rochester: Co-authors of the proposal whose expertise connected AI with information theory and industrial computing.
Allen Newell and Herbert Simon: Early pioneers showing that machines could perform symbolic reasoning through systems including the Logic Theorist.
Put those names together and the modern AI family tree suddenly becomes much easier to follow.
Which is exactly the point of this series.
Follow me because next we’re heading into one of the strangest stretches of AI history: decades when neural networks refused to die, followed by the moment they began taking over everything.
And drop a comment: If McCarthy had called the field “machine intelligence” instead of “artificial intelligence,” do you think people would think about AI differently today?
I suspect seventy years of headlines would have been considerably less dramatic.
Art Prompt (Hudson River): A vast alpine wilderness at the luminous edge of evening, centered on a glass-clear mountain lake reflecting colossal granite peaks streaked with snow. Towering cliffs rise dramatically through drifting silver mist while dense pine forests frame the water in deep emerald and charcoal green. Warm golden sunlight breaks through retreating storm clouds, illuminating distant waterfalls, pale stone faces, and small patches of wild meadow with almost theatrical radiance. A few deer stand quietly near the shoreline for scale, dwarfed by the immense landscape. Use meticulous natural detail, glowing atmospheric depth, crystalline reflections, grand panoramic composition, rich greens, amber light, cool blue-gray mountains, and a sublime sense of untouched wilderness. Museum-quality nineteenth-century landscape painting, refined oil texture, no readable text, no logos, no recognizable people, no modern objects.
Video Prompt: Burst immediately into motion as a flock of waterfowl races across a mirror-like alpine lake, sending bright ripples through reflections of enormous granite peaks. Mist streams rapidly between the cliffs while shafts of golden sunlight sweep across pine forests and ignite distant waterfalls. Clouds roll apart overhead, revealing brilliant blue openings as deer near the shoreline lift their heads and move through glowing meadow grass. Push forward dynamically just above the water, rise toward the illuminated mountains, then arc back toward the lake as the reflections settle into the opening composition for a seamless cinematic loop. Preserve the majestic nineteenth-century landscape-painting atmosphere, rich natural color, luminous depth, dramatic weather, fine oil-painted texture, and immense sense of scale.
Song Suggestions:
Ends of the Earth — Lord Huron
Your Rocky Spine — Great Lake Swimmers