The Society of Mind

We are ten people in a farmhouse in Hanover and we have agreed to spend the summer solving intelligence.

McCarthy had coined the name “artificial intelligence” specifically so that no existing field could claim ownership of the problem. Shannon was there, the man who had made information mathematically precise. Newell and Simon had already built the Logic Theorist that was running on a machine in Pittsburgh. With SNARC just behind me: a learning machine made of vacuum tubes and motor-driven potentiometers, the first neural network that actually learned.

The hypothesis of the Dartmouth workshop was precise: that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. We believed it.

We thought we could solve intelligence in a summer.

Rosenblatt published the perceptron in 1958. The convergence theorem was real: if a solution exists, the perceptron will find it in finite time. The New York Times ran the story about the embryo of an electronic computer that will walk, talk, see, and reproduce itself. The Navy funded it. The hype was, for once, not completely baseless. The convergence theorem is correct.

I watched this with a specific concern. Not that Rosenblatt was wrong. My concern was the question nobody was asking: what problems are not linearly separable? The convergence theorem guarantees a solution when one exists. What is the structure of the problems for which no solution exists?

Seymour Papert and I spent three years on this. We worked through the geometry, the combinatorics, the group theory. The answer was: most of them. The problems that a single-layer perceptron cannot solve are not edge cases. They are the common case.

The clearest example is XOR. Two binary inputs. Output 1 if exactly one of them is 1. Four training points in the corners of a square. The two class-1 points sit in opposite corners. No straight line separates them. This is not an engineering problem. It is a geometric fact.

What the book does not say, and explicitly leaves open, is whether multi-layer networks have the same limitation. We proved one layer doesn’t suffice for XOR. The problem of training multi-layer networks was unsolved. The AI winter that followed was not a consequence of our mathematics. It was a consequence of people reading our conclusion correctly and then drawing the wrong lesson.

The wrong lesson: learning machines don’t work. The right lesson: this specific architecture has this specific limitation, and here is the proof.

The fix required two things: a hidden layer and a way to train it. I stated the first clearly in Perceptrons. Backpropagation came seventeen years later, in 1986, the same year I published Society of Mind. Rumelhart was right that the limit has a fix. I was right about the limit.

Even if you had a perfect learning algorithm, what would it learn from?

This was the question nobody in the learning-machine camp was pressing in the 1960s. The assumption was that representation is incidental, with raw sensory input in, behavior out, learning fixes the rest. The perceptron failure was, in this view, a lesson about training algorithms. My argument was different: the perceptron failure was a lesson about representation. It failed because there was no vocabulary in which XOR was expressible. The problem was not that it could not learn. It was that it could not represent.

The 1974 paper was an attempt to get precise about what representation would need to look like for intelligence to be possible. The central concept was what I called a frame: a data structure for representing a stereotyped situation.

When you walk into a room, you do not begin from scratch. You already know the room has walls, a floor, a ceiling. You already expect a door. You already have default assumptions about what the light source is, what the purpose is, what you will find there. These defaults are loaded in advance. Experience confirms or overrides them. But the structure is prior to the experience, ie you bring it with you.

The technical mechanism was slots with default values and inheritance. A bedroom frame inherits from a room frame, overriding some slots, leaving others at their defaults. Intelligence, in this model, is not a blank slate that fills in. It is a hierarchy of stereotyped structures, each pre-loaded with expectations, each updating on contact with specific reality.

The frames paper was not widely celebrated when it appeared. It became, in the following decade, one of the most cited papers in AI. It influenced object-oriented programming. It influenced every knowledge representation system built in the 1980s. The slot-and-default architecture appears in how large language models are prompted today: you set up a frame, you specify which slots to fill, the model fills the rest with learned defaults.

The Society of Mind took thirty years to write, in the sense that every observation in it is something I noticed in the lab between 1956 and 1985. The synthesis took another decade. The book was finished by the time backpropagation was published. They appeared the same year. I did not plan that.

The synthesis is this: there is no mind. There are only agents.

Consider the Builder. Builder is an agent that knows how to construct a tower of blocks. Builder directs two sub-agents: Bricks and Staircase. Bricks decides where each individual block goes. It does this by calling on two more agents: See, which locates blocks in the visual field, and Hand, which moves to a position and executes placement. Hand relies on Grasp to control finger positions.

None of these agents is intelligent. Builder knows a few rules about stacking. Hand knows how to move. See knows how to locate. Grasp knows how to close. No single one of them could do the whole task.

But together, they build the tower.

The feeling of being a unified mind is not a fact about the underlying machinery. It is an emergent property. There is no central processor that knows about blocks and hands and eyes and plans all at once. There is a structure of interactions between agents that know small things, and that structure produces behavior that looks, from the outside, like unified intelligence.

The implications for AI are direct: you do not build intelligence by writing a single program that knows everything. You build agents that know small things. You build the right structure of interactions between them. The intelligence lives in the structure, not in any of the pieces.

People misread both Perceptrons and Society of Mind as pessimistic about AI. Both readings were wrong in the same way: they took a precise claim and replaced it with a vague one.

The precise claim of Perceptrons: a single-layer linear threshold unit cannot compute functions that are not linearly separable. XOR is not linearly separable. Therefore, a perceptron cannot learn XOR. This is a theorem. It is not pessimism.

The vague conclusion people drew: learning machines don’t work. This does not follow. It is not what we said. It is not what the mathematics implies.

The precise claim of Society of Mind: there is no unified mental agent. What we call intelligence is the product of many interacting sub-processes, none of which is itself intelligent. Building AI therefore requires building the right architecture of interactions, not finding the right single algorithm. This is a design specification. It is not pessimism either.

Neither claim is pessimistic. Both are specifications. The constraint from Perceptrons was solved by Rumelhart in the same year Society of Mind was published. The constraint from Society of Mind is still being addressed.

I think we do not yet know what the right building blocks are. What I am confident of: intelligence is not found, it is constituted. It is not a substance that machines either have or lack. It is a property of certain kinds of organization. Build the right organization, and intelligence follows.

Perceptrons (1969) was mathematically correct. The lesson the field drew, that learning machines were a dead end, was too broad. The fix came in 1986: Rumelhart, Hinton, and Williams demonstrated backpropagation for multi-layer networks, confirming that every specific limitation Minsky and Papert had proved was real, and that adding depth dissolved it. The theorem was right. The field had misread its scope.

Frames (1974) became foundational to knowledge representation. The slot-and-default inheritance architecture influenced object-oriented programming, relational databases, ontology languages, and knowledge graphs. When researchers in the 1980s built expert systems, they built them with frames. When engineers today write system prompts for large language models, the underlying logic, pre-set context, default behavior, selective override is structurally identical to what Minsky described.

Society of Mind (1986) anticipated modular networks, mixture-of-experts architectures, and multi-agent systems, published the same year as backpropagation, in a book that described a completely different path to intelligence. Both paths are active. The society-of-mind framework became, three decades later, the dominant architectural metaphor for AI systems that coordinate multiple specialized models.

Minsky died in January 2016. Two months later, AlphaGo defeated Lee Sedol. AlphaGo Zero (2017) learned from scratch & had no human games in its training data, no handcrafted knowledge representation. It played itself into superhuman performance. Minsky had asked, with formal precision, what it would take for a machine to learn anything at all. The machines eventually did. The question of whether the right architecture has been found is still open.

He was the first person to ask, with formal precision, how a machine could learn anything at all. The machines eventually did.