The First AI Was a Reasoner: Newell & Simon’s Human Problem Solving

The book

In 1972, Newell and Simon published Human Problem Solving, the consolidation of a program nearly two decades old. Its central claim: problem solving is search through a problem space — a representation of the task as states, operators that transform one state into another, and knowledge for evaluating progress. The solver is somewhere; the goal is somewhere else; the moves are enumerable. Nothing mystical remains — skill lives in how the space is represented and how the search through it is controlled.

The control method they made famous is means-ends analysis. Compare the current state to the goal state and name the differences. Select the operator relevant to reducing the largest one. If that operator cannot yet be applied — its preconditions unmet — make the precondition a subgoal and recurse. The General Problem Solver was this loop as a running program: difference tables, a goal stack, working logic problems and puzzles step by recorded step. And the theory was grounded empirically, in think-aloud protocols — transcripts of human solvers talking through cryptarithmetic, symbolic logic, and chess, analyzed move by move. The book’s claim, stated with the authors’ characteristic care, was that human solving in these domains could be tracked as search through a problem space, organized to a striking degree around means-ends structure. The framework became foundational for cognitive psychology, protocol analysis became a standard method, and later work qualified the details without displacing the frame.

The oldest architecture in AI

The historical point is worth as much as the theory. The founding era of AI did not build statistical predictors; it built engineered reasoners. GPS’s reasoning was not an emergent property of anything — it was the program: explicit states, explicit operators, an explicit list of differences, a goal stack you could print and read. When GPS chose a move, there was a place where the choice happened and a record of why. The statistical turn — learn behavior from data, let the reasoning be implicit in the parameters — arrived decades later and earned its place on capabilities the symbolic programs never approached. But the genealogy runs the way it runs: artificial reasoning — reasoning that is engineered and inspectable — is not a retrofit bolted onto the LLM era. It is the founding tradition of the field, and next-token prediction is the historical newcomer to it, not the default it displaced.

From GPS to a governed graph

Our means-ends analysis graph is in a direct line of descent from GPS, and says so. The loop is Newell and Simon’s: state the current and goal states, produce the difference list, generate the operator for the largest difference, subgoal when preconditions fail, re-measure. What the graph adds is governance GPS never needed, because GPS never faced an examiner. The difference list is an artifact — recorded at each iteration, referred to by item, auditable after the run. Every operator must name the difference it reduces, so busy-work cannot be smuggled in as progress. And a progress gate halts the loop the moment a full pass closes nothing, routing the case to a human as stuck rather than iterating in place. One further change is the deepest: where GPS’s every operator and difference was hand-coded, the graph’s judgment inside each step is executed by a model. That division — engineered structure, learned judgment — is the response to the limit that follows.

Limits, honestly stated

GPS earned its results in well-structured domains — logic theorems, cryptarithmetic, classic puzzles — where states, operators, and the goal test can be crisply specified in advance. Ill-structured real-world problems resisted it, and the received assessment is fair: where the problem space itself has to be discovered, hand-coding the representation was the wall, the same wall later expert systems hit at scale. Newell and Simon largely scoped their claims to the domains they studied; their successors sometimes did not. The modern division of labor is a response to this criticism rather than a dismissal of it: the model proposes differences and operators inside the engineered loop, and the gates check that the proposals ground in evidence and demonstrably close the gap. Whether that division holds up on regulated workloads is an empirical question — the kind a backtest answers — not something six decades of pedigree settles.

The takeaway for a compliance reader

When a vendor implies that machine reasoning is inherently opaque — that inspectability is a research aspiration — the historical record disagrees. A running program could show its current state, its goal, and the exact difference it was working on before most of today’s examiners were born. The question to put to any reasoning system is Newell and Simon’s: where does the reasoning live? If there is a place — a state, a difference list, a subgoal stack — it can be examined. If the answer is “in the weights,” then what you are being shown is not the reasoning; it is prose about it.