Why Did GOFAI become Gone AI?

---

GOFAI vs. the Crushing Realities of Scale

When I was an AI researcher in the 1980s, the dominant paradigm was knowledge representation. AI, we assumed, was essentially the art of representing knowledge in a formal language - predicate calculus, or some applied version of it - and then using that representation via inference rules, plus  heuristics to stop the whole thing from exploding combinatorially.

There was a huge and optimistic research programme around this: fuzzy logic, probabilistic and Bayesian variants, and a menagerie of ad hoc representational structures - frames, scripts, semantic nets - all trying to bottle “commonsense” in forms that could be manipulated in increasingly powerful ways. 

In the mainstream, nobody cared about neural nets. That community existed, of course, but it was marginal stuff in a distinctive physics-based mathematical paradigm we didn't really understand, and which we thought useless.

In hindsight, this now looks extraordinarily short-sighted. The whole approach has been shoved into the cupboard labelled Good Old-Fashioned AI - GOFAI - as if it were a Neanderthal cousin: earnest, ingenious perhaps, but doomed. I assure you it did not feel that way at the time; it felt like the right approach.

Why? Because the way we consciously think about thought is through language. When you ask someone why they did something, they give you reasons. Those reasons come out as propositions. And if you take propositions seriously, the best formal machinery we have for them is logic. So we formalised. We inferred. We drafted and polished the rules. We argued about representation. We believed we were closing in on the core of intelligence: meaning made explicit, knowledge made inspectable, inference made principled and then effectual.

If this was such a plausible programme, why did it fail so completely?

Some people will tell you that GOFAI was about “meaning” and neural nets are about “statistics”. But that opposition is uninformed and usually tendentious. When you ask a large language model a question, you ask something with meaning and you get back something with meaning. If it were all just statistical noise, it would be useless - and it obviously isn’t. So the question is not whether artificial neural net systems have semantic understanding, but how they acquire it and where it lives.

To see the difference, it helps to focus on what we were really trying to do in the 1980s. We were trying to declare meaning in advance. We tried to build a world model by writing down the world: concepts, relations, constraints, defaults, exceptions, and the rules for moving between them. We aimed for explicitness because explicitness feels like the true essence of knowledge. If the machine “knows” something, we should be able to point at it. If it reasons, we should be able to justify it. If it is wrong, we should be able to debug it.

Those are admirable ambitions. They are also, it turns out, ruinously expensive.

The world is not just insanely large; it is also densely structured. Human life is soaked in tacit knowledge: the thousands of micro-regularities that never make it into explicit articulation because they don’t need to. Such commonplace utterances as:

• “If someone says X in that tone, they probably mean Y.”
• “If this tool is in that drawer, it implies the last person who used it was doing such-and-such.”
• “If the kettle is boiling and the milk is out, the next event is likely tea.”

We do not store these things as neat propositions. We carry them as an accumulated, largely unspoken competence - our multi-decade store of condensed experiences.

GOFAI tried to drag that whole submerged continent into the daylight, label it, and file it. And every time you write a proposition you commit yourself: the predicates have sharp boundaries; the ontology has cliff edges; the exceptions multiply. Brittleness is not an engineering accident in that world - it is structural. Logic, by its nature, deals in closed systems: a statement is true or false; a condition holds or it doesn’t; an entailment follows or it doesn’t.

Human cognition, by contrast, lives on slopes. We cope with partial fit, family resemblance, analogy, and improvisation. When our models are wrong, we often degrade gracefully rather than crash. In a certain sense we live by operational not denotational semantics: praxis or dasein, if you like.

The deeper problem, though, was not merely brittleness. It was bandwidth.

Even if you grant that explicit representations are, in principle, adequate - even if you believe you can model “meaning” propositionally - you still have to get enough of it into the machine to matter. That is where the knowledge representation research programme hit the wall. The amount of world-structure a useful system needs is not “a lot”; it is effectively astronomical.

Doug Lenat’s Cyc project is the heroic proof of concept and the cautionary tale in one. It was an attempt to win by scale within the symbolic paradigm: build a vast commonsense knowledge base by hand, assertion by assertion, rule by rule, over decades. It turns out the scale of what a lifetime of expert labour can encode is orders of magnitude below what is required for the kind of fluent, flexible competence we casually expect from “intelligence”.

It is not that Cyc was trivial; it is that the world is not.

So what did the neural approach do differently? It did not “abandon meaning”. It abandoned the fantasy that meaning must be pre-declared by a human engineer.

Large language models are trained on corpora so vast that they function, in practice, as a proxy for civilisation’s accumulated linguistic trace. That trace is not random. It is massively redundant, massively structured, and shot through with regularities about the world - causal regularities, social regularities, narrative regularities, and the implicit ontologies that language users share through use without ever explicitly listing them.

Training is then a brutal compression process. The system is forced, by optimisation pressure, to extract the stable patterns that make text predictable. In doing so it precipitates a semantic geometry: a high-dimensional space (thousands of dimensions!) in which meanings are not discrete atoms with hard edges, but regions and directions - similarity structure, entailment gradients, pragmatic association, contextual modulation. It is representation, certainly, but not the sort that sits obediently on a page as “facts”. It is meaning condensed into weights.

This is why the modern paradigm scales and the older one didn’t. Not because logic is false and vectors are true, but because the neural method exploits a source of structure GOFAI could never match: the pre-existing structure in the world’s textual data dump, harvested at industrial scale.

Symbolic AI tried to build an ontology from the top down, by explicit design. LLMs inherit an ontology from the bottom up, by statistical consolidation. The difference is not merely technical; it is ecological. Neural methods outsource much of the labour of world-modelling to the culture that produced the data. They do not ask a small priesthood of knowledge engineers to type out reality, line by line. They take what humanity has already written - the messy, contradictory, redundant bulk of it - and compress its median regularities into an internal structure that can generalise.

There is, of course, a price.

GOFAI’s great virtue was epistemic. Even when it failed, it failed in the open: you could inspect the rules, challenge the ontology, argue about the premises. Modern systems give you competence first and legibility only as an afterthought. They are astonishingly capable, but they do not come with built-in notions of truth or justification. They optimise. Sometimes that aligns with truth; sometimes it aligns with plausible rubbish. You gain power and lose a certain kind of accountability.

Still, the historical lesson seems clear enough. The decisive factor was not that symbolic representations are “wrong”, but that explicit hand-built representations cannot reach the density required for general intelligence-like behaviour. GOFAI was working with teaspoons; the world required oceans.

In the 1980s we tried to formalise intelligence as propositions plus inference. The 2020s arrived with a different discovery: intelligence scales with the compression of structured experience. Meaning can be precipitated, not merely declared. And once you accept that, the dominance of neural methods stops looking like a betrayal of reason and starts looking like a humiliating empirical fact about where the informational mass of the world really sits.

In the knowledge lies the power, the old knowledge engineers used to say; in the world's data lies the knowledge is the new mantra of the foundation model engineers.

---