Writing in the Artificial Intelligence Review in 1989, Jordan Pollack got the big thing right and the important thing wrong. He said the connectionist revival would keep growing, and it did, beyond anything he set out. What he expected it to grow into was a field he called connectionist fractal semantics, in which the distributed pattern would become a new kind of symbol, one with internal structure you could reason about, capable of pointing back to the larger structure it came from. He had a name ready for the object: a supersymbol. Nothing of that sort was ever built. The prediction failed in the strangest possible direction, by being fulfilled in volume and refused in kind.

Between 1988 and 1992 artificial intelligence occupied a strange interval between promise and afterlife. One system was fading and the other returning, and for a few years both were audible at once. Symbolic AI still carried the authority of the old dream: intelligence made explicit, filed, named, formalised, replayed. A bureaucracy of mind, in which every concept has its proper label and every conclusion can be traced back along a chain of reason to where it came from. The machine would not learn the world so much as be instructed in it, and for a while this seemed less like one option among several than like the only description that could possibly be true.

By the end of the decade the dream had gone hollow. Expert systems, sold as the industrial future of the field, turned out to be brittle contraptions, impressive inside narrow corridors of competence and expensive to keep upright outside them. They worked when the world behaved like the system's map of the world, and ordinary reality refused that containment, going on producing exceptions, ambiguities, tacit meanings and half-known contexts that no rule base absorbed.

Symbolic AI didn't disappear, though. It lingered the way an institution lingers after its purpose has become uncertain, and the great knowledge-engineering projects of the period have a melancholy grandeur about them. If machines lacked common sense, then common sense could be entered by hand, one assertion at a time, as though the entire background of human life were a document awaiting transcription. Douglas Lenat began Cyc in July 1984 at MCC in Austin on roughly that premise. By the end of its first six years the project had entered over a million assertions, and Lenat's own estimate of what remained was about two person-centuries of further work to reach the hundred million he thought necessary before the system could begin learning usefully on its own. Two centuries of clerical labour, budgeted, to arrive at the starting line. The more the project encoded, the more clearly it showed the abyss underneath knowledge: the residue of habit, embodiment, memory and practical familiarity that people draw on constantly without knowing they are doing it.

Connectionism was coming back at the same time, out of an earlier obscurity. The two Parallel Distributed Processing volumes landed in 1986 and offered a different picture, cognition as pattern emerging across many small adjustments rather than the manipulation of clean symbols. Meaning spread across weights instead of filed in a drawer. The claim that mattered for what followed was not that this worked better. It was that the distributed pattern was supposed to remain readable: microfeatures standing for something, a representation you could open.

The revival had a spectral quality, because none of it was new. Pollack described it plainly as the rebirth of a programme that thrived from the forties through the sixties and was severely retrenched in the seventies. The ideas had been there near the beginning and were pushed aside by the prestige of symbolic reasoning; the key training algorithm had already been written down in 1974 and left to sit. What surfaced in the late eighties was a path not taken, resurfacing precisely as the official future began to decay.

The symbolic order got one more authoritative statement, and it was a good one. Jerry Fodor and Zenon Pylyshyn published their critique in Cognition in 1988, opening with the observation that connectionist models were catching on, that there were conferences and new books nearly every day, and that the fan club included the most unlikely collection of people. Their argument was systematicity: anyone who understands "John loves the girl" understands "the girl loves John", and a classical architecture explains that symmetry for free. What looks like a concession in their paper isn't one. You can reconcile the two, they write, all that's required is that you use your network to implement a Turing machine, which is to say a network only gets systematicity by becoming the thing it claimed to replace.

David Waltz, a year earlier, had doubted you could take a large randomly wired network, show it enough raw sensory input and desired output, and get intelligence out the other end. The learning space for vision and audio was astronomically large, he said, and learning to perceive by feedback seemed cognitively and technically unrealistic. On the narrow question of whether the method scales, he was wrong, and the last fifteen years are the refutation.

The asymmetry between those two objections is the thing I'd point at. Waltz was answered. Fodor and Pylyshyn were not answered; they were outrun. Nobody demonstrated that distributed representations achieve systematicity without implementing a classical architecture underneath. The models simply got large enough that the question stopped being asked, which is a different outcome from being settled, and it left the philosophical objection intact and unattended somewhere behind the industry.

So the lost future of that interval isn't symbolic AI. Symbolic AI failed legibly, and a legible failure can at least be mourned on schedule. The loss that goes unmarked is the connectionism imagined in 1989: the readable microfeature, the supersymbol with inspectable internal structure, the representation you could open and reason about. That research line didn't die so much as get overtaken by systems whose representations nobody can read. We have interpretability now as a field, staffed and funded, which is itself the admission. It exists because the thing Pollack expected to come built in has to be excavated after the fact, from the outside, with uncertain results.

Cyc, meanwhile, never stopped. The transcription continued for decades, and by 2017 the knowledge base held something like 24.5 million assertions, roughly a quarter of the way to the line Lenat had named as the beginning. The project outlived its own future, which is a quieter fate than collapse and much harder to put a date on.

Sources: