Minsky's own 'the perceptron project was really dead' testimony is a documented instance of the 'retrospective illusion' Dreyfus's 1988 paper warns the victors will produce

dreyfusminskyconnectionismperceptronssociology-of-sciencehistoriographyphilosophy-of-aihistory-of-ml

Dreyfus and Dreyfus's 1988 Daedalus essay makes a specific prediction about how the history of AI would be narrated by the winners:

"Thus, it looks to the victors as if symbolic information processing won out because it was on the right track, while the neural network or connectionist approach lost because it simply didn't work. But this account of the history of the field is a retrospective illusion."

Olazaran's 1993 interviews preserve that exact narration in the mouth of the person most credited with producing it. Minsky, describing SRI's switch to LISP and the Shakey robot, concludes: "By that time the perceptron project was really dead." The "dead" verdict is delivered by the co-author of Perceptrons (1969) — the critique conventionally credited with killing connectionism — describing his own side's victory in precisely the "it simply didn't work" register Dreyfus says a victor will reach for.

The pairing is why the two vault notes it draws on are more than thematically adjacent. claim-dreyfus-1988-connectionism-was-vindicated-not-target supplies the general claim — the program was "never really refuted," only repressed — and Minsky's testimony, captured in claim-sri-neural-group-retooled-for-arpa-funding's source, supplies a documented instance of the myth-making that claim argues against, quoted from the myth-maker himself. This is the historiographic engine behind myth-perceptrons-book-killed-connectionism: the "manifest destiny" reading of symbolic AI's win is not a neutral record but a winner's account, and here the winner is on tape giving it. That the death he names actually predates his own book by four years is the chronology carried by claim-sri-retooling-predates-published-perceptrons-refutation. See claim-dreyfus-critique-targeted-symbolic-ai-not-neural-nets and moc-backpropagation-origins.