Ideas from L'intelligence artificielle : mythes et limites

By Hubert L. Dreyfus

Back to L'intelligence artificielle : mythes et limites

346 ideas

Sample Ideas

  • James Culbertson’s claim that one could, in principle, build robots from flip‑flops that reproduce any human behavior by meeting arbitrary input‑output specifications simply assumes, without evidence, that human inputs and outputs can be isolated and correlated like digital signals, an assumption undermined by neurophysiological considerations.
  • The emerging field of artificial intelligence, as Dreyfus characterizes it, aims not to build an artificial human organism but to program digital information‑processing machines with heuristic procedures that allow them to compete with humans in disembodied, objectively testable intellectual tasks, especially those relevant to the Turing Test.
  • C. E. Shannon anticipates that efficient machines for pattern recognition and language translation will require a fundamentally different kind of computer whose natural operation is in terms of patterns, concepts, and vague similarities rather than numerical sequences, and Dreyfus adds that, judging from human capacities, such a 'machine' would also need a body and a situation, implying that present digital computers cannot be programmed to behave with human intelligence and that only human–computer cooperation in the short run and, in the long run, nondigital embodied automata offer a plausible path toward dealing with nonformal information.
  • Quillian proposes a compositional model of language understanding in which people (and, by extension, programs) understand new sentences by retrieving stored meanings of isolated words and phrases, combining and perhaps altering them via explicit 'combination rules', and then bootstrapping new word meanings from definitions written in already‑encoded vocabulary.
  • The history of alchemy provides an instructive analogy for AI: like alchemists who, buoyed by partial successes such as distilling quicksilver, persisted for centuries in a fundamentally impossible quest to transmute metals, AI researchers can always justify continuing work despite repeated failures; yet if we attend to predictions versus results, empirical evidence shows current efforts are misdirected at the information‑processing level, and real progress requires stepping back to investigate the deeper structure of the problem rather than pouring more resources into the same paradigm.
  • If, contrary to Dreyfus’s own doubts, the mind is even in part an information‑processing mechanism, then a mechanical, cognitive‑scientific explanation of its capacities in terms of functional components and their interactions would constitute exactly the sort of understanding of the mind one should seek, undermining Weizenbaum’s dismissal of cognitive science as philosophically trivial.
  • Work on automatic speech recognition exhibits the characteristic AI pattern of early success in restricted domains followed by failure to scale, because current systems treat speech as classification of fixed sound patterns in a feature space, whereas human hearing is guided by Gestalt perception in which the overall meaning of an utterance determines how individual sounds and phonemes are perceived.
  • Dreyfus concludes Phase I by arguing that the fundamental limitation of Cognitive Simulation is the need to make all alternatives explicit: in game playing, problem solving, language translation, and pattern recognition this leads to combinatorial explosion, structuring problems, and unresolved ambiguity, so that early successes have been followed by failure on more complex tasks, Simon’s grand predictions have not been fulfilled, and his optimism exemplifies Bar‑Hillel’s 'fallacy of the successful first step.'
  • Dreyfus clarifies that his critique targets the explicit and implicit philosophical assumptions of AI leaders such as Simon and Minsky, not their technical work, which he regards as important and valuable both for AI’s limited achievements and for other areas of computer science.
  • Dreyfus contends that there is no evidence humans solve analogy problems via heuristic search over explicit transformation rules, and that descriptive psychological evidence instead indicates a quite different, Gestalt‑like mode of reasoning, so the optimistic expectation that Evans’s techniques can be generalized rests on an unsupported hypothesis about human cognition.