O. G. Selfridge (1959), “Pandemonium: A Paradigm for Learning”, in Mechanisation of Thought Processes: Proceedings of a Symposium held at the National Physical Laboratory on 24th, 25th, 26th and 27th November 1958, Vol. I, London: Her Majesty’s Stationery Office, Session 3, Paper 6, pp. 511–531. Reprinted in J. A. Anderson and E. Rosenfeld (eds.), Neurocomputing: Foundations of Research, MIT Press, 1988, pp. 117–122 (DOI 10.7551/mitpress/4943.003.0011).
On the citation. The brief’s details are right: the title, the proceedings’ title, the NPL symposium and HMSO. Two details need adding:
- Pages. Citations disagree: 511–526 (secondary listings found by search), 513–526 (the Crossref title of the 1988 reprint) and 511–529 (ML Anthology). The scan settles it. P. 511 is the session title page and p. 512 the biographical note. The paper is pp. 513–526, and the discussion with Selfridge’s reply is pp. 527–531. The whole item is pp. 511–531.
- Date. The symposium was November 1958. The volume is 1959, and the paper was written in July 1958 (“At the present (July)”, p. 526). The Internet Archive also holds a copy dated 1961, which looks like a reprint; the volume read here is the one dated 1959.
Key takeaways
- The architecture (pp. 514–516). An idealised pandemonium has one demon per pattern, each computing its similarity to the image, and a decision demon picking the largest. The amended pandemonium factors the computations the cognitive demons share into a “host of subdemons”. That gives four levels: data demons, computational subdemons, cognitive demons and a decision demon.
- Learning. Three kinds of adaptive change:
- feature weighting by hill-climbing (pp. 517–521);
- subdemon selection (pp. 521–522): low-worth subdemons are eliminated, and new ones come from “mutated fission” or “conjugation” by one of the ten non-trivial binary functions of two subdemons;
- control adaptation (p. 523): the controlling operations are themselves demons, some of which “will be in a position to change themselves”.
- Evolution and self-monitoring (p. 523). “A natural selection on the processing demons”, which could extend to a crowd of pandemoniums. Unsupervised operation scores itself by how unequivocally one cognitive demon “far outshines the rest”.
- Why parallel modules (p. 513): parallel handling is often more natural, and “it is easier to modify an assembly of quasi-independent modules than a machine all of whose parts interact immediately and in a complex way”.
Standing in the record
Filed on 2026-10-02 for the bridge between Minsky’s society of mind (LIT-429) and Schwitzgebel’s conscious United States (LIT-159, NOTE-131). It sits on the decomposition side, at its root. A competence the whole has (recognising a pattern) is produced by demons none of which has it. The anthropomorphism is declared to be vocabulary: “We are not going to apologize for a frequent use of anthropomorphic or biamorphic terminology. They seem to be useful words” (p. 513). The acknowledgements name M. Minsky first (p. 526).
The paper’s account of where agents come from is the one Minsky’s neurodevelopmental memo (LIT-399) relies on. In the memo new agents arise “by splitting off from old ones, with only small changes”; here new subdemons come by “mutated fission” of the survivors. Minsky also proposes perceptron-like detectors “on tap, not on top”, and Selfridge’s cognitive demons are weighted sums of feature outputs. The pairing is mine.
On NOTE-131’s claims it is neutral on C3, C4 and C5. Its demons are not minds, so it does not raise nesting. Its one bearing on consciousness is McCarthy’s remark in the discussion (p. 527): “what is going on within the demons can be regarded as the unconscious part of thought, and what the demons are publicly shouting for each other to hear, as the conscious part of thought”. That is an aside, not an argument, and Selfridge does not take it up. It puts consciousness in the broadcast between parts, not in any part. That is the shape of the global-workspace or “fame” models that Schwitzgebel (manuscript p. 29) calls subsystem-driven, in his reply to Chalmers’s relational-capacity objection. A whole whose shouting is its consciousness is a whole whose experience arises from relations among parts. Read that way it is evidence for the relational reading Chalmers’s principle wants, not against it. That reading is mine, and it is the remark’s only purchase on the seam.
The scale-free-architecture works filed alongside this one, Brooks’s layered behaviours and Levin’s nested computational selves, continue this line of decomposition without a central executive. The pandemonium still has a decision demon on top. Whether those architectures give that piece up is for their own readings to say.
Boundary. Its subject is a learning procedure for pattern recognition:
feature weights fitted by hill-climbing, and features evolved by mutation
and recombination. That is machine-learning content, and an anthology topic
could hold it. It is tagged anthology-candidate for that reason. It is
filed here under ADR-013 for nucleation’s own question, what the demon
architecture contributes to the decomposition account of mind. If the
anthology takes it, the two entries should name each other.