David J. Chalmers, Robert M. French, Douglas R. Hofstadter (1992), Journal of Experimental & Theoretical Artificial Intelligence 4(3):185–211; preprint CRCC Technical Report 49, Indiana University, March 1991 — DOI-10.1080/09528139208953747
Key takeaways
- A model that is given its representations ready-made has been given the answer: the “20–20 hindsight” charge against BACON and SME.
- Representation-building cannot be split off into a front-end module, because analogy-making is part of perception and the right representation depends on the task in hand.
- Copycat is the group’s demonstration that perception and mapping can be run together, by many small stochastic agents biased by a network of concepts.
Standing in the record
Filed on 2026-10-01 at the owner’s request for Douglas Hofstadter’s work, and the Fluid Analogies Research Group’s programme in particular. NOTE-359 is the close reading, and it placed the work: Active. It is the group’s methodological manifesto in one paper: why a theory of cognition must start from the system building its own representations, and why analogy is a perceptual process rather than a special-purpose reasoning tool. It argues; it does not measure.
Its bearing on the record:
- Copycat. §4 is the most accessible account of Copycat, and it cites
the Physica D paper (LIT-435, filed beside it) for the details. Read
together, this paper is the argument and that one the program. The
emergence claim — “higher-level understanding emerges” from many local,
parallel processes, under top-down bias from the Slipnet — is why it
carries
complex-systems, and it bears on the record’s emergence readings (LIT-141, LIT-150) as an engineered instance, not a theory. - Minsky. Frames (LIT-410) are among the representational formats the paper names (“frames and scripts”), and its “data do not come prepackaged as slots and fillers” is aimed at that tradition: the critique is not that frames are the wrong format but that a format leaves the filling unexplained. Copycat’s many small agents also stand near Minsky’s society of agents (LIT-429), though the paper does not cite it.
- Scientific discovery. The BACON section and the closing reading of
Copycat’s re-perception as a “scientific revolution” in a microdomain are
claims about how discovery works, read against Kuhn. That is why it
carries
philosophy-of-science: someone browsing that topic for computational models of discovery should find this critique of them. - Understanding in present AI. The paper’s 1992 “meaning barrier” is the ancestor of the record’s question whether deep networks understand (LIT-131) and of Hofstadter’s 2026 essay (LIT-416). Its §2 names connectionist models with context-dependent distributed representations as “a step in the right direction”, which a reader of today’s learned representations will want to set against its demand that perception and task interact.
- Boundary. The paper carries a methodological instruction for AI
modelling: do not hand-code representations; build perception into the
model. It is aimed at symbolic cognitive modelling, not machine-learning
practice. The nearest anthology topic,
analysis-and-evaluation(“which comparisons are unsound”), would hold only its charge that BACON’s and SME’s successes are artefacts of pre-built inputs, and that charge is about cognitive models, not learned ones. It is not taggedanthology-candidate.