David J. Chalmers (1996), Synthese 108(3):309–333. DOI 10.1007/BF00413692. Author’s copy at https://consc.net/papers/rock.html.
Key takeaways
- Putnam’s argument (Representation and Reality, 1988, appendix): every ordinary open system realises every inputless FSA, by mapping disjunctions of its successive maximal states onto the formal states.
- What it gets wrong (§3). Implementation needs strong conditionals: if the system were in a state mapped to P, it would go to one mapped to Q, however it got there. That has to hold for every transition in the table, including those not run. Putnam’s construction captures one trace, not the automaton.
- What survives (§§4–5).
- Any system with a clock and a dial implements every inputless FSA. So inputless FSAs are the wrong formalism for a mind.
- Any system with the right input–output dependencies, an input memory and a dial implements every FSA with those dependencies. A zero-outputter so equipped implements a primality tester. With FSAs as the formalism, functionalism “still almost reduces to behaviorism”.
- Combinatorial-state automata (§6). A state is a vector of substates. Each component must be realised in a distinct physical region, and each transition rule must hold as a strong conditional over the components. That demands fine-grained causal structure few systems have. A blind Putnam-style implementation must grow by a factor of n per step and outruns the universe within a fraction of a second of brain time (§7). FSAs, Turing machines, cellular automata and programs all translate into CSAs.
- Open problems (§7). Astronomically large false implementations remain logically possible, and a uniformity or causal-relevance clause is left as a challenge. The one-region-per-component condition may be too strict, for instance for virtual memory.
- Searle (§8). Computational descriptions are not observer-relative in any way that threatens vacuity. A system implements many automata, and whether it implements a given complex CSA is objective. A system running two AI programs would host two minds, so minds belong to (system, CSA) pairs. In the Chinese room there are two CSAs and “we should expect two distinct minds”.
Standing in the record
Filed on 2026-10-02 at the owner’s request, with the other Chalmers works, for its bearing on Block’s China brain (LIT-407) and on THEORY-043. NOTE-342 is the close reading. Active: it is the reference account of implementation that later work on computational individuation starts from. Its main limitations are ones Chalmers states himself.
Primary subject left unsaid by the vocabulary. The paper is about
computation: what it is for a physical system to implement an abstract
automaton, and whether that relation is trivial. The vocabulary has no
word for it. cognition is justified, because the paper’s stated aim is
to secure the computational foundation of cognitive science. metaphysics
fits the realisation question, and mereology the decomposition of a
system into independent components and the individuation of two minds in
one hunk of matter. But none of the four is the paper’s subject.
mathematics would stretch a neighbour: the automata theory is used, not
studied. This is a gap to close by decision, not by these tags.
On Block’s China brain (LIT-407). Chalmers does not discuss Block’s 1978 paper; he cites only Block’s 1981 “Psychologism and behaviorism” (LIT-642), for the lookup table. But his account decides what the China system implements, and the answer is sharper than “yes”.
- Block’s homunculi each execute one quadruple of a machine table, and the current machine state is shown on a card or by satellite (NOTE-366). That system implements an FSA with input and output, whose states are monadic. Chalmers’s §5 shows that this formalism is too weak to carry a mind. Anything with the right input–output dependencies, an input memory and a dial implements every FSA with those dependencies. So, on Chalmers’s terms, Block’s machine-table China system makes a weaker case against functionalism than it appears to: a thesis of computational sufficiency stated for FSAs is close to behaviourism already.
- The version Block sets aside in n. 17, a billion people each simulating one neuron, is the one that implements a CSA. Each person is a distinct physical region carrying one component. The radio links carry the strong conditionals, if they are reliable. On Chalmers’s account a nation so organised implements the brain’s CSA. Whether it has a mind then turns on computational sufficiency, which Chalmers here leaves to “the harder part” (§9). That pairing is mine, not the paper’s.
On THEORY-043. The account’s functionalist premise is that a whole’s mental properties come from how its parts are organised. This paper is where “organised the right way” gets implementation conditions.
- A CSA’s components are positions, not minds. The account requires only distinct physical regions and reliable dependencies, and is indifferent to whether a component is a neuron, a chip or a person. So it treats Minsky’s mindless agents and Schwitzgebel’s citizens alike. It supplies no anti-nesting principle.
- It allows two minds in one system (§8). Chalmers puts mental properties on (system, CSA) pairs. A human who facilitates another CSA’s dynamics, as in the Chinese room, does not prevent a second mind. He expects two. That runs directly against Putnam’s and IIT’s exclusion postulates, and it is the 1996 statement of a nesting-permissive view. It is argued from the CSA account plus an analogy to a computer running two programs. It is not defended against an anti-nesting principle, which the paper does not mention.
- It does not state NOTE-131’s correspondence principle. The CSA account requires complex causal interaction among separate parts. It says nothing about whether a system’s capacities must come mainly from the relations among its subsystems rather than from their own capacities. A CSA whose components are humans each doing most of the work is still a CSA.
On the extended mind (LIT-097). The paper does not discuss extension. But its implementation conditions ask only for distinct physical regions with counterfactual-supporting dependencies, and they let the theorist “move the boundary inward” (§4) by redescribing inputs. So nothing in them ties a cognitive system to the skull. That is the computational side of the parity claim Clark and Chalmers state two years later. The extended-mind programme being filed alongside this note is where that link would be tested. The connection is mine; the 1996 paper does not draw it.
Neighbours. Dennett’s “Real Patterns” (LIT-220) makes a parallel move about intentional descriptions: many descriptions of one system, each objective. Chalmers, French and Hofstadter (LIT-414) is the earlier Chalmers paper in the record on AI methodology; it argues for emergent representation in Copycat and does not touch implementation. It carries no instruction for machine-learning practice.
The room it answers. Searle’s “Minds, brains, and programs” is filed as LIT-400, read in full from the BBS preprint.