Contribution
The paper gives a precise, information-theoretic sense to the claim that an abstract pattern discerned from a stance is “really there”. It uses that sense to defend an intermediate realism about propositional attitudes, placed among Fodor’s “industrial-strength Realism”, Davidson’s “regular strength realism”, Rorty’s irrealism and Churchland’s eliminativism (p. 30).
Key insight
Reality for patterns is compressibility. A description that beats the bit map captures something that is there in the data, whether or not it is perfect, and whether or not the underlying process “contains” it in pure form. Because patterns trade compactness against noise, several real patterns may coexist in the same data, and there may be no fact about which is “the” right one.
Assumptions
- NOA (Fine): the dispute is conducted within the natural ontological attitude. The question is whether beliefs are “as real as electrons or centers of gravity”, not the ultimate metaphysical status of either (p. 30).
- Chaitin’s definition of randomness: “A series of numbers is random if the smallest algorithm capable of specifying it to a computer has about the same number of bits of information as the series itself” (fn. 10, p. 32).
- Folk psychology is highly predictive. All parties to the debate share this premise (p. 42 and fn. 19).
- Quine’s indeterminacy of translation is accepted and extended to radical interpretation (p. 46).
Key results
- Centres of gravity are abstracta whose realism should be judged on the “scientific path” (whether they are “good abstract objects”) and not the metaphysical path. The contrast cases are the “center of population” and “Dennett’s lost sock center” (pp. 28–29).
- Frames A–F (pp. 31–33): bar code (ten rows of 90 dots, ten black then ten white) with noise ratios A 25%, B 10%, C 25%, D 1%, E 33%, F 50%. F’s pattern is “indiscernible”. Transmitting a 900-dot frame verbatim costs 900 bits. Any description shorter than that captures a real pattern.
- Definition (p. 34): “A pattern exists in some data – is real – if there is a description of the data that is more efficient than the bit map, whether or not anyone can concoct it.”
- Proper names are not a cheat (fn. 12): a naming scheme for all 900-dot frames needs names of up to 900 bits.
- Three attitudes to noise (p. 35): exact reproduction, noise statistics only, or ignoring the noise.
- Jones and Brown (pp. 35–36): a simple, noisy pattern and a complex, less noisy one can both “get rich” when betting. Which to use is a trade-off made by evolution and culture, and it produces the manifest image. Idealised models in science (Newtonian mechanics, single genes “for” traits, perfectly rational market agents) make the same trade (pp. 36–37).
- Game of Life (pp. 37–42): the physical level has one exceptionless law and perfect prediction. The design level has gliders, eaters and so on, “a distinct ontological shift” to persisting moving objects, with generalisations that need “provided nothing encroaches” clauses (p. 39). A universal Turing machine built in Life is on the order of 10¹³ pixels, a screen about 3 km across (p. 40). Treating it as a chess player lets one predict its output glider streams from the intentional stance (p. 41).
- Intentional patterns (pp. 42–45): a hidden, pure pattern plus noise, as Fodor needs, is not the only explanation of a visible pattern. The pattern can also arise “thanks to the statistical effect of very many concrete minutiae producing, as if by a hidden hand, an approximation of the ‘ideal’ order” (p. 43). Fig. 4 is contrasted with fig. 1 to make the point.
- Indeterminacy (pp. 46–49): two intentional interpretations of “Ella” can be reliable, compact, long-run predictors that nevertheless disagree on crucial cases. So “there could be two different, but equally real, patterns discernible in the noisy world” (p. 49). This is the ground on which Dennett is “less of a realist than Davidson”.
- Against Rorty (p. 50): folk psychology, unlike astrology, lets people “get rich by betting on the patterns”.
- Against Churchland (pp. 50–51): neuroscience will explain more of the variance in the limit, but a pattern “scarcely an improvement over the bit map” gives no reason to eliminate. “Physical-stance predictions trump design-stance predictions, which trump intentional-stance predictions – but one pays for the power with a loss of portability and a (usually unbearable) computational cost” (fn. 32).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | A pattern is real iff some description of the data is more efficient than the bit map | definition | stipulated via Chaitin, pp. 32–34 |
| C2 | Different perceivers or theorists can discern different real patterns in the same data, and these need not be disagreements | informal argument | perception and expertise examples (de Groot’s chess), pp. 34–35 |
| C3 | Choosing between a compact, noisy pattern and a complex, precise one is a pragmatic trade-off, and it is ubiquitous in science | informal argument | idealisation examples, pp. 36–37 |
| C4 | The Game of Life’s design level is a real ontology that permits risky but vastly cheaper prediction | informal argument | Poundstone’s constructions, pp. 38–41 |
| C5 | Folk psychology describes a real pattern even if nothing in the brain realises it in pure form | informal argument | the fig. 1 / fig. 4 contrast, Anscombe, Churchland’s “projection” image, pp. 43–45 |
| C6 | Two equally reliable interpretations can disagree with no deeper fact settling which is right | informal argument | extension of Quine’s indeterminacy; the Ella case, pp. 46–49 |
| C7 | Finding a predictive pattern is ipso facto finding a causal power | assertion | fn. 22, “Free Lunch” sign example |
| C8 | Fodor’s Realism and Churchland’s eliminativism are “gratuitously strong forms of materialism” | informal argument | p. 51 |
Method
Conceptual argument by graded examples: dot frames, then the Game of Life, then folk psychology. A formal definition of randomness (Chaitin) is borrowed and used informally. There are no theorems. The noise ratios and pixel counts are illustrations, not data.
Concepts
- Bit map: the verbatim, uncompressed description of the data, which is the baseline.
- Real pattern: C1.
- Stances: physical, design and intentional, ordered by predictive power and by computational cost (fn. 32).
- Mild realism: beliefs are real because a real, noisy pattern is measured by belief talk, understood as “indirect ‘measurements’ of a reality diffused in the behavioral dispositions of the brain (and body)” (p. 45).
Connections
It is the base of the real-patterns programme: Ross 2000 “Rainforest Realism” (not reachable), Ladyman & Ross 2007, r2 (Ladyman 2024), LIT-184 (Ladyman & Lorenzetti) and LIT-151 (Wallace’s instantiation-based reconstruction). It relies on Chaitin (1975), Sellars’s manifest image, Quine’s indeterminacy and Churchland’s projection metaphor. The 2026 MIT volume Dennett’s Real Patterns in Science and Nature reprints the paper as its ch. 1 (per Crossref: 10.7551/mitpress/15550.003.0005, pp. 33–58). That version was not read.
Bearing on the record
It supplies the primary text for the “Dennett-style real patterns” that rainforest realism builds on. It lets the record separate Dennett’s lossless, observer-free compression criterion from the projectibility-and-perspective version of Ladyman and Ross. It carries nothing for ML practice. The compression criterion resembles minimum-description-length ideas, but the paper makes no claim about learning systems.
Limitations
- The criterion is stated for finite data frames with a fixed encoding. How to fix the baseline “bit map” for continuous or open-ended data is not addressed. Ladyman–Ross’s later scale- and perspective-relative reworking exists to fill that gap.
- The definition makes almost every non-random dataset contain real patterns. It does not rank them except through pragmatic trade-offs, which is by design (C3, C6) but leaves realism thin.
- The argument for the reality of intentional patterns (C5) is by analogy with the visual examples. The empirical premise, folk psychology’s predictive success, is assumed.
- The paper is not a general ontology. Dennett is “rightly sceptical” of metaphysical uses of it, as r2 (p. 1) records.
Open questions
- Does compressibility “whether or not anyone can concoct it” make real patterns observer-independent after all, and if so what remains of the “link to observers”?
- How should lossy, statistical compression (r2 thesis 14) be scored against the lossless bit-map baseline?
- When two real patterns conflict (C6), is either one’s ontology real, or only the data?
Corrections to the seeded skim
- There is no dossier or skim, so these are corrections to the record’s existing accounts.
- What Dennett’s criterion is, exactly. NOTE-038 (on LIT-045) glosses a real pattern as “an entity … that features non-redundantly in projectable regularities … often glossed as compressibility”. NOTE-129 (on LIT-184) glosses it as “a compression criterion relative to physically possible perspectives”. Dennett’s own criterion is narrower and plainer. A pattern is real iff some description of the data is more efficient than the verbatim bit map, “whether or not anyone can concoct it” (p. 34). It is an existence claim about a shorter description, not about any perspective actually having it. The “projectibility” and “physically possible perspective” clauses are later additions, not Dennett’s. On r2 (fn. 5, thesis 15) and NOTE-129 they belong to Ross 2000 and Ladyman–Ross 2007. That attribution was not checked against Ross 2000, which was not reachable. So the record should not attribute them to Dennett 1991. Dennett does tie patterns to observers: a pattern “is ‘by definition’ a candidate for pattern recognition”, which is a “loose but unbreakable link to observers or perspectives” (p. 32). But that link is an intuition he makes precise with the observer-free compression test.
- Real patterns in Dennett are pluralist and can conflict. Two schemes can both be real patterns in the same data while disagreeing about “which parts of the world were pattern and which were noise”, “and yet nothing deeper would settle the issue” (p. 49). The choice “would indeed be up to the observer, a matter to be decided on idiosyncratic pragmatic grounds”. None of the record’s OSR accounts (LIT-045, LIT-184, LIT-151) mentions this. It is the feature that makes Dennett’s view “mild” realism, and the one rainforest realism has to domesticate (r2 thesis 15; LIT-151’s “best precisification” is one way of doing so).
- The Game of Life’s role. Dennett uses it as a “stepping-stone” between dot frames and folk psychology (p. 37), not as a model of physics. r2 (p. 3) lists its disanalogies with real science. A reader of both should note that the lossless, one-level structure Ladyman criticises is exactly why Dennett chose it: “nothing is hidden from view … no hidden variables” (p. 38).
- Causation. Some readers took Dennett to make beliefs epiphenomenal. Footnote 22 (pp. 43–44) says that finding a predictive pattern “has ipso facto discovered a causal power”. This anticipates r2 thesis (13)‘s higher-level difference-making causes.
- LIT-170, LIT-173, LIT-196 and LIT-152 are not engaged. The paper predates OSR. Floridi’s informational objects (LIT-152) and Dennett’s compressible patterns are separate uses of information theory, and neither cites the other here.