Claude finished the response ---------------------------------------------------------------- summarize ronald loui's review of brian smith's objecs ---------------------------------------------------------------- 1:56 PM I don't have any context about "Ronald Loui's review of Brian Smith's Objects" -- this isn't something I recognize, and I want to make sure I'm finding the right thing before summarizing. Searched the web Searched the web Found it. This is Ronald P. Loui's 1998 review (in Artificial Intelligence, vol. 106) of Brian Cantwell Smith's book On the Origin of Objects (MIT Press, 1996). Let me pull the actual review text to give you a proper summary rather than relying on secondhand commentary. Fetched: On the Origin of Objects (towards a philosophy of computation) Fetched: On the Origin of Objects (towards a philosophy of computation) This is Ronald P. Loui's 1998 review of Brian Cantwell Smith's book On the Origin of Objects (1996), published in Artificial Intelligence 106(2): 353-358. The book itself is Smith's attempt to build a comprehensive theory of computation, which ends up pulling him into full-blown metaphysics. His core claim, as Loui summarizes it, is that there's one real world but our ways of symbolizing it seem arbitrary, and he treats this gap between "metaphysical monism" and "ontological pluralism" as a problem needing a fix. Smith's proposed solution is a "middle distance," a pre-theoretic layer of "registration" that lets an observer (human or AI) make contact with the world before language or symbolic reference enters the picture. Blogger Loui's reaction is sharply two-sided. He calls it an important, even beautiful book that confirms Smith as one of the deepest thinkers in computing, but also found it intellectually uninteresting and frustrating to read. His complaints break into two parts: first, that the project is purely metaphysical with no practical payoff, unlike epistemology, which actually has to ground real criteria for knowledge and belief; and second, that Smith's writing style (which Loui likens to Nietzsche's) buries the argument rather than stating it, making the book a poor match for an AI audience expecting more directness. Blogger Blogger Despite the criticism, Loui's closing line is essentially backhanded praise: Smith's book had, for the moment, made him computing's leading philosopher, with the important caveat that the philosophy of computing shouldn't be confused with the philosophy of mind. ---------------------------------------------------------------- but registration is clearly representable in many formal and informal language. the only question is whether one can see the snow without having the many terms for it. clearly one can. though having concepts helps see the 16th kind from the 17th. and if you don't have the concept at all, you end up saying things like plinian magma is a pillar of fire and smoke. a category error. so there are concepts and then there is syntax. ---------------------------------------------------------------- 2:00 PM You're making a point that actually cuts pretty hard against Smith's whole setup. If "registration" -- the pre-theoretic contact with the world before symbols enter -- is something we can already represent formally (truth-conditional semantics, situation theory, even just indexicals and demonstratives in ordinary language), then it's not clear what work the "middle distance" is doing as a novel metaphysical layer rather than just an underdescribed semantic phenomenon dressed up in new vocabulary. Loui's complaint about Smith not getting to the point may partly be this: the formal tools to talk about pre-conceptual reference already exist (think Kaplan on demonstratives, or Barwise and Perry's situation semantics, which long predates Smith's book and is squarely about an observer registering a situation without yet fixing it under a description). Your snow example draws the line precisely. The Sami having dozens of terms for snow types doesn't mean a non-Sami speaker can't perceive snow, or even perceive a particular snow's qualities; it means the non-Sami speaker is now individuating less finely, lumping what could be split. That's concepts doing discrimination work on a perceptual manifold that's available either way. It's a quantity-of-categories problem, not an access problem. The Plinian magma case is the more interesting failure mode, though, because it's not just coarse-graining, it's a category error: describing a thing using the wrong ontological kind ("pillar of fire and smoke" for what's actually pyroclastic density current, ash, and superheated gas) rather than just a cruder bucket within the right kind. That suggests two distinct ways concepts interact with registration: (1) resolution -- how many bins you cut the continuum into, which doesn't gate perception, and (2) kind-assignment -- what basic sortal you're slotting the phenomenon into, which can be simply wrong and actively mislead inference (people died at Pompeii partly because "fire and smoke" suggested escape strategies suited to a structure fire, not a several-hundred-degree gas surge). If that's right, Smith's "registration" prior to symbols would need to already do sortal work to be useful at all (you have to register that's a magma cloud, not smoke to act correctly), which means it can't actually be prior to conceptual content in the way he wants. The "level" he's positing may be doing covert conceptual labor while claiming to precede concepts. ---------------------------------------------------------------- probably situation semantics and demonstratives are a stretch there. i would look at scientific theory terminology. it shifts, it emerges, it gets used or unused. we communicate in different theoretical language with different entailments (consider a queuing model of humans at an airplane gat vs. a thermodynamic model or even an economic model). we hold onto pre-linguistic phenomena prior to any one of those descriptions. if you come up through ontology thinking in logic, it's a magical level, that pre-theoretical phenomena. but if you like animals, it's no mystery. rabbits understand threat vs non-threat, and may even speak rabbit-language, a body language no doubt, to baby rabbits. but threat-with-intent vs threat-by-accident seem to escape them, predicate or not. just not in their conceptual discernment. language has nada to do with this. perhaps there is a language of thought, but if so, there is also a pre-language of perception. ---------------------------------------------------------------- 2:09 PM Right, that reframes the whole thing much better -- the queuing/thermodynamic/economic-model case is the real test, because there the same airport-gate phenomenon supports multiple incompatible formal vocabularies, each with different entailments (a queuing model gives you wait-time distributions and server-utilization, a thermodynamic model gives you something like pressure and flow under aggregation, an economic model gives you incentives and willingness-to-pay for priority boarding), and none of them is "the" description the phenomenon was secretly waiting for. The phenomenon doesn't individuate itself into one of these; we choose the formalism for the entailments we want. That's a genuinely different kind of pre-theoretic than "pre-linguistic perception," and it's the one that should worry a philosopher of computation, since computation is exactly the business of picking a formal theory and propagating its entailments. And the rabbit case dissolves the mystery on the other end. If a rabbit can register threat vs. non-threat, and signal it, but cannot register threat-with-intent vs. threat-by-accident, the limitation isn't syntactic, there's no missing word the rabbit is groping for. It's a representational/inferential ceiling: intent requires modeling another agent's internal states as a separate variable from the outcome, and that's a cognitive architecture problem, not a vocabulary gap. So "pre-theoretic" registration is just perception and primitive inference running in whatever representational format the organism's brain uses, no predicate calculus required, but also no special metaphysical "middle distance" required either. It's just cognition before it's verbalized, which is unmysterious to anyone who's spent time with animals or, frankly, with prelinguistic children solving object permanence tasks. So you get two genuinely separate things that Smith's "registration" seems to be running together: The animal/perceptual case is just non-linguistic cognition, no metaphysics needed, it's an empirical question about what inferences a given nervous system can support. The scientific-theory case is the more profound one, because the "pre-theoretic" phenomenon there isn't sub-symbolic perception, it's a phenomenon that's already complex and structured enough to admit multiple competing symbolic theories, and underdetermination among them is a real fact about scientific representation (this is basically Quine and van Fraassen territory, modulo the constructivist flourishes). But that's a story about theory choice and pragmatics of formalization, not about some primordial register that precedes objecthood. If Smith collapses those two into one "level," that would explain why Loui found him hard to pin down: he's gesturing at something true in both domains while implying they're the same insight, when one is biology and the other is philosophy of science. ---------------------------------------------------------------- i would accept "cognitive architecture" limit only under protest that it can change. you can learn the difference between aa and pahoehoe magma flows. but that's not a question of what "a given nervous system can support" unless you mean that in terms of perceptual refinement through neural reorganization. maybe that's apt. but it's speculative. in fact, when i pull dandelions each spring, i do reorganize my perceptual acuity (seems pre-linguistic too!). but my nervous systems supported that last spring and the spring before that. so don't tell me i lacked the capacity all winter. just didn't care. ---------------------------------------------------------------- 2:14 PM That's a fair correction, and it catches a sloppy move I made: "what a nervous system can support" smuggled in a capacity claim when what actually changed was attention and salience, not hardware. You had the wiring all winter. What you didn't have was a standing perceptual habit of sorting dandelion-rosette-versus-grass-blade at a glance, and that habit only forms under repeated, motivated engagement -- pulling them, caring whether you got the root, noticing the ones you missed last week. Nothing about your visual cortex was incapable in January. It just wasn't running that discrimination because there was no task recruiting it. That's actually a cleaner case than the AA/pāhoehoe one, because it's not even conceptual learning in the sense of acquiring a new category from instruction. You didn't read a glossary of dandelion-rosette morphology. The discrimination sharpened through what's sometimes called perceptual learning -- closer to how a birder starts seeing field marks they couldn't pick out a season before, or how a chick sexer's hands learn a tactile distinction nobody can fully verbalize. It's pre-linguistic in your sense, prior to any predicate you'd assert, but it's also clearly acquired, which is the part that makes "cognitive architecture limit" the wrong frame entirely; architecture suggests a ceiling fixed by the kind of nervous system you have, when what's actually flexing is tuning within a system that was already capable. So the rabbit case and the dandelion case may not be the same kind of limit after all. The rabbit's failure to register intent-versus-accident might be a real architecture ceiling (no representational slot for another agent's hidden mental state, full stop, no amount of motivated attention fixes it), while your dandelion case is pure perceptual learning with no ceiling in sight, just disuse. If so, "cognitive architecture" should be reserved for the rabbit-style cases and kept away from cases like yours, where the right description is closer to skill acquisition or perceptual tuning, something that was always within reach and just hadn't been exercised. That actually sharpens the original point rather than undoing it. Pre-linguistic discrimination is real and graded, not gated by language, but it comes in at least two flavors: hard ceilings set by what the system can represent at all, and soft ceilings set by what's been trained up through use. Smith's "registration," if it's meant to cover both without distinguishing them, is flattening a distinction that actually matters. ---------------------------------------------------------------- certainly though my perceptual discrimination and granularity are improved by linguistic frameworks and exchange. by the way, let me help you out. (well, gemini's assist) ---------------------------------------------------------------- AI Overview ‘A‘ā and pāhoehoe are the two primary types of basaltic lava flows, both named from the Hawaiian language. Pāhoehoe (pronounced pah-hoy-hoy) features a smooth, billowy, or ropy surface, while ‘a‘ā (pronounced ah-ah) is characterized by a rough, jagged, and clinkery surface. [[1](https://www.nps.gov/articles/000/lava-flow-forms.htm), [2](https://pubs.usgs.gov/gip/hawaii/page37.html)] 2:18 PM Good, that confirms the framing rather than complicating it: the smooth/ropy versus rough/clinkery distinction is something you could in principle see and feel without ever hearing the Hawaiian terms (lava that drapes and folds looks and behaves visibly differently from lava that crumbles and tears), but having the two words lets you fix the distinction, transmit it, and then refine it further than solo perception probably would on its own. That's the loop you're pointing at: pre-linguistic discrimination is real and gradable, but language doesn't just label an already-finished perceptual category, it feeds back and sharpens the category, the way naming pāhoehoe's ropy texture might make you start noticing gradations of rope-tightness you'd have lumped together before you had the word. That's a more interesting two-way relationship than either "concepts are required for perception" or "concepts are irrelevant to perception." Your dandelion case showed perceptual learning running with zero linguistic input. The lava case shows the reverse current: language as a tool that recursively retunes perception once it exists, probably by giving you a stable category to test new instances against, which is exactly the discrimination-by-comparison-set mechanism that lets a geologist eventually split aa into sub-types a layperson would never separate. So registration, if it's a real pre-theoretic layer at all, seems to be something that linguistic and conceptual frameworks can reach back and restructure, not just report on. Smith might have had a version of that intuition. The trouble, per Loui, is that whatever Smith's "middle distance" is supposed to explain, it's not obviously doing more work than the ordinary, fairly well-understood story of perceptual learning interacting with category acquisition, just relabeled with heavier metaphysical machinery. ---------------------------------------------------------------- another example that is not binary but noisy: mbti categories. some don't see them. s types in particular. others see them clear, infjs famously. but most can see the i vs e axis. and other axes with margins of error or wide confidence ranges. that's not an ideal revision of conceptual space but it may yet be useful in probabilistic inference. most observations come with error, and you have to deal with them. loui loves kyburg's framework for theory-formation, -revision, and -use-under-probabilistic-acceptance. quine+carnap+reichenbach. ---------------------------------------------------------------- 2:22 PM That's a genuinely different register from the lava case, and it's worth being precise about why. AA versus pāhoehoe is noisy at the boundary (some flows are transitional, "transitional lava" is a real category), but the underlying physical variable, viscosity and cooling rate, is doing real causal work and the categories track something mind-independent. MBTI axes are something else: they're a forced low-dimensional projection of personality variation that the field's own best evidence (the Big Five literature, factor-analytic work going back to Costa and McCrae) suggests isn't carving real joints at all, especially on dichotomous axes like S/N where the underlying trait is probably continuous and not even bimodal. So when you say S types often don't "see" the categories while INFJs famously do, that's not just measurement noise around a real signal, it may be partly an artifact of which axis happens to align with a trait that has more natural bimodality (E/I shows up reasonably reliably in factor analyses as something like a real continuous dimension with decent self-recognition) versus an axis that's imposing a binary on what's actually a unimodal continuum (S/N, where forcing two types onto one bell curve will obviously produce a lot of people near the middle who don't recognize themselves in either pole). That's exactly the setup Kyburg's framework is built for, though, and I think your instinct to bring him in here is sharp. Kyburg's whole project (the Lottery Paradox, his work on evidential probability and statistical inference) is about how to rationally accept claims that come with explicit error bars, without collapsing into either dogmatic acceptance or total skepticism, and the apparatus is exactly suited to "I'm 70% confident this person is an introvert based on three correlated but imperfect indicators." That's a different epistemic object from a clean kind-membership claim like "this is pāhoehoe." You're not discovering a sharp natural kind with fuzzy edges; you're doing probabilistic classification against a model you already suspect is a low-resolution compression of the real, higher-dimensional trait space (which is presumably closer to the Big Five, or whatever the next refinement turns out to be). So if Loui is drawn to Quine-Carnap-Reichenbach-Kyburg as the right machinery, I'd guess his objection to Smith isn't "metaphysics is bad" so much as "you've gone looking for a single foundational layer (the middle distance) where what's actually needed is a theory of how we accept, revise, and act on imperfect statistical models of a world that may not have the joints our categories assume." That's a research program with actual content (acceptance rules, confidence levels, when revision is warranted) rather than a metaphysical posture. It also explains the AI-relevant complaint underneath Loui's review: for someone doing knowledge representation and uncertain reasoning, "objects" aren't a metaphysical puzzle to be resolved once, they're provisional commitments held under a confidence level that you're prepared to revise when the next observation comes in. Smith's book, on that read, is solving a problem (how do objects get their unity) that working AI researchers had already moved past, in favor of the more tractable problem of how to reason well despite never getting that unity guaranteed. ---------------------------------------------------------------- well i would have a lot of energy to defend mbti from a method view, not just a sordid and sociology of psych profs view. but you point out that epistemology (mbti) and ontology (pahoehoe) are graded in different ways. the former because errors mount, esp when fitting putative observation to theoretical constraints of meaning and inference. the latter because the boundaries are fuzzy because we want 9 icd codes per group, not 900. convenience and economy come into play in language choice. there's actually a mix of epistemics and ontics: too many categories, no error, but less inference; too few categories, too many boderline coin-flip assignments, too many errors as well. can you see ron loui's paper on medical records and annotating with asterisks, hashtags, marginalia, and parentheticals when epistemological or ontological doubt is present? ---------------------------------------------------------------- 2:29 PM Searched the web Searched the web That search came back empty for what you're describing. Let me try a more targeted search since this is likely a specific Loui paper, possibly in a medical informatics or AI-in-medicine venue rather than general AI. Searched the web Searched the web Found it. The paper is titled "An Easily Adopted Markup Discipline for Annotating Electronic Medical Records with Ontological and Epistemological Qualification: Toward Disciplined Asterisks, Parentheticals, Marginalia, Footnotes, and Hashtags." academia It's exactly the synthesis point you were making, made concrete as a practical proposal. Loui's premise is that structured data and controlled vocabularies are common aspirations in medical records, but getting clinicians to actually adopt them widely and eagerly probably requires something more modest and incremental rather than a wholesale imposed ontology. academia His move is to split the markup into two separate jobs rather than one. The first is a light, voluntary discipline of annotation using simple, dynamic text affordances that clinicians could adopt without retraining. The second, and more interesting one, is to discipline the meta-vocabulary instead of the clinical vocabulary itself: marking up epistemological attitudes (how probable a claim is, the degree of uncertainty, where the claim came from) separately from ontological criteria (how fuzzy or vague a category is, which version of a definition is being used, which reporting standard applies). That's your S types not registering an axis versus pāhoehoe's fuzzy real boundary, made into an annotation schema: don't conflate "I'm not sure this is true" with "this category itself is poorly carved." academia The motivating case is exactly your mounting-errors point: medical records typically have multiple authors contributing at different times under different criteria, so a summary view assembled from all of that is liable to silently average over incompatible epistemic and ontic statuses unless the doubt is flagged at the point of entry. He also flags the ICD-code-style economy problem directly, framing the asterisks-and-hashtags approach as a hedge against the all-or-nothing choice between staying unstructured and imposing a heavy, one-shot ontology: since it may be too difficult to impose ontologies on so many different record-keepers all at once with no room for customization, a more organic, evolutionary path is to start by getting people to agree on shared annotations instead. academia academia So the paper is basically Kyburg's acceptance-under-uncertainty framework operationalized as punctuation. Instead of forcing every clinical claim through a fixed-confidence, fixed-category structured field, you let the asterisk carry "I'm not confident in this," the hashtag carry "this is using definition-version X," and the parenthetical carry the provenance, so the categories-versus-confidence problem you raised with MBTI gets handled the same way in medicine: keep the few necessary categories (your ICD-9-codes-per-group economy point) but attach a cheap, low-friction sidecar that records exactly where epistemic or ontological doubt is hiding, rather than pretending the structured field already resolved it. ---------------------------------------------------------------- it is actually philosophy of language, epistemology, and ontology applied to a real problem. ---------------------------------------------------------------- 2:32 PM Right, and that's what separates it from Smith's book, which is also philosophy of language/epistemology/ontology, but stays at the level of asking what objects fundamentally are. Loui's paper takes the same three areas and asks a much more answerable question: given that clinical claims arrive from multiple authors with different confidence levels and different category schemes, what cheap notational device lets you keep that information from collapsing into a single undifferentiated assertion. The philosophy-of-language part is the choice of marker as speech-act modifier. An asterisk or hashtag isn't adding content to the claim, it's doing something closer to what hedges and evidentials do in natural language (the difference between "the patient has pneumonia" and "the patient apparently has pneumonia, per the night resident, unconfirmed by imaging"). Most clinical documentation throws that away under the pressure of brevity and structured fields, and Loui's proposal is to keep the modifier without forcing full natural-language hedging back into the note. The epistemology part is the asterisk/confidence-level slot, which is just your MBTI point turned into infrastructure: don't force every claim to the same level of acceptance. Kyburg's whole framework exists because rational belief under evidence isn't binary, and a flagging system that distinguishes "established," "probable," "rumored," "contradicted-by-later-note" is evidential bookkeeping made cheap enough that a tired resident will actually do it. The ontology part is the hashtag/version-tag slot, which is your pāhoehoe point: which definition of "diabetes" or "sepsis" is in play, since these categories themselves get revised (sepsis criteria changed substantially between Sepsis-2 and Sepsis-3, with real clinical disagreement about which cases counted under the old versus new definition). Without a tag, a downstream reader or algorithm doing outcomes research can't tell whether two notes calling something "sepsis" are even using the same category. The thing that makes it more than a notation scheme, and genuinely philosophy doing work, is that it refuses to pretend those two kinds of doubt are the same kind of doubt. A structured EMR field with a dropdown and a confidence percentage tends to flatten "I'm uncertain whether this is true" and "I'm using a category that's inherently fuzzy or contested" into one number, which is precisely the conflation you flagged a few turns back when you separated epistemics-mounting-errors from ontics-driven-by-economy-of-categories. Loui built the distinction into the markup itself rather than leaving it implicit, which is the part that would interest someone who'd spent a dissertation under Kyburg thinking about exactly this gap.