> [!info] Legacy > New remembered items are created as individual tagged notes in `Notes/`. See [[Recent Captures.base]] or the Project Captures view on [[Sessions/Generating Philosophy]]. This note is preserved as historical archive. ### 2026-03-09 — Four Pigliucci claims with implications and open problems > 4. Candidate claims from Pigliucci > Here are four candidate claims, each with the three-part format. > > Claim 1 > Text says: > Philosophy is concerned with the world, and its starting points are empirical data from experience and science. > > Possible implication: > The anti-LLM objection can be sharpened. The issue is not just whether LLMs can imitate philosophical prose, but whether they can inherit or generate the worldly constraints from which philosophical inquiry begins. > > Open problem: > Are philosophy’s operative empirical starting points available in sufficiently articulated textual form, or does philosophy depend on a kind of experience that text cannot transmit? > > Claim 2 > Text says: > Philosophy is not best understood as producing “theories” in the scientific sense, but “accounts.” > > Possible implication: > This may help detach philosophical evaluation from overly science-shaped expectations like prediction, convergence, and decisive empirical testing. It may also support a more artifact-level conception of philosophical assessment. > > Open problem: > What exactly distinguishes a good “account” from a bad one? Pigliucci hints at this, but he does not provide a worked-out evaluative theory in the chapter. > > Claim 3 > Text says: > Philosophy can make progress through the refinement of aporetic clusters rather than convergence on one final answer. > > Possible implication: > Lack of full consensus need not show failure. A discipline may improve by eliminating bad positions, sharpening surviving ones, and mapping the structure of disagreement more clearly. > > Open problem: > How stable is the line between principled pluralism and an uncomfortable near-relative pluralism? Pigliucci wants the former, but whether he fully secures it is not obvious. > > Claim 4 > Text says: > Philosophy is closer to mathematics and logic than to science in that it explores conceptually evoked spaces, but unlike mathematics and logic it remains constrained by the world. > > Possible implication: > This gives you a powerful intermediate characterization of philosophy: neither empirical science nor free invention. That in turn may be useful for framing what an LLM would have to be doing if it were to count as producing philosophy. > > Open problem: > Can an LLM genuinely participate in empirically informed evocation, or only in secondary articulation and recombination within already-evoked spaces? > > If you want, the next useful step would probably be one of these two: > > I can take just Claim 1 and work through the strongest version of the objection and the strongest possible reply. > I can take these four claims and sort them into “promising for the paper,” “interesting but risky,” and “probably a distraction.” --- ### 2026-03-09 — Zahavy's E→A jump restated in Pigliucci's terms > That seems to be the strongest connection to Zahavy in particular. His E→A point can be read, in Pigliucci-ish vocabulary, as saying: if axioms or starting points come from experience, then a system with no experience may be unable to originate the relevant sort of philosophical move. --- tags: - integration-queue - generating-philosophy created: 2026-02-08 --- # Integration Queue — Generating Philosophy Passages and ideas to work into the project. Banked from conversations via /remember. > [!warning] Claude: Nothing here is endorsed > These are working materials — ideas being explored, not positions being held. Do not infer that Nick believes, endorses, or is committed to any claim in this queue. Entries capture thinking-in-progress for later evaluation. Treat everything as provisional and unranked. --- ### 2026-02-08 — Gaut fn. 23 — mechanically generated metaphors still guide audience, supports provenance irrelevance > **Gaut — footnote 23 is gold.** He concedes that even mechanically generated metaphors: > >> "would still guide their audience imaginatively to link together two domains, and if the metaphors were successful, to discover original and apt connections between them and perhaps to elaborate the metaphors further. They would thus guide those who understood them through a process akin to the process of creative imagination that could have, but did not, produce them." > > The output's *structure* does real cognitive work for its audience regardless of production history. This directly supports provenance irrelevance and the appearance-reality collapse: if an argument guides a competent reader to genuine philosophical insight, it has performed its function. Gaut also separates good chess from creative chess — Deep Blue plays objectively good moves that aren't creative moves. That's exactly the parallel you need for LLM philosophy. > > His account of metaphor value is audience-directed: "A good metaphor doesn't so much prompt thought, as guide thought... and its standard of success isn't the volume of thought it causes to gush from us, but the quality of that thought." Quality of guided thought, assessed from the artefact. Not quality of production process. --- ### 2026-02-08 — Zahavy's two concessions: deductive capacity granted + domain-specificity admission restricts critique to physical sciences > Yet Zahavy makes a concession that deserves attention. He grants that modern LLMs could plausibly handle the deductive phase of the scientific process — the phase in which, given a set of foundational axioms, one derives their consequences: > >> "we posit that a modern LLM, initialized with the specific physical assumptions available to Einstein in 1915, could plausibly derive General Relativity" (Zahavy, 2026) > > This is not a trivial concession. Zahavy points to AlphaProof's silver-medal performance on International Mathematical Olympiad problems as evidence that LLM-adjacent systems are already capable of substantial deductive work, with successor systems reaching gold-level performance. The picture that emerges is selective: LLMs can reason deductively, perhaps impressively so, but they cannot perform the abductive leap that provides deduction with its premises. The question this raises for philosophy — a question Zahavy does not ask — is whether philosophical work falls on the deductive or the abductive side of this line, or indeed whether the line maps onto philosophical practice at all. That question will need to be addressed directly. > > There is a further concession, and it is the one that matters most for the argument of this paper. In his conclusion, Zahavy restricts the scope of his critique: > >> "we emphasize that this proposal is specifically tailored to the physical sciences, where the object of study is external material reality. In abstract domains such as Mathematics or Computer Science, the Sense Experience (E) may be grounded in high-dimensional topology or have other goals such as generality or minimality." (Zahavy, 2026) > > Zahavy does not mention philosophy, but the restriction is telling. His architectural bottleneck — the requirement for embodied simulation to ground the E→A Jump — applies where the object of study is external material reality. In the physical sciences, the foundational intuitions that drive abductive leaps are intuitions about what it is like to be a body in a physical world: the feeling of free fall, the experience of acceleration, the phenomenology of spatial orientation. The E→A Jump, as Zahavy describes it, runs through embodied experience because the domain demands it. But in philosophy, the object of study is not external material reality. The materials of philosophical reasoning are concepts, arguments, inferential relationships, and the logical space of possible positions — materials that are, as Zahavy himself notes of mathematics and computer science, grounded differently. Whether this difference is sufficient to dissolve the architectural bottleneck is a question I shall return to. --- ### 2026-02-10 — Abduction as multiple phenomena — disaggregating Peirce, Lipton, Floridi, Williamson yields different LLM verdicts > We have at least three different conceptions of abduction across these texts: > > - **Peirce** (via Zahavy): abduction as hypothesis *generation* — the creative leap from surprise to candidate explanation > - **Lipton**: IBE as an inferential *process* with two stages, governed by explanatory virtues > - **Floridi**: abduction as a high-level reasoning *pattern* that can be mimicked by stochastic processes > > And Williamson mostly talks about "IBE" as a *method* — a way of evaluating philosophical theories — without committing to a detailed cognitive account. > > The question: Are these the same phenomenon? If not, which one are we asking whether LLMs can do? > > This isn't just a taxonomic exercise. The answer genuinely affects the LLM question: > > - If abduction is *Peircean generation*, LLMs probably can't do it (Zahavy's argument). > - If abduction is *Liptonian IBE* (generation + selection), LLMs might manage the selection part but not the generation part. > - If abduction is a *reasoning pattern* (Floridi), LLMs can reproduce the pattern without instantiating the reasoning. > - If abduction is a *method* (Williamson), then the question becomes whether LLMs can follow the method — apply theoretical virtues, respect evidential constraints — regardless of their internal process. > > Different conceptions, different verdicts. > > A potentially productive hypothesis: Maybe the apparent disagreements between these papers are partly *terminological* — they're each talking about something slightly different when they say "abduction." If we can disambiguate, we might find that they're all right about their respective targets and the disagreement dissolves. --- ### 2026-02-10 — Transitive calibration — can loveliness-truth tracking be inherited through training data? > When an LLM produces what looks like an inference to the best explanation, its "loveliness standards" aren't calibrated by successful past inferences in the way Lipton describes. They're calibrated by statistical regularities in training data. But — and this is important — the training data *encodes* the outcomes of millions of human inferences to the best explanation. So there's a question: is the calibration transferred? When humans write down their abductive conclusions, and LLMs learn the patterns, have the LLMs inherited the loveliness-truth tracking relationship, even if they haven't participated in it directly? --- ### 2026-02-10 — Loveliness encoded via training data — selection effects, transitive calibration, standards vs patterns, implications for Floridi and Voltaire > **The selection-effect argument.** Philosophical training data isn't a random sample of philosophical attempts. It's a *filtered* sample. Papers get published, taught, anthologised, and cited in rough proportion to their perceived quality — and "quality" in philosophy is substantially constituted by the theoretical virtues Lipton calls loveliness (elegance, unification, simplicity, explanatory power). So the corpus that LLMs train on is *enriched* for lovely explanations. Not perfectly — there's noise, there are fashions, there are bad papers that get cited for sociological reasons — but the signal is there. Lovely explanations are statistically over-represented relative to their base rate among all possible explanations. > > This means that when an LLM learns to produce philosophy-like text, it's learning from a sample that has already been *pre-filtered by human loveliness judgments*. The LLM doesn't need to have its own loveliness detector. The training data has already done the filtering. The LLM just needs to learn the *distribution* of what survived the filter. > > **The transitive-calibration question.** Lipton's answer to Voltaire is that our loveliness standards are calibrated by a feedback loop: we make IBE inferences, check them against evidence, and adjust our standards of loveliness accordingly. Over time, this feedback loop ensures that loveliness tracks truth (or at least: that our loveliness standards aren't wildly miscalibrated). > > Now, this feedback loop is *encoded in the corpus*. The philosophical tradition *is* the record of this feedback loop — centuries of philosophers proposing explanations, testing them dialectically, refining their standards, discarding what failed, building on what survived. When an LLM trains on this record, it's absorbing the *outcomes* of the calibration process. It hasn't participated in the process, but it's ingesting its results. > > The question is whether this is enough. An analogy: imagine a student who has never done any experimental science but has read every published paper in physics. They'd have excellent judgment about which hypotheses are considered well-supported, which experimental designs are considered rigorous, which theoretical virtues physicists actually deploy. They'd lack *first-person calibration* — they've never felt the surprise of a failed prediction, never experienced the feedback loop directly. But their judgment would be informed by the outcomes of everyone else's feedback loops. > > Is that student's judgment *reliable*? In many cases, probably yes — they'd make the same calls a seasoned experimenter would. But there might be edge cases where the *reason* for a standard matters, where understanding *why* simplicity is a virtue (not just *that* it is) makes a difference. And those edge cases might be precisely the novel, boundary-pushing cases that matter most for scientific progress. > > **The standards-vs-patterns distinction, developed.** This is the sharpest way to frame the question. Consider two models: > > *Model A:* The LLM has internalised something like a *norm* — "prefer simpler explanations" — and applies it as a criterion when generating outputs. It has learned the standard. > > *Model B:* The LLM has learned that certain argument structures (which happen to be simple) produce higher next-token-prediction scores because they're more frequent in approved philosophical text. It has learned the patterns that *result from* the standard without learning the standard itself. > > These are empirically very hard to distinguish, because they produce identical outputs in standard cases. The divergence comes in *novel* cases — cases where the standard needs to be *extended* to a new domain or *balanced against* competing standards in an unfamiliar way. If the LLM has the standard (Model A), it can generalise. If it only has the patterns (Model B), it can only reproduce what it's seen. > > But here's a complication: how many philosophical cases are genuinely *novel* in the relevant sense? Philosophical argumentation is highly conservative in its *forms* — the same moves recur across very different content areas (counterexample, distinction, reductio, analogy, dilemma). If these forms are what "loveliness" looks like in philosophy, and these forms are well-represented in training data, then Model B might be extensionally adequate even without genuine norm-internalisation. The forms transfer across content domains because they're the *same forms* — the LLM doesn't need to extend the standard to new territory because the territory isn't new at the level of argumentative structure. > > **What this means for Floridi.** Floridi's position is essentially that LLMs are stuck in Model B — they have patterns, not standards. His "zeroth-order abduction" and "stochastic core / abductive appearance" framings both insist on this. But if the argument above is right, Model B might be *sufficient for philosophy* in a way it isn't sufficient for empirical science, precisely because philosophical loveliness is structural. Floridi's diagnosis might be correct (stochastic pattern-matching, not genuine reasoning) while his *evaluation* of that diagnosis (therefore the outputs are unreliable) doesn't follow — because in this particular domain, pattern-matching gets you the right answer for structural reasons. > > **What this means for Voltaire's objection applied to LLMs.** Lipton's defence against Voltaire relies on the feedback loop: loveliness tracks truth because we've calibrated our standards through experience. For LLMs, the defence would be: loveliness tracks truth because the training data was produced *by agents who had calibrated their standards through experience*, and the calibration results are encoded in the distributional properties of the text. The LLM is borrowing its calibration rather than earning it. Whether borrowed calibration is sufficient depends on whether you ever need to understand *why* the standard works (in which case borrowed calibration fails in novel cases) or merely *that* it works (in which case borrowed calibration is fine as long as you stay within the distribution). > > And here again, philosophy might be special. The *reason* that simplicity is a virtue in philosophy (avoiding ad hocness, over-fitting, unprincipled epicycles — Williamson's point) is itself a *structural* reason, fully expressible in the same text that exemplifies the virtue. Unlike empirical science, where the reason simplicity tracks truth might ultimately be something about the structure of physical reality (not expressible purely in text), in philosophy the reason is itself part of the argumentative tradition. So even the "why" is in the training data. --- ### 2026-02-10 — Lipton two-stage framework as diagnostic — three questions, separability of generation and selection > Three questions instead of one: > > Can LLMs recognise when to abduce? (Floridi: no — over-abduction) > Can they generate genuinely novel candidates? (Zahavy: no — need embodied simulation) > Can they select among given candidates? (Probably yes — loveliness-ranking is learnable) > Open question: Is the selection stage genuinely separable from the generation stage, given that Lipton himself says the feedback loop connects them? For single inferences, probably yes. For sustained philosophical inquiry, maybe not. --- ### 2026-02-10 — Lipton's actual vs. potential explanation — LLM outputs as paradigmatic potential explanations, process irrelevance grounded in Lipton > Lipton distinguishes actual explanations (what causally produces belief) from potential explanations (what would explain the phenomenon if true). IBE should be understood in terms of potential explanation — we infer the hypothesis that would provide the most understanding if true. > > LLM outputs are paradigmatically potential explanations — they're hypotheses that would explain things if true, but nothing in the LLM's causal history constitutes genuine understanding. Floridi would seize on this: the LLM has potential explanations without actual understanding. > > But Lipton says it's potential explanation that matters for IBE evaluation. The ranking procedure cares about intrinsic properties (loveliness), not causal history. So the LLM's lack of actual understanding is irrelevant to the evaluation of its outputs. > > I think this point is underexploited and potentially quite powerful. It provides textual grounding from Lipton himself for the "who cares about process?" intuition. If the locus of evaluation is the potential explanation — the hypothesis as it stands, assessed by its loveliness — then the process that generated it genuinely doesn't matter for evaluative purposes. Floridi's insistence on the stochastic core is a claim about process; Lipton's framework evaluates product. --- ### 2026-02-10 — Sokal comparison — surface/depth distinction holds only where evaluative norms are impressionistic, not in analytic philosophy > The Sokal hoax did work — but in a domain (postmodern cultural studies) where the evaluative norms were, arguably, less publicly checkable. One way to read the Sokal affair is as evidence that the surface/depth distinction can hold in domains where the evaluation is impressionistic rather than argument-checkable. The Sokal hoax probably couldn't have worked in a top analytic philosophy journal, precisely because the referees would have checked the arguments. This isn't because analytic philosophers are smarter, but because the evaluative norms of the discipline are specifically the kind that operate at the artefact level — publicly checkable, argument-by-argument. --- ### 2026-02-10 — Unmasking LLM philosophy reduces to artefact-level critique — show the mechanism or accept the output > Which means: if someone tried to 'unmask' LLM-generated philosophy by saying 'this works by pumping intuitions rather than making valid arguments' — fine, show the mechanism. Identify the specific feature of the text that does the pumping. That's a philosophical contribution. And the LLM output either survives the critique or it doesn't. But the critique happens at the artefact level, same as always. --- ### 2026-02-11 — LLMs as the occasion for metaphilosophy — these questions arise NOW because of the technology + reflexive self-test `f81df7cf-13a5-43d0-a68a-665428a24bcb` > The paper should make clear early on that these metaphilosophical questions — about philosophy's relationship to its textual medium, about what makes philosophical work good, about whether the process behind a text matters for its evaluation — are arising NOW specifically because of the rise of LLMs. The technology is what forces the question. Philosophy hasn't needed to think about whether "the text IS the thing" until something non-human started producing texts that look like philosophy. The LLM is the occasion for the metaphilosophy, not the other way around. Also connected: the paper is itself a test of its own thesis (reflexive point) — a philosophical argument about philosophy's nature, advanced in a text. That self-referentiality should be noted, at least implicitly. --- ### 2026-02-12 — Dellsén's understanding is domain-general — nothing specifically "philosophical" about it that would resist LLM production `6d1dff2e-473e-4c54-a720-7471c1e85b1f` > This is actually a really important point for Nick's paper. If understanding is domain-general, and it's constituted by accuracy and comprehensiveness of dependency models, then there's nothing specifically about philosophical understanding that would resist LLM production. The understanding is the same kind as scientific understanding — just applied to philosophical phenomena. --- ### 2026-02-15 — Self-evidencing explanation gives "textual all the way down" a precise explanatory-theoretic articulation `38096cdf-0110-4fda-8a52-eb2e3f49e591` > From the Ch. 2 and Ch. 4 notes. Lipton's self-evidencing explanations are cases where the explanans explains the explanandum, and the explanandum provides the evidence for the explanans. The tracks-in-snow case: the tracks require explanation (the explanandum) and provide the evidence for the explanation (someone passed on snowshoes). The circularity is benign. > > The mapping: philosophy is pervasively self-evidencing. A philosophical text presents an argument; the argument explains why its conclusion holds; and the only evidence that the argument is good is the text itself. There's no laboratory result or physical observation that independently confirms the argument's force. The text is both the explanation and the evidence for the explanation's adequacy. > > The Ch. 2 note calls this "underexploited in the current drafts" and argues it would give "textual all the way down" something it currently lacks: a precise explanatory-theoretic articulation rather than just a metaphor. If self-evidencing explanations are "ubiquitous" and benign (Lipton), then an LLM that produces a self-evidencing philosophical text — one that presents an argument and simultaneously provides the textual evidence of its own cogency — is doing something explanatorily legitimate, not merely circular. --- ### 2026-02-15 — Squash analogy — levels-of-description fallacy, "just statistics" as confusing mechanics with technique `38096cdf-0110-4fda-8a52-eb2e3f49e591` > The squash analogy appears in the Ch. 7 note (Bayesian Abduction). Here's what happens: > > Lipton is addressing the objection that IBE must be wrong because Bayesianism is right — that since belief revision is governed by Bayes's theorem, there's no room for explanatory reasoning as a separate inferential engine. His response is the analogy: > >> "arguing that Inference to the Best Explanation is wrong because Bayesianism is right is like arguing that thinking about technique cannot help my squash game because the motion of the ball is governed by the laws of mechanics. Even if Bayesianism gave the mechanics of belief revision, Inference to the Best Explanation might yet illuminate its psychology" > > The point: the fact that a process can be *described* at one level (mechanics, probability) doesn't mean that a different description at another level (technique, explanatory reasoning) is idle or wrong. Both descriptions can be true simultaneously — they're operating at different grains of analysis. The ball obeys mechanics whether or not you think about technique, but thinking about technique genuinely helps you play better, and describing your game in terms of technique captures something real that the mechanical description misses. > > Now, here's why the Ch. 7 note flags this as interesting for the project. The note draws the following parallel: > > The most common dismissal of LLM philosophical competence runs: "LLMs are just doing statistical pattern-matching over token distributions, therefore they can't be doing anything genuinely explanatory or philosophical." This has the same structure as "the ball is governed by mechanics, therefore thinking about technique can't help your game." The fact that LLM outputs are *generated by* probability distributions over tokens doesn't mean that describing those outputs in terms of philosophical structure — as exhibiting explanatory virtues, tracking dialectical obligations, satisfying argumentative constraints — is an idle or mistaken description. The probability distribution is one level of description; the philosophical structure is another. Both can be true. > > The Ch. 7 note develops this further. It says (and I'm quoting the note's own analysis): > >> The structural parallel at the centre of this chapter's relevance to the project is this: Lipton argues that explanatory reasoning is a cognitive mechanism that *realizes* Bayesian probability updating, and LLMs are themselves systems whose outputs are determined by probability distributions over tokens. The question is whether Lipton's realization framework helps characterize what is happening when an LLM generates philosophical text. > > And later: > >> Lipton's compatibilism provides a framework for understanding how learned evaluative dispositions could inform probability-governed outputs without requiring that the system "do" explicit probabilistic reasoning or explicit explanatory reasoning in any introspectable sense. The system's probability distributions are the formal constraint; the learned patterns from philosophical text are the process that shapes those distributions into philosophically structured outputs. > > The Conclusion note also picks this up, calling the compatibilist position one of the book's three distinct resources for the project, and the most underexploited one: > >> If explanationist reasoning is a psychological realisation of Bayesian inference, then the fact that LLMs operate by statistical inference over token distributions does not automatically exclude them from the explanationist framework. The "just statistics" dismissal — which Floridi's critique relies on when contrasting "stochastic core" with "abductive appearance" — assumes that statistical processing and explanationist inference are categorically distinct activities. Lipton's compatibilism directly challenges this assumption. > > So the interest of the squash analogy for the project is threefold: > > - **It names a fallacy** — confusing levels of description and concluding that because one level is operative, another must be idle > - **It has a direct application** — the "just statistics" objection to LLM philosophy commits exactly this fallacy, treating the token-probability description as exhaustive > - **It comes from a philosopher with no AI agenda** — Lipton is making a point about the relationship between Bayesian formalism and human cognition, not defending AI. The analogy's applicability to the LLM case is a structural bonus, not a motivated construction > > The Ch. 7 note does flag an important disanalogy in its divergences section: Lipton's compatibilism is about *cognitive agents with beliefs*, and he moves freely between "degrees of belief," "inquirers," and "psychology." LLMs don't have beliefs in any uncontroversial sense. So the realization relation — explanatory reasoning as the psychological process that realizes Bayesian constraint-satisfaction — may not transfer straightforwardly to systems that lack a psychology. The note suggests the paper would need to either argue that the realization relation can hold without a traditional cognitive substrate, or restrict the analogy to the structural level (the *patterns* are analogous, not the *processes*). --- ### 2026-02-17 — GPT-5.2 physics breakthrough quotes — AI contribution to theoretical physics motivates the question about philosophy > ## GPT-5.2 Physics Discovery (Feb 2026) — Quotes for Introduction > > Paper: "Single-minus gluon tree amplitudes are nonzero" (arXiv: 2602.12176, 13 Feb 2026) > Authors: Alfredo Guevara (IAS), Alexandru Lupsasca (Vanderbilt/OpenAI), David Skinner (Cambridge), Andrew Strominger (Harvard), Kevin Weil (on behalf of OpenAI) > > ### What happened > An internally scaffolded version of GPT-5.2 spent ~12 hours reasoning through a problem in theoretical physics (gluon scattering amplitudes), independently arriving at a formula and producing a formal proof. The result shows that single-minus tree-level n-gluon amplitudes, long presumed to vanish, are nonvanishing for certain "half-collinear" configurations. The AI identified a regime the human physicists had not explored. > > ### Quotes > > **Andrew Strominger (Harvard)** — via Greg Brockman tweet, 13 Feb 2026: > "It is the first time I've seen AI solve a problem in my kind of theoretical physics that might not have been solvable by humans." > > Note: "chose a path no human would have tried" is widely attributed to Strominger across secondary sources but the primary source (OpenAI blog) couldn't be fetched to verify exact wording. > > **Nima Arkani-Hamed (IAS)** — via The Quantum Insider, 13 Feb 2026: > "The physics of these highly degenerate scattering processes has been something I've been curious about since I first ran into them about fifteen years ago, so it is exciting to see the strikingly simple expressions in this paper." > "To me, finding a 'simple formula' has always been fiddly, and also something that I have long felt might be automatable by computers." > > **Nathaniel Craig (UC Santa Barbara)** — via The Quantum Insider, 13 Feb 2026: > "This is clearly journal-level research advancing the frontiers of theoretical physics, and its novelty will inspire future developments and subsequent publications." > "This preprint felt like a glimpse into the future of AI-assisted science, with physicists working hand-in-hand with AI to generate and validate new insights." > > **Lupsasca (co-author, on Hacker News):** > The researchers "believed that a simple formula should exist but had not been able to find it despite significant effort." > > ### Why this is useful for the paper > 1. **Motivation**: Makes the question "can AI contribute to philosophy?" urgent rather than speculative — if AI is producing genuine results in *theoretical physics*, the question about other disciplines is live > 2. **Dialectical pressure on Zahavy**: Zahavy argues LLMs can't make the E→A Jump (from empirical to abductive). GPT-5.2 appears to have done something like that in physics — Zahavy's own domain > 3. **The paper's pivot**: Nick's argument is that philosophy's textual nature makes it *more* amenable than physics. So the physics result sets up: "even in the hard case, AI is making contributions; philosophy is the easier case for specific structural reasons" > 4. **Strominger's quote** is the money quote — "might not have been solvable by humans" is exactly the register needed for the introduction > > ### Sources > - OpenAI blog: https://openai.com/index/new-result-theoretical-physics/ > - arXiv: https://arxiv.org/abs/2602.12176 > - The Quantum Insider: https://thequantuminsider.com/2026/02/13/ai-scientist-spots-what-physicists-missed-in-gluon-scattering/ > - Brockman tweet: https://x.com/gdb/status/2022446135655436431 > - Hacker News discussion: https://news.ycombinator.com/item?id=47006594 --- ### 2026-02-23 — Philosophy's evaluative standards as internal to the practice — no external yardstick, encoded in corpus/training > What's distinctive about philosophy's evaluative standards is that they're internal to the practice. There's no external yardstick (like prediction success). What makes philosophy good is what competent practitioners recognize as good — and this is encoded in the corpus, in peer review, in graduate training. --- ### 2026-02-21 — Unpacking janus's "simulator" — what is being simulated, three nested readings, language as the special case > ## Unpacking janus's simulator definition > > ### "Simulator" as a framing choice with theoretical load > > janus tries other categories — agent, oracle, tool, genie, behaviour cloner — and finds them inadequate. "Simulator" evokes a specific structure: a rule that takes initial conditions and evolves them forward in time, producing processes distinct from the rule itself. > >> "It suggests an ontological distinction between the simulator and things that are simulated, and avoids the fallacy of attributing contingent properties of the latter to the former." (line 354) > > ### "Predictive loss" does the heavy lifting > > The model is trained to predict what comes next — not to maximise reward, not to fool a discriminator, but to match the actual transition dynamics of the training data. janus calls this the "simulation objective" because: if you can predict what comes next from any context, you can generate rollouts that are statistically indistinguishable from real data. That IS simulation. > >> "Guessing the right theory of physics is equivalent to minimizing predictive loss." (line 385) > > Training is the process of discovering the rules. Once you have the rules, you can simulate. > > ### "Invariant to data type" — formally true, semantically uneven > > The abstract structure — θ: T\* → ΔT, sample, append, repeat — is the same whether T is words, pixels, amino acids, or chess moves. But what the model must learn to predict well varies enormously: > > - Game states → game strategy > - Proteins → folding constraints, evolutionary pressures > - Code → syntax, semantics, design patterns > - Natural language → essentially everything that produces human text > > Language models are uniquely powerful because the scope of what they must model is unbounded. > > ### The "what is being simulated?" ambiguity > > Three nested readings: > > 1. **The distribution.** Token sequences obeying the same probability distribution as training data. Formally correct, explanatorily thin. > 2. **The generative processes.** The dynamics that produced the training data — human writing, reasoning, arguing. The model compresses the generative rule, not just the outputs. > 3. **Configurations / scenarios.** Specific situations animated forward from initial conditions (the prompt). The richest reading. > > These are levels, not alternatives. The model operates on tokens (1), but to predict well it must model the processes generating tokens (2), which involves modelling the world those processes are about (3). > > ### Where the levels collapse and where they diverge > > For non-language data types, the levels collapse. A protein model's tokens ARE proteins. A chess model's tokens ARE games. No representation gap. > > For language, the levels come apart. The tokens are signs pointing beyond themselves. A physics paper's tokens are not physics. This is the displaced reference problem (Jan's fn23), and it's why "what is being simulated?" is genuinely ambiguous for language in a way it isn't for proteins or chess. > > ### The deepest tension > > janus's invariance claim holds formally — same θ, same concept. But it conceals the fact that for language specifically, the simulation acquires an extra layer of semiotic complexity. The tokens are signs, the transition rule is an interpreter, and the trajectories are meaningful texts that refer to a world. The simulator concept is invariant; the nature of the simulation is not. --- ### 2026-03-02 — Comprehensive metaphilosophical survey from NotebookLM sources — 15 positions with AI compatibility analysis > ## Comprehensive Survey of Metaphilosophical Positions and AI Compatibility > > Survey derived from NotebookLM notebook "Generating Philosophy - Source Materials" (fde06bb2), containing 80+ sources including The Future of Philosophy chapters, Bengson et al.'s Philosophical Methodology, Williamson's Widening the Picture, Dellsén et al. on philosophical progress, Floridi on LLM reasoning, Zahavy's "LLMs Can't Jump," and others. > > ### COMPATIBLE (AI can do genuine philosophical work) > > **1. Enabling Noeticism (Dellsén, Firing, Lawler, Norton)** >> "[P]hilosophical progress consists in putting people in a position to increase their understanding, where 'increased understanding' is a matter of better representing the network of dependence relations between phenomena." > > Explicitly "for-whom rather than by-whom" — progress is about the public utility of the *product*, not the internal states of the producer. > > **2. Conceptual Analysis (Bengson et al.)** >> "Method of Analysis: When constructing a theory about a given domain, theorists ought to identify a set of theses about the domain that provide analyses of its central terms, concepts, or properties, where such analyses meet some sufficiently high standard." > > If evaluation concerns "properties of texts," LLMs can produce texts with those properties. > > **3. Evocation of Conceptual Landscapes (Pigliucci)** >> "...philosophy -- the way I see it -- attempts to clarify things, or to analyze in order to bring about understanding, not really to discover new facts, but rather to evoke rational conclusions arising from certain ways of looking at a given problem or set of facts." > > Maps directly onto "semiotic physics" where GPT explores "a world of semantic possibilities inferred and extrapolated from human linguistic traces." > > **4. Concept-Mapping (Hellie)** >> "...philosophy is more like mathematics than science: what philosophers do is map and develop conceptual space concerning the range of possible treatments of a topic, as per Benj Hellie's evocative description of philosophy as 'the neo‐natal intensive care unit of theory'..." > > LLMs are fundamentally statistical mappers of conceptual and semantic space. > > **5. Accessible Public Engagement (Cherry)** >> "Coming out of the shade is the act of philosophers leaving their philosophical bubbles. It entails: (1) making philosophical work accessible in form and in medium..." > > LLMs excel at stylistic translation and rendering complex academic text into accessible discourse. > > ### HOSTILE (AI cannot do genuine philosophical work) > > **6. Abductive Philosophy / IBE (Williamson)** >> "I propose that philosophy should use a broadly abductive methodology. Indeed, to some extent it already does so. I propose that it should do so in a bolder, more systematic, more self-aware way." > > Zahavy: "While Large Language Models have mastered the inductive compression of data and the deductive verification of theorems, they are structurally incapable of the abductive 'jump' required for scientific invention." > Floridi: "The model does not understand what an explanation is, but it produces text that follows the typical phrasing and structure of explanations." > > **7. Identity-Conferring Conversation (Jones)** >> "A necessary characteristic of a philosopher is that she has entered into an evolving dialogue or conversation that takes place within a community of individuals." > > LLMs have no personal identity, cannot form authentic values, are not community members. > > **8. Intellectual War of Values (Sorgner/Nietzsche)** >> "According to Nietzsche, philosophers are creators of values." > > LLMs have no biology, drives, or psychophysiology to express. > > **9. Science-Tethered Truth-Seeking (Boghossian & Lindsay)** >> "Philosophy has already worked its way to the correct set of rules for making sense of the world, and it named them 'science.'" > > LLMs "do not possess an inherent concept of truth or verification beyond what their training data provides." > > **10. Empirical Problem-Solving (Kamber)** >> "More than any other goal, seeking to solve philosophical problems is what sets philosophy apart from other disciplines." > > LLMs cannot gather novel real-world data — "confined to the logical deduction of existing texts." > > ### CONDITIONAL (Depends on specific factors) > > **11. Theoretical Understanding (Bengson et al.)** — Depends on whether AI or human must "fully grasp" the theory > > **12. Reflective Equilibrium (Bengson et al.)** — LLMs can simulate structural output without internal equilibrium > > **13. Model-Building (Williamson)** — LLMs can describe models but may not invent genuinely novel ones > > **14. Therapy (Wittgenstein)** — Hostile if therapy requires internal process; compatible if AI acts as therapist for human > > **15. Egalitarian Conversation (Green)** — Conditional on overcoming training data bias > > ### Key finding > Positions focused on *textual products* and *public utility* are compatible with AI philosophy. Positions requiring *internal states*, *embodied experience*, or *genuine abductive leaps* are hostile. --- ### 2026-03-02 — Full case against anti-AI argument from abduction — six prongs with textual evidence > ## The Case Against the Anti-AI Argument from Abduction > > The anti-AI argument from abduction has this structure: > > 1. Williamson says philosophy should use an abductive methodology > 2. Abduction requires generating genuinely novel explanatory hypotheses via an "abductive leap" (Zahavy's E→A Jump; Floridi's critique of "zeroth-order abduction") > 3. LLMs cannot perform this leap — they're stochastic pattern-matchers lacking embodied simulation and genuine understanding > 4. **Therefore:** LLMs cannot do Williamsonian abductive philosophy > > I've argued that **Premises 2 and 3 either don't hold or don't apply to philosophy** in the way the argument requires: > > ### Prong 1: Williamson's Method is Primarily Ranking, Not Generation > > Williamson's explicit criteria concern **evaluating** theories, not **generating** them: > >> "Apart from its relation to E, the more T has the **intrinsic virtues of a good theory**, the better (ceteris paribus). It should be elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated." > > And crucially: > >> "Of course, we rank only those potential explanations **that have been thought of**." > > The method operates on pre-existing candidates. Generation is presupposed but not specified. **Implication:** If LLMs can propose candidates (even via recombination from the corpus) and those candidates can be ranked using Williamson's criteria, LLMs participate in Williamsonian abduction. > > ### Prong 2: Zahavy's E→A Jump Doesn't Apply to Philosophy > > Zahavy is explicit: > >> "Finally, we emphasize that **this proposal is specifically tailored to the physical sciences**, where the object of study is external material reality. In abstract domains such as Mathematics or Computer Science, the Sense Experience (E) may be grounded in high-dimensional topology or have other goals such as generality or minimality." > > Philosophy's object of study is not "external material reality." It's concepts, arguments, logical relationships, the space of possible positions. **Implication:** The E→A Jump concerns grounding theoretical axioms in embodied experience of physical reality. Philosophy's starting points are grounded differently — in linguistic usage, conceptual relationships, the dialectical situation. These are textually available. > > ### Prong 3: Philosophical Novelty is Novel Content in Familiar Forms > > Walton: argumentation schemes are "the historical descendants of Aristotle's topics." The forms are ancient (counterexample, distinction, reductio, thought experiment, analogy). What changes is **content** — which specific counterexamples, which specific distinctions. > > **Example:** Gettier. Novel form? No — counterexamples to proposed analyses are ancient. Novel content? Yes — those specific scenarios. But the novelty is **combinatorial**. **Implication:** If "genuine novelty" means novel forms, philosophy rarely requires it. If it means novel content, LLMs can produce it through combinatorial reconfiguration. > > ### Prong 4: Process-Level Critiques Don't Entail Product-Level Failure > > Floridi: > >> "Internally, it involves random sampling guided by probabilities; externally, **it can produce answers that align with human reasoning norms**." > > Williamson's criteria are about the product: > >> "the more T has the **intrinsic virtues** of a good theory, the better" > > The Lipton/squash analogy applies: "arguing that Inference to the Best Explanation is wrong because Bayesianism is right is like arguing that thinking about technique cannot help my squash game because the motion of the ball is governed by the laws of mechanics." **Implication:** Floridi's process-level critique and Williamson's product-level criteria operate at different levels of description. Both can be true. > > ### Prong 5: Williamson's Cognitive Requirements Concern Navigation, Not Evaluation > > Williamson says abduction requires "a strong aesthetic sense" and "good judgment, honed by experience." But he's describing the capacity to **navigate toward** good theories. The aesthetic sense helps you **find** elegant solutions. It doesn't **constitute** the elegance. > > **Implication:** Even if humans need an aesthetic sense to find elegant theories, the elegance is a property of the theory itself, assessable from the text. If LLMs arrive at elegant theories via different means, the elegance is unaffected. > > ### Prong 6: Transitive Calibration > > The corpus encodes the results of millennia of human aesthetic calibration. LLMs trained on this corpus inherit the calibration transitively. And in philosophy (unlike physics), the *reason* that simplicity is a virtue is itself a *structural* reason, fully expressible in the text. The "why" is in the training data. > > ### The Bottom Line > > The anti-AI argument has **two steps**: > > **Step 1:** LLMs can't perform genuine abduction (generation of novel hypotheses via embodied experience). > → *Probably true* for Zahavy's strong, physics-specific sense. > > **Step 2:** Williamson's abductive philosophy requires this kind of abduction. > → *Probably false.* Williamson's method is primarily ranking using intrinsic textual properties. Zahavy explicitly restricts his argument to physical sciences. > > **So:** Zahavy/Floridi's critique and Williamson's methodology are likely **orthogonal**. They're talking about different things when they say "abduction." --- ### 2026-03-03 — Three layers of response to Floridi — text-internal evaluation, corpus saturation, levels-of-description > There are three distinct layers of response to Floridi's "stochastic core / abductive appearance" critique operating in the paper's materials. They're complementary but do different argumentative work: > > **Layer 1: Text-internal evaluation (Section 1's argument).** Even if Floridi is right that LLMs have a stochastic core, it doesn't matter for *evaluation*, because evaluation is text-internal. The standards concern the output — coherence, handling of objections, theoretical virtue. Whether the producer "really reasoned" is beside the point. This layer says: *grant the diagnosis, deny its significance*. > > **Layer 2: Corpus saturation / transitive calibration (Section 3).** The training data encodes the results of millennia of evaluative practice. The corpus is filtered by peer review, citation, and anthologisation — enriched for what Lipton calls "loveliness." The stochastic core isn't random; it's calibrated by the tradition's feedback loop. This layer says: *the stochastic core is better than you think, because it's been shaped by the very standards you care about*. > > **Layer 3: Levels-of-description / Bayesian compatibilism (Lipton Ch. 7).** The "just statistics" dismissal commits a fallacy — confusing the mechanical level of description with the philosophical level. Lipton argues explanatory reasoning *realises* Bayesian constraint-satisfaction (squash analogy). Statistical processing and explanatory reasoning are compatible, not competing. If explanatory reasoning is a cognitive process that realises formal probabilistic constraints in humans, then LLM outputs shaped by distributional patterns over philosophical text might realise philosophical structure in an analogous way. This layer says: *the stochastic core and the philosophical structure are not even in competition — they're at different levels*. > > Layers 1 and 2 are discussed in the Nick/Enrico transcript (3 March 2026). Layer 3 is in the integration queue (squash analogy entry, 15 Feb) but wasn't discussed as a distinct prong. > > **Structural decision needed:** whether to keep these integrated in a single response section, or separate them. The interleaved structure (Floridi's argument → all three layers of response; then Zahavy's argument → his specific responses) might be cleaner than the current draft's separation of objections from responses. --- ### 2026-03-05 — CEV: Theoretical virtues as latent structure in LLMs — full Section 4 development > ## Coherent Extrapolated Volition: Theoretical Virtues as Latent Structure in LLMs > > ### What the idea is > > The argument you are reaching toward has two parts. The first is an *encoding* claim: that the evaluative standards governing philosophical quality — what Williamson calls the "intrinsic virtues of a good theory" (elegance, unity, non-ad-hocness, combining simplicity with strength) along with the broader dialectical competences (handling objections, drawing distinctions at the right places, illuminating subject matter) — are embedded in the training data of an LLM, not as explicit rules, but as statistical regularities reflecting centuries of filtering by those very standards. The second is an *elicitation* claim: that these latent standards can be drawn out through prompting, and that the prompter's skill consists in writing text whose most probable continuation, given the virtue-shaped distribution the model has learned, is itself good philosophy. > > I think the right umbrella term is broader than "theoretical virtues" in the narrow Kuhnian sense (where it refers specifically to elegance, simplicity, fruitfulness, etc., in theory choice). What the paper needs — and what Sections 0–3 have set up — is the full set of evaluative standards that the philosophical corpus encodes: theoretical virtues in the narrow sense, *plus* dialectical competences (anticipating objections, repairing arguments under pressure), *plus* what Bengson et al. call "understanding-enabling features." The paper already draws on all three vocabularies. The right label for Section 4 purposes might be "the evaluative norms of the discipline" or simply "philosophical quality" — but "theoretical virtues" works as a synecdoche provided the reader understands it in the broad Williamsonian sense where abduction governs the whole of philosophical methodology, not just theory choice. > > ### The encoding claim: how the virtues get in > > The philosophical corpus is not a random sample. It is the output of a filtering process that operates at multiple levels: > > 1. Peer review selects for handling of objections, engagement with the literature, non-trivial contribution. > 2. Citation selects for arguments that prove useful — arguments other philosophers find themselves needing to address, refine, or build upon. > 3. Teaching and anthologising select for clarity, illumination, and pedagogical power. > 4. Sustained philosophical attention selects for depth — works that reward re-reading because their arguments have structure worth unpacking. > > What survives this multi-stage process is text exhibiting properties that track Williamson's criteria: the elegant and unified theories get published, taught, and cited more than the arbitrary, gerrymandered, ad hoc ones. The corpus is therefore not a neutral repository. It is a *virtue-filtered* repository. > > When an LLM is trained on this corpus, it learns the distribution of text that has survived these filters. The learned probability distribution is shaped by the theoretical virtues — not because the model has been instructed in those virtues, but because texts exhibiting them are overrepresented in the training data relative to texts that lack them. The virtues are, in a precise sense, *latent* in the model: implicit in the statistical regularities of the learned distribution, recoverable from the model's outputs, but not explicitly represented as rules or criteria the model applies. > > This gives us a way to reframe Floridi et al.'s diagnosis. They are correct that LLMs are "engines of generative plausibility" — they generate the most probable continuation given learned associations. But in a corpus filtered by theoretical virtues, "plausible" and "exhibiting theoretical virtues" are not independent properties. They are correlated, because the filtering shaped what counts as plausible. The LLM does not evaluate its output against criteria of elegance and unity. It does not need to. The criteria are already baked into the distribution it draws from. > > There is a Williamsonian passage that supports this framing. Williamson writes that "the more T has the intrinsic virtues of a good theory, the better (ceteris paribus). It should be elegant and unified, not arbitrary, gerrymandered, ad hoc, or messily complicated" (2024, p. 354). He is describing criteria for ranking theories. The claim here is that these criteria are not merely evaluative standards applied from outside — they are structural properties of the corpus, properties that an LLM trained on the corpus has absorbed as features of the probability landscape. > > A useful analogy, which the paper could develop: The relationship between the LLM and the theoretical virtues is like the relationship between a language model and grammar. A model trained on grammatical text produces grammatical outputs without having been taught grammar as a set of rules. The grammatical patterns are latent in the distribution. Similarly, a model trained on philosophically filtered text produces outputs tending toward philosophical quality without having been taught the evaluative criteria as rules. The quality patterns are latent in the distribution. > > This is not a claim that every LLM output is good philosophy, any more than every output is grammatical. The claim is about the *tendency* of the distribution — the direction in which the probability landscape slopes. > > ### What "latent" means here > > "Latent" earns its keep in two ways: > > First, technically: in the machine learning sense, the theoretical virtues are latent features — features of the training data that are not directly represented in the model's architecture but can be recovered from its behaviour. Just as a language model represents syntactic structure without having explicit syntactic representations, an LLM trained on the philosophical corpus represents the evaluative structure of the discipline without having explicit evaluative representations. > > Second, philosophically: the virtues are latent in the sense that they require *drawing out*. They are present in the model's distribution but not automatically expressed in every output. Whether a given output exhibits the theoretical virtues depends on *what the model is prompted to generate*. This is where prompting enters. > > ### The elicitation claim: drawing arguments out > > If the encoding claim is right, then the question is not whether the LLM "has" the theoretical virtues, but under what conditions its outputs will *exhibit* them. The conditions are largely set by the prompt. > > A prompt determines a region of the continuation space. The probability distribution the LLM has learned extends over an astronomically large space of possible continuations. The prompt constrains which region the model generates in. Different prompts access different regions, and these regions differ in how reliably they exhibit the theoretical virtues. > > Here is where Enrico's distinction between problem-oriented and solution-oriented prompts (from today's recording) is useful. A bare question — "What is consciousness?" — sets up a continuation space whose most probable occupants are surveys, hedges, and textbook summaries. These are probable because they are common in the corpus: there are many more survey-type discussions of consciousness than there are original arguments about it. The theoretical virtues of the most probable continuation, given this prompt, are low — not because the model cannot produce better, but because the prompt activates a region of the distribution dominated by cautious, generic text. > > A more structured prompt — one that lays out a position, identifies its vulnerability, and gestures toward a repair — accesses a different region. The most probable continuation of such a prompt is not a survey but a philosophical *move*: the next step in the dialectic. This is what the "obvious move" technique captures. The continuation space for a dialectically structured prompt has a higher concentration of text exhibiting theoretical virtues, because the texts in the corpus that follow such setups tend to be the substantive philosophical moves. > > Three modes of prompting, each accessing a different region of the virtue-landscape: > > ### Mode 1: Dialectical framing (one-shot, problem-oriented) > > You pose a question embedded in dialectical context. Not just "what is X?" but "given these considerations, what follows?" or "the obvious objection is Y; address it." The good continuation is a dialectical response — and the training data is densely populated with such responses at the appropriate points in the argumentative structure. > > This connects to Walton, Reed, and Macagno's argumentation schemes: the "critical questions" licensed by each scheme are exactly the pressure points that a well-trained model will respond to, because the corpus contains thousands of instances of exactly this kind of exchange. > > ### Mode 2: Solution-gestured prompting (one-shot, solution-oriented) > > You write a paragraph that points toward a solution without fully articulating it. The good continuation is the next step in developing that solution. This is richer than mode 1 because the prompt itself contains philosophical content — it is not just asking a question but *beginning* an argument. The model continues the argument in the direction the setup indicates, and the theoretical virtues of the continuation are higher because the prompt has done work to constrain what "good continuation" means. > > Enrico described this well in the recording: "you may write something, which there are already ideas that point toward the solution, and the good continuation is another step toward the solution but then you can add another beat and a good continuation." The iterative version — multiple beats, each building on the last — produces richer arguments because each step further constrains the continuation space. > > ### Mode 3: Conversational iteration (multi-turn) > > The most powerful mode. You and the model produce philosophy together in an iterative process: you write something, the model continues, you refine, the model develops, you object, the model repairs. Each turn constrains the next. The theoretical virtues of the emerging argument increase with each round because each round further specifies what the good continuation is. > > This is where the autonomy question Enrico raised becomes interesting. On a continuum from "the prompter does all the work" to "the LLM does it independently," conversational iteration sits in the middle. The prompter provides direction, constraints, and editorial judgment. The LLM provides dialectical moves, articulation, and pattern-completion over the space of philosophical arguments. Neither is doing philosophy alone. What they are producing *together* is a text exhibiting theoretical virtues. > > ### The connection between "plausible continuation" and "exhibiting theoretical virtues" > > This is the argument's load-bearing joint. The claim is not that *any* plausible continuation exhibits theoretical virtues. It is that in the right region of the distribution — the region accessed by philosophically structured prompts — the most probable continuations *tend* to exhibit theoretical virtues, because the corpus that shaped the distribution was filtered by those virtues. > > Here is where Floridi et al.'s "zeroth-order abduction" diagnosis and Williamson's evaluative criteria meet. Floridi et al. say the LLM produces "plausible" text without performing genuine inference. Williamson says philosophy is evaluated by intrinsic virtues of the theory. The connection: what counts as "plausible" philosophical text — what the model has learned to produce as probable continuation — is shaped by what has survived the discipline's filtering. And what has survived the discipline's filtering is text exhibiting Williamson's virtues. > > So the process is: > 1. The discipline produces text. > 2. The filtering (peer review, citation, teaching, anthologising) selects text exhibiting theoretical virtues. > 3. The filtered text becomes the training data. > 4. The LLM learns the distribution of the filtered text. > 5. The LLM's "plausible continuation," in the right context, therefore tends to exhibit the theoretical virtues encoded in the distribution. > > This does not require the LLM to *understand* the theoretical virtues, or to *apply* them as criteria, or to *evaluate* its outputs against them. It requires only that the training data was shaped by those virtues — which it was, because that is what philosophical filtering consists in. > > ### What about novelty? > > The obvious worry: if the LLM is producing continuations shaped by existing filtered text, can it produce anything genuinely *new*? Doesn't this reduce LLM philosophy to sophisticated recombination? > > The response draws on Williamson's own point about philosophical innovation. He notes that "enumerative induction is inadequate for systematic philosophical theorizing, which often requires introducing new distinctions at a more abstract level not given in the data" (p. 353). The examples he gives — Dummett's distinction between assertoric content and ingredient sense — are conceptual innovations that reconfigure existing materials at a higher level of abstraction. > > The claim is not that LLMs introduce wholly unprecedented frameworks. It is that they can produce the kind of novelty Williamson describes: new distinctions at a higher level of abstraction, new ways of organising existing materials. The model's learned distribution includes not just particular arguments but patterns of argumentative *structure* — patterns of how distinctions are drawn, how arguments are constructed, how positions are developed. These structural patterns can be instantiated in novel ways, producing arguments that do not appear verbatim in the training data but follow the patterns that the training data established. > > Is this enough for serious philosophical contribution? It depends on what percentage of philosophical contributions consist in this kind of reconfiguration versus genuinely unprecedented framework-introduction. I suspect the honest answer is that most good philosophy — most published, cited, taught philosophy — consists in careful articulation, rigorous argument, and sophisticated engagement with existing positions, all of which are forms of reconfiguration at varying levels of abstraction. The rare framework-introducing genius (a Kripke, a Lewis) may be beyond what current LLMs can produce. But the bulk of what the discipline values — and what gets published in good journals — does not require that kind of genius. > > ### The Hitchhiker's payoff > > The paper opens with Deep Thought receiving a question and providing an answer that, while correct, is useless because the question was wrong. Section 4 inverts this: the LLM has the *answers* latent in its distribution (the theoretical virtues, the dialectical patterns, the argumentative structures), and the skill consists in asking the right *questions* — prompts whose good continuation draws out what the model has absorbed from the corpus. Deep Thought's problem was that humanity asked an underspecified question and got an uninterpretable answer. The prompter's art is to specify the question so that the model's answer exhibits the philosophical quality latent in its training. > > This payoff works narratively and philosophically: the paper's argument about prompting *explains* why Deep Thought failed. It was a prompting problem. > > ### The self-proving dimension > > Enrico mentioned that in a previous version, you "presented it as self-proving — if you think this paper is good..." This is still available and perhaps stronger now. If the reader finds the arguments in this paper clear, the distinctions illuminating, the engagement with objections substantive — then the paper is an instance of what it describes. The theoretical virtues exhibited by the paper are partly the product of LLM-human collaboration (to the extent that any of the argument development used LLM assistance). The paper does not need to *demonstrate* LLM philosophy as a separate exercise. It *is* a demonstration, presented for blind review, evaluated by the very criteria it discusses. > > This move is risky — it can seem cute rather than substantive. But it is philosophically legitimate: the paper's own quality is evidence for the claim that the process it describes can produce philosophy meeting the evaluative criteria it articulates. > > ### What this section is NOT > > It is not a how-to guide for prompt engineering. It is not a cookbook. It is a philosophical argument about the relationship between (a) the evaluative structure of the discipline, (b) the statistical structure of models trained on the discipline's corpus, and (c) the role of prompting in accessing the right region of that structure. The practical dimension — "this is how you get LLMs to do good philosophy" — follows from the theoretical argument rather than preceding it. > > ### Possible structure of the section > > I am grouping these as potential subsection beats, not committing to an order: > > 1. The corpus as virtue-filtered repository. (Quick, since Section 3 already established "dialectical saturation" — this extends it from dialectical moves to evaluative standards.) > 2. What "latent" means: the grammar analogy. Virtues are to the philosophical corpus what grammatical structure is to natural language — implicit in the distribution, recoverable from outputs, not explicitly represented. > 3. Prompting as region-selection. The prompt determines which part of the distribution the model generates from. Skillful prompting accesses the virtue-dense regions. > 4. Three modes of prompting (dialectical, solution-gestured, iterative), with brief examples or illustrations. > 5. The novelty question: reconfiguration at higher levels of abstraction (drawing on Williamson on new distinctions). > 6. The autonomy continuum and the collaborative picture. > 7. The Hitchhiker's payoff: the prompt is the question, and now we know what the question is. > > ### Tensions and limitations I see > > The encoding claim is strong but faces a potential objection: even if the corpus is filtered by theoretical virtues, the relationship between statistical frequency and philosophical quality is imperfect. Bad philosophy gets published too. Popular but mediocre work gets cited more than excellent but obscure work. The filtering is noisy. The response: noisy filtering is still filtering. The *tendency* is toward virtue, even if individual data points deviate. This is analogous to saying that a language model trained on mostly grammatical text will produce mostly grammatical outputs, even though the training data contains some errors. > > There is also a question about whether the encoding claim is *too* strong — whether it proves too much. If the corpus already encodes the theoretical virtues, and the LLM has learned the distribution, then *any* LLM output in the philosophical domain should tend toward quality. But this is clearly false: unprompted LLM outputs are often generic, hedging, and lacking in philosophical quality. The response is the elicitation claim: the virtues are latent, not automatically expressed. The prompt is needed to access the right region. > > This means the argument requires both claims — encoding and elicitation — working together. Neither alone is sufficient. The theoretical virtues are in the model (encoding) but need the right prompt to come out (elicitation). The philosophical contribution of the paper is in showing that these two claims jointly explain something that otherwise looks mysterious: how a system that "merely" produces statistically probable text can produce text that meets evaluative criteria it has never been taught. > > ### How this relates to the rest of the paper > > Section 1 establishes that philosophical evaluation concerns text-internal criteria. Section 4 answers: and those criteria are what the corpus encodes. > > Section 2 presents Floridi et al.'s objection that LLMs do not perform genuine abduction, and Zahavy's objection that they cannot make the E→A leap. Section 4 responds: what philosophical abduction requires is not the *process* of evaluating alternatives (Floridi) or the *experience* of embodied simulation (Zahavy), but the *production of text exhibiting theoretical virtues* — and the virtue-filtered corpus provides exactly the material from which such text can be generated. > > Section 3 addresses novelty worries. Section 4 deepens this: the kind of novelty philosophy values — new distinctions at higher levels of abstraction, new ways of organising existing materials — is precisely the kind of novelty that pattern-completion over a virtue-filtered distribution can produce. > > The whole paper then has the structure: (1) philosophy evaluates text; (2) some say LLMs can't produce the right kind of text; (3) but they can, because the relevant standards are encoded in the corpus; (4) here is how the encoding works and how prompting accesses it. Section 4 is the constructive payoff — it answers "if they can, show us how" — and the Hitchhiker's Guide joke lands. --- ### 2026-03-05 — Terminology decision: "intrinsic virtues" (Williamson's phrase) as the label for evaluative standards latent in LLMs > "Intrinsic virtues" — Williamson's own phrase: "the more T has the *intrinsic virtues* of a good theory, the better" (p. 354). Distinguishes from extrinsic/relational properties (fit with evidence). "Intrinsic" does philosophical work: it marks that these are properties of the theory *itself*, assessable without knowing the production process — which is exactly the point the paper needs the term to carry. --- ### 2026-03-06 — Pigliucci's "evoked truths" framework — rigidity of philosophical objects supports text-internal evaluation and corpus-filtering > ## Pigliucci's "Evoked Truths" and What They Do for the Paper > > ### The framework (from Ch. 6 of Blackford & Broderick 2017) > > Pigliucci borrows a 2×2 taxonomy from Unger and Smolin (2015). Two axes: did the object exist prior to human thought? Does it have rigid properties? Four cells: discovered (planets — prior existence, rigid), invented (Sherlock Holmes — no prior existence, not rigid), fictional (biological species concepts — prior existence, not rigid), and evoked (chess — no prior existence, rigid). The evoked cell is the one that matters. > > Smolin on chess: > >> "When a game like chess is invented a whole bundle of facts become demonstrable, some of which indeed are theorems that become provable through straightforward mathematical reasoning. As we do not believe in timeless Platonic realities, we do not want to say that chess always existed — in our view of the world, chess came into existence at the moment the rules were codified. This means we have to say that all the facts about it became not only demonstrable, but true, at that moment as well ... Once evoked, the facts about chess are objective, in that if any one person can demonstrate one, anyone can. And they are independent of time or particular context: they will be the same facts no matter who considers them or when they are considered." (Unger and Smolin 2015, 423) > > Pigliucci's move: philosophy, like mathematics, deals in evoked truths. But with a qualification — unlike pure mathematics, philosophy is "inherently concerned with the state of the world." Philosophy's starting assumptions (its equivalent of axioms) are empirical data about the world. So philosophy does "empirically informed evoking" — its conceptual spaces have rigid properties AND are anchored by real-world constraints. > > Pigliucci explicitly distinguishes philosophy from fiction on two grounds: (1) fiction involves "inventing" (no rigid properties — even the constraints could have been otherwise), philosophy involves "evoking" (rigid properties once assumptions are made); (2) philosophy concerns this world, not arbitrarily invented ones. > > He replaces "theory" with "account" for philosophy: "philosophy — the way I see it — attempts to clarify things, or to analyze in order to bring about understanding, not really to discover new facts, but rather to evoke rational conclusions arising from certain ways of looking at a given problem or set of facts." > > ### Why this matters for the paper > > The paper argues that philosophical evaluation concerns text-internal properties. Pigliucci's framework gives you a way of saying what those properties are properties OF. They're properties of evoked objects — structures that have rigid, assessable, objective characteristics once the starting assumptions are in place. The "rigidity" part does the work: once you adopt the framework (say, Kripke's apparatus of rigid designators, or Dellsén's understanding-based account of progress), the consequences are constrained. Good philosophy maps those constraints accurately. Bad philosophy gets the rigid properties wrong. And this assessment can be made by anyone examining the text, because evoked truths are "independent of time or particular context: they will be the same facts no matter who considers them or when they are considered." > > This connects to the Section 4 argument about filtering. If the corpus consists of texts that evoke and map conceptual landscapes, and if the properties of those landscapes are rigid (not arbitrary), then what the corpus encodes is not just stylistic patterns but the structure of evoked objects. An LLM trained on that corpus absorbs that structure. When it produces a "plausible continuation," the plausibility tracks the rigidity of the evoked landscape — because that's what the training data was shaped by. > > The "evoked truths" framework also helps with the chmess worry. Chmess is invented (no rigid properties — it's an arbitrary game variant). Real philosophical work is evoked (rigid properties, empirically constrained). The difference between productive philosophy and chmess is precisely the difference between evoking and inventing — between exploring a space with genuine constraints and tinkering with arbitrary stipulations. > > ### How it integrates with existing sources > > Pigliucci would sit alongside Dellsén (progress as enabling understanding), Williamson (intrinsic virtues), and Bengson (evaluative criteria for theory construction) as a fourth independent voice supporting the text-internal evaluation framework. Each provides a different angle: > > - Dellsén: what matters is whether the text puts someone in a position to understand (for-whom, not by-whom) > - Williamson: the text is evaluated by intrinsic virtues (elegance, unity, non-ad-hocness) > - Bengson: the text is evaluated by criteria of accommodation, explanation, integration, virtue > - Pigliucci: the text maps an evoked conceptual landscape whose properties are rigid and objective once the starting assumptions are in place > > The convergence across four independent frameworks — none of which was designed with AI in mind — strengthens the paper's text-internal evaluation thesis. It shows the thesis isn't idiosyncratic but follows from multiple accounts of what philosophy is. > > ### Aporetic clusters and the Bourget-Chalmers data > > Pigliucci also draws on Rescher's "aporetic clusters" — families of alternative solutions to philosophical problems. From Rescher (via Moody 1986): "in philosophy, supportive argumentation is never alternative-precluding. Thus the fact that a good case can be made out for giving one particular answer to a philosophical question is never considered as constituting a valid reason for denying that an equally good case can be produced for some other incompatible answers to this question." > > He uses the Bourget-Chalmers 2013 survey to provide empirical evidence that the structure of philosophical space is clustered and coherent, not random. Professional philosophers form structured positions with internal correlations (moral realists tend to be aesthetic objectivists and Platonists). Principal components analysis reveals three axes: Anti-naturalism, Objectivism/Platonism, Rationalism. > > This bears on the corpus-filtering argument: if the discipline's output is structured into coherent clusters refined over centuries, then the training data encodes not random text but the architecture of philosophical space itself. An LLM trained on this data has absorbed the cluster structure — which positions cohere with which, which moves are available within each cluster, where the pressure points are. --- ### 2026-03-05 — Novelty implicit in Williamson's virtues via informativeness and generality — not a separate criterion > On novelty in Williamson's intrinsic virtues: novelty is *not* listed as an intrinsic virtue. His list is: elegant, unified, not arbitrary/gerrymandered/ad hoc, informative, general, combining simplicity with strength. He mentions "fruitful" in passing — "discriminate fruitful from pointless definitions, promising conjectures from dead ends" (p. 367) — but as a feature of the aesthetic sense needed to *navigate toward* good theories, not as a virtue of the theory itself. > > However, novelty is implicit in a different way. A theory that merely restates what's already known scores low on informativeness and generality. The kinds of innovations Williamson values — Dummett's assertoric content / ingredient sense, new distinctions at higher levels of abstraction — score high on intrinsic virtues *precisely because* they organise materials in new ways. So novelty is a *consequence* of scoring well on the virtues, not a virtue itself. A paradigm shift would be a theory that scores exceptionally high on unification and informativeness while being elegantly simple — the virtues don't need a separate "revolutionary novelty" criterion because such a theory would already max out the existing criteria. > > This is actually useful for the paper: it means you don't need to argue that LLMs can produce paradigm-shifting novelty (a hard sell). You just need to argue they can produce texts scoring well on intrinsic virtues — and whatever novelty that produces is novelty enough, by Williamson's own standards. --- ### 2026-03-07 — Evocation framework gives text-internal evaluation thesis an ontological ground — "found not created" = "process-irrelevant" > The text-internal evaluation thesis (philosophical quality is assessable from the text, regardless of production process) and the evocation framework's credit analysis (arguments are rigid consequences of evoked landscapes, ownerless, found not created) are the same claim stated in two registers — methodological and ontological. "Arguments are found not created" = "evaluation is text-internal." The evocation framework connects them: because arguments are rigid consequences of evoked structures, their quality is fully manifest in the text that shows them. Pigliucci gives the text-internal thesis an ontological ground: process-irrelevance isn't a methodological preference, it's a consequence of the ontology of philosophical objects (evoked truths with rigid properties). This emerged from a late-night session working through Pigliucci's Ch06 (Philosophy as the Evocation of Conceptual Landscapes) and Wilson's Ch07 (Three Barriers to Philosophical Progress), both from Blackford & Broderick 2017. The full chain: Smolin/Unger's evocation taxonomy → Pigliucci's application to philosophy → rigid inferential architecture of evoked landscapes → credit distribution (tradition: structural credit; prompter: directorial credit; LLM: articulatory acknowledgment; argument: ownerless) → convergence with text-internal evaluation thesis. --- ### 2026-03-15 — Zahavy three-component decomposition — for prose paragraph in Zahavy section > Zahavy's argument has three components, not one: (a) The source of axioms is sensory experience — Einstein's raw sensation of what freefall would feel like; (b) The mechanism is embodied simulation — imagining bodily scenarios, not manipulating symbols; (c) The precondition is access to physical referents — the things symbols are about. LLMs fail on all three: no sensory experience, no embodied simulation, no access to physical referents. Therefore (Zahavy argues) they can't make the jump from experience to axioms. This should be included as a paragraph somewhere in the Zahavy section (not in list format, written in proper prose). --- ### 2026-03-15 — Coarse-grained phenomenology is linguistically encoded, not just described — Chalmers "spread over the surface" example > The LLM's access to coarse-grained phenomenology isn't just "it's read descriptions of experiences." It's deeper than that — the phenomenological facts are structurally embedded in how people use language, and the LLM has learned those patterns. Coarse-grained phenomenological facts are linguistically encoded as presuppositions of ordinary language. You can't speak a natural language without implicitly committing to certain phenomenological claims (colours are spatial, sounds have locations, time passes, etc.). An LLM that has learned natural language has therefore learned these phenomenological facts as implicit content. This means the LLM's phenomenological knowledge at this level isn't second-hand (descriptions of experience) but is structurally encoded in the very medium it works with (language). The Chalmers example from "Perception and the Fall from Eden" (2010, p. 398) illustrates: "Phenomenologically, it seems to us as if visual experience presents simple intrinsic qualities of objects in the world, spread out over the surface of the object." This is presupposed by every sentence about coloured objects — "the red book," "the blue wall." The LLM has absorbed this phenomenological fact through the linguistic patterns, not through reading phenomenological reports. --- ### 2026-03-15 — Fine-grained phenomenology: LLM can't originate or evaluate novel claims — Merleau-Ponty touching-touched example > After articulation, fine-grained observations enter the corpus too. Merleau-Ponty published it; it's been discussed, taught, cited. An LLM trained on the current corpus has it. But it couldn't have originated it. And — this is the nuance — it might not be able to evaluate a NOVEL fine-grained phenomenological claim, because evaluating such a claim requires being able to perform the relevant act of attention yourself, to check whether the description is accurate. > > The concession: LLMs cannot perform original phenomenological investigation, cannot originate novel fine-grained phenomenological observations, and cannot evaluate novel fine-grained phenomenological claims by checking them against experience. This is a real limitation. It is not dissolved by the grain argument — it is bounded by it. The limitation applies to one specific type of philosophical work (origination of new phenomenological starting points) and does not extend to the conceptual analysis, argument construction, theory evaluation, and thought experiment work that constitutes the bulk of the discipline within the text-focused tradition. --- ### 2026-03-15 — Two Machery moves: dissolution (intuitions are textual) and judo (unreliable intuitions favour corpus) > 1. The dissolution move: philosophical intuitions, as they function in the literature, are propositional and textual. They're in the corpus. The LLM has them. > > 2. The judo move: Machery's own work shows intuitions are unreliable. So depending on them is a weakness of human philosophy, not a strength. The virtue-filtered corpus has already done the work of sorting reliable from unreliable intuition-driven arguments. ---