[▶ Audience fullscreen](https://htmlpreview.github.io/?https://gist.githubusercontent.com/nickneek/103bc1ec9a1d895e46890f96e1d44348/raw/generating-philosophy-audience-deck-FINAL.html) · [▶ Speaker fullscreen](https://htmlpreview.github.io/?https://gist.githubusercontent.com/nickneek/953ba2942fa444af47e29264fc15e5a1/raw/generating-philosophy-speaker-deck-FINAL.html) · [▶ Split view](file:///Users/nickyoung/My%20Obsidian%20Vault/Attachments/tools/talk-split-view.html?left=https%3A%2F%2Fhtmlpreview.github.io%2F%3Fhttps%3A%2F%2Fgist.githubusercontent.com%2Fnickneek%2F103bc1ec9a1d895e46890f96e1d44348%2Fraw%2Fgenerating-philosophy-audience-deck-FINAL.html&right=https%3A%2F%2Fhtmlpreview.github.io%2F%3Fhttps%3A%2F%2Fgist.githubusercontent.com%2Fnickneek%2F953ba2942fa444af47e29264fc15e5a1%2Fraw%2Fgenerating-philosophy-speaker-deck-FINAL.html&sync=gpth-ny-lgk-20260423-r7f3c9) · [[Generating Philosophy with AI — Argument Moves (Lingnan–Genoa–Kobe, 2026-04-23)|moves deck]] · [[Sessions/Generating Philosophy|project]] # Generating Philosophy — Talk Speaker Notes Speaker-facing notes for the 40-minute lecture. 29 slides total (title + §0 scene + 4 section dividers + 22 content scenes + close). Slide IDs match the counter top-right of the hybrid deck (`01/29`, `02/29`, …). --- # Pre-talk checklist (3 min before 08:00) - Water within reach. - Split view open on Mac in Safari, fullscreen (launcher or Split view link above). - Speaker-side deck open on second device in Safari, fullscreen (launcher or Speaker fullscreen link above). - Zoom: share the browser *window*, not the whole screen; confirm audio input; turn off notifications. - Breathe once. Smile. Go. # Running schedule (wall-clock) - **08:00** — start, slide 01 (title) - **08:02** — §1 begins, slide 03 (DIV) - **08:09** — §2 begins, slide 08 (DIV) - **08:14** — §2 mid-check, slide 12 (Lipton) - **08:19** — §3 begins, slide 16 (DIV) - **08:25** — §3 mid-check, slide 20 (the reframe) - **08:31** — §4 begins, slide 25 (DIV) - **08:37** — close, slide 29 - **08:38** — end; Q&A to 08:40 (or per chair) Content budget ≈ 37:45. Buffer ≈ 2:15 for transitions and modest elaboration. Pace rule of thumb: if you're late by more than a minute at any check, go to the catch-up levers at the bottom of this note. --- # 01 — TITLE · **08:00** - ON SCREEN — "Generating philosophy with artificial intelligence." / "Can LLMs produce philosophy worth reading?" / Nick Young & Enrico Terrone · Thursday 2026-04-23 - SAY - Thank the organisers briefly. - Note that this is joint work with Enrico Terrone. - State the question verbatim. - Route-map in two sentences: one constitutive challenge (§1), two capacity challenges (§2 [[abduction]], §3 [[phenomenology]]), and a speculative coda (§4). - NOTE — Enrico has not seen the latest iteration of these slides; some framings may not yet reflect his sign-off. If anyone asks about the joint authorship, the long-form paper is the co-authored version; the talk is a working shape. - TIME — ~40 s --- # 02 — § 0 · the question - ON SCREEN — "What would it take for an LLM paper to be worth reading?" / diagram: reader → reads → a text → yields → time well spent (with "if not this, time wasted" arc) / two minimal ledes: an everyday ordinary understanding of "worth reading"; journals strive to publish what is worth reading, with varying success. - SAY - Let the diagram animate in before you speak. - Clarify: "worth reading" is not being given a technical definition. It is the everyday, ordinary understanding philosophers already work with. - Journals are its institutional expression — they aim at it, and sometimes succeed. - Resist any temptation to elaborate the concept. The audience has it already; elaborating makes it sound like a term of art, which it isn't. - TIME — ~80 s --- # 03 — §1 · DIV · **08:02** - ON SCREEN — big §1 megnum / "Authorship." / lede: person-only domain - SAY - State the challenge: philosophy is a person-only domain, no LLM text can be philosophy because no philosopher stands behind it. - Register: not explicitly defended in print, but captures an intuition a fair number of philosophers hold. - Structure: this is the *constitutive* challenge — not about capacity, but about what counts. - TIME — ~60 s --- # 04 — §1 · the intuition, made precise ([[Davies]]) - ON SCREEN — "The work is the philosopher's sustained activity — not the text she leaves behind." / attribution: davies 2004 · performance theory, transposed - SAY - The intuition has an analogue in art: many deny a purely AI-generated image is an artwork because no artist lies behind it. - Davies makes this precise for art: the *work* is the artist's generative performance; the canvas is the product. - Transposed to philosophy, the work would be the philosopher's sustained activity — her working through of a problem, her formulating and revising of arguments. - On this view, a text indiscernible from a philosophy paper would fail to *be* philosophy if no philosopher stood behind it. - QUOTE (read if time, otherwise gloss) > "The work — what the artist achieves — is the process eventuating in that product. … They are, rather, intentionally guided generative performances that eventuate in contextualized structures or objects." — Davies, *Art as Performance*, p. 98 - TIME — ~110 s - NOTE — set-up slide; the next slide is where it falls. --- # 05 — §1 · where the analogy breaks - ON SCREEN — "In art, surface underdetermines the work. In philosophy, it does not." / pair: art (forgery vs Rembrandt) · philosophy (two type-identical papers) / lede: "What we ask of a philosophical text concerns only what the text itself says." - SAY - The feature driving Davies in art has no philosophical analogue. - Art: indistinguishable-surface cases — canvas-from-a-washing-machine, Danto-style red squares, molecule-identical forgery. Surface leaves the work underdetermined. - Philosophy: two type-identical papers make the same arguments, face the same objections, admit the same evaluations. - If a list is wanted in the room: are premises defensible? inferences valid? distinctions real? counter-examples apt? — keep these in your delivery; they are no longer on the slide. - TIME — ~90 s --- # 06 — §1 · the norm, already institutionalised - ON SCREEN — "Journals strip authorship before review." / redacted paper-card figure with three fig-note sub-beats (blind review / Dellsén / Sokal) - SAY - The text-focused norm is already enacted across the discipline. - Blind review: journals strip author information as a matter of principle — what is assessed is what the paper says, not who said it. - [[Dellsén]] et al. 2024 makes this normative: philosophical progress is a *for-whom* rather than *by-whom* matter. - Gloss the Dellsén point carefully: progress consists in putting readers in a position to increase their understanding. It happens paradigmatically through philosophical content being made publicly available. What matters is the accessibility of the arguments, distinctions, counter-examples — not the cognitive states of whoever produced them. - So "for-whom" = the audience whose understanding the work advances; "by-whom" = the producer. Dellsén's claim: progress turns on the former, not the latter. - Why this matters for today's talk: if philosophical progress is a for-whom matter, then whether the producer was a philosopher or an LLM is beside the point; what matters is whether the text puts a reader in a position to understand. - Sokal is recognisable as a breach precisely because the background norm is that the text is what matters: what went wrong was that *Social Text* had assessed Sokal's standing rather than the argument on the page. - TIME — ~110 s - NOTE — read Dellsén's *for-whom / by-whom* phrase slowly; it recurs in §2. Gesture at the three sub-beats on the paper-card fig-note. --- # 07 — §1 · payoff - ON SCREEN — "The work is in the text." / two short ledes - SAY - The work consists in the text. Philosophical evaluation goes no further. - What remain are two capacity challenges. §2 takes up abductive reasoning. §3 takes up conscious experience. - Question in each case: can an LLM without the relevant capacity produce a text with the corresponding properties? - TIME — ~60 s --- # 08 — §2 · DIV · **08:09** - ON SCREEN — big §2 megnum / "Philosophy without abduction?" / lede: Floridi's zeroth-order abduction - SAY - State the first capacity challenge. - [[Floridi]] and colleagues diagnose LLMs with *zeroth-order abduction*: plausible continuations by pattern-matching, without comparing hypotheses. - TIME — ~50 s --- # 09 — §2 · the charge - ON SCREEN — "LLMs produce the appearance of abductive reasoning." / full pullquote (Floridi) / attribution - SAY - Read pullquote aloud. - Floridi's example: LLM asked why a car won't start on a cold morning says "dead battery, cold weather reducing efficiency" — plausible, but without having compared alternatives. - Zeroth-order abduction = a plausible continuation produced on learned associations, with no stage at which competing hypotheses are generated and compared. - Two things Floridi is careful to say: first, the output can look indistinguishable from reasoned explanation; second, the diagnosis is not that the output is bad but that the process is not abductive — it reproduces the form of explanation without the work. - Floridi's own framing: the issue is the cognitive history of the text, not its surface. Worth noting in the room, because it sets up the move I make later — that philosophical evaluation has nothing to say about the cognitive history either. - Lipton's two-stage account of IBE is in the background here: background beliefs generate a shortlist of plausible hypotheses; a selection is then made from among them. Floridi says LLMs collapse both stages into a single step. - QUOTE (fuller, on slide) > "LLMs seem to perform a kind of zeroth-order abduction: given a prompt, they generate a plausible continuation (a hypothesis or explanation) based purely on learned associations. … The model does not understand what an explanation is, but it produces text that follows the typical phrasing and structure of explanations. It does not reason about causes from scratch but outputs typical causes for typical effects observed in the training data." — Floridi et al., p. 9 - TIME — ~90 s - NOTE — slow on the quote; this is the charge being laid. If someone challenges whether philosophy is really abductive, that's §2.10 (Williamson) territory — park it for that slide. --- # 10 — §2 · [[Williamson]] · abduction in the armchair - ON SCREEN — "Philosophy assesses theories by their intrinsic virtues — simplicity combined with strength." / attribution: williamson 2024, pp. 354, 358, 368–69 - SAY - Williamson's claim is that contemporary analytic philosophy already proceeds partly by abduction *from the armchair*. - Philosophers defend a view by comparing it to rivals, weighing how each would, if true, explain the evidence. - "Intrinsic virtues of a good theory" is Williamson's phrase — roughly, simplicity combined with strength. Not a defined term; a shorthand for the cluster of qualities that makes one theory more worth holding than its rivals. - Put together: Floridi's charge becomes a capacity challenge. Humans can do this; LLMs, on Floridi's diagnosis, cannot. - TIME — ~80 s --- # 11 — §2 · the page - ON SCREEN — "Philosophy is conducted in writing. The LLM is trained on that same writing." / two ledes + Floridi sub-quote - SAY - Philosophy isn't thought in the head and written up afterwards. It is conducted *in* writing: formulated by writing, revised by reading, submitted to other philosophers who respond in further writing. - Hypothesis-weighing, objection-handling, virtue-ranking — the comparative work Floridi says LLMs cannot do — is done *in the prose itself*. - Floridi himself grants this in the paper's abstract (see quote). - QUOTE (on slide) > "This effect is due to the model's training on human-generated texts that encode reasoning structures." — Floridi et al., Abstract - TIME — ~90 s - NOTE — the pivotal move. Pace it. Do not rush. --- # 12 — §2 · [[Lipton]] · likeliness, loveliness · **08:14** - ON SCREEN — "Likeliness speaks of truth. Loveliness of potential understanding." / pullquote + attribution - SAY - Lipton's distinction is between two properties an explanation can have. - *Likeliness* is about probability: given the evidence, how probable is this hypothesis? - *Loveliness* is about explanatory power: IF the hypothesis were true, how much understanding would it give us? How elegantly does it unify? How much does it illuminate? - The two come apart. A conspiracy theory that posits a coordinated agent behind many otherwise-puzzling facts is lovely — it would, if true, make everything make sense — but not very likely. A banal truth can be highly likely but explain nothing. - Lipton's own move: loveliness can serve as a guide to likeliness. The lovelier explanation tends, in practice, to be the more probable one. That is how IBE works — we infer to the best explanation because "best" correlates with "true". - Why this matters for LLMs: over an *arbitrary* corpus, statistical probability has no connection to explanatory power. But over the *philosophical* corpus — filtered by what philosophers have found explanatorily valuable — statistical likelihood ends up approximating loveliness. - So the key move of §2 is: an LLM's likeliness-tracking (which is all next-token prediction does), applied to the philosophical corpus, tracks what has already survived loveliness-tracking evaluation. Pattern-matching over this particular corpus inherits the evaluative filtering that produced it. - TIME — ~60 s - NOTE — §2 mid-check. You should be here by 08:14. If you're at 08:16+, skip the §2.13 elaboration on the next slide (lede 2 only) to catch up. --- # 13 — §2 · the corpus is self-evaluating - ON SCREEN — "Surviving patterns are the patterns later philosophers took forward." / filter-svg with animated particles + two ledes - SAY - An LLM's continuation is the likeliest in a distribution-theoretic sense. Over an arbitrary corpus that tells us nothing about loveliness. - The philosophical corpus is not arbitrary. - *Filtered* by loveliness-tracking evaluation: work that fails to meet the discipline's standards for simplicity, explanatory reach, handling of objections, is less cited, less assigned, less anthologised, less likely to survive. - *Self-evaluating*: every article in the corpus was written by a philosopher reading other articles in the corpus; each is in turn read and engaged with by further philosophers writing further articles. The process runs on itself. - Iterated IBE, concretely: over decades, a paper has been written against the background of others, cited by others, pressed against objections, revised or overtaken. The surviving texts are what has held up across many rounds of such pressure. - The LLM trained on this corpus inherits its *distillate*: the patterns left standing after that process. Not a static record; the residue of an active filter. - **Distillate vs distillation** — worth the paragraph-long gloss here because it returns in §4: - The *distillate* is the result: the surviving patterns of argument, the handling of objections, the ranking habits, already baked into the statistical structure of the text. - The *distillation* is the process that produced it: temporal, distributed across many minds, generating-testing-revising over time. - An LLM trained on the corpus has the distillate as pattern. It does not perform the distillation in any single completion — a single completion is not an iteration across time. (This is the §4 handoff.) - TIME — ~120 s - NOTE — gesture at each gate in the filter figure as you name the filters. Let the particles do work; pause a beat. --- # 14 — §2 · the mechanism - ON SCREEN — "Argument lives at the level of discourse — where next-token prediction operates." / lede + markers pills / closing lede - SAY - Argumentative structure shows up in philosophical prose above the sentence level — at what linguists call the *discourse* level. Nothing technical is meant by the word here; it just names the layer at which clauses cohere, objections get handled, and hypotheses are compared. - Realised in the kinds of markers on screen — "however", "the stronger reading is", "one might press the objection", "consider the cost of denying". They encode the comparative work of IBE; they recur because that work recurs in recognisable forms across philosophers writing in response to one another. - Next-token prediction picks up discourse-layer regularities by the same mechanism it uses to pick up syntactic regularities. The mechanism doesn't care which layer it's operating on. - Worth noting: the markers aren't decorative. They are the traces left in prose by the comparative-evaluative work Floridi says LLMs can't do. An LLM trained on enough of this prose absorbs the shapes those traces take. - TIME — ~70 s --- # 15 — §2 · payoff - ON SCREEN — "Pattern-matching over this corpus can carry philosophical quality." / three short ledes: what remains / §3 takes up / §4 returns - SAY - §2 upshot: pattern-matching over this corpus can carry philosophical quality. - Two threads remain. - §3 takes up the capacity challenge from [[phenomenology]]. - §4 returns to the fact that the model inherits the *distillate* of this process without performing the *distillation*. - TIME — ~60 s --- # 16 — §3 · DIV · **08:19** - ON SCREEN — big §3 megnum / "Philosophy without phenomenology?" / lede: philosophy draws on conscious experience an LLM has not had - SAY - State the second capacity challenge. - Philosophy often draws on what it is like to see red, to feel time passing, to touch one's own hand. - An LLM has had none of this. - Natural worry: philosophy of this kind is not something it can do. - TIME — ~55 s --- # 17 — §3 · [[Zahavy]]'s Einstein - ON SCREEN — "Some arguments need the felt experience an LLM has never had." / lift-in-deep-space figure + fig-note with Zahavy gloss - SAY - Zahavy calls this kind of reasoning *manipulative abduction* — inference via simulated sensory experience, not by manipulating symbols. - What Einstein was using the lift thought experiment for: to arrive at the **equivalence principle** — the claim that the physical effects experienced inside a uniformly accelerating frame are indistinguishable from those experienced in a gravitational field. It is from this equivalence that General Relativity was later developed. - The maths for it didn't yet exist when Einstein had the thought. So he couldn't derive the axioms; he had to *simulate* them from the felt experience of being inside the accelerating box — released objects appearing to fall with identical acceleration, the floor rushing up, etc. - Zahavy: the simulation was "not a permutation of symbols, but a manipulation of perceptual experience." Pause on this phrase. - QUOTE (if time — otherwise the fig-note gloss carries it) > "He envisioned a physicist inside an elevator being uniformly accelerated through deep space. Inside this enclosure, the sensory experience reveals a specific pattern: when objects are released, the floor rushes up to meet them. To the physicist, the objects appear to fall with identical acceleration, regardless of composition. Thus, the simulation here was not a permutation of symbols, but a manipulation of perceptual experience." — Zahavy 2026, §5 - TIME — ~120 s - NOTE — anchor for §3. Don't rush. The lift image does work. --- # 18 — §3 · Chinese Rooms, transferred - ON SCREEN — "LLMs are high-dimensional Chinese Rooms." / pullquote + attribution / lede: Mary / Bengson / agency / time - SAY - Background for the phrase. Searle's original Chinese Room (1980): a person in a room follows rules to shuffle Chinese symbols, producing output that looks to an outside observer like understanding — but there is no understanding inside. Point: symbol manipulation by itself is not meaning. - Harnad's extension (1990), the *symbol grounding problem*: symbols get meaning by being tied to sensory-motor experience with their referents. A system that only manipulates symbols, without ever having grounded them in perceptual experience, cannot mean anything by them — it can only shuffle them. Meaning comes from grounding, not from combinatorics. - Zahavy's extension to LLMs (2026): LLMs are "high-dimensional Chinese Rooms" because they manipulate the language of physics — and of any other domain — without ever having had perceptual access to the physical referents that give those words meaning. The sophistication of the shuffling doesn't change its underlying groundlessness. - Why this matters for the talk: this is the strongest form of the phenomenology challenge. Not just "LLMs lack experience" but "LLMs are constitutively cut off from the referents philosophy sometimes needs." If the challenge holds, an LLM couldn't originate any argument that depends on those referents. - Zahavy's own framing is restricted to physics; the move in §3 is whether the same concern applies to philosophy. The transfer is via Jackson's Mary: philosophy has arguments whose central thought-experiment requires the felt quality of an experience (seeing red). If Mary works, something like the phenomenology challenge bites in philosophy too. - Parallel cases: phenomenology of agency (what it is like to will an act), phenomenology of time-passing, phenomenology of having an intuition (Bengson 2015 is the reference here). Each could yield a Mary-type case. - Transfer to philosophy: just as Einstein needed the felt difference between weight-shift and free fall, *Jackson* needed the felt quality of seeing red. - Mary isn't the only case: phenomenology of agency, of time passing, of having an intuition ([[Bengson]] 2015) — each could yield a Mary-type case. - TIME — ~75 s --- # 19 — §3 · [[Pigliucci]] · the world, used differently - ON SCREEN — "Science and philosophy use the world differently." / pair: science (thought experiment gave Einstein the equivalence principle as hypothesis; physics waited on Eddington + Mercury) · philosophy (the thought experiment's conclusion is where the work lands) / closing lede on starting points. - SAY - Pigliucci's observation: the world figures differently in each discipline. - Return explicitly to Einstein here — this is where the audience needs to see the role-distinction. - Einstein's thought experiment was the *origin* of the equivalence principle; in physics it gave him a hypothesis, not a settled result. The principle still had to face empirical verification — Eddington's 1919 eclipse, the already-observed precession of Mercury's perihelion. - In philosophy the shape is different. A philosophical thought experiment's conclusion is where the work lands. No further experimental verification needed, and no further verification available. - So the thought experiment plays a different role in each discipline: in science it is a way-station en route to empirical testing; in philosophy it is the destination. - Consequence (the Pigliucci move): what philosophy takes from the world is not experimental confirmation but articulated starting points — data in propositional form. The next slide cashes this out. - TIME — ~80 s - NOTE — the pair figure does the work. Read across science-cell, then philosophy-cell; land on the "no further verification demand" lede. --- # 20 — §3 · the reframe · **08:25** - ON SCREEN — "What philosophy takes from the world can be articulated — not raw felt experience." / two ledes + Pigliucci pullquote + attribution - SAY - The challenge reframes: no longer about whether an LLM has phenomenological experience, but about whether it has access to articulated descriptions of it. - Articulated descriptions are exactly what text corpora contain. - QUOTE (on slide) > "the basic parameters philosophers use as their inputs, the starting points of their philosophizing, their equivalent of axioms in mathematics and assumptions in logic, are empirical data about the world." — Pigliucci, Ch. 6, p. 79 - TIME — ~90 s - NOTE — §3 mid-check. You should be here by 08:25. If at 08:27+, trim elaborations on §3.21 sources slide. --- # 21 — §3 · the reply · the corpus is saturated - ON SCREEN — "Articulated phenomenology is everywhere in the training data." / sources list with descriptors (perception · how things look, etc.) / two ledes on literature and saturation - SAY - Sub-disciplines work constantly with articulated phenomenology — philosophy of perception, temporal experience, aesthetics, philosophy of emotion — each with its own descriptor on screen. - Literary adds: Proust on the madeleine; Woolf on ordinary thought in *Mrs Dalloway*; Henry James on social perception; Nabokov on visual particulars. - All of it in any LLM's training data. Saturation is demonstrable, not a hope. - NOTE — the four literary examples are real authors with the specific associations claimed: - *Proust*, *À la recherche du temps perdu*, vol. 1 (*Du côté de chez Swann*, 1913): the madeleine-and-tea passage that sets off involuntary memory. Canonical phenomenological trigger-scene in 20th-c. literature. - *Virginia Woolf*, *Mrs Dalloway* (1925): Clarissa's walk through London, the interior monologue tracking moment-to-moment perception. One of the textbook examples of stream-of-consciousness prose attending to phenomenal texture. - *Henry James*'s late novels (*The Ambassadors*, *The Wings of the Dove*, *The Golden Bowl*): dense with the tracking of social perception, inference, and interior response to other minds. - *Vladimir Nabokov*'s prose (across *Speak, Memory*, *Pnin*, *Pale Fire*, *Lolita*): hyper-specific visual attention, particulars of objects and rooms rendered in precise descriptive detail. - I have not pulled verbatim quotes into these notes — the characterisations are defensible from the works themselves. If an audience member wants specifics, name a work and the specific passage type; do not improvise quotes. - TIME — ~80 s --- # 22 — §3 · the reply, through cases - ON SCREEN — "An LLM has never felt weight-shift. But the corpus is thick with it." / two short ledes: examples + anecdatum - SAY - Apply the reply to Zahavy's own Einstein case. An LLM has never felt weight-shift or free fall, but ordinary English is thick with articulated descriptions of these sensations — every elevator passage in fiction, every description of a roller-coaster drop, every astronaut's memoir of zero-gravity. - Anecdatum: ask a current frontier LLM about design and colour — palettes, complementary colours, contrast, harmony — and it will discuss the phenomenology of colour with a sophistication indistinguishable from a knowledgeable speaker. - TIME — ~90 s --- # 23 — §3 · the narrow limit - ON SCREEN — "One thing the corpus cannot give: previously undiscovered aspects of experience." / horizontal bar figure (long teal segment "articulated — available to the LLM", thin red segment at the right "previously undiscovered") · Merleau-Ponty callout / lede on self-touch / concession lede - SAY - A narrower case the reply does not cover: a philosopher sometimes articulates a *previously undiscovered aspect of phenomenology* — a structural feature of conscious experience that is present in everyone's lives but has not been explicitly described. - Paradigm: [[Merleau-Ponty]] on self-touch. When one fingertip touches another, one finger plays the role of toucher and the other of touched; the two can reverse roles but cannot simultaneously both be toucher. - An LLM cannot originate such an observation — there is nothing in prior text to recombine into it. - But the limit is narrow: most of philosophy proceeds from phenomenology already articulated. Gesture at the proportion on the bar figure — the unavailable region is a small sliver, not a large chunk. - TIME — ~100 s - NOTE — end §3 on the concession. Don't rush the Merleau-Ponty sentences — let them land. --- # 24 — §3 · payoff - ON SCREEN — "Wherever inputs have been articulated — most of the discipline — the corpus supplies them." / lede: what remains for §4 - SAY - §3 upshot: the corpus supplies the phenomenological inputs wherever they have been articulated — which, across most of the discipline, they have. - Handoff: if the capacity §§2–3 have defended is really present, why are we not seeing its products? - TIME — ~50 s --- # 25 — §4 · DIV · **08:31** - ON SCREEN — big §4 megnum / "Coda." / lede: "If the capacity §§1–3 argue for is there — *why* are we not seeing more LLM-produced philosophy worth reading?" (with gold-italic emphasis on "why") - SAY - §§1–3 have argued that the in-principle obstacles do not hold up. - In practice: we are not seeing LLM-produced philosophy worth reading at the rate one would expect. - §4 floats some speculative ideas about the gap. - TIME — ~50 s --- # 26 — §4 · what the prompt asks for - ON SCREEN — "The simplest diagnosis: what is the LLM being asked for?" / three short ledes - SAY - Simplest diagnosis: a matter of what the LLM is being asked for. - Prompted generically ("write on free will", "explain the Mary argument"), the model outputs what is most probable in its distribution — for philosophical topics that is summary text: survey-style, hedged, balanced, non-committal. - A paper worth reading is *distinctive argument* for a specific conclusion against specific alternatives. That is what the prompt has to specify. - TIME — ~85 s --- # 27 — §4 · distillate ≠ distillation - ON SCREEN — "The corpus is the distillate. Not the distillation." / dist figure (rings → prompt → problem/position/iteration box) / two ledes - SAY - The philosophical corpus was produced by many rounds of argument and counter-argument. Each paper was written by someone who had read earlier papers, and was in turn read and responded to by later philosophers. Weak work drops out of the discipline's attention — less cited, less assigned, less anthologised. Strong work survives and gets taken forward. - The corpus you end up with is the *distillate* of this process. Not a random sample of what has ever been written about philosophy; the residue of what has held up to iterated criticism. - An LLM trained on this corpus inherits the distillate. It has the surviving patterns — the ways of handling objections, the comparative moves, the ranking habits — already baked into the statistical structure of its training data. - But inheriting the distillate is not the same as performing the *distillation*. The distillation is a temporal process, running across many minds and many years: generate a claim, have it tested, revise, iterate. A single LLM completion runs nothing of the sort. It picks up the patterns that survived; it does not repeat the survival. - That is where the philosopher prompting the LLM comes in: performs, in the prompt, the iteration the discipline performs across generations. Draft. Press the draft against its strongest objection. Redraft. - Worth slowing on the distinction: "distillate" = the product; "distillation" = the process that produced it. Both are real; only the first is carried in the corpus. - TIME — ~90 s - NOTE — climax of §4. Let the *distillate* / *distillation* distinction sit. The same paragraph is mirrored on slide 13 so you have it twice if you need it. --- # 28 — §4 · a specialist LLM? - ON SCREEN — "A specialist LLM?" / lede on the candidate response / full Sellars pullquote / attribution / closing lede on the generalist claim - SAY - One might think that a way to get good philosophy out of an LLM is to train a specialist one, as has been done in mathematics (proof-assistants, specialised maths models). - A reason for thinking not, from Sellars. - Read the pullquote aloud — it's long, take it slow, let the inventory of items ("cloth, ships, and sealing-wax … numbers, duties, possibilities, finger snaps, aesthetic experience, and death") do its work. - Gloss: if philosophy is about how things in the broadest sense hang together, a general-purpose model — exposed to the broadest range of human description — is, arguably, more useful to the philosopher than one narrowed to philosophy proper. - Close the slide with that claim: a general LLM is more valuable to a philosopher than a specialist one. Held tentatively; offered as a reason against the specialist intuition. - QUOTE (on slide — full opening paragraph of "Philosophy and the Scientific Image of Man", 1962) > "The aim of philosophy, abstractly formulated, is to understand how things in the broadest possible sense of the term hang together in the broadest possible sense of the term. Under 'things in the broadest possible sense' I include such radically different items as not only 'cloth, ships, and sealing-wax,' but numbers, duties, possibilities, finger snaps, aesthetic experience, and death. To achieve success in philosophy would be, to use a contemporary turn of phrase, to 'know one's way around' with respect to all these things, not in that unreflective way in which the centipede of the story knew its way around before it faced the question, 'how do I walk?', but in that reflective way which means that no intellectual holds are barred." — Sellars, "Philosophy and the Scientific Image of Man" (1962) - TIME — ~100 s - NOTE — source-work flag: the quote above is the standard opening paragraph of Sellars's 1962 essay, widely reproduced. I have not verified it line-for-line against the original 1962 print. Before tomorrow's talk, check it against a reliable reprint if you can — the paragraph breaks and the centipede-clause wording are the points most likely to vary between editions. If in doubt, deliver only the first two sentences (to "finger snaps, aesthetic experience, and death") — those are unambiguous. --- # 29 — close · in sum · **08:37** - ON SCREEN — "In sum." / four short ledes, one per section: §1, §2, §3, §4 - SAY - Summarise each section plainly. Four sentences. - §1: the challenge from authorship does not go through. What philosophical evaluation tracks is the text. - §2: an LLM trained on the philosophical corpus inherits patterns the discipline has filtered for their explanatory virtues. - §3: the phenomenological inputs philosophical reasoning draws on are, in most cases, already articulated somewhere the corpus preserves. - §4: what the LLM does not itself perform, in any single completion, is the iteration that produced the corpus — that is where the philosopher's work now sits. - Thanks · invite questions. - TIME — ~40 s --- # Q&A ( ≈ 08:38 onwards) - Before fielding: take a breath, thank the audience. - If an aggressive-sounding question comes: restate it in your own words first. That buys you 5 seconds and guards against mishearing. - If you don't know the answer: say so plainly, then say what you'd need to know to answer. - If a question tries to pull you into defending an adjacent claim you didn't make: name the move ("that's not quite the claim I'm making — I'm claiming …"), then address the nearest thing you did say. - Anticipate and have ready: - "Isn't this just a defence of prompt engineering?" — No. §§1–3 are in-principle; §4 is speculative about practice. The in-principle argument doesn't depend on the coda. - "What about [non-analytic conception of philosophy]?" — Acknowledged; the paper assumes a text-focused conception (Dellsén, Williamson), and the argument runs within that. - "Couldn't a specialist model do X better?" — §4 flags this is the least-supported claim; Sellars is the reason, and it's a bet, not a proof. - "Merleau-Ponty edge case — isn't this a bigger concession than you're letting on?" — Acknowledge its narrowness is a claim, not a proof; most of the discipline works from already-articulated phenomenology, that's the scope claim. --- # Catch-up levers (if running behind) In order of cost-to-cut (cheapest first): - At **08:14** — if past 08:15, trim slide 11 (the page) to the first lede + Floridi quote (skip the "hypothesis-weighing, objection-handling…" elaboration). - At **08:19** — if past 08:21, trim slide 21 (sources saturated) to display + sources list (skip the literary elaboration + saturation closer). - At **08:25** — if past 08:27, trim slide 23 (the narrow limit) — read the Merleau-Ponty sentences once, skip the "reverse roles / cannot simultaneously both be toucher" elaboration. - At **08:31** — if past 08:33, trim slide 27 (distillate ≠ distillation) to display + first lede only. The figure still carries. - At **08:37** — if past 08:38, go straight to the three-beat close without preamble. Emergency 30-second trim: skip slide 07 (§1 payoff). The §1 handoff to §2 can be done in the §2 DIV opening. --- # Timing check (cumulative) - 01 title ≈ 40 s → 08:00:40 - 02 §0 ≈ 80 s → 08:02:00 - 03–07 §1 ≈ 60+110+90+110+60 = 430 s → 08:09:10 - 08–15 §2 ≈ 50+90+80+90+60+120+70+60 = 620 s → 08:19:30 - 16–24 §3 ≈ 55+120+75+80+90+80+90+100+50 = 740 s → 08:31:50 - 25–28 §4 ≈ 50+85+90+100 = 325 s → 08:37:15 - 29 close ≈ 40 s → 08:37:55 Running total ≈ 37:55. Buffer ≈ 2:05 before 08:40. Comfortable unless you stall twice. --- # Afterwards - Don't take the laptop offline until any hallway / DM questions have slowed. - Note down any objection or question you couldn't answer cleanly — capture in vault for next revision. - If any source was disputed: flag for re-check against the extracted markdown in Learning/generating-philosophy/.