# opus[1m] ## Retrieval Notes - Session id: `c90ba0dc-bcde-4c0b-8d7f-7d32f8051d94` - Last activity: `2026-03-20T15:28:34.372Z` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User opus[1m] ### User take a look a look at this transcript of my paper along with the corresponding longform poject that is being discussed (secs. 0 -3). I would like you to write a complete report on all of the issues that were raised in the conversation, this is not an easy task, please take it seriously. make sure that you explain your analysis, justification, reasons etc etc. for your answer BEFORE giving me your final answer. Please give me a long, deep, and detailed answer, with lots of options/ideas: as we are in the 'kicking ideas around' stage, I want as much feedback as you can give me so that i have a lot to consider. ### Assistant Let me read the transcript and the manuscript sections first. ### Assistant Now let me read all four manuscript sections in parallel. ### Assistant Let me also read the active Section 2 file (the Index shows "2. Likeliness, Loveliness, LLMs" as the active one, not "2. LLMs and Abduction"). ### Assistant Request timed out ### User you got cut off, please continue from where you left off. ### Assistant I have all the material now. Let me write the full report. Let me work through this transcript systematically, because the issues raised are layered — some are local editorial problems, some are structural proposals about how to reorganise sections, and some are genuine philosophical difficulties that Enrico is flagging as unresolved. I need to disentangle all three layers. Starting from the top of the transcript. ## The Douglas Adams quotation (lines 30–56) The first substantive issue. Nick opens by saying the Adams quotation "is not super clear" in the role it plays — he noticed this when presenting at a workshop. Enrico's response clarifies what the quotation is actually doing: introducing the idea that in 2026 you can ask machines philosophical questions, and the problem was that the question was not asked properly (i.e., a prompting issue). Two options emerge: 1. Drop or cut back the quotation in the introduction, replace with a brief rhetorical contrast between the science-fiction case and the present situation, then move the detailed Adams example to a later section on prompting. 2. Keep it but only if there's a section on prompting. They converge on option 1: a brief reference in the intro ("in science fiction you might ask an AI something absurd; in 2026 we can actually ask machines philosophical questions"), with the detailed Adams example moved later. Nick suggests moving it to the conclusion. Looking at Section 4's final bullet, I can see that the Adams quotation has indeed been placed at the end of the moves outline — "Humanity asked a computer to do philosophy. It received the answer '42'..." — so this has been partially implemented already. But the introduction (Section 0) still has the full quotation and extended discussion, plus a `%%maybe drop the quote, and cut or cut back the whole example%%` comment. So the intro hasn't been revised yet. There's also a further development: by line 260–270, they discuss whether the final section should be about prompting or whether prompting should be deferred to another paper. They settle on using the Adams quotation as a conclusion rather than as an introduction, with the prompting section deferred. This makes the paper "almost finished." ## Missing references and factual errors (lines 58–68) Enrico flags: - A missing reference in "other updates" - "Stoljar" has no date - "crystallographic data" may be the wrong term for the Watson-Crick example — Enrico says crystallographic is "not biological" Looking at Section 1, line 4: "the same crystallographic data" — this is indeed the Watson-Crick passage. Now, actually, Watson and Crick did use X-ray crystallography data (Rosalind Franklin's Photo 51), so "crystallographic" is technically correct here — X-ray crystallography was the method used. But Enrico's concern seems to be that "crystallographic" sounds like it belongs to inorganic chemistry rather than biology. This is worth noting: the term is defensible but might mislead readers who associate crystallography with minerals rather than with the technique used to study biological structures. ## The dependence claim (lines 70–76) Enrico suggests adding "or not" to the dependence formulation: "depends or not." Looking at Section 1, line 8: "more accurately or more comprehensively representing the dependence relations in which a phenomenon stands to others" — there's already a `%%or not, add the negative%%` comment. So this is flagged but not yet fixed. ## "Two speakers on Twitter" (lines 78–80) Enrico says it's one speaker, not two. Looking at Section 1, line 8: "Two speakers on Twin Earth" — wait, Enrico says "two speakers on Twitter" but the text says "Two speakers on Twin Earth." Let me re-read... Enrico's point at line 78 is about a different passage. Actually, re-reading more carefully, the transcript at line 78 says "two speakers on Twitter" but the manuscript says "Two speakers on Twin Earth." It's possible the manuscript was revised after this conversation, or Enrico misspoke — but regardless, the Twin Earth thought experiment involves two speakers (one on Earth, one on Twin Earth), so "two speakers on Twin Earth" as currently written may indeed be inaccurate. The thought experiment has one speaker on Earth and one on Twin Earth — they're not both on Twin Earth. Hmm, but actually, Putnam's original formulation does describe Oscar on Earth and Twin Oscar on Twin Earth. So "two speakers on Twin Earth" is wrong — it should be "two speakers, one on Earth and one on Twin Earth" or just "speakers on Earth and Twin Earth." The %%not correct description of thought experiment%% comment in the manuscript at line 8 of Section 1 seems to flag exactly this. ## Anaphoric reference clarity (lines 82–88) Enrico says "we have some call this overfitting" has unclear anaphoric reference — "the last sentence" doesn't clearly connect. Looking at the current manuscript, the Williamson overfitting passage in Section 1, line 18, does now include a `%%this does not work well given what precedes immediately%%` comment. So this issue is recognized but unresolved. ## "Philosophical corpus" sentence too enigmatic (lines 90–95) At the end of Section 1, there's a sentence about philosophical corpus being "shaped by repeated judgements." Enrico says this is unclear at this stage and needs more unpacking because the concept is explained later. The current manuscript has exactly this comment: `%%this sentence is unclear at this stage, it fits better with what comes later%%` at line 22 of Section 1. ## Missing transition between sections (lines 96–98) Enrico flags that the transition from Section 1 to Section 2 is "too abrupt" — "Floridi and others argue that..." appears without any bridge. There's a `%%there needs to be a signpost bridge here or at the beginning of the next sections%%` comment at line 25 of Section 1. Not yet fixed. ## "External feedback loop for posterior evaluation" (lines 100–108) In the Floridi section, Enrico says the phrase "external feedback loop for posterior evaluation" is "not understandable" without explanation. Nick suggests cutting it. Looking at the active Section 2 (Likeliness, Loveliness, LLMs), line 10: the phrase is still there — "without what Floridi et al. call 'an external feedback loop for posterior evaluation' (pp. 5–6)" — with a `%%this is not understandable by itself, cut it or elaborate on it.%%` comment. Still unfixed. ## "Strong work may be overlooked" (lines 110–112) Enrico suggests qualifying the corpus filtering claim: "strong work may be overlooked" or "sometimes overlooked." The active Section 2, line 16 has: "weak work survives and strong work is overlooked" with `%%hedge this slightly%%`. Still needs the qualification. ## Williamson recap needs signposting (lines 114–118) Enrico says that when Williamson reappears in Section 2, it reads as though new material is being introduced, but it's actually a recap. It should say "as said above" or similar. The active Section 2, line 18 has: `%%some redundancy here. if we need a recap it should be signposted%%`. Still unfixed. ## Lipton recap needs signposting (lines 122–124) Same issue: "Lipton distinguishes the likeliest from the loveliest" in Section 2 is a repeat from Section 1. Should be introduced as a recap. The active Section 2, line 20 has: `%%redundancy from previous section%%`. Still unfixed. Wait — actually, this is interesting. The manuscript has been reorganised since the transcript conversation. The Index shows the active Section 2 is "Likeliness, Loveliness, LLMs" and the old "LLMs and Abduction" is in ignored files. So some of the Lipton material has been moved from Section 1 into Section 2. The old Section 2 has a preserved block of Lipton material in a comment at the top. So the reorganisation has happened but the redundancy/recap signposting hasn't been sorted out. ## Walton sentence unclear (lines 126–132) Enrico flags a sentence about Walton that is unclear. Nick says he wants to remove Walton anyway because there are too many people in the paragraph. The active Section 2, line 18 has Walton struck through: `~~As Walton et al. observe...~~` with `%%i thought i wanted to remove walton. and this a long paragraph already%%`. So the deletion has been marked but not executed. ## Two senses of "likeliness" (lines 132–148) This is a significant conceptual point. Enrico identifies that "likeliness" is being used in two different senses: 1. Statistical likeliness (Floridi's sense) — the most probable continuation 2. Lipton's sense — the most warranted explanation, which is different from statistical probability The opium/dormative virtue example illustrates Lipton's sense: the dormative virtue explanation is "likely" in Lipton's sense (almost guaranteed true) but useless because it says nothing. This is not statistical likeliness. The active Section 2, line 20 has `%%there are two senses of likeliness here, make sure they are clearly distinguished%%`. This is a conceptual issue that affects the argument's coherence — if the paper trades on an ambiguity between statistical probability and Lipton's evaluative notion of likeliness, that's a problem. Actually, this is quite a deep issue. Let me think about it more carefully. The paper's argument seems to be: in a corpus filtered for loveliness, the statistically likeliest continuation also tends to be the loveliest. But "likeliest" in Lipton's sense doesn't mean "statistically most probable" — it means "most warranted" or "most probably true." So there are three things in play: 1. Statistical probability (what the LLM optimises for) 2. Lipton's likeliness (epistemic warrant) 3. Lipton's loveliness (explanatory depth) The paper's argument conflates (1) and (2) when it says "the likeliest continuation, given that data, will tend to be a lovely one." This is actually the crux of the paper's positive argument, and if the equivocation isn't resolved, it undermines the whole thing. ## "Properties theorists describe" (lines 150–170) Enrico can't understand the sentence: "A model trained on philosophical corpus is exposed to a distribution systematically skewed toward the properties theorists describe." He doesn't know what "the properties" are or who "the theorists" are. This is in the active Section 2, line 18 (last sentences of that long paragraph). The child-grammar analogy also doesn't clearly draw the parallel. Nick agrees it's overcomplicated. The active Section 2, line 18 still has this problem — the last two sentences about the child and the model are flagged: `%%these last two sentences are weird together...%%` ## "One may still insist" paragraph (lines 176–198) Several issues here: - "The reason a standard..." is a strange expression (should be "theory" or similar) - The whole paragraph feels like a conclusion without setup — the idea that philosophy's justifications for its standards are internal to the corpus needs more groundwork - "Blind review would be a defective practice, but it is not" — Enrico suggests softening ("surely it is not") The active Section 2, line 22 has `%%not a great paragraph, fix it%%` and `%%not how i write%%` markers. The blind review point appears at line 24 with `%%not how i write%%`. ## Lipton quotation too abrupt (lines 200–204) The Lipton squash quotation appears "out of the blue." The active Section 2, line 24 has `%%this change to lipton is too abrupt%%`. Nick says they're introducing the quotation badly. ## NOW THE BIG STRUCTURAL/PHILOSOPHICAL ISSUES ## Section 3: Chinese Room placement (lines 206–218) Enrico says the Chinese Room material ("rule-following without comprehension") should be cut or moved from its current position because: 1. It anticipates something that belongs later 2. Searle's Chinese Room can't be reduced to "rule-following" — the central feature is the feeling of not understanding while inside the room 3. He's not sure whether to use both Searle and Einstein or just one Looking at Section 3, line 12: the Chinese Room passage is already struck through: `~~The same holds for the Chinese Room...~~` with `%%this is argued better later, maybe not pre-empt that here%%`. So this has been partially addressed. ## THE CENTRAL PHILOSOPHICAL PROBLEM: Physics vs. Philosophy distinction (lines 220–242) This is the deepest issue Enrico raises, and it's not just editorial. He says: "I do not see that physics is so different from philosophy in a way strong enough for your point." The paper argues that philosophical thought experiments operate through described scenarios (textually available), whereas Einstein's physics required embodied perceptual experience (not textually available). But Enrico pushes back: 1. Einstein's experiences (feeling acceleration) can also be described. If descriptions are sufficient for philosophy, why not for physics? 2. If you supplement physics training data with descriptions of experience, you might not need Einstein either. 3. Mary's case (Jackson) involves a scenario not in any prior description — a scientist in a black-and-white room. That's structurally like Einstein: a novel scenario requiring imagination beyond prior descriptions. 4. The Chinese Room is also like this — it involves introspection/first-person experience just as Einstein does. Enrico sees two options: 1. Challenge Zahavi (accept that the distinction between physics and philosophy isn't as sharp as claimed) 2. Find a genuine reason why physics works differently (perhaps physics training sets are "only data" while philosophy training sets include descriptions) But even option 2 has problems: physics notebooks (like Einstein's reflections) also contain descriptions. And philosophy's thought experiments are precisely about finding cases that AREN'T in prior descriptions. This is flagged in Section 3, line 17 with several comments: `%%enrico is not sure whether both chinese room and einstein%%`, `%%enrico also doesn't see how physics is so different to philosophy in this respect%%`, etc. The current Section 3 attempts to address this with the Pigliucci material (philosophy as "empirically informed evoking" with propositional starting points) and the spectrum from coarse phenomenological facts to Merleau-Ponty's fine-grained observations. But Enrico's challenge hasn't been fully met. The paper currently says philosophy's starting points are "often already available as public descriptions" — but "often" is doing a lot of work, and the thought experiment cases (Mary, Chinese Room) seem to be exactly the cases where this isn't true. ## Too many examples in Section 3 (lines 330–338) Enrico flags: - The Dummett example doesn't seem like a big discovery — on one reading, it's obvious - The Islamic distinction is unfamiliar and doesn't seem to require experience - Nick admits the Islamic example came from the LLM, not from him - There may be too many examples overall Looking at Section 3, I see the Dummett example at line 28 but I don't see an Islamic example — it may have been removed already, or it might be in a version I haven't read. ## THE MAJOR STRUCTURAL PROPOSAL (lines 288–364) This is the most important part of the conversation for the paper's architecture. Enrico proposes reframing the entire paper around a unified structure: Section 1: Philosophy is in the text (the continental-inflected point). Section 2: Can we have philosophy in the text without abduction in the mind? (Floridi objection) Section 3: Can we have philosophy in the text without phenomenology in the mind? (Zahavi objection) The two objections share the same form: "you cannot have good text without the proper psychological process." The replies are structurally similar but different in detail: - For Floridi: the forms of abductive reasoning are already at work IN the corpus — the statistical process through the corpus extracts the right patterns. - For Zahavi: descriptions of phenomenological processes are in the corpus — even Einstein/Jackson turn their phenomenological insights into descriptions before they enter arguments. Enrico frames the unified reply as: "LLMs don't literally do abduction or have phenomenology, but they have abduction-star and phenomenology-star — enough to generate the same kind of text." This is a significant reorganisation. Looking at the current active files: - Section 2 is titled "Likeliness, Loveliness, LLMs" — which doesn't yet have the framing "can we have philosophy in the text without abduction in the mind?" Wait — actually, looking more carefully at the active Section 2, line 4: `## can we have philosophy in the text without abduction in the mind` — the heading IS there. And Section 3, line 4: `## can we have philosophy in the text without phenomenology in the mind`. So the unified framing HAS been implemented at the heading level. But Enrico says the current Section 2 still reads as "Floridi is wrong because we care about the text" — an easy objection. He wants it reframed as the harder version: "even if value is in the text, you cannot have valuable abduction in the text without abduction in the mind." Let me re-read the active Section 2 to see if this harder framing is in place... Lines 6-10 introduce Floridi's criticism. Lines 12-13 present the "statistical echo" worry. Line 14 begins the rebuttal — "statistical plausibility is always plausibility relative to a body of training data." But this goes directly to the corpus-quality argument without first steelmanning Floridi's position in the stronger form Enrico proposes. Enrico's proposed steelmanning: Floridi should be read not as saying "abduction is in the mind, not the text" (easy to rebut) but as saying "you cannot have valuable abduction in the text without abduction in the mind" (harder). This connects Floridi to the paradigm-shift/creativity concern — it's about whether there's genuinely creative abduction, not just whether there's abduction at all. This hasn't been fully implemented yet. ## The prompting section (lines 258–280) They discuss whether to include a section on how to prompt for good philosophy. They decide to defer this to another paper and use the Adams quotation as a conclusion. But the CEV (coherent extrapolated volition) example Nick mentions is interesting — a prompt hack that forces the LLM to produce the best version of an idea. Looking at Section 4, the moves outline does include material on prompting (dialectical framing, solution-gestured prompting, conversational iteration) at lines 26-32. So the prompting material IS in the paper currently. The question is whether it should be cut or kept. ## Experience sedimented in text (lines 282–290) Enrico raises an unresolved philosophical concern: can experience be sedimented in text? He uses grief as an example — grief is always about something, and descriptions of grief exist in literature and philosophy. But some philosophers would say that for innovative philosophy, mere descriptions aren't enough. This connects back to the physics/philosophy distinction problem. The mushroom-eating reference suggests there's another example they've discussed (perhaps psychedelic experience as something that resists textual capture). ## Enrico's proposed reply structure (lines 340–356) Two different mechanisms for the two objections: 1. For phenomenology (Zahavi): descriptions of phenomenological processes are in the corpus. Even Einstein turned his phenomenological insight into a description before it entered his argument. What really enters the argument is the description. 2. For abduction (Floridi): the forms of reasoning, the comparisons between possibilities, are already at work in the corpus. The statistical process extracts these patterns. Nick summarises: "for one of them we're saying something about the training data, specifically descriptions in the data. For the other we're saying something about the mechanism itself." This is a really important distinction. The Zahavi reply is about content (the right descriptions are in the data). The Floridi reply is about form (the right reasoning patterns are in the data). Both work through the corpus, but in different ways. Let me now think about what the current manuscript state looks like relative to all these issues... The current state seems to be: - Introduction: Still has the full Adams quotation (should be cut to a brief reference, with the detailed example moved to conclusion) - Section 1: Mostly in good shape but needs the enigmatic "philosophical corpus" sentence unpacked, missing transition at the end, and some reference fixes - Section 2: Has been reorganised with the new heading but still reads as the "easy" version of the Floridi objection; needs the stronger steelmanning; has numerous %%comments%% flagging voice issues, unclear sentences, and structural problems; the old Section 2 material has been partially absorbed - Section 3: Has the physics/philosophy distinction problem unresolved; Chinese Room struck through; too many examples; Enrico's challenge about whether physics is really so different not fully addressed - Section 4: Exists as a moves outline, seems to contain a lot of the material that's also (less effectively) in the current Sections 2 and 3 Actually, now I notice something important. Section 4 (the moves outline) contains much of the positive argument — the corpus filtering claim, the grammar analogy, the prompting modes, the Lipton convergence argument. But a lot of this same material also appears in the active Section 2. There seems to be significant overlap/duplication between Section 2 and Section 4. Section 2 may need to be stripped back to focus on the Floridi objection and reply, with the constructive case consolidated in Section 4. Let me also note the voice issues. The old Section 2 ("LLMs and Abduction") is riddled with %%not how i write%% comments and voice problems. The active Section 2 has fewer of these but still has some. This suggests the rewrite hasn't fully resolved the voice problems. Okay. I think I have a comprehensive picture now. Let me organise this into a report. # Report: Issues from the Draft Discussion (Sections 0–3) This covers every issue I can identify in the transcript, cross-referenced against the current manuscript state. I've grouped them into three tiers: architectural/philosophical (the ones that require thinking, not just editing), structural (section-level reorganisation), and local (editorial fixes, unclear sentences, voice). --- ## I. Architectural and Philosophical Issues These are the issues where Enrico is flagging something that hasn't been resolved by revision — where the paper's argument still has a gap or an equivocation. ### 1. The physics/philosophy distinction is not established (lines 220–242) The single hardest problem in the paper. The argument in Section 3 depends on a distinction: physics requires embodied perceptual experience (Einstein feeling acceleration), while philosophy works on described scenarios already available in text. Enrico challenges this directly: > "I do not see that physics is so different from philosophy in a way strong enough for your point." (line 226) His reasons: - Einstein's experiences of acceleration *can be described*. If descriptions suffice for philosophy, they might suffice for physics too — just supplement the physics training set with descriptions of experience. - Philosophical thought experiments (Mary, the Chinese Room) involve *novel scenarios not present in prior descriptions*. Mary's case requires imagining a scientist raised in a black-and-white room — that's not in any corpus of prior description, and it's structurally like Einstein imagining the elevator. - Both philosophy and physics seem to require a leap from described experience to something new. The distinction between them isn't as clean as the paper needs it to be. Enrico identifies two options (lines 236–238): - Option A: Accept the distinction is leaky and challenge Zahavi — concede that the limit applies to some philosophy too. - Option B: Find a genuine reason physics is different. Perhaps: physics training data is "only data" (numbers, measurements), while philosophy training data includes rich descriptions. But physics notebooks also contain descriptions, so this line is fragile. The current Section 3 attempts an answer through Pigliucci ("empirically informed evoking" with propositional starting points) and the spectrum from coarse phenomenological facts to Merleau-Ponty's fine-grained observations. But the manuscript acknowledges the problem without fully resolving it: "The phenomenological leading edge — where new experiential distinctions are first articulated from lived attention — remains a genuine limit, and we do not want to minimise it" (Section 3, line 30). This is honest, but it concedes ground that the paper might need. The %%comments%% in Section 3 (line 17) flag this: `%%enrico also doesn't see how physics is so different to philosophy in this respect%%`. It remains the paper's open wound. Possible directions to consider: - Draw the distinction not at the level of *descriptions* but at the level of *what enters the argument*. In physics, the axiom (equivalence principle) was abduced from a perceptual episode that hasn't been previously formalised. In philosophy, the "axioms" (our shared judgements about cases) are already in circulation — already formalised, debated, revised. The difference isn't that philosophy doesn't need experience; it's that philosophy's experience has already been processed into propositions by the time it enters philosophical argument. - Concede the limit more gracefully: say that philosophy and physics sit on a continuum, and that the cases where philosophy requires genuinely novel phenomenological attention (Merleau-Ponty's self-touching) are real but represent one end of a spectrum, while most philosophical work operates at the other end. - Use the Enrico-suggested framing: physics training sets lack descriptions of experience; philosophy training sets are *constituted by* descriptions of experience. The asymmetry isn't about whether experience matters but about where it shows up in the corpus. - Nick mentions someone (unclear audio) who writes about the difference between science and philosophy — "philosophy is a vocation of concepts" — this might provide the distinction needed. ### 2. Steelmanning Floridi (lines 288–316) Enrico argues that the current Section 2 treats Floridi as making an easy objection ("abduction is in the mind, not the text") and then dispatching it with the reply "we care about the text, not the mind." He proposes a harder reading: > "Even if value is in the text, you cannot have valuable abduction in the text without abduction in the mind." (line 304) This is the difference between: - Weak Floridi: Abduction isn't in the text. (Reply: yes it is, because we evaluate texts.) - Strong Floridi: Abduction may be in the text, but *good* abduction in the text requires *real* abduction in the mind behind it. Without the psychological process, you get surface-level imitation. (Reply: harder — needs the corpus-filtering argument.) The current active Section 2 has the heading "can we have philosophy in the text without abduction in the mind?" — which is the strong framing. But reading through the section, the argument still largely takes the weak-Floridi route: "statistical plausibility is relative to training data" → "the corpus is filtered" → "so probable continuations tend to be good." The stronger version would first construct Floridi's best argument — that the filtered corpus was produced *by* human abductive reasoning, and without that reasoning in the production process, you can't get the same quality out — and *then* rebut it by showing how the corpus itself encodes the patterns of abductive reasoning. This connects to the unified framing Enrico proposes (see Structural Issue 1 below). ### 3. The two senses of "likeliness" (lines 132–148) Enrico identifies an equivocation in the paper's use of Lipton. There are (at least) two senses of "likely" in play: - Statistical probability: what the LLM optimises for (next-token prediction) - Lipton's "likeliness": the most *warranted* or most *probable-to-be-true* explanation These are not the same thing. The dormative virtue example shows why: "opium has a sleep-inducing power" is the *likeliest* explanation in Lipton's sense (almost certainly true) but it's explanatorily empty. It's not the *statistically most probable* sentence that would follow a prompt about opium. The paper's argument trades on the convergence of statistical probability and philosophical quality in a filtered corpus. But if "likeliness" means different things in different parts of the argument, the convergence claim is shakier than it looks. The active Section 2, line 20 flags this: `%%there are two senses of likeliness here, make sure they are clearly distinguished%%`. This needs careful disambiguation. You might need three clearly labelled concepts: - Statistical probability (what the LLM does) - Lipton's likeliness (epistemic warrant / probable truth) - Lipton's loveliness (explanatory depth / understanding) The paper's argument is that in a corpus filtered for loveliness, statistical probability tracks loveliness. Lipton's likeliness is a separate thing and shouldn't be confused with either. ### 4. Experience sedimented in text (lines 282–290) Enrico flags an unresolved philosophical concern that cuts across Sections 2 and 3: can experience genuinely be "sedimented" in text? He uses grief as a test case — grief is always about something, and descriptions exist, but some philosophers would say the descriptions aren't enough for innovative philosophy. The current Section 3 handles this with the Austin passage about surfaces looking different under different lighting (line 16) and the grief passage (line 18, with `%%too quick%%` comment). But the treatment is acknowledged as insufficient. This connects to the physics/philosophy problem: if experience can't be fully sedimented in text, then philosophy has the same problem as physics, and the paper's distinction collapses. --- ## II. Structural Issues ### 1. Unified framing for Sections 2 and 3 (lines 308–364) Enrico proposes the most important structural change in the conversation. The paper should be organised around a shared question with two variants: - Section 1: Philosophical value is in the text (the ground-clearing move). - Section 2: Can we have philosophy in the text without abduction in the mind? (Floridi) - Section 3: Can we have philosophy in the text without phenomenology in the mind? (Zahavi) The unified reply: LLMs don't literally perform abduction or have phenomenology, but they have "abduction-star" and "phenomenology-star" — substitute processes (statistical learning over the corpus) that produce the same kind of text. But the two replies differ in mechanism: - Against Floridi: the *forms of reasoning* (comparisons between possibilities, selection among hypotheses) are already at work in the corpus. The statistical process extracts these patterns. This is about *form*. - Against Zahavi: *descriptions* of phenomenological processes are in the corpus. Even Einstein turned his perceptual insight into a description before it entered his argument. This is about *content*. Nick's summary (line 368): "for one of them we're saying something about the training data, specifically descriptions in the data. For the other we're saying something about the mechanism itself." Current status: the headings have been updated (Section 2: "can we have philosophy in the text without abduction in the mind"; Section 3: "can we have philosophy in the text without phenomenology in the mind"). But the body text hasn't been fully rewritten to reflect this unified framing. The introduction still lacks a roadmap (`[ROADMAP TO GO HERE]` at line 30 of Section 0). This unified structure should be stated there. ### 2. Adams quotation placement (lines 30–56, 260–268) Decision: Brief rhetorical reference in the introduction (science fiction vs. reality), move the detailed Adams example to the conclusion. Currently, the introduction still has the full quotation (Section 0, lines 9–18). Section 4's final bullet uses Adams as a conclusion. The introduction needs to be cut down. ### 3. Prompting section (lines 258–268) Decision: defer the practical prompting section to another paper. Use Adams as conclusion instead. But Section 4 currently contains substantial prompting material (dialectical framing, solution-gestured prompting, conversational iteration at lines 26–32). If prompting is deferred, this material may need to be cut or reduced to a brief gesture. Alternatively: Enrico says at line 262 that they could "at least allude to how these things could be used" — which would justify keeping a lighter version of the prompting material. ### 4. Overlap between Sections 2 and 4 The active Section 2 and Section 4 (moves) cover much of the same ground: the corpus-filtering argument, the grammar analogy, the Floridi rebuttal, the Lipton convergence claim. If Section 4 is meant to be the constructive case ("can LLMs actually produce such texts?"), then Section 2 should be stripped back to focus on the Floridi objection and its rebuttal, with the positive argument consolidated in Section 4. ### 5. Too many examples in Section 3 (lines 330–338) Enrico flags: - The Dummett example (assertoric content vs. ingredient sense) doesn't read as a big discovery — "a ten-year-old child can say yes, you sent this as the meaning of this song, but if you put it in a bigger sentence it contributes to the meaning of the bigger sentence." Either elaborate it or replace it. - The Islamic distinction example is unfamiliar to Enrico and doesn't seem to require experience. Nick admits it came from the LLM, not from him. This should be removed. - There may be too many examples overall — the section needs distilling. I don't see the Islamic example in the current Section 3, so it may have been removed. The Dummett example remains (line 28) and could use either more development (to make clear why the distinction is non-obvious) or replacement. --- ## III. Local/Editorial Issues ### Reference problems (lines 58–60) - Missing reference in "other updates" - "Stoljar" has no date - These are straightforward fixes once the references are tracked down. ### "Crystallographic data" (lines 62–68) - Technically correct (X-ray crystallography was the method), but may mislead readers. Consider "X-ray crystallographic data" or "X-ray diffraction data" to make the biological application clearer. ### Dependence claim needs the negative (lines 70–76) - Section 1, line 8 has `%%or not, add the negative%%` twice. The formulation should include "whether or not" a phenomenon depends on others. ### Twin Earth description (lines 78–80) - "Two speakers on Twin Earth" is inaccurate — one is on Earth, one is on Twin Earth. Section 1, line 8 has `%%not correct description of thought experiment%%`. Fix to "speakers on Earth and Twin Earth" or similar. ### Anaphoric reference to overfitting (lines 82–88) - The transition to Williamson's overfitting passage needs clearer connection to what precedes it. Section 1, line 18: `%%this does not work well given what precedes immediately%%`. ### "External feedback loop" (lines 102–108) - Cut or explain. Currently flagged in Section 2, line 10: `%%this is not understandable by itself, cut it or elaborate on it.%%` ### "Strong work is overlooked" needs hedging (lines 110–112) - Add "sometimes" or "may be." Section 2, line 16: `%%hedge this slightly%%`. ### Williamson and Lipton recaps need signposting (lines 114–124) - When these figures reappear in Section 2, introduce with "as argued above" or similar. Flagged at Section 2, lines 18 and 20. ### Walton should be removed (lines 128–132) - Already struck through in Section 2, line 18. Execute the deletion. ### "Properties theorists describe" (lines 150–170) - Unclear who "theorists" are and what "properties" are. The child-grammar analogy doesn't clearly draw the parallel to LLMs. Flagged at Section 2, line 18. Needs rewriting — probably simplify by using the LLM directly rather than going through the child analogy, and specify what the properties are (the theoretical virtues Williamson describes). ### "One may still insist" paragraph (lines 176–198) - "The reason a standard..." → "The reason a theory..." - The paragraph needs more setup before the conclusion that philosophy's justifications are internal to the corpus - "But it is not" about blind review → soften to "but surely it is not" or similar - Flagged as `%%not a great paragraph, fix it%%` in Section 2, line 22 ### Lipton squash quotation too abrupt (lines 200–204) - Needs a smoother introduction. Flagged: `%%this change to lipton is too abrupt%%` in Section 2, line 24. ### Chinese Room placement (lines 210–218) - Already struck through in Section 3. Execute the deletion. Enrico suggests either using Searle OR Einstein but not both, since they function similarly (both involve first-person experience/introspection). ### Voice issues throughout Section 2 - The old Section 2 ("LLMs and Abduction") is saturated with `%%not how i write%%` markers — triplet examples, formulaic transitions, literary vocabulary. The active Section 2 is better but retains some of these problems (lines 22, 24). A dedicated voice pass is needed once the structural issues are resolved. --- ## Summary of Priorities The issues aren't all equal. Here's how I'd rank them by impact on the paper: 1. The physics/philosophy distinction (Issue I.1) — this is the paper's philosophical weak point and Enrico is clear it's "not just an editorial issue" but "a philosophical problem." Until this is resolved, Section 3 doesn't work. 2. Steelmanning Floridi (Issue I.2) — the current Section 2 dispatches a weak version of the objection. Making Floridi's argument harder makes Section 2 worth four pages; the current version doesn't justify the space. 3. Unified framing (Issue II.1) — the headings are in place but the body text and introduction don't yet reflect the parallel structure. This is the organising principle of the paper. 4. The likeliness equivocation (Issue I.3) — if this isn't disambiguated, a careful reader will spot the equivocation and the convergence argument won't land. 5. Everything else is local and can be fixed in a revision pass once the architecture is settled. One remaining uncertainty: Section 4 (moves outline) seems to contain the constructive case that would follow from Sections 1–3. Its relationship to the current Sections 2 and 3 — how much material migrates where, whether the prompting material stays — depends on the architectural decisions above. The manuscript is close, but the architecture needs to settle before the prose can be polished. *Il problema più spinoso resta la frontiera tra fisica e filosofia — se la distinzione non regge, l'intero impianto del terzo capitolo va ripensato da capo.*