# I'd like to talk and just kick around some ideas about my generating philosophy paper. To get thi... ## Skills Used - epistemic-discipline - source-work ## Retrieval Notes - Session id: `c84a5338-0e4a-4253-b2e8-223f6b4a9530` - Last activity: `2026-02-10T12:05:33.513Z` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User I'd like to talk and just kick around some ideas about my generating philosophy paper. To get things clear right from the start, I am not interested whatsoever in progressing the draft today. I'm interested in deepening my understanding of the papers that I'm working on. In particular, I would like to work on the following papers with you. I think they're all in the folder, but I'm not 100% sure. LLMs Can't Jump, the Floridi paper and section 9.2 of Widening the Picture by Williamson, maybe 9.3 as well. You know what I'm talking about? ### Assistant I'll start by loading the relevant skills and getting oriented with your project context. ### Assistant Good — let me get oriented. I'll read your session note for the generating philosophy project, today's daily note, and check what's in the project's academic references folder. ### Assistant Good — I can see all three papers you mentioned are in the folder. Let me now extract them so we can work from the actual text. I'll start with the source extraction for all three. ### Assistant All three are manageable. The Williamson is 91 pages (the full chapter from *The Philosophy of Philosophy*), but you only want sections 9.2 and 9.3, so I'll extract a range. Let me extract all three now. ### Assistant Good, all three extracted. Let me now read them — I'll read Zahavy and Floridi in full, and search Williamson for sections 9.2 and 9.3. ### Assistant Good — I can see sections 9.2 and 9.3 start at lines 1493 and 2305 respectively. Let me read those sections now. ### Assistant Good. Let me read the rest of 9.3. ### Assistant I now have all three source texts extracted and read. Let me confirm what I've got and respond to your question. ### Contemplation Hmm, let me think about what Nick is asking. He wants to kick around ideas — no draft work, just deepen understanding of these three papers. He's asking whether I know what he's talking about. So let me first confirm that, and then open the space for conversation. What do I have? **Zahavy, "LLMs Can't Jump" (2026, Google DeepMind)**. A position paper that uses Einstein's formulation of General Relativity as a case study. The argument structure: three types of inference — induction, deduction, abduction (following Peirce). Zahavy's claim is that LLMs have mastered induction (statistical pattern matching) and are "rapidly conquering" deduction (formal proof, citing AlphaProof etc.), but lack the mechanism for abduction — the generation of novel explanatory hypotheses. The specific bottleneck is what he calls "manipulative abduction" (from Magnani et al.): embodied simulation that grounds abstract symbols in physical sensation. Einstein's "happiest thought" — the falling observer — as the paradigmatic abductive jump. Zahavy calls LLMs "high-dimensional 'Chinese Rooms' (Harnad, 1990)" that manipulate language without access to physical referents. His proposed solution: physically consistent world models that enable counterfactual simulation. **Floridi et al., "What Kind of Reasoning (if any) is an LLM actually doing?"**. The argument: LLMs have a "stochastic core" and an "abductive appearance." They do not perform abductive inference; their outputs *resemble* abduction because they were trained on human texts that encode reasoning structures. The appearance is systematic, not random — but it's still appearance, not genuine reasoning. They distinguish weak abduction (hypothesis generation) from strong abduction (selecting the best explanation). LLMs seem to perform "zeroth-order abduction" — generating plausible continuations based on learned associations. The paper is careful: they don't say LLMs are "stochastic parrots" full stop, but that the appearance of reasoning is useful (as brainstorming assistants, hypothesis generators) while the stochastic core means outputs must be critically examined. Five objections and replies. Interesting concession in Reply 1: "the output structure of LLMs often resembles that of an abductive reasoning process... this resemblance is not random but systematic, resulting from training on human explanations." **Williamson, "Widening the Picture," section 9.2 (Abductive Philosophy)**. The argument for using abduction/IBE as a legitimate method in philosophy. Opens with the Dummett anecdote — Dummett thought IBE was illegitimate in philosophy; Williamson found this surprising. Then a sketch of abduction: ranking potential explanations by intrinsic theoretical virtues (elegance, simplicity, strength, non-ad-hocness). Evidence base for philosophy: total evidence, including all knowledge from sciences and common sense. The comparison with mathematics — abduction justifies first principles even there (Russell 1907). The contrast between deductivist and abductivist methodology in philosophy: deductivism tends to stalemate, abductivism bypasses deadlocks by accumulating evidence and developing better explanations. Abduction rewards bold, precise theories. The over-fitting point from Forster and Sober applied to philosophy (Gettier literature as philosophical over-fitting). **Section 9.3 (Model-Building in Philosophy)**. Philosophy progresses through building better models — hypothetical examples described precisely enough to be formally tractable. Models as a fall-back when the target is too complex for exceptionless generalizations. The human world is paradigmatically complex. Models in epistemology (lotteries, epistemic logic), philosophy of language (Carnap's possible worlds semantics). The method: explore the model deductively, then compare with target phenomenon. Counterexamples play a smaller role — what defeats a model is a better model, not a counterexample. Simplicity, idealization, isolating factors. Now, how do these connect to Nick's paper? Looking at his session note: - Section 1 of his paper steelmans the Floridi critique but extracts a concession (structure IS learned from training data) - Section 2 uses Williamson on abductive methodology — intensifies the concern, then pivots to the intrinsic nature of theoretical virtues - The Zahavy paper seems relevant as another voice in the "LLMs can't do abduction" space — but it's specifically about *scientific* abduction (embodied, physical), which may be a different claim from what's at stake for *philosophical* abduction So yes, I know what Nick's talking about. Let me tell him. ### Final Answer Yes, I know exactly what you're talking about. All three are in the folder and I've now extracted and read them in full. Here's what I've got open: 1. **Zahavy, "LLMs Can't Jump" (Jan 2026, Google DeepMind)** — 10-page position paper. Uses Einstein's formulation of General Relativity as a case study to argue LLMs are structurally incapable of abduction. Three-part framework: LLMs have mastered induction, are conquering deduction, but lack the "Jump" — manipulative abduction grounded in embodied simulation. Proposes world models as the solution. 2. **Floridi, Morley, Novelli & Watson, "What Kind of Reasoning (if any) is an LLM actually doing?"** — 26-page paper. The "stochastic core, abductive appearance" thesis. LLMs don't reason abductively; their outputs *resemble* abduction because training data encode human reasoning structures. That resemblance is systematic, not accidental, but it's still statistical pattern completion, not inference. Five objections considered. 3. **Williamson, "Widening the Picture" §9.2 (Abductive Philosophy) and §9.3 (Model-Building in Philosophy)** — from *The Philosophy of Philosophy*. 9.2 argues philosophy should use abductive methodology more boldly and self-consciously — opens with the Dummett anecdote, sketches abduction, argues it's legitimate for philosophy as it is for mathematics and science. 9.3 argues philosophy makes progress by building better formal models rather than discovering exceptionless universal generalizations. Where do you want to start? I'm ready to work through any of these with you — whether that's close reading of particular passages, working out how they relate to each other, or thinking through what they mean for the generating philosophy project without touching the draft. *Ho estratto tutti e tre i testi e li tengo aperti: possiamo attraversarli riga per riga se vuoi.*