# Section 1 — What LLMs Aren't Doing > [!info] Working note > Paragraph-by-paragraph plan for Section 1, based on the detailed plan from [[Notes/Generating Philosophy - Integration Queue|the session of 2026-02-11]]. Block quotes to be verified against source extractions. This note is for drafting — final version goes into the longform project scene file. --- ## Plan **Overall function:** Present the two best general critiques of LLM reasoning (Floridi: epistemological; Zahavy: architectural). Frame them as serious and domain-specific — *not* as claims about philosophy. Set up the question: what happens when we ask about philosophy specifically? --- ### ¶1 — Opening: the question, framed through a concrete observation - These metaphilosophical questions — about philosophy's relationship to its textual medium, about whether the process behind a text matters for its evaluation — are arising now because of LLMs. The technology forces the question. - Two recent critiques — one epistemological, one architectural — present serious challenges. Neither targets philosophy specifically, but both would generalise if their assumptions hold. - **Style note:** Short, orienting. Don't preview the thesis. Preview the *structure*: "Two critiques, from different directions. The question is whether their assumptions hold for philosophy." - **Style note:** Open with something concrete — not an abstract philosophical thought experiment. Consider what a real, grounded opening looks like here. --- ### ¶2 — Floridi's central claim: stochastic core, abductive appearance - The duality between mechanism and output. LLMs are probability-distribution samplers; their outputs resemble reasoning. - Block quote needed: Floridi et al. on the "conceptual space between" stochastic processes and human-like abduction - Block quote needed: Floridi et al. Abstract on "stochastic core / abductive appearance" - **Note:** Present cleanly. No editorialising yet. Floridi's position should feel strong and precise. --- ### ¶3 — Why the outputs look abductive: training on human reasoning - The abductive appearance isn't accidental — it arises because training data encodes reasoning structures. This is Floridi's own point. - Block quote needed: Floridi et al. on Wikipedia articles, Q&A forums, scientific papers encoding IBE patterns - **Seed planted here:** The training data encodes reasoning structures. This seed grows in Sections 2 and 3. Don't highlight it; just let Floridi say it. --- ### ¶4 — What Floridi says LLMs lack: truth-tracking, verification, grounded semantics - Objective function is next-token prediction, not truth. Hallucinations reveal the architecture. LLMs cannot lie because they don't know what's true. - Block quote needed: Floridi et al. on generating text "without grounding them directly in truth, semantics, verification" - Include the hallucination/over-abduction point. - Grounding problem: LLMs lack sensorimotor verification. But Floridi concedes they have "a vast repository of commonsense causal knowledge, although not explicitly structured." --- ### ¶5 — Floridi's own question: does process matter? (THE HINGE) - The remarkable passage where Floridi himself wonders whether the distinction matters. - Block quote needed: Floridi et al. p. 12–13 — "if an AI can generate the same explanatory hypothesis a human would, does it matter that the process was different?" - Also the concession about commonsense knowledge encoded in training data. - **This is the hinge.** Floridi raises the question but doesn't develop it. The paper will develop it. Don't telegraph this — note the question and move on to Zahavy. --- ### ¶6 — Transition to Zahavy: a different kind of critique - Floridi's critique is epistemological (what are the outputs worth?). Zahavy's is architectural (what can the system do in principle?). They share a Peircean framework but attack from different angles. - One paragraph. Brief. Establishes the contrast. --- ### ¶7 — Zahavy's central claim: the E→A Jump requires embodied simulation - Peirce's triad. LLMs have mastered induction, are conquering deduction, but lack abduction. The E→A Jump — from sense experience to axioms — is the creative moment, and it requires embodied simulation. - Block quote needed: Zahavy on "mastered Induction... conquering Deduction... lacks the mechanism for Abduction" - Einstein case study: the "happiest thought." Not data analysis but imagining what it would be like to be a physical subject in a particular situation. - Block quote needed: Magnani's "manipulative abduction" — "active interaction with mental models" --- ### ¶8 — Why compression fails for physics: no error signal - Zahavy's anti-Schmidhuber argument. The Newtonian framework was empirically adequate. No dataset to compress against. The creative leap came from imagining, not error minimisation. - Block quote needed: Zahavy on "near-zero loss function" - **Seed planted here:** Philosophy has *abundant* error signals (dialectical saturation). Don't water this seed yet — it pays off in Section 3. --- ### ¶9 — Zahavy's two concessions: deductive capacity + domain restriction - The deductive concession: LLMs could derive GR given the axioms. AlphaProof as evidence. - Block quote needed: Zahavy on domain restriction — "specifically tailored to the physical sciences, where the object of study is external material reality" - The domain restriction is the opening. Zahavy doesn't mention philosophy. But the restriction opens the door. Note this. Don't walk through it yet. --- ### ¶10 — The shared assumption: both critiques ask about the producer - Taken together, the critiques look like they cover everything. One says the outputs aren't what they seem; the other says the architecture can't produce them anyway. - But both share an assumption: they frame the question as whether LLMs can perform abductive inference *internally*. Floridi asks whether the internal process is abductive. Zahavy asks whether the architecture supports abduction. - Neither asks whether this is the right question for every domain — and in particular, neither asks what happens when the domain is philosophy, a discipline whose relationship to its textual medium is unlike any of the cases they discuss. - **Closes Section 1. Opens Section 2.** --- ## Notes to self - [ ] Extract and verify all block quotes from source PDFs - [ ] Find a concrete opening (not the smoke/fire example — that was too abstract and inaccurate) - [ ] Make sure characterisation of LLMs is technically defensible (avoid "statistical pattern-matching" oversimplification) - [ ] Floridi and Zahavy are foils, not opponents — generous treatment throughout - [ ] The word "abduction" is being used differently by Floridi and Zahavy — this feeds Section 2's disaggregation. Let the tension sit here without resolving it.