# \\\\\"Zahavy, \\\\\"LLMs Can't Jump\\\\\" (2026, Google DeepMind). A position paper that uses E ## Skills Used - session-search ## Retrieval Notes - Session id: `6bc0ff9d-6f54-406c-9f71-9fa2fa005951` - Last activity: `2026-02-10T12:17:46.294Z` ## My Notes <!-- Add your notes here. This section is preserved across syncs. --> ## Conversation ### User "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." you'll see a conversation has just started in Claude Cody and Claudidian in which this paragraph comes up. I want you to find the chat and then I want you to go online and download me the following paper. Harnad 1990 ### Assistant I'll start by finding the chat you're referring to and then search for the Harnad 1990 paper to download. ### Assistant Let me try a more targeted search while also looking for the Harnad paper online in parallel: ### Assistant Good — I found the paper. It's **Harnad, S. (1990). "The Symbol Grounding Problem." *Physica D*, 42, 335–346.** There's a direct PDF hosted at Oxford. Let me check on the session search and download the paper. ### Assistant The session search is taking a while — it's scanning through many session files. Let me proceed with downloading the paper while that runs, and also try a faster search approach: ### Assistant The session files are large and grep is taking time. Let me try a faster approach — searching just today's sessions: ### Assistant The path indirection hit the wrong directory. Let me fix: ### Assistant Good, I can see recent sessions. Let me search the most recent ones for "Zahavy" and simultaneously download the Harnad PDF: ### Assistant I found the paper and multiple sessions mentioning Zahavy. The PDF is downloaded — 15 pages, ~120KB, confirmed title "The Symbol Grounding Problem". Now let me identify which session is the active one (the current conversation in "Claude Cody and Claudidian") and move the PDF to Learning: ### Assistant Let me check the most recent session — `6bc0ff9d` (49KB, just updated at 13:16) — and also the current one (`c7e942c8`, which is tiny, 375 bytes — that's likely this conversation). The `6bc0ff9d` is probably the one you're referring to: ### Assistant That's *this* session — `6bc0ff9d` is the current conversation. Let me look at the other sessions that mention Zahavy, specifically the ones that are clearly different conversations: ### Assistant Found it! There are multiple sessions from the same conversation — you ran `/contemplate` to kick around ideas about your generating philosophy paper, discussing "LLMs Can't Jump" (Zahavy), the Floridi paper, and Williamson's *Widening the Picture* §9.2–9.3. The most substantial one looks like `c84a5338` (538KB) or `50bdd64a` (536KB). Now let me move the Harnad PDF to your Learning folder with proper naming: ### Assistant Done. Now let me also check — since this is related to your generating philosophy project, should it go in the project-specific folder? ### Assistant The generating-philosophy folder has the Zahavy paper and all the others you mentioned. Let me copy the Harnad there too since it's directly relevant to that project: ### Assistant All done. Here's what I found and did: **The session:** Your `/contemplate` conversation about the generating philosophy paper — discussing Zahavy's "LLMs Can't Jump," Floridi, and Williamson §9.2–9.3. Multiple session files from the same conversation (the largest are `c84a5338` and `50bdd64a`, both ~530KB, suggesting a long and rich exchange). **The paper:** Downloaded **Harnad, S. (1990). "The Symbol Grounding Problem." *Physica D*, 42, 335–346** — the paper Zahavy cites when calling LLMs "high-dimensional Chinese Rooms." 15 pages, sourced from the [Oxford CS e-library](http://www.cs.ox.ac.uk/activities/ieg/e-library/sources/harnad90_sgproblem.pdf). It's now in two places: - `Learning/The Symbol Grounding Problem by Harnad 1990.pdf` (main Learning folder) - `Learning/generating-philosophy/The Symbol Grounding Problem by Harnad 1990.pdf` (project folder, alongside the Zahavy, Floridi, and Williamson PDFs) Note: Zahavy actually cites Harnad 1990 for the "Chinese Room" framing, but the paper is really about the *symbol grounding problem* — a distinct (though related) argument. Searle's Chinese Room is from 1980; Harnad's contribution is showing that even if you escape the room (by adding a causal connection to the world), you still face the grounding problem. Worth keeping that distinction in mind for the generating philosophy project. *Il problema del radicamento dei simboli resta irrisolto anche quando i modelli diventano sempre più potenti — la potenza computazionale non sostituisce il contatto col mondo.*