LLMs write code without compilers, could they do philosophy without logic? (lywald.github.io)

Large language models routinely produce working code despite never having seen a compiler run or a machine execute instructions: they learn syntax, semantics and debugging patterns as statistical regularities in text. That surprising success exposes a deeper question for AI and ML — how much of human expertise is just stacked, inheritable patterns rather than something grounded in lower-level, operational experience? If models can “skip floors” of the abstraction ladder for programming, could the same happen for domains like philosophy, where most discourse is rhetorical rather than formal? Technically, this hinges on learning from distributional patterns: LLMs infer token co-occurrences that map to valid algorithms, race-condition reasoning, or plausible philosophical argumentation without physiological or executional grounding. The implication is twofold. Practically, it challenges evaluation and safety: text-only competence can look correct but still be brittle under distributional shift or fail when embodiment or causal reasoning matters. Conceptually, it revives the classic mimicry vs. understanding debate — and shows we lack decisive tests to tell simulated mastery from genuine grounding. For researchers, that means developing benchmarks and architectures that probe causal, operational, or embodied grounding (execution checks, interactive environments, interpretability) rather than relying solely on surface-level fluency.