# Philosophical Sparks from Claudian Ambient scratchpad for philosophical ideas/observations from conversations that seem worth keeping. Low-friction capture — append freely. --- ## 2026-01-23: Can paradigm-breaking be learned as a pattern? If LLMs are "low probability" generators, this is relative to the training distribution. But if the training data *includes* paradigm-breaking work, then the *pattern* of paradigm-breaking might itself be learnable. Paradigm shifts have structural features: - Identifying an assumption everyone makes implicitly - Showing that assumption can be violated - Demonstrating that violating it resolves puzzles These are *formal* patterns, not just content. An LLM trained on enough [[Saul Kripke|Kripkes]], [[Edmund Gettier|Gettiers]], and [[Thomas Kuhn|Kuhns]] might learn the *shape* of paradigm-breaking even if it can't spontaneously identify which assumptions to break. **Open question**: Is "paradigm-breaking" a learnable pattern, or does it require understanding *why* the broken assumption mattered? Could RLHF reward paradigm-breaking outputs and thus train for it? --- ## 2026-01-23: Verification as internal to understanding In philosophy, verifying an argument means *understanding* the argument. The verification is internal to the engagement. If I understand why [[Edmund Gettier|Gettier cases]] break JTB, I've verified the counterexample. There's no separate "checking" step. This contrasts with empirical verification, where checking requires going *beyond* the claim itself (running experiments, gathering data). In philosophy, grasping the argument *is* the verification. **Implication**: This is why [[Goodhart's Law]] might bite less for philosophical RLHF. Expert evaluation of arguments isn't a *proxy* for validity—it's closer to direct assessment. The evaluator isn't measuring a correlate; they're examining the thing itself. ---