Notes on the Symbol Grounding Problem (dstrohmaier.com)

A new set of notes reframes the classic Symbol Grounding Problem for large language models, distinguishing an "easy" and "hard" grounding problem. The easy problem is engineering: getting systems to integrate multimodal, sensorimotor data so their internal representations correlate with the world. The hard problem splits into two parts — (a) building systems whose representations and outputs carry intrinsic meaning, and (b) giving an epistemic explanation for why any computational representation could have intrinsic meaning. The author argues these are entangled: you can’t validate the engineering success without answering the epistemic question, and whether the hard problem is really “hard” depends on whether its explanation can be given in purely computational-functional (Turing-equivalent) terms. Technically, the notes tie this debate to meta-semantics, semantic externalism, and the hard problem of consciousness: one route to hardness is showing intrinsic meaning requires phenomenal consciousness; another is showing externalist conditions that functional descriptions don’t guarantee. For AI/ML this matters for evaluation and aims — solving multimodal integration may not suffice to claim models “understand.” If we want systems to reproduce human-like judgments about meaning (an epistemic ability), that becomes a target for AI research. The essay therefore reframes grounding as both a technical and philosophical challenge with direct implications for model design, benchmarks, and what counts as genuine semantic competence.