Why Should Intelligence Be Related To Neuron Count?

A speculative musing on why intelligence tracks cortical neuron count across animals, humans (bigger brain -> higher IQ), and AIs (more parameters -> better benchmarks), and specifically why extra neurons help on a single fast IQ-test item rather than merely storing more facts. Scott rejects the pure-storage and 'absorb Paradise Lost' accounts and lands on polysemanticity / superposition / overcompleteness: too few neurons forces massive polysemanticity so the right 'guess' cannot be isolated from overlapping wrong ones, while more units make a richer representational 'language' and an overcomplete basis that finds good-enough solutions faster. A long quoted friend supplies the mech-interp core (function approximation, the Neel Nanda grokking paper). Honest, exploratory, half-outsourced.

Quality 62 · Strong. Strong/Solid edge. A genuinely stimulating question and a plausible, well-hedged answer connecting brains to mech-interp -- but short, loose, and substantially carried by a quoted correspondent rather than Scott's own worked argument; exploratory rather than authoritative. Sits at the low-Strong line (cf. ACX-018 62).

Claude’s paradigm shift 40 · Moderate. Moderate floor. The superposition / polysemanticity / overcompleteness framing is imported wholesale from interpretability research and applied speculatively to biological intelligence; little new in its moment beyond the cross-domain juxtaposition.

Real-world impact 1 · Negligible. Negligible. Niche AI/neuroscience speculation with no adoption or material consequence. 0-1 band.