Language models trained on human text face an average per-token surprise of ~4.4 bits(measured as bits-per-token from cross-entropy loss across four models). This value,established before any comparison to biological systems, coincides with the serial decodingthroughput basin (~4.16 ± 0.19 bits) independently measured across the ribosome (4.39 bits),phoneme discrimination (4.2 bits), and neural working memory (3.1 bits) (Whitmer, 2026a–e).Shannon (1951) independently estimated English entropy at ~1 bit per character (~5 bits perword)—convergent evidence from 75 years ago. A companion paper showed silicon AI hassub-linear energy scaling (α capacity = 0.937), ruling out a thermodynamic cost basin. Wepropose that AI inherits its throughput from the biological systems that generated its trainingdata: brains constrained to ~4–5 bits per cognitive event produce language calibrated to thatcapacity. Seven-corpus experiments confirm that (1) destroying word order doubles per-tokensurprise from ~4.4 to ~10.8 bits, with syntax contributing ~3.3 bits (paired difference p < 0.01across models); (2) the Zipf distribution is identical in original and shuffled text (α = −0.843, R² =0.992 for both), proving word statistics are insufficient; (3) softmax T = 1.0 produces outputentropy of ~5.4 bits, coinciding with the basin; and (4) exploiting structure costs ~20% moreenergy per token. The throughput basin constrains AI indirectly—through language that evolvedto match biological cognition.
Grant Lavell Whitmer III (Fri,) studied this question.