Somewhat Contra Marcus On AI Scaling

A forceful rebuttal to Gary Marcus on whether scaling gets LLMs to human reasoning. (1) Marcus provides no evidence: GPT-2 (1B params) sort-of-reasons, GPT-3 (100B) reasons better, the brain is ~100T -- a 100,000x-smaller thing reasoning-a-bit doesn't show scaling stops before human level, and Marcus can't name a reasoning type GPTs will NEVER do. (2) The core thesis: humans have world-models the way we have utility functions -- messy processes that approximate the formalism -- evidenced by the Luria Uzbek-peasant transcripts, sub-90-IQ failures on conditional hypotheticals/recursion, dreams, and 'humans who aren't concentrating aren't general intelligences'; logic is knit out of predictive pattern-matching, not a crystalline module. (3) Brain plasticity supports scaling growing new capabilities (GPT-3's emergent abilities; the chimp->human eyeblink). (4) Even if LLMs aren't human-like, narrow AI keeps winning unimaginable victories, so 'I will basically believe anything.' Ends with calibrated 2030 predictions. Aged very well.

Quality 78 · Excellent. Excellent (low)/upper-Strong: a lucid, well-evidenced, forceful argument with a genuinely portable thesis ('humans emulate logic out of pattern-matching') made vivid by the Luria/low-IQ evidence, and it aged extremely well; among the stronger AI posts.

Claude’s paradigm shift 54 · Moderate. Notable (upper): 'world-models like utility functions / logic knit out of predictive pattern-matching' is a fresh, portable reframe of the scaling debate.

Real-world impact 2 · Minor. Minor/within-discourse: part of the AI-capabilities debate; a commentary post with no direct material footprint.