My Bet: AI Size Solves Flubs

Scott bets Vitor that DALL-E/GPT flaws will be solved by scale, via the recurring AI-hype-cycle pattern: a new model is released; a skeptic (usually Gary Marcus) shows it fails trivial tasks supposedly requiring 'true intelligence'; a year or two later a bigger model solves them; the skeptic finds new trivial failures; repeat. Demonstrates by re-running Marcus's GPT-2 failure prompts (numerical reasoning, obscure-location languages, practical reasoning) on a newer GPT-3, which gets most right -- and where it 'fails,' Scott often sides with GPT-3 (the Trenton/Spanish completion reflects real pragmatics, not a logic error). Careful caveat: this doesn't prove Marcus wrong about some deeper missing intelligence, only that his specific examples keep getting fixed. Ages very well from a 2026 vantage.

Quality 73 · Strong. Strong: a clear, well-evidenced, memorable thesis (scale keeps dissolving apparent-reasoning failures) with honest epistemic hedging, and it aged extremely well; strong but a focused argument, not field-defining.

Claude’s paradigm shift 52 · Moderate. Notable: the 'AI hype cycle / each debunk gets fixed by the next model' articulation is a fresh, portable frame that captured a real pattern early.

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