I Won My Three Year AI Progress Bet In Three Months

Scott recounts winning a three-year bet in three months. In mid-2022 he'd bet commenter Vitor (who, citing Gary Marcus, argued compositionality -- combining elements accurately, like 'a woman in a library with a raven on her shoulder with a key in its mouth' -- was AI-complete and wouldn't be solved by 2025) that image models could nail 3/5 such hard prompts by 2025. DALL-E2 got 0/5, but Google's Imagen (announced May 2022) got 3/5 when he finally got access three months after the bet -- meaning the bet was probably already won before it was made (they had to swap humans for robots due to Imagen's trust-and-safety no-human-form rule). The point: AI progress is faster than expected, and plain scaling + normal progress can sometimes crack problems that look fundamentally hard, increasing his confidence that scale suffices for difficult capabilities.

Quality 67 · Strong. Strong/Solid edge. A clean, punchy data point about the surprising speed of AI progress, with concrete compositionality examples and the nice 'won the 3-year bet in 3 months (maybe before it started)' hook. Low-Strong: short and narrow, a single anecdote rather than a developed argument.

Claude’s paradigm shift 44 · Moderate. Moderate, low. The empirical 'compositionality fell to scaling faster than skeptics predicted' point is a fresh data point, lightly developed.

Real-world impact 2 · Minor. Minor. Within the AI-progress discourse; conceptual. 2.