We didn't move the AGI goalposts: the real missing piece is meta-learning
laoma1255 · reddit · 2026-09-08
A long Reddit essay rebuts the "moved goalposts" complaint about AGI: the issue is conflating accumulated skill with actual intelligence, echoing Chollet's ARC-AGI premise that intelligence is the sample efficiency of acquiring new skills on unfamiliar problems.
Key argument: human engineers are the model's "external meta-learner" — when a model fails a novel benchmark, humans analyze failures, build simulators, design curricula and generate targeted synthetic data, while the model just runs SGD on human-built scaffolding. This repeats the AlphaGo fallacy (trained on millions of games yet couldn't play on a 20x20 board without retraining). Brute-force curriculum design yields over-engineered specialists, not general learners. True AGI must close the loop itself: autonomous diagnosis of blind spots, autonomous curriculum/data generation, and self-directed skill acquisition — of which today's RSI efforts in coding/math loops are only primitive versions. Demanding this isn't moving goalposts; it was the definition of general intelligence from day one.
More from AGI Musings
- AI accusations shake literary prizes from Caine to Goncourt: who decides what's human? — TuhinChakr · 2026-10-26
- Tech discovers wargaming: applying to be Anthropic's tabletop exercise lead — round · 2026-10-26
- RSI is the culmination of the industrial revolution: a factory building a faster factory — tobias_rees · 2026-10-26
- NYT roundtable asks five experts what the rules for AI companies should actually be — ShakeelHashim · 2026-10-26
- Anti-AI content is algorithm gold and very profitable, quips commentator — annetgriffin · 2026-10-26
- Silicon Valley breakdown: five Stanford-area giants worth ~$18T — FinanceYF5 · 2026-10-26