"We already have AGI — Artificial Goodhart Intelligence," on verifiable-domain progress
jessi_cata · x · 2026-10-18
The author offers a satirical take: AI advances rapidly in verifiable domains like Lean proof discovery while having modest real-economy impact because we've already achieved "AGI" — Artificial Goodhart Intelligence, models that excel at optimizing whatever is easily measurable. The joke lands a real point about the gap between benchmark-driven capability gains and real-world value.
More from AGI Musings
- Peter Diamandis: optimism is a conclusion I keep re-deriving from evidence — PeterDiamandis · 2026-10-18
- Thomas Dietterich: proving humans did the real authorship may need signed AI logs — tdietterich · 2026-10-18
- Vercel Ship 2026: agents now trigger 60% of deploys and have crossed human traffic — schwentker · 2026-10-18
- "Cuteness solves AI alignment" — and the brutal rebuttal about pugs — zetalyrae · 2026-10-18
- Dev: after reading AI slop, I cleanse my palate with pre-internet writing — PeterHndrsn · 2026-10-18
- Predicting GPT-7/8 as AGI by 2027-28, but hallucinations still a barrier — koltregaskes · 2026-10-18