Gary Marcus's 2-year-old o3 predictions hold up: gains in verifiable domains, little elsewhere

GaryMarcus · x · 2026-11-22

Gary Marcus revisits his three predictions about o3: people would initially be amazed, find it unreliable once they dive in, and see it work best in closed domains like math but less so in open-ended real-world reasoning. The retweet notes that two years on, models show significant improvement in verifiable domains but little in non-verifiable ones — he was largely right.

Original post →

More from AGI Musings

AGI Musings channel →