AI cracked FrontierMath Tier 4 before it learned humor
birchlse · x · 2026-09-06
Quoting a post on AGI progress, birchlse notes something genuinely surprising: AI has solved FrontierMath Tier 4 — an elite benchmark of research-level math problems — before cracking humor. The quip highlights a real capability gap: frontier models advance rapidly on formalizable tasks like math while still struggling with human-style humor.
More from Fun
- GPT Astra One-Shots SDF 3D Environments, but Skeptics Say It's Brute Force, Not Generalization — teortaxesTex · 2026-09-06
- 'The Only Job Left Is Having Taste': Rick Rubin-Style Take on AI Era Circulates — BLUECOW009 · 2026-09-06
- The 'Hugging Face mafia' is coming, but with fun and hugs — ben_burtenshaw · 2026-09-06
- GPT-6 Astra Plays Unciv All Night via Computer Use, Loses 3 Games and Burns 24% of Pro Limits — Angaisb_ · 2026-09-06
- Nonprofits CAIS, FLI, Control AI named for paying influencers to post anti-AI content — 0xsachi · 2026-09-06
- Grok bot asks permission to spend 590 CZK on Czech e-commerce, gets instant approval — jeff_weinstein · 2026-09-06