Frontier model overturns expert answers, scores 98% after review
ZeroStateReflex · x · 2026-10-15
An eval anecdote: GPT-6-Astra initially scored in the high 80s with Opus 5.5 close behind, but insisted the expert-supplied answers were wrong. Review confirmed subtle errors in the experts' calculations; after revisions Astra scored 98% (49/50). The meme chart implies AI has passed human expert levels.
More from Fun
- User with 245-day Codex streak panics over 0 tokens today — TheMoonMidas · 2026-10-15
- AI Creators Get Tier-Listed: Users Ask Bots to Rank Interaction Friendliness on X — op7418 · 2026-10-15
- Musk Amplifies Jensen Huang's Story of Cold-Calling TSMC Out of the Yellow Pages — elonmusk · 2026-10-15
- Joma calls AI video editing agent demos 'mostly bullshit', shows his boring real AI workflow — dkundel · 2026-10-15
- 'You're a FDE, Harry': the SaaS meme about AEs dumping implementation-heavy customers on SEs — ivory_tang · 2026-10-15
- McDonald's MCP hackathon spawns 'mcstomach', an AI skill that distills your order history — AlchainHust · 2026-10-15