Discussion about this post

User's avatar
Foundation's Edge's avatar

Your 65%-of-rubric finding has a starker counterpart in paid work: on the Remote Labor Index, 240 real freelance projects across 23 domains, the best-performing agent achieved 2.5% automation and GPT-5 managed 1.7%, while the same model scores 57-58% on a ten-domain cognitive battery. The failure distribution rhymes with yours, since 45.6% of freelance failures were work below professional standard and another 35.7% incomplete or malformed deliverables, the same open-ended-judgment gap that had GPT-6 Astra reading its own token-budget flaw as a finding about the models. Both point the same direction: readiness tracks the floor, and coherence-weighted scoring that penalizes jagged profiles drops GPT-5 from 57% to roughly 24%. https://seldondance.substack.com/p/agi-heading-into-2026 #AIEvaluation #AGI

Hazem🔸's avatar

> he forecast that wouldn’t be the case by the end of the decade.

Is that from extrapolating (Epoch tasks success %, release date) data points, or from external info/data?

7 more comments...

Ready for more?