Dwarkesh Patel Interviews Dario Amodei: OSWorld Scores Hit 65-70% But Can't Automate White-Collar Work Yet
Zvi Mowshowitz / Substack·medium signal
In a 2026 Dwarkesh Podcast interview, Dario Amodei revisited his 2023 prediction that AI would be indistinguishable from well-educated humans in conversation. Dwarkesh pushed back: while OSWorld benchmarks jumped to 65-70% from 15% a year ago, the systems still can't automate large parts of white-collar work. Amodei acknowledged the gap between benchmark scores and real-world autonomy.