Tom Dietterich: AI safety hinges on novel, ill-defined tasks AI can't optimize

tdietterich · x · 2026-09-09

In a debate sparked by Melanie Mitchell's "Deep Blue–Kasparov moment" remark, Tom Dietterich agrees AI will beat humans on well-defined tasks within a decade, but argues novel and ill-defined situations lack benchmarks and training data, making them hard to optimize for — and safe AI requires solving them too. The counterparty cites METR's time horizon data to argue the crossover will come faster.

Original post →

More from AGI Musings

AGI Musings channel →