Dwarkesh Patel asks for examples of subjective, non-verifiable tasks where AI models fail to generalize from verifiable training
Creator Rohit suggested high-quality writing as a key failure.
——0——
@dwarkesh_sp Writing well.
Dwarkesh Patel@dwarkesh_spWhat is the most compelling example of a task in a non-verifiable domains where models really struggle? That might hint at lack of generalization from verifiable to non-verifiable domains.
7:04 PM · May 29, 2026 · 38.3K Views
8:05 PM · May 29, 2026 · 840 Views
@dwarkesh_sp Persuasion
At best we have proxies for persuading someone to feel or think something, via buying stuff or engagement metrics
But actually incepting an emotion or thought and checking if it worked is hard if not impossible
Dwarkesh Patel@dwarkesh_spWhat is the most compelling example of a task in a non-verifiable domains where models really struggle? That might hint at lack of generalization from verifiable to non-verifiable domains.
7:04 PM · May 29, 2026 · 38.3K Views
8:23 PM · May 29, 2026 · 350 Views