Reaction
WeirdML v3 introduced as an 11-task, fully agentic benchmark
In a September post, htihle described tasks requiring models to work with unfamiliar data and limited feedback.
TLDR
On Sept. 18, 2026, htihle introduced WeirdML v3 as a fully agentic benchmark with 11 hand-made tasks. The post describes models exploring unfamiliar data, building machine-learning and data-analysis pipelines, and producing results despite limited data, unspecified goals or little feedback. teortaxesTex later offered a model-performance comparison, but that is the commenter's assessment, not a result given in the introduction.
Combined views
1.9K
1 Source, first seen ago
