AI tools score high against current benchmarks but often do not work the way we want them to. Do we need new ways to evaluate them?

→ View original post on X — @stanfordhai