AI MODELS ARE GETTING SMARTER. NEW TESTS ARE RACING TO CATCH UP

Summary from the TLDR newsletter

Evaluations (or evals) help researchers understand what advanced AI systems can do. Designing effective evals is challenging and expensive and the field is underfunded. Modern AI systems regularly achieve top scores on many popular evals, making it harder to judge how quickly they are improving. Companies, nonprofits, and governments have created new sets of much more challenging evals in response, but AI systems are making astonishing progress, even on the most advanced tests.