AI isn't hitting a wall, it's just getting too smart for benchmarks, says Anthropic

As their self-correction and self-reasoning improve, the latest LLMs find new capabilities at a rate that makes it harder to measure everything they can do.