It’s only a very short letter, but it’s good. They argue that AGI claims are based on shifting the definition of AGI to something easier to meet, that advocates rely on hitting artificial (gameable) benchmarks rather than dealing with novelty, and that providing answers to questions doesn’t mean they’re doing cognition as we’d normally understand it, especially as LLMs don’t express doubt (they don’t say this, but I’ll add: this is because they can’t do metacognition: they can’t think about their own thought processes, as we can)