The G in AGI means (to me) the human-equivalent competence at thinking through a novel situation for which there is no pre-trained understanding. That is where first principles thinking must be applied to new perceptions of reality, which then must be abstracted inductively into concepts (some new).
Benchmarks that measure whether a LLM can respond as well as a human could for solving some specialized problem or regurgitating specializes knowledge when prompted only exercise rote learning, which is what pretraining enables. Human-equivalent rote learning is unsurprisingly suited to machine automation. Human-equivalent epistemology in the face of novelty is something entirely different.
Applying rote learning to do as well as a trained human for some well-understood problem domain like law, software development, or medicine is so "2025", as far as I am concerned. It is impressive that a machine can do as well as a book-smart human. Book-smart humans are different animals from explorers, inventors, and entrepreneurs, who can see the future and manifest the never-before-imagined into reality.
