I'd argue that it's because intelligence has been treated as a ML/NN engineering problem that we've had the hyper focus on improving LLMs rather than the approach you've written about.

Intelligence must be built from a first principles theory of what intelligence actually is.

The missing science to engineer intelligence is composable program synthesis. Aloe (https://aloe.inc) recently released a GAIA score demonstrating how CPS dramatically outperforms other generalist agents (OpenAI's deep research, Manus, and Genspark) on tasks similar to those a knowledge worker would perform.