Right.

We can certainly get much more utility out of current architectures with better engineering, as "agents" have shown, but to claim that AGI is possible with engineering alone is wishful thinking. The hard part is building systems that showcase actual intelligence and reasoning, that are able to learn and discover on their own instead of requiring exorbitantly expensive training, that don't hallucinate, and so on. We still haven't cracked that nut, and it's becoming increasingly evident that the current approaches won't get us there. That will require groundbreaking compsci work, if it's possible at all.