I don't think there is any "humans are metaphysically superior to LLMs" subtext to this talk, it's just a technical/educational observation.

Access to some forms of evaluation and selective retention is inherent to humans and it's not inherent to LLMS. But it can be somehow bolted on and that's when they work best. It makes sense that more focus on those principles can yield better AI. I think the retention part is the real limitation of LLMs, because it's limited to stuffing things in context window.