Well this statement is simply not true. Agent systems based on LLMs have made original discoveries on their own, see the work Deep Mind has done on pharmaceutical discovery.
I recall the recent DeepMind material science paper debacle. "Throw everything against the wall and hope something sticks (and that nobody bothers to check the rest)" is not a great strategy.
I also think that Dwarkesh was referring to LLMs specifically. Much of what DeepMind is doing is somewhat different.
