Great insights and thought provoking analogy, and fully agree on an advice on keeping humans in the loop and acquiring data. My observations is that while 2D quadrant plot is nice and insightful, it misses many more axes, for one, reward. There is a huge reward for going robust/general in self driving not so much with LLMs, thanks to the non-time-critical/asynchronous nature of decision making in LLMs (no one minds reviewing, say, reversible ticket buying agent decisions 3 times in a day and correcting it, reviewing self-driving irreversable decisions are not so appealing).
Great insights and thought provoking analogy, and fully agree on an advice on keeping humans in the loop and acquiring data. My observations is that while 2D quadrant plot is nice and insightful, it misses many more axes, for one, reward. There is a huge reward for going robust/general in self driving not so much with LLMs, thanks to the non-time-critical/asynchronous nature of decision making in LLMs (no one minds reviewing, say, reversible ticket buying agent decisions 3 times in a day and correcting it, reviewing self-driving irreversable decisions are not so appealing).