The article was talking about Level 5 as a long way off but it also spent time talking about a lack of consistency in how these levels are applied [1]. Some sys

The article was talking about Level 5 as a long way off but it also spent time talking about a lack of consistency in how these levels are applied [1]. Some systems might state they are Level 4 but that's only true for well mapped good weather locations so one might legitimately wonder if such a system is truly Level 4. The article is mainly about: Level 5 as difficult, inconsistent application of Levels, the fact that rare difficult situations carry most of the valuable information for a learning system, preventing mismatched expectations, untenability of systems which require human attention and vigilance. The final point made was, instead of focusing on self-driving cars, focus on cars which complement human ability. For example, humans are good at theory of mind and inference in rich world models while machines are better at total situational awareness. I'll add another. Self-Driving cars need models which at the very least, are able to capture uncertainty over predictions. This makes them more robust to glitching catastrophically in parts of the state space that are novel to them. [1] > Yes, so NHTSA had 1 to 4, and SAE was very smart, they said we’ll take Level 4 and we’ll split it into two things, 4 and 5, and the difference is that Level 5 is everywhere at any time, and 4 is only some places at some times. But otherwise they’re basically the same. The SAE levels are fine, but people keep making mistakes. And so Level 2 systems are often called Level 3, which is wrong. A key thing of Level 3 is that the person does not need to supervise the autonomy. So no guarding of the machine by the human being.