> And some work with remote recovery interfaces (something that Tesla doesn't seem to be focusing on) would get it to L4 in the kind of constrained environments

> And some work with remote recovery interfaces (something that Tesla doesn't seem to be focusing on) would get it to L4 in the kind of constrained environments Waymo and Cruise are operating in reasonably quickly, IMHO. It takes more than that. Just think, Uber's driver had lidar, safety driver, and still killed a woman. There's _a lot_ going on under the hood that separates the L4 from L2/L3 drivers. One thing I think Tesla's missing is a repeatable evaluation framework for how their software performs (a.k.a what you need to convince yourself of the safety). As I understand it, they mostly test autonomous software by deploying it in ghost mode (current/future miles) instead of simulating on past miles. A lot of long tail problems (which is what's hard about autonomous driving) are only visible on the order of millions of miles. Some cases you may never encounter at a high enough rate, and you may need to build synthetic data or use other tricks to teach your driver how to act. Consider construction zones: where are the lanes formed by the cones? Where should I enter, and where should cars come shooting out? Every construction zone is quite nuanced and contextual -- can you prove the autonomous driver can handle situations reasonably? I think Tesla's software isn't close to being able to crack those driving situations. I mean they make basic mistakes with traffic lights still (which are very challenging in and of themselves; how about those lights where there's an arrow? light that is out?) If you lack infra to save difficult cases and no ability to generate synthetic situations and resimulate, it gets real dicey when you try to draw conclusions about the real-world performance. Their simulation capabilities are also limited by the fact that they only use cameras; depth estimation / object contours are not accurate enough to do the kinds of sims the Cruise and Waymo do. Without good ground truth it's really hard to do proper safety evaluation. I suspect Tesla would either need a huge breakthrough in vision technology (maybe they can use something like this https://waymo.com/research/block-nerf https://waymo.com/research/block-nerf) or at least map places their cars are driving so they can properly assess rates to give themselves, regulators, and customers confidence that they are safe for L4. Otherwise, their "good performance" will just be a product of survivorship bias, and bad performance will really hurt people (this is kind of happening now).