As someone who has had FSD for the past 2 months, I would never trust it with left turn merges into fast lanes, any weird situation (I’ve seen it deal with cons

As someone who has had FSD for the past 2 months, I would never trust it with left turn merges into fast lanes, any weird situation (I’ve seen it deal with construction fine but once an excavator was blocking the road and it kept trying to inch forward in the smallest of gaps even though it was clear that the only solution was to wait till it was off the road) and extreme turns (>150 degrees). My drive to work is 20 miles and a really good stretch of that is just 1 lane where I’m stuck behind a car or the road is empty. In those cases I have no problem letting FSD handle everything, the road is extremely hilly, bendy, lane lines vanish every once in a while and it’s fine so I do find it useful at times, though not worth 200$ a month by a long stretch. I’ll probably cancel the subscription soon. Its problems appear very clearly whenever it needs to accurately predict what other cars on the road will do to make its decision. It is honestly horrible at that and while a few times I tried to not intervene, I realized I was just creating a nuisance and now stopped allowing FSD to deal with those situations. It is also bad if there are wrong lane lines (happens at one final stretch of my daily drive) Finally, as someone working in AI, I’m actually very optimistic about their next approach of just learning an end to end model to drive the car. I think this is the only scaleable way to reach human level driving. Any set of rules based + optimization based planner will have those edge cases that ruin the cars capability, in fact it turns out the long tail is really long. When a human is not drunk and focused he is a remarkably good driver. On the other hand, if you just need to capture the edge cases data and add it to your training set to solve it, you can capture all the possible driving conditions that come up on road. This reminds of how there were hundreds of failure modes and jailbreaks for ChatGPT early on while now there are close to None, presumably because they just added them to the training set. If a good human level self driving car appears, I’m convinced it will be an end to end approach, whether the Tesla team are capable of this, who knows. Also for those who think, OpenAI’s scale based approach only works for language, I have access to ChatGPT Vision and its capabilities astound me. Scale for sure works