Much like ChatGPT, the capabilities of FSD-S are incredible. But also like ChatGPT, it gets stuff wrong.
The "supervised" part of its name is not just a suggestion, you absolutely need to supervise it like you'd supervise a teenager driving your car. The problem, really, is that FSD-S v12 drives very well 99% of the time. I can see that lulling people into a false sense of security. 99% isn't good enough for letting your attention wander; you need to assume it's going to try to kill somebody the other 1% of the time. I've only had v13 for a few days but it's dramatically better. Maybe now it's 99.9% awesome and 0.1% murder machine.
There's an open question as to whether FSD - Unsupervised is even possible or not. I'm not sure it is, especially using cameras only. However for most driving tasks it does a stunningly good job, and it has the potential to be better than a human driver because it's using a 360° view of its environment. It's possible that it could significantly decrease crash rates for daily driving. But when things get weird—construction zones with complicated routing, unusual obstructions in the road, traffic doing strange things—I don't think it'll figure things out as well as an attentive human driver.
One other thing: you can use FSD-S like a copilot on long drives. Drive along yourself until your stamina starts fading, turn on FSD-S for a while, then take over again when you want to. You still have to pay attention the whole time, but it lets you recover a bit.
Anyway, I pay $99/mo for FSD-S. At the moment I consider it worth the money even though I'm hand-driving most of the time. (Teslas are a ton of fun to drive.) With the version 13 software I may start letting it drive more frequently. At some point I might pay the $8000 flat fee instead of the monthly subscription, but I haven't decided yet.
