CurtisVL
Member
By your logic the vNew FSD would lead to the same behaviors as the vOld FSD. An existing, tested model executes an existing, tested action. If either of those statements is false, either the model or the action being untested, then by default the chain is untested.
You're absolutely right that the model will benefit in object recognition by all the billions of miles of real world data. That model will be further enhanced by the ability to recognize and track objects across multiple camera views, as Elon's said publicly. That improved model has not been tested by billions of real world miles.
Similarly, you have logic executed based on those models. Today, you will see phantom breaks due to any number of interpreted signals from that model. Elon has suggested that no longer happens. That new, changed behavior has not been tested by billions of real world miles.
This is in contrast to today, where adding one additional piece of functionality doesn't invalidate the entirety of AP, because the change is supplemental and iterative, vs. a fundamental replacement.
If you work in software, which it sounds like you do, you know that the con of a rewrite is that it has the potential to introduce far more bugs than a single, incremental update to an existing codebase.
Though, having said that, vOld FSD was at a point where training wasn't making massive changes. It's very likely the re-write will release in a funky state and in the coming weeks/months it'll rapidly improve, especially with the additional data it'll be working with due to using all cameras instead of just 5, and with the '4th dimension' (time) for a bit of context to what its seeing.
Some issues will be fixed for sure, for example, long vehicles such as trucks (semi's) can cause issues for the car as they switch from one camera to another, the car confuses the depth so the position flickers around a lot.
Regardless, I'm excited to see what's changed! I do expect the first times we're shown it that it'll be pretty buggy, but it should improve pretty fast once it starts getting worldwide data rather than just parts of California data mainly.