Incredible human-like behavior from FSD ver 12.3: The naysayers have lost
Tesla’s end-to-end neural net camera-only AI FSD ver 12.3 self-driving system is stunning seasoned FSD beta drivers.
And us AI guy watchers.
Ver 12 has essentially no-hard coding. It just learns from video from GOOD drivers.
The learning is so ‘deep’ that such systems generalize to never-before-seen scenarios. It’s because neural nets, given enough data, build internal models of the underlying ‘causes’ of the data.
That means AI DRIVR – and many others – have treated us to awesomely long intervention-free drives and fascinating human-like scenarios.
FSD beta will now, without specific coding:
- anticipate pedestrian intentions almost supernaturally
- do U-turns
- park kerbside
- park in bays
- pulls up slowly when pedestrians, especially with kids, are walking across an intersection. And then pulls up to the line only after they have passed the bonnet. (MY FAV)
- wait for delivery vans to do quick letter box drop-offs if on single-lane blind corners but over-take, including over double lines, if not on a curve (very human-like)
- not perform right hand turns through red if not allowed by sign
- negotiate hideous unprotected lefts
- quickly cut across 5 lanes to perform a right then left dog-leg
In this photo below, although it’s normally healthily aggressive in the interest of not confusing other drivers, FSD leaves plenty of room to the guy trying to enter from the right.
Just because that’s what good drivers do.
Ver 11 would have blocked the guy.

It’s actually just as us neural net AI engineers expected.
But awesome to see it in real life, at that scale.
The naysayers, sorry to say, are so ignorant, careless or disingenuous, it’s hard to fathom.
Yes regulatory approval will be a very hard ask.
But Tesla is committed.
And up to it.
Progress is now so rapid between ver 12 sub-releases, we can be confident FSD will become an approved product in the next 18–36 months IMO.






