I believe that. I'm no expert. My undergrad engineering degree did focus on robotics, but it had nothing to do with AI. Plain ol programming & control systems.
I LOV when people just throw around L3, L4, L5, like it is just a series of baby steps.
So we just took a trip up to BB using FSD 14.1.3 on the way up and installed FSD 14.1.4 while at BB and then all the way home. Well over 1000 miles of driving. I literally hardly ever had to intervene the entirety of the time we were up in BB - including multiple day trips to Fort Ticonderoga, Burlington VT, Woodstock VT, etc. Keep in mind this is mostly rural mountainous driving with windy back roads with fading paint (in some cases no paint at all), bad lighting, fall foliage season (leaves all over the place), etc. Literally not a single critical safety intervention the entirety of over 1000 miles - and in most cases - the vehicle not only drove itself everywhere - but parked at the conclusion of each drive in a parking space with, again, zero intervention. I had a few what are termed "encouragements" where I hit the go pedal a few times because the system was being overly cautious - it's still an early release version of v14.1.x - so this is to be expected - and I came across a couple of edge cases where I had to intervene - like the dirt parking lot at Fort Ticonderoga - where FSD wasn't sure what to do. The edge cases are almost myriad in number and difficult to train on. I'd say we're solidly at L3 now with FSD v14. This is all because Tesla has six billion miles of real-world data to train the neural networks, and synthetic data to train on RL for edge cases. Show me another system in place, right now, that can come anywhere close to this and do so without geofencing and ultra-high-definition mapping (which is what every other system is using to get around gaps in their real-world datasets today - and why systems like Waymo can only function within certain geofenced areas). My argument is based on real world experience, using a real-world system that we literally use in our vehicle every day. My wife literally uses FSD 14.1.4 every day to drive back and forth to work. Literally door to door now - parking and everything. The only thing FSD won't do for us - is pull into our garage at home - it parks in our driveway at present - and tries to back into the garage space - but we have two individual garage doors - and it's just not quite there yet when attempting to pull into a two car garage with older smaller garage doors (8x7 not 9x8 like most newer homes have now) or back into a single garage space - it's trying every day though. My buddy in comparison has a single car garage - and it pulls right into it every time - zero issues - and backs out of his garage every time - zero issues. FSD pulls right out of our garage bays though - zero issues. So we're halfway there already. That's pretty impressive really. I suspect by 14.3 it'll have it figured out when "reasoning" is added into the FSD NN stack with respect to my garage configuration. FSD v13.2.9 was lacking what I'd term "last mile" type edge case management. FSD v14.x largely resolves these last mile items. IMHO we are solidly in L3 territory right now with FSD v14.x - whether Tesla pursues this officially or not remains to be seen. I'd personally rather they certify FSD v14 as L3 so I can stop having to pay attention when driving - and only get my attention when really needed.
