Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

FSD tweets

This site may earn commission on affiliate links.
FSD 12.4 first impressions: feels more capable, but needs a lot more “polish”.
  1. no wheel nag is incredible, really makes you feel like you’re being driven around and truly transforms the experience. I feel like I’m doing something naughty by not touching the wheel but it’s so nice. nags are still present but looking forward at the road satisfies it. never touched the steering wheel once except to park. will do more testing with sunglasses in day time
  2. feels much more assertive / human like while turning in general, especially U-turns, stop signs, and parking lots. not any hesitancy or odd creeping in any of the intersections I’ve gone through so far, it just goes for it
  3. slowed down when a police officer turned on his lights beside my car, I know for a fact 12.3 didn’t react to emergency lights at all
  4. “auto” speed has regressed a bit, at least at night time. it was going ~5mph under the speed limit most of the time and never went 2mph over where 12.3 was always 5-6mph over on the same roads
  5. huge regression in random lane changes, it feels very indecisive about which lane it wants to be in it was going back and fourth between two lanes in short periods of time on multiple roads around where I live
  6. “lane dancing” still a thing in a couple of intersections, but seems to be more decisive. 12.3 was definitely worse, but it’s not completely fixed yet
  7. it spotted some pedestrians crossing the street I never saw and I doubt 12.3 would have waited for
  8. doesn’t park itself when it reaches the destination, will literally just circle the parking lot over and over instead of just stopping at the navigation pin. still need to disengage to auto park
  9. vision auto park feels much faster. shifts are super quick and it feels like it knows what it’s doing. I might actually use it a bit now
I will try to upload some clips later, but the spring update messed with the visualizations so I need to redo my overlay which takes time. I’ll stay up late though

AI DRIVR uploaded this extreme example of the lane indecisiveness:

 

1717887627161.png
 
  • Informative
Reactions: JB47394
HW3 = x times better than HW2
HW4 = 5 times better than HW3
HW5 = 10 times better than HW4
HW6 = 20 times better than HW5 :p

Performant != reliable. 20x more performant doesn’t necessarily mean 20x fewer mistakes, especially if the reliability becomes perception-limited. I’ll be really interested to see the enhancements to the sensor suite that go along with HW5. (Er, “AI5!”)

IMG_1232.gif
 
  • Like
Reactions: DanCar
Performant != reliable. 20x more performant doesn’t necessarily mean 20x fewer mistakes, especially if the reliability becomes perception-limited. I’ll be really interested to see the enhancements to the sensor suite that go along with HW5. (Er, “AI5!”)

View attachment 1056570
Tesla has stated in the past they have bigger models that perform better. There is the commonly known scaling law. Will be interesting if Tesla switches to a bunch of different models that can complete and compete against each other. An interesting example is if they can incorporate this 1.3B parameter depth estimation model: Depth Anything V2

There was some chatter last year about Tesla wanting to incorporate LLM into the car. Would improve voice input tremendously.
 
  • Like
Reactions: Ben W
Performant != reliable. 20x more performant doesn’t necessarily mean 20x fewer mistakes, especially if the reliability becomes perception-limited.
unless there's some sort of emergent property breakthrough (and that would be on training data size first) I conjecture first order scaling is going to be logarithmic in compute performance:

like log(mistake_rate) = A - B*log log flops


And you need the clean dataset size and model memory size to scale up with flops. You might have to take 2-3 days to download your model upgrade on a slowish wifi, as many are being in the garage far away from the router/wifi modem.
 
  • Like
Reactions: DanCar and Ben W
Might be a repost. Elon at Shareholder conference talks for 10 minutes about ai5 and FSD. 1 hour 17 minutes and ~25 seconds into the video:
Elon says that in 18 months Optimus, the Tesla Bot, will switch to ai5. ai5 can draw up to 700 to 800 watts of power, but is usually in more power efficient modes.
Around end of year HW4 cars will get their own optimized version of FSD. Elon says HW4 is about 5 times better than HW3. (smirk)