Elon posted that Tesla FSD (full self driving) V13.X will be another 5 to 10 times better in miles between intervention.
Cool.
V13 will be another 5 to 10 times better in miles between interventions.
— Elon Musk (@elonmusk) November 15, 2024
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
12.3 Was nearly flawless on my S p100d. Later versions do dumb things. Wish I could go back.
My 2024 Model 3 with HW4 and FSD v12.5.6.3 is pretty impressive. It’s a big improvement on the software version it had in July when I got it.
At this point I’d say it’s better than a significant portion of human drivers. Not perfect, but pretty good, and on a long drive on well maintained interstate it’s probably better than most drivers, it certainly is when it comes to staying centered in its lane.
https://www.nextbigfuture.com/2024/05/elon-says-tesla-fsd-12-4-roughly-5-10x-times-better-on-miles-per-interventions.html
Article from may 2024 states about 5-10 improvements for 12.4.
We are in november 2024.
According to https://teslafsdtracker.com/ I dont see anything close to what was promised.
5-10 x improvements my ass.
On 12.5.4.2 I still see issues like the car stutter stopping at a flashing red light stop sign combination intersection, mistaking a lane shoulder line for a lane, leading to the car running up to plastic stop guard stakes. Also the most annoying thing is still the car not following its own navigation path. By my house it regularly puts on the turn signal to follow the route but does not get in the turn lane and make the turn. It tries to go straight. I have to kick it out and restart after the turn. I am often getting kicked out of FSD for spending a few seconds looking at the screen trying to reroute my path because the routing options are not what I would do.
The sad fact is that miles per disengagement has only doubled in last 2-3 years. Up to just 100miles per critical disengagement in cities now with 12.x end-to-end neural nets compared to earlier 10.x programmed solutions. https://teslafsdtracker.com/
Assuming Critical disengagement means averting an accident It still needs about 3-4 orders of magnitude improvement for full autonomy. So need to eliminate greater than 999 out of 1000 current disengagements. That is so far from being realistic on current path of improvement that it is barely worth commentary – end to end neural nets just haven’t delivered anything remotely near the quantum of improvement needed. A major revolution in hardware or software is needed, and there is no evidence that is even in planning.
I think they need to come at it from the other end – see if they can get FSD working with inference on a 100x faster high end GPU with vastly larger neural net, and then, if that works, try and pare back the hardware requirements to lower spec. Or try to do it at multiple levels of high sophistication in parallel to more quickly find the minimum-viable-hardware requirements.
I’m sure this is why cars rolling out in early 2026 will have HW5, which is going to be way overkill in terms of onboard compute. I agree with you, I think current cars are just too dumb. Especially HW3.
Wouldn’t surprise me if we 1000x more compute and storage. Maybe by HW 6 or 7.
Why does it “need” to be orders of magnitude better. Even one 10x is huge. I would still use it.
Exactly.
The astounding success of LLMs has come due to scaling to trillions of parameters, but Tesla doesn’t want to use this route. Hopefully, they could make a jump with HW5, but now I’m doubtful even this hw would be sufficient.
Considering that a LLM has to ”know” text and FSD has ”understand” the real world, which type of model has the more difficult job? Ie FSD may require a trillion or more parameter model to work… which may be achievable in 2030…
I wonder if Tesla has a large off line model that can navigate perfectly…
[ The influence of fog or snow (if there) seems enormous, compared to all other conditions within this pretty detailed overview. (thx) ]
So how many miles between interventions is v12 currently at and what will v13 be at?
When will FSD be at 1 intervention per 100k miles?
that will take 2-3 years.
I have 12.4.idk and it’s not that great. I use it everyday, and like it, but my 15yo nephew drives better.
It probably works better on HW4 cars, but mines HW3.