Nvidia Corp unveiled autonomous driving technology at the Consumer Electronics Show (CES) 2026.
“What they will find is that it’s easy to get to 99% and then super hard to solve the long tail of the distribution,” [lead architect] said in the post.
(99% of autonomous driving miles are incident-free)
Sounds like someone speaking from experience who is hanging out at 99.9 or 99.99 but needs to get to 99.9999.
Nvidia is probably re-thinking everything after this flash of enlightenment.
How does 99.9% correct compare to the human driver?
Perfection isn’t required.
free kink:
https://wapo.st/49q23Gq
intercst
Perfection doesn’t exist except in certain rock n roll songs from the 70s.
For ballpark, we can say 1 incident per million miles (depends on details), so six 9s.
Since the average driver goes 12,000 miles a year, I think that means he could go 11,988 before having an accident.
Every year.
People who drive less could go longer, of course.
I was expecting something different. lol
Me, too. At the moment at least, I’m not interested in “kink” whether free or not.
Pete
Pretty terribly. Humans are really amazing at driving. The average U.S. driver goes 500K-700K miles between accidents. Yes, we get tired and drive drunk and look at our phones - and even with that, accidents are very rare events. A typical U.S. driver will go multiple decades between crashes. So AV cars will have to be very good indeed before they will be safer than a typical American human driver.
Very related:
“Elon’s own lawyers say Tesla shareholders shouldn’t listen to him, calling his statements “mere corporate puffery.” That’s an actual quote.”
Nvidia has resources to investigate all sorts of opportunities for its chips. Self driving vehicles has been on their radar for a few years. They probably have a group working on it. One day it could be a growth opportunity for them.
Not an earth shattering development but ok to update from time to time.
Yeah, but it’s when you start labeling your products with attributes which contribute to dangerous practices that it gets dicey.
“Full Self Driving” is not full, it is not self. It probably will be some day, but you don’t get to call something something it isn’t, otherwise we’d see a raft of “Healthy Cigarettes” from RJR based on the fact that they’re researching it, and maybe someday they’ll have it.
I understand “puffery” but when you make a specific claim you should be able to defend it. At least I’ve learned that the Tesla lawyers like their jobs enough not to object to such obvious legal peril in packaging.
More talk on the long tail of rare driving scenarios that make AI driving a hard problem:
”Roughly 10B miles of training data is [are] (the word miles is plural) needed to achieve safe unsupervised self-driving. Reality has a super long tail of complexity."
Is the long tail now top of mind?
Only now?
So, 10 billion miles of training data are needed to train Tesla’s AI driver, apparently.
How did Waymo achieve unsupervised AI driving while Tesla runs their human taxi business?
Tesla has billions of miles from their retail fleet, Waymo has no such fleet (but they are clever with AI and do lots of real-world-based simulations).
The problem is always the unexpected that no one anticipated. Most recently school bus stopped to load/unloaded w stop sign extended. Earlier shadows on the road. Snow, rain, icy road, road debris, bicycles, pedestrians, etc. Flooded roads.
The multi billion miles finds these situations. Not difficult to deal w once recognized. Bottom line is pull over and ask for instructions when encountering the unknown.
As McNamara pointed out, there are sometimes when you know it’s unknown. But there are other times when you don’t realize it’s unknown at all. Maybe the AI thinks it’s a bird. Or a pennant flag on a car. Or a shadow. Or maybe somebody’s arm waving. Maybe it thinks it knows, but doesn’t.
To be fair to the technology that’s about to rule us all, humans make those kinds of mistakes too, so we shouldn’t expect perfection - but I guarantee the news media is going to make a story of it simply because it’s “new”, which is the foundation of “news”.
Do you mean Rumsfeld?
DB2
Yeah, that guy. Been doing a lot of that lately. Have to be more careful, I guess.
Be sure to have somebody keep track of the old cognitive decline thing…
DB2
Every year in the annual. Giraffe, chair, church, table, shoe.
As long as they keep using the same one I’m golden.
Oh, no. I think I messed up big time. What I remember is camel, rug, temple, lantern, sandal.
Pete
The very useful recognition that there are
things that we know we do NOT know, and that those things are quite different conceptually from
things that we do not even know that we do NOT know.
The statement was first made near where I grew up at North American Aviation in Los Angeles, probably during the design of the F-86 Sabre. I learned it in 3rd grade from the son of one of their engineers.