A question I posed to Grok:
“Tesla claims 7x fewer accidents with FSD. Is that correct?”

Grok’s response:
“Partly — the 7x figure is Tesla’s own comparison to a constructed U.S. national average, not an apples-to-apples measure of FSD vs. comparable modern cars.
……
Against the same modern Teslas driven manually, the gap is much smaller: about 2.7x fewer major collisions. That is the comparison safety researchers such as Phil Koopman say is the only reasonably matched one inside Tesla’s own tables.”

2.7x fewer major collisions is still quite significant. And I’d bet that the multiplier is higher for Senior Citizen drivers, e.g., me.

I work from home, as rarely have to go to the main office, 5 hours away.
I was bitching (light heartedly) to the boss about long drives as I get older.
He tells me how he doesn’t like it either, and recently got a Model Y with FSD.
He leaves the DC area weekly for a meeting in Philly.
Door to door and back home, never touches the steering wheel or pedals once.

I did ask about pot hole avoidance , which he said is an ongoing issue, and probably one of the last hurdles to solve.
Here in southern WV, that’s the only thing that gives me pause.

Despite the fact that the last accident I had on public roads was in 1958; now, at 85, I worry about my reaction time!

I drive less than 100 miles per month, never on I-4, and still worry! I always drive accompanied by a caregiver and ask them to evaluate my driving… I believe them when they say I am a safe driver. I also avoid driving at night - limited visual field and delayed reaction time.

In my younger (under 60, before my wife’s stroke) days I loved riding motorcycles and took the Motorcycle Safety Foundation rider course every two or three years. That taught me paranoia - that every vehicle on the road would try to kill me! That habit has carried over to my driving today.

All that said, I look forward to not having to drive any more, even loving driving both motorcycles and race cars in my youth.

My mom is 76 and hates driving. This year we got her a new model Y and she loves it. Now she is a woman about town with the car doing 90% of the driving. I have tested it and it’s very impressive. If Tesla had more compelling family cars I would get one as my daily. I’m 47 and have a 1 and 3 year old. I often get like 3 hours of sleep a night. That lack of sleep doesn’t preclude me from having tons of drive places. Supervised seems to be the answer for us tired drivers. I have tried the Rivian R2 with Autonomy + as well as Ford Blue Cruise and GM Super Cruise and if Tesla FSD is an IPhone 18 Pro those systems in comparison are an IPhone 3G.

quote:

Originally posted by ElKabong:
I did ask about pot hole avoidance , which he said is an ongoing issue, and probably one of the last hurdles to solve.
Here in southern WV, that’s the only thing that gives me pause.

And there you go...Confirmation that I am a better driver than Tesla FSD!

quote:

Originally posted by Dusty78:
<snip>
If Tesla had more compelling family cars I would get one as my daily.
<snip>

Member sjtill has ordered a Model Y L, I believe it’s called. Three rows of seats.

https://www.tesla.com/event/modelyl-cfg

quote:

Originally posted by Pipe Smoker:

quote:

Originally posted by Dusty78:
<snip>
If Tesla had more compelling family cars I would get one as my daily.
<snip>

Member sjtill has ordered a Model Y L, I believe it’s called. Three rows of seats.

https://www.tesla.com/event/modelyl-cfg

Yes but the 2nd row seems to lose some leg room and the second row seats have a one piece design making installation of kids car seats tricky (since there is no pass through for the bend the seat latch point.) I would want like a Rivian R1S/R1T size or bigger with Tesla FSD. I would even go with something like the Chevy Silverado EV if it had FSD. The Cyber Truck is just too out there for me. Unfortunately all of these other companies are decades behind Tesla in regards to self driving.

The Model Y L, with the 2nd row being captain chairs?

quote:

Originally posted by Expert308:
I spent the last 35 years of my career, until I retired 3 years ago, programming computers. No way I'd ever trust one to control my car. No way. There are just too many things that even the best programmers can't foresee happening.

quote:

Originally posted by crue-dell:
Tesla FSD is real world ai, not traditional software coding. Its learned/trained how to drive by watching actual driving scenarios by the best drivers. Its making decisions based on a 360 degree view. There is no other assisted driving system that is comparable. Tesla has spent over $10 billion on training compute. We are in the bottom of the 1st inning.

^^^ Correct, crue-dell.

Grok query:
‘How does Tesla FSD know how to drive?’

Grok response:
‘Tesla FSD does not “know how to drive” the way a human does, and it does not follow a giant book of if-then traffic rules. It is a large neural network that has learned a mapping from what the car sees to what the car should do, by watching an enormous amount of real human driving.

Photons in, controls out
Modern FSD (from v12 onward, refined through later versions) is end-to-end. The network takes:
• raw video from the car’s cameras (typically eight cameras covering 360°)
• vehicle motion data (speed, IMU, etc.)
• audio
• maps and navigation
and directly outputs the next steering and acceleration commands. Tesla describes this as compressing on the order of two billion input tokens per second down to two control outputs.

There is no separate “detect cars, then predict paths, then plan a trajectory” pipeline that a human engineer wired together. Intermediate ideas like “that’s a pedestrian” or “that’s a lane line” can still exist as auxiliary heads the network is trained to predict (the old HydraNet idea, now often described as a video foundation model). Those heads help the network learn a useful internal world model. The thing that actually drives the car is the end-to-end path from pixels to controls.

Why Tesla chose this instead of a modular stack
Most of the industry still splits the problem into perception, prediction, and planning. Tesla’s argument is that those module boundaries are poorly defined, errors compound, and many human values (courtesy, intent, “that flock of geese is not going to move”) are almost impossible to write as rules. In an end-to-end net, gradients flow from the steering wheel all the way back to the cameras, so the whole system can be optimized for the actual job: driving well.

A favorite Tesla example: the net can treat a chicken crossing the road differently from geese standing still, without anyone writing a chicken-intent module.

Where the “knowledge” comes from
The training data is Tesla’s fleet.
Millions of cars record video while people drive. The system also runs in shadow mode: the network silently predicts what it would have done and compares that to what the human actually did. Interventions, near-misses, rare events, and “interesting” clips get uploaded. Tesla has said the fleet produces on the order of hundreds of years of driving experience per day. That data is filtered, labeled (often auto-labeled), and used to retrain the network on large clusters (Cortex GPUs; Dojo has been part of the story as well).

Training is a mix of:
• imitation — copy good human driving at scale
• reinforcement learning — reward safer, smoother, more human-like behavior
• co-training on extra tasks (occupancy, objects, signs, etc.) so the shared backbone learns richer features
New models are evaluated in simulation and on recorded fleet clips, then rolled out over the air.

What runs in the car
Inference happens on Tesla’s custom Autopilot / AI computer in the vehicle, at high frame rate and low photon-to-control latency. The car is not phoning home for each steering decision.

In short: FSD knows how to drive the way a very large video model knows how to continue a sentence. It has seen enough examples of “what a competent human did next in this visual situation” that it can produce the next action itself.’