I agree that v9 seems like a great improvement, but this... isn't how machine learning works. Sure, you're feeding data to Tesla constantly when EAP is in use. But expecting "exponential" improvements in autonomy is far fetched. Neural network machine learning is just fancy pattern matching. Yes the more patterns you have to train a specific model more granularly the more often you'll get high confidence results FOR THAT MODEL, but it's not like learning in a human sense where you master one thing and then build experientially off of that previous knowledge to learn a whole new thing.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
I agree that v9 seems like a great improvement, but this... isn't how machine learning works. Sure, you're feeding data to Tesla constantly when EAP is in use. But expecting "exponential" improvements in autonomy is far fetched. Neural network machine learning is just fancy pattern matching. Yes the more patterns you have to train a specific model more granularly the more often you'll get high confidence results FOR THAT MODEL, but it's not like learning in a human sense where you master one thing and then build experientially off of that previous knowledge to learn a whole new thing.
That's been there for the last couple of V8 releases.
Cirrus MS100D
Member
Yes
Except that there are recent examples where “exponential improvement” was exactly what took place. Thinking specifically of the last few iterations of the game “Go” how it improved... exponentially from one to the next.
To add to this point: v9, from what we’ve seen so far, is technically capable of being level 3 on the freeway. It’s currently disabled from being so, but that’s just caution on Tesla’s part. Full level 3 would be a matter of flipping 2 simple switches: hard code lane change confirmation to “yes” and disable the nag timer.
This was because humans developed new machine learning techniques to tackle this problem. They then iterated the reinforcement learning thousands of times to create each "iteration" of the system that you're thinking of.
One very important thing to note here is that as far as I understand it, Tesla uses machine learning only for performing detection on the camera data. I do not think machine learning is involved in lane changing to any significant extent, other than detecting lane lines and vehicles/pedestrians, and it's already probably nearly as good at that as it's going to get -- no room left for exponential improvments until they bring in HW3 which will allow them to run much larger models, maybe do pixel-level labeling instead of just bounding boxes.
8.0 was released in November, 2016. It's been out for literally 23 months. If you're going down the "could be part of a point release!" path, then you're saying that it could happen in the next 23 months. Or 7 years if Tesla decides to not call any future release 10.0, because that's an arbitrary human decision.
Yes, but the problem is that once you do that, it's Tesla's fault if it hits anything. There's no data yet if it can avoid cars at an acceptable rate. The fact that Tesla has been unwilling to allow it to even cruise in a lane on a highway by itself without nagging every 60 seconds is a huge hint about how much the trust lane holding, much less lane changing.
Any car with no driver is capable of autonomy if you put it in drive and point it down a road. Just matters how long it will be until that autonomy fails
That was AI learning to play a specific game with known rules, with a insanely minimal input data set (positions of pieces). It also had very binary outputs (win/loose).
It also doesn't kill someone when it looses, so a non-zero loss rate is acceptable.
You're asking AI to look out cameras and decide when a car is going to be in your way and if you can change lanes now. With no hard data afterwards if this was OK. Maybe the human just didn't want to change lanes then. Do you mark that in your learning as a time you couldn't change lanes? How many times do you get a solid "bad decision" answer? Only when you crash, which hopefully never happens. What if the car behind you has a blinker on? Lane is merging ahead?
There's so much to combine here, it's hard to see how a real world system self-learns, which is the only way computers go exponential, and why Waymo is focusing on simulation, not real world learning, where they can drive millions of miles a minute, and can fail without consequence.
Yes, but the problem is that once you do that, it's Tesla's fault if it hits anything. There's no data yet if it can avoid cars at an acceptable rate. The fact that Tesla has been unwilling to allow it to even cruise in a lane on a highway by itself without nagging every 60 seconds is a huge hint about how much the trust lane holding, much less lane changing.
Any car with no driver is capable of autonomy if you put it in drive and point it down a road. Just matters how long it will be until that autonomy fails
That's not a matter of trust/faith in the system as much as how much liability they can afford. Even Audi's supposed "level 3" system never explained how it could understand potholes, puddles, road debris, or sudden and poorly marked construction closures on whitelisted roads. Just a month ago, I-5's repaving just simply used a sign to say "USE SHOULDER" and redirected both lanes of travel off into the soft shoulder while construction crew and large trucks sat where the road used to be.
Let's not equate the nag interval with how much the system is "trusted" by the manufacturer. Officially, you're always supposed to be vigilant for situations that can develop in an instant. That's never changed. The nag reduction was an effort to keep drivers engaged at that level, whether or not we believe it to work.
And it's almost unthinkable that any US company that has a quarter million customers is willing to take on the liability of higher levels of autonomy in completely customer-controlled vehicles. Even Waymo/Uber mainly just assumes liability for a limited number of employees and a by-invitation-only small preview program.
dark cloud
Active Member
I was waiting for someone to answer this. What is the size of the hard drive in these things?
I had chicken tonight. Thought I would help you out a bit. Didn't know they had a Pagan version.
Dude this is my pet peeve, and you need to stop.
There is absolutely no correlation or connection between SAE level 3 and making a lane change / deciding when to make a lane change.
Noooooo! Stop. L3 is NOT about capability!
Adding new features in v9 has nothing to do with bringing it closer to L3.
There can be a system that is only capable of single lane keeping and is L3. There can be a system that is not even capable of lane keeping at all or doesn't even drive in lanes that could be an L3 system.
I do agree that v9 is bringing major advancements forward, and laying the ground work for many more features to come. And v9 definitely is moving towards making Tesla AP MORE CAPABLE... which is not related to SAE level.
There is absolutely no correlation or connection between SAE level 3 and making a lane change / deciding when to make a lane change.
Noooooo! Stop. L3 is NOT about capability!
Adding new features in v9 has nothing to do with bringing it closer to L3.
There can be a system that is only capable of single lane keeping and is L3. There can be a system that is not even capable of lane keeping at all or doesn't even drive in lanes that could be an L3 system.
I do agree that v9 is bringing major advancements forward, and laying the ground work for many more features to come. And v9 definitely is moving towards making Tesla AP MORE CAPABLE... which is not related to SAE level.
rush6410
Member
But, in my opinion, very significant improvement. The difference between when the driver reads the environment and makes a decision vs. when the car makes that decision is extremely more advanced. So, I emphatically agree with your assessment.
I totally get where you're coming from, but I'll just point out that if the "feature" they were adding was the ability to look ahead approximately 10 seconds in time to know when it's about to encounter something it can't handle, then that would be a feature bringing it closer to L3.
rush6410
Member
If that makes you feel better, then keep thinking that. But, in actuality, AP2 has been much better for a long time minus the gimmicks; vehicle rendering on adjacent lanes, speed sign identification. In fact, for the important items that involve keeping the car between the lines in difficult scenarios has been much better six months after release. I have roads that lead to my house that AP1 to this day cannot handle. AP2 has been handling the same route flawlessly for at least the last six months.
Hm. Does "technically capable" include being at least as safe as a human driver even when the person in the driver's seat is completely not paying attention, like maybe is reading a book or watching a movie or composing a sonnet? Does "technically capable" include being able to react, at least as well as a human, to a deer running across the road or debris falling off a vehicle in front of it, without relying on the driver to take over immediately? Does "technically capable" including being able to handle a tire blowout, again without requiring the human driver to take over? Does "technically capable" including being able to know at detect that it's going to encounter a situation it can't handle at least 10s before it actually can't handle the situation so it can alert the driver to take over?
Similar threads
