The Tesla full self-driving thread

[Rolfe]

I'm hearing great things about this new version of Tesla's self-driving software. I honestly thought it could be decades before this was feasible, but they seem to be pretty damn close with the version 12 the Tesla boys in Merika are raving about.

Will it take all the fun out of driving?
Who will be liable it it goes banjax and kills someone?
Will we get cheaper insurance premiums for using it if it's proven to be significantly safer than a human driver?
Will we even need a car of our own if we can just call one to our door and ask it to take us somewhere?
Will buses, trains and even short-haul flights become redundant?

I'm interested because I've recently turned 70 and I live alone. Nothing wrong with my driving, but as someone remarked in the insurance thread, age itself bumps up your premiums. And there does come a point for many people when they do have to give up. My mother had to surrender her driving licence on the order of her ophthalmologist when her glaucoma progressed too far. This does feel as if we're on the edge of a step-change for people who might have problems driving.

I saw a YouTube video where someone predicted that Tesla was going to kill all other manufacturers stone dead with this. Although they will license the software to other manufacturers, it will be years before they have made their hardware compatible with it, and it won't be back-compatible. In contrast, every Tesla on the road is capable of running the software, perhaps with a hardware upgrade in the case of some older cars. The speed that this seems to be improving at (someone estimated that they're 90% of the way there now, and a lot of people are now driving to work and back without touching a control) suggests that within a few years it could be so desirable to a hell of a lot of people that buying a Tesla will be the only way to go.

On the other hand, will it take all the fun out of driving in the long term?

 

[salty]

On the other hand, will it take all the fun out of driving in the long term?

Yes.

 

[tsedge]

They are not that close at all.

While it is an amazing achievement, the driver still has to sit there and be ready to intervene. There are still plenty of cases where it fails to work.

We are something like 6 years after he promised it would be released.

Even when it works well enough, there will be years of regulatory hurdles to get it approved before there can be no driver.

And this is with it optimised for US roads, it will need training all over again for the UK, including on narrow lanes where it will need to reverse (currently it can't).

It may or may not take decades but it is still years and years away before it is a reality and there must still be some doubt if it will ever get good enough.

I'm a fan of the technology but this is a phenomenally hard problem to solve.

If it works, the destination is clear: it will become illegal to drive manually except on special private roads as a retro leisure activity.

 

[Bricktop X PWR]

There is no doubt in my mind that it will get here eventually, but don't hold your breath for a UK roll out.

We don't have the more simple grid system of road layout, as per a lot of the USA making the FSD task much more difficult.

My TM3 was excellent at staying accurately in lane, at 70mph, on my commute to work.

Truly mind blowing when experienced for the first time.

Yesterday on my way out to the main road, I had to back up for 2 tractor/trailers in my narrow country lane.

There's no self driving package I'm aware of that could deal with that.

I consider that a simple operation for a Human to carry out, but definitely beyond a computer and lines of code.

Not saying it couldn't be done eventually, but not at the speed and accuracy that a human (some humans ) can deal with the issue.

When it's good enough, I can see it having to be switched on when you join a motorway, but we are a generation away from it taking away the fun of driving I.M.O.

 

[tsedge]

Let me expand a bit on my doubts about the current technology.

Current AI is amazing but flawed and it isn't the same as how humans learn.

Humans can do the following:

• Understand and follow rules
• Consciously conceptualise and update rules from experience
• Learn consciously and unconsciously from experience

We don't learn to drive by watching thousands of hours of videos of people driving and then jump in the car.

We've moved from an era of automation that was rule-based and brittle (unable to adapt and cope with different experiences), with solutions like Waymo; which uses limited roads, maps and rules and additional sensors backed up by human remote operators who can intervene, as an example of the best of these.

To current AI which is experience-driven only with reinforcement training. This can cope with a much wider variety of scenarios. BUT it is essentially a black box, the rules have gone, we no longer know how it works internally and there's no consciousness that can communicate with us and explain what it is doing.

The problem with these systems is they can go wildly wrong - as we see with ChatGPT et al, where they can invent and spew nonsense because they have no real understanding of what they are doing: it is all pattern-matching and extrapolation. When you hit the edges you get unpredictable behaviour, which you don't want in a car. There are no longer any rules.

We are really waiting for true AI, which we will know because it will have consciousness and will be able to reason and explain itself to us (not simply imitate responses). Then you can train it in a similar way to humans: give it the rules, a ton of experience, test it and have it continue to learn and adapt.

When it's good enough, I can see it having to be switched on when you join a motorway, but we are a generation away from it taking away the fun of driving I.M.O.

It will also be a liberator for some people. I'm hoping it comes in soon enough for me to be able to rely on it in my old age, that would considerably expand independence for people with infirmities and mental decline that might otherwise preclude them having a car.

 

[Archev]

Let me expand a bit on my doubts about the current technology.

Current AI is amazing but flawed and it isn't the same as how humans learn.

Humans can do the following:

• Understand and follow rules
• Consciously conceptualise and update rules from experience
• Learn consciously and unconsciously from experience

We don't learn to drive by watching thousands of hours of videos of people driving and then jump in the car.

We've moved from an era of automation that was rule-based and brittle (unable to adapt and cope with different experiences), with solutions like Waymo; which uses limited roads, maps and rules and additional sensors backed up by human remote operators who can intervene, as an example of the best of these.

To current AI which is experience-driven only with reinforcement training. This can cope with a much wider variety of scenarios. BUT it is essentially a black box, the rules have gone, we no longer know how it works internally and there's no consciousness that can communicate with us and explain what it is doing.

The problem with these systems is they can go wildly wrong - as we see with ChatGPT et al, where they can invent and spew nonsense because they have no real understanding of what they are doing: it is all pattern-matching and extrapolation. When you hit the edges you get unpredictable behaviour, which you don't want in a car. There are no longer any rules.

We are really waiting for true AI, which we will know because it will have consciousness and will be able to reason and explain itself to us (not simply imitate responses). Then you can train it in a similar way to humans: give it the rules, a ton of experience, test it and have it continue to learn and adapt.

That’s the guy down the road isn’t it??

 

[Bricktop X PWR]

That’s the guy down the road isn’t it??

I thought that was @OriginalBigAl ?

He'll be along in a moment

 

[Gadget Geek]

I have a little anecdote from my mother, who was a primary school teacher. She told me about this Asian girl, who was the best reader in the class. However, something made my mother suspicious, so one day she started to ask her simple questions about what she was reading, and the girl hadn't got a clue. She just knew all the words, without knowing what they meant.

I think we are approaching the point where technology meets philosophy, and we will have to work out how we determine if a car with A.I "knows" what it is doing.

 

[Archev]

It will also be a liberator for some people. I'm hoping it comes in soon enough for me to be able to rely on it in my old age, that would considerably expand independence for people with infirmities and mental decline that might otherwise preclude them having a car.

I’ll look forward to that then.
The car saying “hello how are you today? Have you taken all your medications? Have your bowels moved? Do you need a wee? OK well go for a little drive now. That’ll be nice for you.”
I can hardly wait.

 

[Gadget Geek]

Let me expand a bit on my doubts about the current technology.

Current AI is amazing but flawed and it isn't the same as how humans learn.

Humans can do the following:

• Understand and follow rules
• Consciously conceptualise and update rules from experience
• Learn consciously and unconsciously from experience

We don't learn to drive by watching thousands of hours of videos of people driving and then jump in the car.

We've moved from an era of automation that was rule-based and brittle (unable to adapt and cope with different experiences), with solutions like Waymo; which uses limited roads, maps and rules and additional sensors backed up by human remote operators who can intervene, as an example of the best of these.

To current AI which is experience-driven only with reinforcement training. This can cope with a much wider variety of scenarios. BUT it is essentially a black box, the rules have gone, we no longer know how it works internally and there's no consciousness that can communicate with us and explain what it is doing.

The problem with these systems is they can go wildly wrong - as we see with ChatGPT et al, where they can invent and spew nonsense because they have no real understanding of what they are doing: it is all pattern-matching and extrapolation. When you hit the edges you get unpredictable behaviour, which you don't want in a car. There are no longer any rules.

We are really waiting for true AI, which we will know because it will have consciousness and will be able to reason and explain itself to us (not simply imitate responses). Then you can train it in a similar way to humans: give it the rules, a ton of experience, test it and have it continue to learn and adapt.

One benefit of the Tesla system is that we humans only have two eyes which concentrate on one thing at a time, whereas it has seven cameras which see in all directions all the time.

 

[tsedge]

One benefit of the Tesla system is that we humans only have two eyes which concentrate on one thing at a time, whereas it has seven cameras which see in all directions all the time.

Very true, although there's a distinction between foveal vision and peripheral vision. That which we focus on is narrow but that which we can respond to is a wide field.

Of course, humans come with lots of flakiness and failings, but any AI system has to be many times safer to overcome the trust problem with adopting new technology - we'd never allow cars if they were invented today on many grounds.

 

[Archev]

One benefit of the Tesla system is that we humans only have two eyes which concentrate on one thing at a time, whereas it has seven cameras which see in all directions all the time.

That’d be a helluva spectacles set up.

 

[Coulomb]

BUT it is essentially a black box, the rules have gone, we no longer know how it works internally and there's no consciousness that can communicate with us and explain what it is doing.

They are talking about integrating Grok (Tesla's equivalent of ChatGPT); that might allow it to explain itself. But I have definite reservations.

I do think that the current crop of machine learning is totally amazing; it performs way better than I thought it would. But it has a definite shallowness, that I suspect strongly will only be properly overcome with true reasoning and logic. There is a combinatorial explosion to be solved before that can be achieved. It's very hard to predict when that break-through will come.

 

[OriginalBigAl]

"Current AI is amazing but flawed," in this font size the difference between a capital I and minuscule l are pretty similar. Nevertheless the message is the same!

 

[Rolfe]

I get the reservations, and that it will be some time before it's available in this country, but the difference between the system as it was only a few weeks ago and what we're seeing now is remarkable. If it goes on improving at anything like the rate we're seeing it should be capable of sorting out the edge cases before too long.

I have seen video of the thing driving dirt tracks in America, and since it can park I very much doubt that it can't reverse.

I've seen a car slowing down for a pothole to drive over it slowly. I've seen a car inching along in a queue of traffic pause to leave a space for another car to come out of a side road. I've seen a car swerve on to a grass verge to avoid a truck that was being driven practically straight at it. I've seen a car move over to the edge of its lane to give a cyclist in an adjcent cycle lane more space.

No, we're not going to see it here for a few years, and it will take years longer than that to be approved, but previously I didn't imagine I'd see it in my lifetime - despite Elon's over-optimistic promises. Now I think there's a decent chance I will.

 

[DaevM]

Ah but. Can it do potholes the size of the grand canyon?
(Apologies - I see it can )

 

[Gadget Geek]

When all vehicles have level 5 autonomy, then everything should go smoothly. The problem is with the unpredictable, irrational, illogical organic life form behind the wheel which causes the problem.

I'd like to see Tesla FSD deal with this lot.

 

[Rolfe]

Having watched a few of these FSD videos, I'd back the Tesla to cope with many of these. If both vehicles were on FSD I don't think it would happen in the first place.

 

[Gadget Geek]

I think the events which will catch everyone out are the "acts of nature"; trees falling, flash floods, landslides, black ice etc.

 

[Rolfe]

Yes, as they do right now. I'd be interested to know how FSD handles black ice, indeed.