Tesla using Full Self-Driving Beta crashes into cycle lane bollard...weeks after Elon Musk's zero collisions claim

MiserableBastard wrote:

Full self-driving requires a computer to be able to instantaneously understand the road environment around it, and to have sufficient theory of mind to model what human road users, both in and out of other cars, are going to do. That's strong artificial intelligence, that is, a computer that literally thinks and understands the world like a human.

Though arguably there are plenty of human drivers already on the roads who can't do that with any great degree of success, so...

Tesla publish their collision stats every quarter. I believe this one is a summary from Q3 2020.

"In the 3rd quarter, we registered one accident for every 4.59 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.42 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 1.79 million miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 479,000 miles.”

As we all know, pulling the first factoid that supports your argument out of the internet's arse is not exactly good data review, and there are any number of confounding factors and caveats in any set of data. We also have to assume that the data that Tesla has presented has gone through some sort of sanity check and peer review and is at least broadly factual. However, from the top line figures it would appear to be demonstrable that AI controlled cars are already considerably better at the task than human drivers and given the ability to rapidly share "knowledge" from every real life incident will almost certainly become even better over time.

'Autopilot' isn't the same thing as self-driving, though - it's a form of driver-assist, no? So would occasions where the car wanted to do something stupid but was overruled by the (human) driver even show up in those figures, given that no accident resulted?

Mungecrundle wrote:

Tesla publish their collision stats every quarter. I believe this one is a summary from Q3 2020.

"In the 3rd quarter, we registered one accident for every 4.59 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.42 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 1.79 million miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 479,000 miles.”

As we all know, pulling the first factoid that supports your argument out of the internet's arse is not exactly good data review, and there are any number of confounding factors and caveats in any set of data. We also have to assume that the data that Tesla has presented has gone through some sort of sanity check and peer review and is at least broadly factual. However, from the top line figures it would appear to be demonstrable that AI controlled cars are already considerably better at the task than human drivers and given the ability to rapidly share "knowledge" from every real life incident will almost certainly become even better over time.

I recall @Rich_cb and I discussing these stats some time ago and there's definitely some cherry-picking going on though I forget the details - something to do with NHTSA's data covering a much wider range of vehicles and situations which made the comparison void.

I think the issues are raised here: https://www.businessinsider.com/tesla-crash-elon-musk-autopilot-safety-data-flaws-experts-nhtsa-2021-4?r=US&IR=T

MiserableBastard wrote:

Self-driving cars are a combination of a techno-utopian fantasy and a massive scam.

I do appreciate your points, but I'm still going to go with the optimism that safe driving is not an unsolvable problem. I think the failings of Tesla are more to do with just using machine vision (via ordinary cameras) rather than building up a true 3-d picture of their surroundings with lidar/sonar/radar etc. Ultimately, it is possible to make a bat's brain (with the help of a mummy bat and daddy bat that love each other very much) that is capable of using sonar to navigate in the dark and that brain is tiny. Now obviously, we're nowhere close to achieving the same kind of performance in silicon, but I don't see why it can never be done.

However, Musk is a showman and I don't believe what he says.

I'm slightly wary of self-driving AI - primarily for other reasons more to do with how humans use this e.g. who benefits, who regulates and what happens when it goes wrong etc.  However your quoted comment from an article somewhere seems to be conflating philosophy with engineering / practicalities.  It may startle that commentor to know that the human brain itself is likely doing exactly what they critique computers for doing a bad job at e.g. tackling "... problems for which the solution can’t really be calculated, only approximated. The whole system is probabilities..."

There's lots of debate on the larger question (e.g. does the "hard problem" even exist or does it arise e.g. due to human framing of our own "beliefs" about what we think we do when we think).  However the more we know the clearer it is that for many human activities we operate a lot of more or less flaky heuristics to achieve our complex behaviours.  True, we almost certainly do the low-level computation differently than computers compute (numbers are likely not a feature) but a "fuzzy computation" it certainly is.  This is readily apparent when things go wrong.

The point is not "but robots!"  As we're aware on this forum humans have an equivocal record for safely travelling on the roads.  Machines are already better at many specific tasks than humans including some complex ones.  They also tend to avoid some of the common human failings (forgetting, being distracted, emotional or tired).  It's more "how can we be sure this handles enough of the vagaries of the 'open world' environment we're letting them free in?"

"Brute simplicity" was certainly a valid critique of early generations of AI.  But that study had to start with a very reductionist view of something so complicated just to get a handle on the questions to ask.

MiserableBastard wrote:

Full self-driving requires a computer to be able to instantaneously understand the road environment around it, and to have sufficient theory of mind to model what human road users, both in and out of other cars, are going to do. That's strong artificial intelligence, that is, a computer that literally thinks and understands the world like a human.

While humans certainly have a keen interest in what others think / believe and have complex systems dedicated to that * I'm not sure that's proof you need any kind of detailed model of a human brain in the system.  I'm not convinced that human drivers are doing something irreducably complex to do with understanding of others when driving.  If you ask them - when they have time to think - what they thought their decision-making process was they may explain it as that.  I suspect this just shows the limits of our introspection into our own processes.

We likely rely on "prediction" - or rather our experience, which could be said to be our own "machine learning" - to a greater extent than computers do simply because our reaction and processing times are much longer.  Certainly engaging higher-order conscious reasoning would result in results arriving too late to help in many cases.

* The details are fascinating - very young children seem to have some mental machinery primed to start teaching itself about others.  Some of this may be shared or have evolved independantly in other animals too.

You know all those times a human misunderstood you?

With a human driven car someone just died.

Self driving cars don't have to be perfect, they just have to be safer than humans.

Every mile they drive they get better and will continue to do so. I would argue it's unethical to not develop the technology given the potential harm reduction it offers.

Fully self driving taxi services are now running in two US cities so I'm not sure they're fantasy any more either.

Secret_squirrel wrote:

All it shows is the tester is a twat.

Well, no - it also shows that allowing things to be public beta tested on public roads isn't a great idea.

mdavidford wrote:

Well, no - it also shows that allowing things to be public beta tested on public roads isn't a great idea.

Everybody has a testing environment. Some people are lucky enough to have a separate environment to run production in.

Isn't this always the problem of "what do you get if you cross {thing} with a computer?  A computer."? e.g. if everything is effectively software then we'll get the practices of that business.  Although I'm sure no company has ever made even its lowest-priority users act as beta (or even alpha) testers ...

Remembering the lead-in-petrol story you highlighted perhaps it's not unique to software.  I think this does change things though because now technological change occurs even closer to the speed of fashion and a "fix as we go" mindset is the norm.  I'm recalling as a counterpoint to that the Brooklyn Bridge where the chief engineer said that due to uncertainties - e.g. what they didn't know they didn't know - he designed some parts with a safety factor of several times the required strength they calculated.  Which turned out fortunate as they were supplied some sub-spec materials and no-one had much idea of the aerodynamics of such structures at that time. Obviously Teslas are not designed to last that long - indeed will have obsolescence built-in!

chrisonatrike wrote:

Isn't this always the problem of "what do you get if you cross {thing} with a computer?  A computer."? e.g. if everything is effectively software then we'll get the practices of that business.  Although I'm sure no company has ever made even its lowest-priority users act as beta (or even alpha) testers ...

Remembering the lead-in-petrol story you highlighted perhaps it's not unique to software.  I think this does change things though because now technological change occurs even closer to the speed of fashion and a "fix as we go" mindset is the norm.  I'm recalling as a counterpoint to that the Brooklyn Bridge where the chief engineer said that due to uncertainties - e.g. what they didn't know they didn't know - he designed some parts with a safety factor of several times the required strength they calculated.  Which turned out fortunate as they were supplied some sub-spec materials and no-one had much idea of the aerodynamics of such structures at that time. Obviously Teslas are not designed to last that long - indeed will have obsolescence built-in!

Before the computer age, iterating designs/production/manufacturing was an expensive and laborious task, so more attention was given up-front (e.g. measure twice, cut once). However, with software, it can be very cheap to throw together a prototype, muck around with it, bin it and then start again. This fits in with the "move fast and break things" ideology as most companies can easily manage having a bit of down-time and will then hopefully prevent that same mistake happening again.

None of that fits in very well with safety-critical engineering which is what Tesla should be doing as there's a world of difference between a website being down and a pedestrian/cyclist being knocked down.

Well we've always prototyped - and still do extensively in some domains (firearms come to mind).  However I think the existing pressures to spend less time at that point is increasing. Also having more than a hazy overview becomes harder as the software increases in complexity and the degree of connection between these systems increases.  Yes, the death of the expert generalist was supposedly almost 200 years ago...

Completely off-topic now but I find this domain interesting.  If you're not already aware Tim Hunkin has some interesting musings about the process of "design" and technology - his and others - as well as some amusing machines. A place to browse around.

I didn't know that!

tbf to Musk on this.  He bought them out 8 months after the company was founded (July 2003-Feb2004)  and before a product had been released.  Thats close enough for me.  They needed piles of money.  He had it.

They still have shareholdings and are likely crying all the way to the bank.

Secret_squirrel wrote:

tbf to Musk on this.  He bought them out 8 months after the company was founded (July 2003-Feb2004)  and before a product had been released.  Thats close enough for me.  They needed piles of money.  He had it.

They still have shareholdings and are likely crying all the way to the bank.

Elon Musk and JB Strauble had been working independantly on putting an electric power train derived from AC propulsion IP into a Lotus Elise before they met Eberhard et al.

It was more of a stealth mode merger than Elon being the first investor in a going concern. 

Eberhard fucked up the development of the Roadster, Elon Musk replaced him as CEO found more funding to the keep the company going and the rest is as they say history.

The "secret master plan mk 1" is also mostly Elon Musk's work, the strategy of sports car, luxuary car, junior executive car, everything else plus the relentless execution is what differentiated Tesla and that is mostly due to Elon Musk's iron will setting the culture.

hirsute wrote:

Awaiting comment from World Bollard Association™

let the hostilities commence!!!

I know what side I'm on 

Surely MIFWARF?  Man In Front Waving A Red Flag.