People still seem to think their fancy cars are fully self-driving Despite years of headlines about driver assistance systems being involved in horrific car crashes, a study by the US-based Insurance Institute for Highway Safety (IIHS) suggests motorists continue to use their vehicles as though they are fully self-driving. One does not have to look back far. Only last month did General Motors …
this illustrates a serious problem with drivers more than it highlights problems with "driver assist" or whatever the manufacturers decide to call it.
Perhaps instead of trying to develop "better" AI the whole idea of "driver assist" ought to be dropped completely until there has been some proper research into driver behaviour in the hope of finding out whether eliminating human error (for which read downright stupidity as often as not) is a realistic prospect.
Whence cometh the clamour for "self-driving vehicles" anyway? I know the concept seems to make politicians salivate but that offers no insight into whether the idea is a good one or not.
Not least it comes from the owners of taxi fleets, bus companies, delivery. Companies and long distance hauliers who would be much happier if they didnt have to pay for all those pesky employees who want things like sleep, and sick leave, and pensions.
Personally I'm a big fan of the idea for my own vehicle because it would allow me to drive to a nice pub in the afternoon, have a few beers with my lunch, and have the car drive me home again.
We are a long, long way from that tho, and the "it's not for hands free driving *wink wink*" approach the current systems have been marketed with is absolutely irresponsible.
"Personally I'm a big fan of the idea for my own vehicle because it would allow me to drive to a nice pub in the afternoon, have a few beers with my lunch, and have the car drive me home again."
You then get old and can't deal with the day after as well as you could even 10 years earlier. If you are just going to the local, a taxi for the next 10 years would be less expensive than what a Level 5 car would cost. It sounds great to have your car drive you home and it's a point lots of people bring up, but is it worth the money?
To my mind, the problem isn't specifically the drivers so much as the terminology. Put an outright ban on all such terms as autopilot, self driving, etc and require them to be described as driving assistance aids and promoted and marketed as such.
If people get given all sorts of marketing promise and buzzwords, they may feel that these "aids" can rescue them from a sticky situation, or take over if they have this one important text to send...
And, furthermore, ban all actual self driving vehicles from public roads until such time as they can be demonstrated to be as good as a reasonably competent driver on test tracks (different ones, not the same one for obvious reasons). The rest of us are not your alpha test crash dummies.
People want to try to blame terminology and marketing, this is a kop out, that’s the easy target.
It seems obvious that if people still believe their vehicles are capable of self driving after all the warnings and efforts to warn them otherwise, then the problem is with those who place too much trust in the autonomy. What I’m saying is there is a bigger problem than marketing words.
And the reason for the clamour for a working self driving solution is safety. Human drivers have a significant number incompetent, reckless, selfish people who continue to cause death to innocent people. I guess that in this incipient self driving era, the two are coming together to create a new class of accident.
What kind of "assistance" are they providing if the driver still needs to be attentive and holding the wheel? If these systems are used as documented, they just change the driver's job from actually turning the wheel to holding it while it turns by itself.
That makes paying attention harder, because the driver is less active and has a reduced role.
I can see the argument for calling features like blind-spot warnings and lane-departure warnings "assistive" (though I have to say that having driven a car with a lane-departure warning a few times, I don't care for it at all; it's intrusive and annoying). Lane-following, not so much.
Well as someone with a disability that makes driving much more difficult even simple lane keeping substantially reduces the cognitive load associated with driving.
If we could get a vehicle that would even just do motorway driving for us that would be a huge benefit, and it would actually be a benefit to get that on all vehicles - motorways would rapidly become even safer, and flow better.
Additionally the roads around motorways would become safer as well, since people would be much fresher and more alert.
Well we do have a Tesla - but I would rather check into rehab than turn on 'self-driving' - in fact, it would probably be a good idea if I found myself turning it on.
However, the idea (well concept) of self-driving cars is great - nothing better than being driven to a restaurant/pub and then rolling back out to be taken home at a time of our convenience. It would also work well fo my Mum - out in the middle of nowhere and self-drive would allow her to live there for longer.
Unfortunately, it won't happen soon.
I used to take a taxi, back when it was still possible to hail a cab on the street. It's like having your own driver, only better: just step out of the office, hail a cab, and ride home.
Except when the driver was drunk, or drug affected, or both, or driving on his cousin's license and turning into the wrong side up a well-known divided 6 lane street
I rode with some exceptionally good and exceptionally knowledgeable drivers. But just 'taking a taxi' doesn't solve the problem of dangerous crashes.
Quite, and not just Scotland!
Some "solutions" to the problems of car ownership/use reflect a very metropolitan viewpoint.
However, self-drive is still facing many challenges on fairly normal roads, I am not sure it would cope with some of the situations often encountered on country lanes - so it is not the solution yet (or for some time). Is anyone aware of any published work in this area?
"FFS, self-driving trains seem to be difficult."
I was watching a show in trains that showed an ore train in Australia that was automated. It was so heavy that it took a really long time to stop so even if a driver could see a car across a level crossing, slamming on the brakes wouldn't do a blind bit of good.
A town I lived in town that dug underpasses at almost all of the crossings. Too many idiots and the driver was required to blow the horn at each one previously, which they took to mean just leaving going half way across town day and night. I never saw the point of blowing the horn. By the time they did it was far too late and they would blow the horn as they went across a crossing which also made no sense. Anybody sitting there would be very flat.
"middle of nowhere"
I used to live in the middle of nowhere, (well, rural Lincolnshire) and I'd love to know what a self driving car would do if faced with a combine harvester on single track road. Those of us used to driving in such places make mental note of each passing place as we go buy, so we know where we havr to reverse to, how is an AI going to do this? Or deal with sheep, cows, horses etc, they sometimes have to share the road with. There are no roadmarkings in these places, no cat's eyes, and the edges of the road are often indestinct and overgrown, it's an overwhelming amount of noise for an AI.
"Whence cometh the clamour for "self-driving vehicles" anyway? I know the concept seems to make politicians salivate but that offers no insight into whether the idea is a good one or not."
Seriously? It may still be a long way off, but it's a wonderful idea. It improves many different areas. Here are a few - note that I'm talking about hypothetical future vehicles with working self-driving:
- Driverless cars make more efficient use of energy inputs
- Driverless cars make more efficient use of road space
- Driverless cars never crash or mow down cyclists
- Driverless freight vehicles cut freight costs by around a third
- Driverless deliveries cut local delivery costs significantly (though you might have to go out to the kerb to pick up your pizza)
And that's without even getting into the personal benefits of being able to sit in traffic or on the motorway and do something other than stare through the windscreen. I strongly suspect the killer app for driverless cars will be the mobile shitter.
Here are a few - note that I'm talking about hypothetical future vehicles with working self-driving:
Interesting that you should include the word hypothetical because not only are the vehicles currently hypothetical but so are the benefits that you listed; they will remain as a wish list until such time as they have been proven to have been realised.
One in particular warrants further comment: Driverless freight vehicles cut freight costs by around a third. I'd like to know how that has been calculated; just the fuel cost for an hour's running is well in excess of what the driver costs.
In addition, the driver is normally responsible for assisting with unloading / loading the vehicle; in his or her absence who will carry that out? Loading is singularly important because it is essential that the load is properly spread to even to distribute the weight on the various axles and that it is secure and cannot move when the vehicle is moving. It is a legal requirement and if he / she gets it wrong the driver can be held liable in Court.
I'll believe that AI really exists when it can reverse an articulated vehicle round corners / into a tight space / etc without mishap. Will this hypothetical LGV be able to refuel itself with either diesel or electricity when the need arises?
"I'd like to know how that has been calculated; just the fuel cost for an hour's running is well in excess of what the driver costs."
The usual rule of thumb in UK haulage is that the driver costs a third or so. Seems about right given fuel costs and total driver pay.
"the driver is normally responsible for assisting with unloading / loading the vehicle; in his or her absence who will carry that out"
Loaders at either end, obviously.
"I'll believe that AI really exists when it can reverse an articulated vehicle round corners / into a tight space / etc without mishap."
Er, that's the easy bit. If that were all that's needed they'd have been ubiquitous at least 20 years ago.
"Will this hypothetical LGV be able to refuel itself with either diesel or electricity when the need arises?"
Obviously, even if only by driving into a bay where a human attends to it.
It's an odd set of questions you have. There's the mountain sized problem of creating a system that can safely interact with traffic, and you're looking at a few grains of sand as significant.
"Loaders at either end, obviously."
You've clearly never worked in the transport industry. Many such loaders are crap, once the vehicle has left it's not their problem anymore. On my own delivery rounds I've witnessed many other drivers who haven't loaded their own vehicle picking goods up off the deck, or more usually bitching about top-heavy loads. Not everything carried on lorries is on pallets, or easily strap-able. For multiple drops there is the added constraint of delivery sequence. An experienced driver should know how to load his vehicle safely and effectively, they will know what might move, and load (and maybe even drive) accordingly.
Some depot's make a point of having nothing whatsoever to do with vehicle unloading, even to the point of not gritting their yard in freezing weather. Getting to the loading bay is the drivers responsibility, as is dragging the load halfway across the warehouse to the specified location. This is how the cheapskate supermarkets keep costs down.
Amazon USA has a massive network of highway vehicles criss/crossing the US between massive Amazon warehouses. Where the loads are loaded / unloaded by warehouse staff.
Point-to-point deliveries will be the last part of trucking that is automated, if ever. The people looking at automated trucking are looking at long-haul container loads.
"Amazon USA has a massive network of highway vehicles criss/crossing the US between massive Amazon warehouses. Where the loads are loaded / unloaded by warehouse staff."
And often trucks have to queue up and wait for the next available dock. Amazon is often moving their goods in container/trailer loads. LTL (less than trailer load) freight often takes an on the ball driver to make sure things are packed properly, secured in the truck correctly and get from point to point without going missing. Amazon and other large companies are the exception in most cases but they do get LTL deliveries and the drivers hate them along with Walmart and Target. If they check-in late, there is a penalty, but it might not matter that there was a queue to back up to the dock or people were on break. There is sometimes no nearby place to wait so they have to deal with city streets and finding a place to park up until their number is called.
There's much more that can be done with trains. Goods used to ship in box cars and would be stacked in the holds of ships. Now most things ship in a container. A little design work and distribution warehouses could be once again sited on railroad sidings and cars with containers on top individually backed in to docking bays using small shunting engines to move them along. Automating Over The Road trucks is like building a machine to flip burgers instead of designing a machine that cooks both sides at once.
You missed
- People will have endless fun stepping into the traffic secure in the knowledge that the self-driving vehicle will always stop
- Anyone still with an older vehicle will be able to pull into the traffic flow more easily as self-driving vehicles with slow down/stop.
I don't think he was being sarcastic. One of the big advantages of self-driving cars will be that they're safer for everyone around - once they're working, remember. They will be far more likely to anticipate/notice someone stumbling into the roadway, and will of course react quicker than human drivers.
One of my biggest worries driving are the people who hide in dark clothes in shadows (at night) and then jump out in front of the car. They think they're crossing the road sensibly, I assume, and are just oblivious to the danger.
This is a discussion we've had here before and I've said before that IMHO a computer (car) has no rights. Pedestrians and drivers interact in ways where both have an idea about how the other is likely to react and also a reluctant acceptance that the other exists. I might know that the gap in the traffic is big enough for me to cross the road but that the car is going to have to slow down/brake/stop, I wouldn't walk into a gap where it would be need a significant avoidance action from a human driver. Replace the human driver and the bloody computer can wait. If the gap is big enough for me I'm going to come over all militant pedestrian and I'm going to exert my right of way and the bloody computer can stop. I don't think I'm alone with this. Try walking round any busy city and you'll find plenty of militant pedestrians. Busy bits of London for example are often an extended game of chicken where any exploitable gap or weakness on one side will be exploited by the other. AVs will not be allowed to play this game and pedestrian will not follow rules, why the .... should they.
They will be far more likely to anticipate/notice someone stumbling into the roadway, and will of course react quicker than human drivers.
Assumes facts not in evidence.
Or are we still laboring under your "self-driving vehicles will be great when they're great" tautological hypothetical?
"- Driverless cars make more efficient use of energy inputs
- Driverless cars make more efficient use of road space
- Driverless cars never crash or mow down cyclists
- Driverless freight vehicles cut freight costs by around a third
- Driverless deliveries cut local delivery costs significantly (though you might have to go out to the kerb to pick up your pizza)"
Not even close. A driverless HGV is a security-less load. If you look at the complaints from drivers that go in and out of distribution warehouses, a lot of what they hate would be impossible for a driverless vehicle to deal with. Plenty of time there is mixed freight and something needs the trailer to have a placard which can also mean a driver that has training for that sort of material. There is often paperwork involved and routes that must be avoided. A driver with multiple pickups and drop offs can also keep track of what is going where so some moron on a forklift doesn't grab the wrong stuff or forget to put a pallet back on that was moved to get to other pallets.
Local deliveries are even worse.
The whole list is based on hopes, not facts. It probably makes more sense to work on railroad automation to cut down on handling time so HGV's aren't days faster.
We have two cars - one with driver assist and one without. The one with driver assist features has RADAR to automatically adjust car separation when on cruise control - the other does not. When changing cars I have to keep remembering that the current does not have driver assist features. So there are two problems:
1. relying on driver assist features too much
2. realizing that not all cars have these features
"When changing cars I have to keep remembering that the current does not have driver assist features"
You also have to know when it's not working correctly. If a RADAR sensor ices up, it doesn't work. It 'should' realize there is something wrong, but....... The same goes for any sensor put on a car. A human driver knows when the windshield has muck on it and they can't see properly. A camera or LiDar might not be able to tell that there is a problem.
From the article:
"the early adopters of these systems still have a poor understanding of the technology's limits"
...which implies, no, they didn't. You'd think "early adopters" would have had them long enough to have learned the limits by now, but it seems many have not. Personally, I'd say they are exactly the sort of people who should not be driving these cars, possible not even allowed to drive at all.
Emergency braking did not understand the concept of corners. Other parts of the car may know the road has a corner. Some cars may even know they intend to go around the corner but do not inform the emergency braking system of the planned route. Even that is only one step in a long path. The next step would be to guess if an on coming vehicle is going to follow the curve of the road.
Today I saw some light reflected from the side of a van and inferred the presence of a vehicle around the next corner. I reduced speed so I would pass that vehicle at one of the wide points on a narrow road. Last month I did something similar because I caught a glimpse of a car shaped shadow moving towards me through a gap in the hedge beside the road. These are low probability events that will barely show up in training data. They require real world experience to interpret and some persistent memory to continue the required response after the evidence has moved out of sight. The only way full self driving will become safer than humans within a couple of decades is if we retro fit humans with Elon levels of foresight.
Inferring hidden vehicles by changes in lighting would be super-tricky.
There are other ways to know about hidden vehicles.
One way would be to use a grid of satellites to monitor all cars on all roads and have the car in communications with that grid to tell it what's near and their vectors.
Hopefully such a system would be prohibitively expensive. But if not, what could possibly go wrong with a total surveillance system.
Another method would be for all cars to enter a radio network of cars within a certain range but that would need to be retro-fitted to all existing cars for it to work - not going to happen.
(I'm not sure why your post was so roundly downvoted.)
The "every vehicle has a transponder" proposal comes up frequently – even CPG Grey has an (older) video proposing it. But it has dreadful failure modes, because a single transponder failure renders that vehicle invisible, or at least anomalous, to its neighbors. And private cars are, of course, maintained by private owners, and often not well.
inferred the presence of a vehicle around the next corner. I reduced speed so
I would have no problem with a self driving vehicle that was this cautious as a matter of course, since you can't rely on 'seeing' around a corner at all times even if you are able to do so sometimes.
Most of us aren't that cautious by nature and if we see no evidence of a car coming around the bend will assume there is no one (unless it is a road that's so busy such an assumption would be stupid) and we figure that if one appears we'll react as needed to avoid an incident and pay a bit more attention around the curve than we do in a straightaway as a result.
An autonomous car should be doing "defensive driving" 100% of the time - the way us meatbags are supposed to drive but mostly don't. Maybe it takes the autonomous car a bit longer to reach the destination, but that's fine. When I'm just a passenger and can do other things I won't care if it takes a bit longer to get where I'm going. I know I drive faster than I should / the law allows a lot of the time, but really that's because driving is a chore and like most chores the sooner is it over and you can get to what you really want to do (i.e. whatever is at your destination) the better.
If your driving style includes putting on the brakes at every slight corner even if there is perfect visibility of the road ahead then the sooner you get relegated to the back seat the better. We all know that back seat drivers are the best (sic)
I've noticed more and more drivers doing this in recent years. It is not necessary at all. There is no need for a dab on the brakes at every bend. If you are slowing down then fine but these dabs are not long enough to change the speed of your vehicle very much.
It is as frustrating as hell to those following you.
I was taught to drive as smoothly as possible. No sudden acceleration, no sudden braking.
Slowing down a bit isn't the same as braking.
You absolutely should be slowing down before many bends, because they're often impossible to see around and so you've no idea if there is a stationary object.
Most drivers don't make sure they can stop within the distance they can see to be clear, they just assume their side of the road is empty around the bend or over the dip, unless proven otherwise.
utting on the brakes at every slight corner even if there is perfect visibility of the road ahead then the sooner you get relegated to the back seat the better. We all know that back seat drivers are the best
This applies to you big time. The sooner we get self driving so I don’t have to share the roads with you the better. If you drive correctly you would not need to be breaking to go safely into where cannot be seen.
If you can't drive, I don't want you sharing the roads with those of us who can. Even Musk's marketing department carefully points out (if in the smallest of print) that his cars aren't really self driving. And even he seems to be coming around to the concept that we don't have that technology, not really ... and probably will not any time soon.
It's all a pipe dream.
I was taught to drive as smoothly as possible. No sudden acceleration, no sudden braking.
They key phrase you are looking for is anticipation , taking your foot off the accelerator before you reach the bend to allow the speed to drop without having to hit the brakes.
Who said anything about applying brakes "even when there is perfect visibility of the road ahead"? This was explicitly about curves where you have blind spots - i.e. the post I was replying to talking about getting hints of oncoming traffic due to reflections or seeing a bit of a vehicle through hedges.
When driving around a corner to somewhere you can’t see you should be driving at a speed where you can deal with anything you encounter anyway. Which is what the autonomy does. It doesn’t assume the way clear because no usual cues, which is why it is safer. Not every obstacle will be emitting light.
Elon levels of foresight.
Um.. I'd to retain my intelligence and non-egotistic personality, thankyouverymuch.
Musk is just a loud mouth who happens to have money and a knack for buying tech companies that he can overhype and make wild promises from. A lot of his ideas have hurt Tesla (and it's self-driving) pretty badly. (like insisting on only using vision for distance and object detection, instead of purpose built sensors that work FAR more reliably.)
I'd like to retain MY levels of foresight instead of his, because I'm doing badly enough without that.
Anybody else get the feeling the author was desperately trying to stitch unrelated fact-points into a coherent story just to prove a point?
Like "the damning statistic that Tesla's Autopilot system is involved in 70 percent of accidents where driver assist software is a factor." Or, 70% of cars with high assistance are Teslas, perhaps?
Tesla doesn't advertise but GM does, so lets see if we can make it look like they both push their messages down your throat? ... Super Cruise ads focus on hands-free capabilities while Autopilot "implies ...
Self driving is coming, and it'll change the trucking industry once people accept that it works. Not much else.
"It also seems to be quite attractive to alcoholics."
I find it easy enough to sort out before I go to the local if I'm having one or lots. If it's the latter, I'll make arrangements in advance if I need to. There's no need to be somewhere that has taxis roaming the streets. In a larger town, most places will be happy to call a taxi for you if you ask. These days it takes some 'splaining so they understand I want a taxi with a licensed and insured driver and not Uber/Lyft. I understand that's not always possible, but I don't 'need' to drink if I don't have a way to get home or back to the hotel.
Tesla's advertising works differently to traditional car manufacturers. Elon writes tweets. Very quotable short tweets about how great full self driving will be real soon now. Those tweets get walked back in longer vaguer less quotable statements. People are good at hearing what they want to hear and the advertising standards authority does not regulate Twitter. After another disaster Elon's lawyers can repeat "Twitter is a home to invective and hyperbole. No reasonable person would consider Mr Musk's Tweets as a source of factual information." Regrettably the internet is not populated entirely with reasonable people.
Sometimes the statistics you want are not easily available, like accident rates per million miles on roads where Tesla's advanced driving assistance features can be activated. Comparing Accident rates with assistance activated to unassisted accident rates on all roads (including the difficult ones) gives a really false impression that has been amplified by social media.
Comparing Accident rates with assistance activated to unassisted accident rates on all roads
Yes I'm sick to death of the false claims that Autpilot is safer than human driving, because people aren't activating it in the sorts of places where humans tend to get into more accidents. Racking up accident free autopilot miles on an expressway proves nothing, because humans rack of lots of accident free mileage on those roads too.
People have most of their accidents in situations with limited vision due to weather conditions or night, and since Musk stupidly refuses to consider LIDAR autopilot will never be very good in such situations either - so owners drive themselves in the exact conditions humans are most likely to have accidents.
"I'm sick to death of the false claims that Autpilot is safer than human driving, because people aren't activating it in the sorts of places where humans tend to get into more accidents."
Even taking that into account, all the current systems are far beyond the point where they are comparable to human drivers. Those are the stats, like them or not - the crash rate per mile is way lower. I
t does not say much about the reliability of self-driving systems, it's just that being better than the average human is a very, very, very low bar. If that's all self-driving can manage, then it won't be allowed on the road (however sensible that might be in a dispassionate analysis).
but their legions of cult members do it for them.
Go to any Tesla Fan Forum and you will see a plethora of posts about how good their self-driving system is.
Cult members are chomping at the bit at being able to set off for work and go back to sleep at the wheel, letting their computer on wheels get you safely to work. Some cult members have been caught sleeping at the wheel. Their 'Elon said...' defence sucks.
Elon promised a totally hands-off coast-to-coast USA trip back in 2018. AFAIK, it still has not been done.
He is a modern snake oil salesman who has learned a lot from Trump.
I have driven a Tesla for 3 weeks when I was in California. That has put me off them for life. Apart from a horrible driving experience, having to mix with fanboi's every time I charged up was awful. They would preach the gospel according to Elon almost every time.
I have my second Tesla and I am pretty pleased with it, however, I doubt that my next car will be a Tesla, other manufacturers are catching up and the charging network is improving. The Xpeng P5 with lidar, radar, cameras, and USS seems to be in the lead when it comes to cost effective "self driving" meanwhile Tesla thinks it can do it all just using cameras, bonkers.
Have I got this straight? A car with autopilot (or whatever its called this week) disables itself if the driver isn't paying attention. Surly that means no driver and no autopilot == more chance of a crash?
boolean chart----
Driver non attentive no autopilot max risk
Driver non attentive autopilot lower risk
Driver attentive no autopilot some risk
Driver attentive autopilot lowest risk
Wont affect me, I use SWMBO for all my autopiloting from the pub !
AIUI, the Tesla computer tries to get the drivers attention with bings and bongs etc and if that fails, pulls over and slows to stop. But I don't think it's clever enough to be sure that where it stops is a safe place, but at least it doesn't just disengage while in motion. The article implies that the computer assist can be disabled and in effect ban the driver from using it. It doesn't say if that's automatic, e.g. if it has to slow and stop three times it stops operating completely or if that's a decision made at the Tesla mother-ship by a human alerted by telemetry,
" the Tesla computer tries to get the drivers attention with bings and bongs etc and if that fails, pulls over and slows to stop."
I'd like to see that "pulling over and stopping" thing demonstrated since there are just a few videos showing Teslas going full speed into crashed trucks, semi-trailers across the road and a smattering of photos of what happens when you smack a stopped emergency vehicle with a Tesla. It's one thing to be a bit safer in some situations but then have the car kill you doing something you would likely never do.
Absolutely. As several have pointed out, it's reckless for car manufacturers to use terms the suggest it can do more for you. "Auto pilot" is the most stupid name for it. Of course people believe it will drive for them - because the manufacturers told them that with the name of the feature.
What matters isn't what "autopilot" in airplanes really does, what matters is what the typical consumer looking to buy a car thinks it does.
That's what Musk's marketing is counting on, he wants people to equate Tesla with "self driving" because he's been stupid enough to continually make promises about Teslas self driving for years when they are no closer than other automakers' technologies like Super Cruise.
I find it hilarious he's now saying they "just" have to solve general AI to make self driving cars. If that's really true then we will not see them in our lifetimes, but he probably thinks AGI is just a few years away (he must, with the ridiculous promises he's making for his robot) He's constantly fooled by the amount of effort required to solve half a problem and thinks the second half will take the same amount of time the first half did.
"He's constantly fooled by the amount of effort required to solve half a problem and thinks the second half will take the same amount of time the first half did."
That shows a lack of engineering experience. I break problems down and solve the easy bits first since that will often give me insight into the remaining parts of the problem, but not always. It stands to reason that the easy stuff is fast to solve and the hard stuff usually takes much longer. It's a great day when you get some insight into the hard parts and only have the easy stuff to do, but that's far from the norm. The real pisser is when I have to redo the easy bits to make the hard stuff possible to solve when I don't get the break-down correct.
Pilots have to undertake extensive training in what their autopilot does and doesn't do before they can use it.
When an aircraft autopilot starts doing the wrong thing or disengages, the pilot usually has a minute or two to figure out why/follow a checklist before it becomes hazardous.
With a road vehicle, the driver usually has a second or less.
Of course, if the pilots work for Air France, they may not remember whether they were cruising or landing, or if the aircraft is a Boeing, they might not have been told that the autopilot exists...
All that will do is make their insurance rates go up.
Which might have the effect of Tesla owners paying a lot more for a car with the "full self driving" service enabled than for one that does not, which will definitely hurt Tesla's ability to sell that expensive uplift if it turns out you are adding thousands to your insurance bill by having that capability due to some misusing it.
"I feel the same about these morons as I do about people who drive off cliffs and quaysides because their SatNav told them to."
There are enough people that will do that to be scary. I remember when Apple's maps first came out and all sorts of exciting errors were found. People were driving themselves way out in the desert on roads best tackled in 4wd and getting stuck when the first clue was that they were going to a nice restaurant not too far away downtown. The second hour of the 20 minute drive is also another dead giveaway.
"What will come first: Practical fusion or functional "strong" AI?"
Practical, sustainable, working fusion will get here, probably in time for my Grandaughter to fully benefit from it. Hopefully much sooner, but I'm not holding my breath.
My Granddaughter's great-great-great-great-grandkids will not see functional "strong" AI. In fact, it may never exist at all.
He's just moving the bar on what he thinks is an easy task. He must think AGI will be solved within a few years given the crazy promises he's making for his robot.
That's probably why Tesla is spending all that money on those AI "dojos". He thinks you can crack AI via throwing computational resources at it. Until we better understand the nature of intelligence, we can only fake it in very narrow problem domains.
And consider that Honda had a (much nicer looking) bipedal robot back in the 90's, the current Optimus that was demo'd was... very underwhelming, especially compared to some of the demos that Boston Dynamics has put out.
Also, Hasbro might have some legal words regarding naming the robot "Optimus"...
""You actually just have to solve real-world AI".
The word "just" does an awful lot of heavy lifting here, Elon..."
Yes, we're completely done painting your house. We 'just' need to finish the trim. Ummm, that means you aren't done. People think that using 'only' or 'just' makes it all ok. I run into that a lot. Contractors have to be asked if they are 1)done 2)everything can be inspected to be done 3)tarps, tools, trash, etc have all been removed and taken away 4)if any of the above isn't true I get the job you did for no charge, right? Sign this piece of paper for me on the bottom and date it, I'll fill out the top later.
"I think he's doing that to himself. Have you seen the "handles" on Tesla cars? They don't really have handles as such on some of the models and those that do have been re-invented to be something other than a door handle. So yeah, I don't think he's got a handle on it :-)"
Did you see the video of the blue Model 3 that died, cut power and the owner was happy to have been able to kick out the window and crawl through? You can hear him telling the person filming the fire department pulling up to extinguish the raging fire. Now Elon wants the cybertruck to have bulletproof glass. I'm not a fan of electric flush mounted door handles. Making them flush does F-All for aerodynamics and making them electric means a helpful bystander would have no hope of finding something to grip to pull the door open should you be unable to do it yourself. There is a manual bypass on the front doors of Teslas, but the people in the rear seat need to read the safety instructions so in case of an accident or water landing, they know to crawl through to the front since there aren't mechanical bypasses on their doors. The other issue is you have to know they are there which supposes you read the manual and need to pull up carpet and find the pull tab. Ever been in a car where all of the airbags have deployed? You aren't finding sh$$.
https://www.autoblog.com/2022/10/08/autonomous-cars-slow-progress-losses-doubt/
My favorite part of the article:
For now, here’s what we know: Computers can run calculations a lot faster than we can, but they still have no idea how to process many common roadway variables. People driving down a city street with a few pigeons pecking away near the median know (a) that the pigeons will fly away as the car approaches and (b) that drivers behind them also know the pigeons will scatter. Drivers know, without having to think about it, that slamming the brakes wouldn’t just be unnecessary — it would be dangerous. So they maintain their speed.
What the smartest self-driving car “sees,” on the other hand, is a small obstacle. It doesn’t know where the obstacle came from or where it may go, only that the car is supposed to safely avoid obstacles, so it might respond by hitting the brakes. The best-case scenario is a small traffic jam, but braking suddenly could cause the next car coming down the road to rear-end it. Computers deal with their shortcomings through repetition, meaning that if you showed the same pigeon scenario to a self-driving car enough times, it might figure out how to handle it reliably. But it would likely have no idea how to deal with slightly different pigeons flying a slightly different way.
"Computers deal with their shortcomings through repetition, meaning that if you showed the same pigeon scenario to a self-driving car enough times, it might figure out how to handle it reliably. But it would likely have no idea how to deal with slightly different pigeons flying a slightly different way."
Well, no. That's not how machine learning works.
That is indeed one possible training strategy, but it's certainly not one that anyone working on large models for real-world tasks is paying any attention to these days. GP's comment was pretty much accurate, in context.
I'm not a fan of, nor optimistic about, full (level-5) self-driving vehicles. I don't think we'll see AGI anytime soon (whether that makes me pessimistic or optimistic is a subjective question). I find the currently popular architectures for ML – transformer-based LLMs and the like – not particularly scientifically interesting. But I do at least follow enough of the research to know that most of the comments about ML in the Reg forums are ignorant hogwash from people with no idea where the field is now.
If you'd care to display any evidence that you know anything about the contemporary work in the area, I'm sure there are plenty of people who'd be willing to discuss it with you.
Show us you've bothered to learn how transformer architectures work, say, or how EfficientZero works.
Uh, my comment was aimed at Dave314159ggggdffsdds, not you. The whole "repetition is not how ML works" thing, which Jake countered with a short and simple "Actually, yes it does", which you had replied to separately.
(Please also note the sarcasm flag in my comment.)
What little information I have about ML is leads me to the opinion that the current systems that are being sold using it are producing garbage results because they are being fed garbage data with no human oversight to correct or re-train it when it makes a mistake; instead it's being treated as a magical black box that's always right, and I've been around in the IT Industry long enough to know that is a bad idea.
What little information I know about AI in general is mahoosivly outdated by at least 30 years (the early CYC days from the late 80's and 90's, and only from an even more ignorant teenager point of view), so I'm absolutely not qualified to get into a detailed discussion except as someone interested in why companies keep wanting to use these systems and treating it's output as gospel truth without having a solid foundation of how it arrived at the output it generated. (I.E., I'm one of the people who have no idea where the field is not making "ignorant hogwash" comments, mostly about the mis-use of this technology that's causing groups in authority (like law enforcement) to make poor decisions on how to police their communities.)
"The whole "repetition is not how ML works" thing"
I didn't say any such thing. I said that ML does not work anything like as specifically as claimed. It's an absurd claim. If you train a computer to recognise pictures of pigeons, it recognises pictures of pigeons, not just the training pictures. That's the whole point.
Obviously, you could train stuff stupidly. But generally, it's obvious when that is the case.
"instead it's being treated as a magical black box that's always right"
The mistake is that people think that if you feed in enough data that the ML system will sort out what is correct. Since that's not a guarantee, there still needs to be human review and lots of fingers crossed.
"People driving down a city street with a few pigeons pecking away near the median know (a) that the pigeons will fly away as the car approaches"
The same thing happens with what I call Suicide birds that wait until a car is a meter away and dart out of the bushes next to the road and fly across. They same thing happens with squirrels, deer, rabbits and mice. I was plucking those stupid birds out of the grille of the caravan I rented on a trip to Black Rock Desert (Rockets, not Burning Man).
It's typical Musk hyperbole that's led to this. Ask the average person in the street, and they'd probably tell you that Tesla i the leader in this type of technology.
Back in the real world, in this household we have a Tesla and a lowly VW Golf. Contrary to popular opinion, the Golf's driver assistance systems are ahead of those on the Tesla, and I wouldn't even trust those for any hands-free driving.
" the Golf's driver assistance systems are ahead of those on the Tesla, and I wouldn't even trust those for any hands-free driving."
Many auto reviewers find that GM's SuperCruise is much better than Tesla's FSD, but they wouldn't rely on it either.
All of them are just good enough to be very dangerous.
There'll come a point where, say, 50% of cars on the road drive themselves, properly autonomously. And if they're doing it properly autonomously, then the OTHER 50% of vehicles will be responsible for basically all of the crashes. At which point - if not before - there'll come a point where the government will, with some justification and probably public support, make it illegal to drive your own car, on the same basis that it's currently illegal to drive with e.g. flat tires, no lights at night or non-functioning brakes. If your car could provably drive better than you can, why would you risk it?
Now: ten years ago, I said something along the lines of "by the time my kids are old enough to learn to drive, they won't have to bother. By the time THEIR kids are old enough, it'll be illegal". My kids are now four and two... and I no longer think self-driving cars will be ubiquitous enough by the time they're driving. In fact, the rate of progress has slowed so much I no longer think I'll live to see a "proper" self-driving car at all (i.e. I don't think it'll happen for at least something like forty years). I expect to see people walking on Mars before I see proper self-driving cars. Not really sure why, now, that I was so optimistic a decade back, about what is obviously a very hard AI problem.
I hope I'm wrong, because proper self-driving car ubiquity would usher in a golden age of motorcycling. Right now, one thing I learned in a month on a major trauma ward was "don't buy a motorbike", because although a bike didn't put me there, it put most of the other people there, almost without exception because they'd been hit by cars/vans/trucks.
"There'll come a point where, say, 50% of cars on the road drive themselves, properly autonomously."
I suspect that more than 50% of the cars will NEED to be autonomous for it to work and the roads will also have to evolve which is the hard part since they can't be updated "over the air" but need works crews and lots of money to change. There is still large parts of Europe and the UK where the existing roads are based on the width of a horse's backside or more likely a team. Much of the US away from cities can be dirt or gravel roads. Where I live is a mix of paving, dirt and gravel and that just the street that runs in front of my house. Road markings are only hinted at on the paved portions and signage is hit or miss. The upside is that speeding deducts a large portion of life out of a car's suspension and the driver's kidneys.
Everything needs to be working close to perfect for the automation to have a hope of doing its thing. Once a bucket of paint falls off a painters truck or leaks a new line down the road, all bets are off.
The roads emphatically will not have to evolve, because they simply won't. The AI can. Just not very fast.
But *in principle* I'm an objectively terrible driver. I have slightly dodgy eyesight (legal, just, y'know, eyesight), a limited field of vision (like all humans) so I'm required to keep switching my attention around to make sure I've not missed anything, and my attention span is limited. I can be distracted by a kid in the back seat or my phone ringing or an attractive woman on the side of the road or fill in any other possible distractions. I might be drunk. I might be tired. I might be angry. I might have a working class chip on my shoulder and not give two shits about the safety of the posh nob who's decided that they'll trot their fucking horse down the road this morning since they don't need to go to work because of daddy's money, and decided I'll pass them two inches out at sixty miles an hour, causing them to stand up on their stirrups and shake their fist ineffectually at my retreating form, and I might get distracted enjoying that view in the mirror and plough into someone who thought jogging on a country road was a good idea. I might not know how to deal with an icy road - they don't teach that in driving school. The list goes on. Machine drivers will have none of these issues. Right now they have others, but they're all in principle solvable, and auto-drivers are not getting worse, only better. Slowly, admittedly, but I still see the inevitability of the end state - humans banned from driving because they're objectively more dangerous than the machines, even (especially) in the situations where 2022-vintage AI egregiously fails what to a human should be blindingly obvious.
"People still seem to think their fancy cars are fully self-driving"
What do you expect of a population which buys--and at a premium--"Smartphones" because they think that either (a) having one makes them "smart", or (b) the phone itself is actually smart...well, it is, actually: smarter than the owner, to be precise.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
