Steve Wozniak: 'I've given up' on Level 5 | Hacker News Reader

Steve Wozniak: 'I've given up' on Level 5

europe.autonews.com

60 pointsclouddrover89 comments

> Autonomous vehicles would fare better, he said, "if we were to modify roads and have certain sections that are well mapped and kept clean of refuse, and nothing unusual happens and there's no road work."

He's talking about trains and other modes of public transit. Which makes sense. Of course, even those have occasional problems.

Personally, I'd love to simply find a way to make "mass transit" more workable.

I wonder if anyone’s considered making a kind of modular way for vehicles to connect to something like a train.

I can think of two examples.

In LA there are busways that are built on former train paths. When they need to divert for whatever reason (road repairs, special event routes, etc) they can just make a turn and act like any other bus. Something like the Hyperloop car carrier could work here or at the very least an expansion of bus ways could provide train like access with a lower infrastructure cost.

Amtrak will also let a certain number of vehicles onto some of their train lines. I don’t know how well that works, but it seems like for a well run popular route you could have a ferry like experience.

Similar to the LA example -- the O-Bahn in Adelaide: https://en.wikipedia.org/wiki/O-Bahn_Busway

Not being sarcastic but cars can go “on” trains.... checkout EuroTunnel for example

like some sort of train that a car could drive onto and be transported?

This has existed for years.

More crazy. Something they either create by linking up to each other or drive onto but can drive off of at speed and any time. Needs to be immediate to join and leave or people won’t use it.

Or it has its own lane, and if you don’t use it, you sit in a traffic jam watching the people who do use it whizz past.

Not everything has to be instant gratification catering to your every need or else you’re buying a helicopter :-)

Road trains don't really give you much savings from a traffic jam unless there is something else that prevents most people from using it. Single passenger cars take up a lot of space. Getting cars/people out of the train will slow the whole thing down where they need to leave the "fast?" lane for the traffic jam to get to their exit.

Unless you are talking about all new freeways usable only by road trains. However at that point putting down steel rail is cheaper and can handle an order of magnitude more people for the investment.

Road trains could give us greater efficiencies from aerodynamics, and if used in dedicated lanes, could also give us something almost as useful as autonomous driving with much less work.

As for steel rail, gosh, I really wish there was a way to make the politics work in North America.

While the theory of road trains is sound, in practice they cannot work as well because of other theory. Getting into/out of a road train is hard. If there is a traffic jam in the other lane, the entire lane is limited to the traffic jam.

Road trains save some space, but they don't solve traffic jams (which already have cars bumper to bumper). There is no way around the fact that cars need a lot of lane space to work well, even in road train model. (steel rail solves that by packing humans tighter, while giving them more space)

Bike and scooter racks on a train. Rental kiosks at stations. This is a solved problem, it just requires investment.

We have those; they’re called parking lots and they’re usually found at commuter train stations.

Snark aside, I don’t think it makes much sense to try to merge cars together at high speed. It’s a very complicated maneuver, requiring all cars behind the intended merge point to back up and make room, and the consequences of an accident have unlimited catastrophic potential. We could see hundreds of cars destroyed in a massive collision.

Besides that, the costs seem astronomical: actually developing the technology, making sure all car manufacturers comply with the standard, and replacing every single car on the road with a compliant model. Wouldn’t it be much easier to put all of that money into building high speed railways? We’ve already perfected that technology.

The biggest issue is trust. I maintain my own cars: I trust myself, but do you trust me? I don't trust some of the professional mechanics that work on cars. In the current scheme we are told to maintain a large enough following distance that this question don't matter: no matter what happens in front of you, you can still react in time to recover. (You will not that the vast majority of drivers do not maintain this following distance) In a raid train situation all cars must be maintained to high enough standards for the most risk adverse person.

In trains we already have standards, and have developed schemes to ensure that cars are maintained to standards.

Note that trust above needs to account for the terrorist trying to kill people. It is debatable what systems works.

Docking wouldn't have to occur at high speed, or moving at all, for this to be useful. For example, I live in Massachusetts and I have a car full of gear that I'd like to have for my vacation in New Hampshire or New Jersey or Michigan. So I drive down to the local train depot, where my car is docked (or even just loaded onto) a train, I get in the passenger compartment, we're reunited at the other end. Already more efficient than driving, and more pleasant too. Now start adding tweaks like loading only the passenger compartment, which is already a separate thing in electric vehicles. Or running rails along one lane of existing highway, with docking/undocking at existing rest areas. Once the basic "dock instead of driving" idea gets established, it's possible to iterate on it.

I recall Popular-Science-ish illustrations of systems like that where you’d drive onto the freeway and your car would hook into a chain of cars to be driven to a “stop,” where you could unhook.

If I understand it correctly, we are just now experimenting with something very like that for trucking, where multiple trucks can convoy, and the systems communicate with each other so that all the vehicles in a convoy adjust their speed simultaneously.

Wasn't this part of the Hyperloop project idea in LA? You drive to a specific parking spot, the vehicle gets lowered to a tunnel, then the tunnel takes you to wherever you need to go.

I think this is a really neat idea. Imagine cars arranging themselves and coupling into a train and then seamlessly decoupling as they need to exit. That would save energy (air resistance) and be a more efficient way to move more cars down the same piece of road faster. It also removes the lost time of intermodal transportation.

No idea if it's practical.

If you need to rely on the customer vehicles themselves for any part of the train's operation (wheels touching the ground, coupling, electrical connections...), it's going to be extremely difficult. Railways have strict requirements for inspections, etc. of customer cars—one malfunctioning car can be catastrophic (if a coupler malfunctions and lets go, for instance).

If you're not relying on the customer vehicles for any part of the operation, you're basically describing existing auto-rail operations. They're not extremely common, probably because they're a terrible waste of space compared to traditional passenger rail cars.

Our goal is to move people and goods around—not extra vehicles.

> Personally, I'd love to simply find a way to make "mass transit" more workable.

It's very workable already. Give it priority and separate lanes where possible, consider giving them exclusive use of roads in some areas, prioritise frequency over speed, maintain the vehicles regularly and replace them eventually. Finally, provide dedicated on-demand transport for people with special needs. The end. It works.

> > Personally, I'd love to simply find a way to make "mass transit" more workable.

> It's very workable already. Give it priority and separate lanes where possible, consider giving them exclusive use of roads in some areas, prioritise frequency over speed, maintain the vehicles regularly and replace them eventually. Finally, provide dedicated on-demand transport for people with special needs. The end. It works.

Many of the things you list are missing in mass transit systems across the US. That deficit being filled would make the system workable for many more people. It's s likely that workable was a poor choice of words and a more apt statement would be something along the lines of, "Mass transit as it is currently deployed across the United States requires a number of fixes to make it a more viable daily alternative."

One more important aspect: cost. In my country mass transit is bizarrely more expensive than private.

I didn't read anything in the article to suggest he's talking about trains here. It looks like a reference to dedicated lanes, etc, for autonomous cars.

> I didn't read anything in the article to suggest he's talking about trains here. It looks like a reference to dedicated lanes, etc, for autonomous cars.

The parent likely means that the modifications required are the trappings of trains.

That is because he isn't visionary enough to recognize what he is talking about. Everybody has their blind spot I guess. (Come to think of it I wouldn't have realized this myself without being told)

Mass transit would be more workable if we simply built more and our autonomous vehicles could drop us at the station then return home.

Indeed this is really one of the best solutions. Light rail, buses, etc. taking people to and from businesses, restaurants, shopping, etc. and have autonomous vehicles traveling from residential areas to and from stations.

Unfortunately, when most people talk about autonomous vehicles they want personal autonomous vehicles. I really don't see the point of this if I'm not actually going to be driving. The fun of a vehicle is driving it, not riding in it. If it is autonomous, I don't want to own it and pay the upkeep.

Personal autonomous vehicles make sense for anyone who will do the majority of their daily travel in an autonomous vehicle. If you can take transit for most daily trips the vehicle is for things like getting a load of lumber that transit can never serve - these are rare enough that renting (at $100/day) is a better deal for most people.

If it is a daily trip, there isn't that much opportunity for car reuse. Rush hour is call an hour for a reason: most people travel at about the same time. Those who are traveling in a different part of the day tend to have different patterns such that they would need a different car (it needs car seats for the kids...)

If you have a luxury car, shared might make sense because you don't care about upkeep costs. However if you are cost conscience a 3 year old car that you keep for ~10 years is well known to be the cheapest way to go. These economics won't change: people who want a luxury car will demand the shared car be fairly new and well maintained. People who want the cheapest will still buy their own because they can guarantee the cheaper ride.

Don't overlook the advantage of knowing your golf clubs are in the trunk waiting just in case you can take an afternoon off for a round. Shared cars won't give you this, if anything like it is for you.

I was thinking more of a standardized type of super-regulated road that was only accessible for self-driving cars that could know about its status (road works, accidents etc) using a standardized API. Only vehicles certified to use the API correctly could enter it.

E.g. on normal roads a Tesla would drive as it does now with some driver assistance, but there would be special stretches here and there for only self driving cars where they could be autonomous.

I don't understand this obsession with achieving level-5. What would deliver tremendous value for many people, is simply the ability to drive autonomously on highways. Highway driving is so radically simpler, and can be achieved in a fraction of the time. In addition, time spent on the highway is the majority of people's time driving. Being able to sleep or read a book for 30-60 minutes everyday, while commuting to work on the highway, would be a tremendous boon. Being able to sleep overnight while your car drives you to your parents' city 500 miles away, would be a godsend. It would be great if we could focus on making highway-autonomous cars mainstream, before worrying about level 5.

I agree with this - I think the way to think about it is not to think of the self driving car as a car at all, but rather a train / bus replacement. It takes the work out of covering the long distances, but you have to solve the last mile problem yourself with walking, taxis, local buses, whatever. The game changer is that you get this whilst being able to pick your own schedule, and none of the other downsides of shared public transport which are numerous.

The lives of the disabled and the impartially abled would be made astoundingly better with such level 5 cars. Greater freedom of movement for those with illnesses, mental disabilities, vision impairments, amputations, etc would greatly improve their lives and thereby the lives of their caretakers and loved ones too. No longer would your legally blind uncle need your help to get to the doctor or to the store.

It's not obvious, but there are medical, familial, and personal liberty implications to such self-driving cars too. I think that these implications should be better championed by their engineers. Such cars can be wheelchairs for the blind, etc. It would hit an emotional nerve better than the meme of people sleeping on their way to work.

Level 5 will be a very very big deal.

In the politest way possible to Mr. Wozniak - this reminded me of the first of Clarke's three laws:

> When a distinguished but elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.

https://en.wikipedia.org/wiki/Clarke%27s_three_laws

He doesn't say it's impossible, just that it's not imminent.

He says in the linked interview:

> "I don't even know if that will happen in my lifetime."

Not that I wish for Woz's death in any way, but he is 69 years old. The average male life expectancy in the US is 78. Now, he's wealthy, so I would expect him to probably live longer than the average, but 15-20 years would be doing good.

15-20 years is also my personal expectation for when we will start seeing truly autonomous self-driving vehicles.

Moreover, I think it will be incremental growth, not immediate massive availability. I would expect the technology to be adopted in certain controlled situations first, and in places where it makes economic sense. And considering how long it can take for people to replace cars and for the technology to filter down to more economic models, 40+ years before the majority of cars are self-driving does not seem like an unreasonable estimate.

> 15-20 years is also my personal expectation for when we will start seeing truly autonomous self-driving vehicles.

Came here to say the same things. Woz is awesome, but not a Spring Chicken. I'm 42, and I tell my 13 son I expect him to see it in his lifetime. If I get it when I'm old and can't drive anymore, I'll be really happy. And if the industry beats expectations, all the better!

Same. I'm 38, and I've told my 7 year old that I would expect that her grandchildren's generation probably won't need drivers' licenses. I think that period from about 20 years from now until then will be a somewhat uncomfortable overlap period where people will still need to know how to drive, but that computers will be doing more and more of the actual day-to-day driving.

The only thing I could see that would shorten the interval is legal approaches. Such as, if the technology proves sufficiently adept, for it to be declared as a mandatory safety feature on par with seatbelts and airbags. But even then there will still be non-autonomous cars on the road (just as there are still cars on the road without seatbelts and airbags) for many, many years. Again, unless you take the extra legal step and ban them from the road.

That doesn't mean it isn't worth doing. Something that we fail at a lot in the modern world is thinking long-term, beyond our own lifetimes. Just because something might not get done during my lifetime doesn't mean that we shouldn't do it, just that others may have to finish the work we started. Ultimately, society will be better for the investment we are making now.

This on the same day Waymo cars seem to be roaming around entirely with nobody in or around them on public streets[1].

[1]: https://www.youtube.com/watch?v=t_1YLHW65M8

Honestly, the Yandex self driving experiments seem much more credible than what Waymo is doing.

Waymo is testing its cars in a situation without any snow, rain, mostly without street parking, without pedestrians, in a city so hostile to pedestrians no one’s walking around.

99% of the work isn’t in getting the self driving to work in perfectly controlled sunshine conditions as Waymo has it working right now, 99% of the work is in getting all those edge cases working.

It’s gonna be decades before waymo cars can handle small mixed use streets with children playing on them, pedestrians, cyclists, and street vendors in a european city.

What they’re doing is amazing given the tech they have — but it’s not enough to reach actual level 5. It’s the same issue why most self-driving solutions only work on highways.

  It’s gonna be decades before...

Citation needed.

We can very accurately identify children, pedestrians, cyclists, and street vendors with machine vision. And with the LIDAR that Waymo is using, very accurately tell where they are and more importantly where they're going. Stopping for those things is pretty low on the "problems that need more work for Level 5" agenda.

That's not something you'll find a citation for. As someone who has worked in the industry it's not as simple as "see child; stop". What's also happening behind you? Beside of you? Do you swerve? How quickly do you stop?

You're very right that OD works great, that doesn't mean you can train a neural network to make the right decision though.

You barely even need NN for this - it's basically pathfinding.

Find the path to where you want to go, avoid the parts with negative weights.

Great idea, maybe you should call Waymo.

It's not a great idea because it's already done.

Again, the hard part is not "here is a human, avoid it". That part is done (in Waymo's case).

I'm afraid asking for a citation for a forward looking speculative statement doesn't come across very well.

But making wild speculation without any evidence comes across better?

Yes, it's just an opinion stated politely and unagressively. Those are valuable qualities that I appreciate.

Except OC was stating it more like fact than opinion, which is exactly why I took issue with it.

I've got no objection to you taking issue with it, you're entitled to your opinion too of course. We're here to talk. It just would have been nice if you had done so politely and unaggressively.

Identifying people is easy, the much harder part is finding the right path to take as result. In a dense european street you’ll have pedestrians, cyclists, and so on all around you.

Imagine driving slowly through a large gathering of people, slowly having them move around your car, etc. Such situations will necessarily become major issues.

Especially if you imagine a snowy dense downtown street, with pedestrians crossing everywhere and snow all around you making the LIDAR useless.

I agree that Waymo will have lots of challenges reaching world-scale (cost, training, etc.), but what they've achieved is undeniable. It seems like it would work extremely well for anyone living in that area.

I see them as providing an invaluable service to the disabled and the elderly who live in appropriate cities. I would move there to have such luxury. But it s not a generic solution for all driving conditions

I've been skeptical for a while now, perhaps overly so, about the prospects for "real" autonomous driving, i.e outside of Californian highways.

I wonder when the public / the press is going to start question the story about full self driving coming "soon".

AFAIK Musk is still saying that the autonomous Tesla ride sharing service is planned for mid 2020, and that full self driving will be at least beta tested by end of this year.

But if these promises are not delivered on, someone ought to start questioning how fully autonomous cars really are progressing.

Hm, "self-driving car roads"-exclusive roads or lanes don't seem like a bad idea, [edit] they could really help in adoption of self-driving cars, and cars supporting automatic chaining into a "train"-like construct[/edit]. Except maybe that roads are expensive and require plenty of space; after all, cars are a highly inefficient mode of transportation :(

It is really interesting to think about how arbitrarily difficult the problem of self driving cars becomes when they must be backwards compatible with existing infrastructure and co-exist with it for quite a long period of time, decades probably.

As Wozniak alludes to here, it might just be intractable (within a few-decades timeframe, probably not on a longer one) without concessions to adjusting the infrastructure to make the problem space smaller and solutions simpler and more practical.

Banning or limiting human drivers on certain roads, repainting surfaces with thick white lines for tracing, automated checking for obstacles and cleaning, sensors along the side of highways, networking self driving vehicles, reducing speed limits -- all things that could vastly reduce the complexity but doing them or getting people to want to spend money on them might be very hard.

I m seeing this as a general trend in investment: they seem to have doubled down on certain moonshots such as robot cars, space travel, quantum computing, virtual reality , lab grown meat to the detriment of more reachable tech such as CRISPR or tele-work or even fusion

> There is simply too much unpredictability on roads, he said, for a self-driving car to manage.

What does an autonomous vehicle need to do in response to a highly unpredicted case on the road? Stop.

Can an AI offer a confidence score as to how well it understands a situation? Yes, that's how they work.

Can an autonomous vehicle brake to a stop in response to an unrecognised situation? Based on the above, yes.

I mean, I wonder how hard Woz has been trying to create an autonomous vehicle prior to announcing that he's given up. Or is he just trying to tear down those that are actually turning this into a reality?

Stopping can be just as dangerous, from causing accidents to blocking an ambulance.

It wouldn't really happen all that often though.

First, not a dig at Woz, but he's 69 and average male US life expectancy is 78.7. Saying he won't see level 5 in his lifetime is not a very bold claim.

Second, I don't understand people's fixation on Level 5 in particular. There are vast numbers of human drivers who cannot meet that criteria, perhaps even the majority of drivers. Despite this, they drive (in situations they feel comfortable), and have a profound impact on how society is structured. Woz will probably see both level 3 and level 4 cars and they'll have similarly profound impacts.

> First, not a dig at Woz, but he's 69 and average male US life expectancy is 78.7

You need to add about 6 years, because he's 69. As you get older, your life expectancy goes up. For an average US male at 69, they can expect 15.09 more years.

> "I don't even know if that will happen in my lifetime."

The man is almost 70, so that's not that unlikely.

Tbf AI hype and failure like this has been around for long enough for even young people to confidently make that assertion.

It is not necessary to be able to deal with unanticipated situations, only to be able to recognise them as such. For self driving taxis, a remote human operator could then intervene.

And what if the connection is down?

Pull off and top, same as would happen if the car suffered a mechanical failure.

What do you think the typical delay would be between identifying an unknown situation, and a remote operator assessing the situation and taking positive control of the vehicle?

If your answer is more than a small fraction of a second, I have some concerns.

Also the remote human operator doesn’t have the same incentives to avoid an accident.

Ostensibly, their career as a human operator?

I refer you to the human operator of an Uber self driving car, who was playing on her phone while the car ran down a pedestrian crossing the road.

You can relax the time requirement if the vehicle stops on unexpected conditions. If it finds roadworks and successfully identifies it as unknown, it stops and requests manual operation for traversing the obstacle.

If it gets out of the way and stops, fine. If it just says “screw it, halting right here”, it’s going to quickly find itself the subject of public outrage and be banned until resolved I’d think. (I think figuring out a possibly creative way to "get out of the way" might be approximately as hard as the rest of the self-driving challenge.)

Imagine every hundredth car during rush hour detour just stopping for no human driver valid reason.

Imagine areas where you’ve driven where you can’t even keep a voice call connected, let alone the data rates needed for remote piloting assistance.

Imagine Boston’s Big Dig and these not-quite-autonomous cars facing an unknown situation en masse inside a tunnel under water...

I’m 2/3rds of Woz’s age and don’t expect level 5 in my or my kids’ lifetime. Level 5 is a very high hurdle.

Many of these are dynamic situations, like some cargo falling off a trailer, or a deer cumping in your path. Either you react right away, or you are dead. Operator is not going to help.

May I remind you are are at a stage where Tesla cars drove into a trailer and into a concrete wall, a mistake no sober human would make.

Okay, but what if the unanticipated situation only gives you 1.2 seconds to react? Even if the car switched control to a remote operator immediately, there's a lot of context required to navigate driving situations (who is around me, what are the road conditions, am I merging, etc.) that would be hard to absorb in a split second.

And what if the car is in a place with a poor network connection?

I'm not saying your idea is untenable; there are just a lot of corner cases that would also make it a very complex challenge.

Being able to handle those new situations is equivalent to "Level 5". You seem to be making the point that Level 4 is enough, but that's not disagreeing with the argument that we won't reach Level 5.

Having little to no responsibility for long periods of time would result in a driver filling the time with other activities, like phone conversations. A frightening number of cabbies blabber on the phone already. It's doubtful their reaction time to take control would be quick enough.

We should be looking at occupations that require people being alert but not engaged for a glimpse into drivers' behavior.

You must not play many online competitive games. I have played them for most of my life and I can tell you that latency of even just 100ms will absolutely wreck your ability to respond and react to dynamic, high speed situations with accuracy. “Save the confused autonomous agent about to plow into some pedestrians” is not a game I want to play at anything slower than the speed of my hands on the real steering wheel.

If you have enough training data cars encountering “new” situations should be very rare. I don’t think it’s a deal breaker for level 5 autonomy.

Problem is that it is not very good pr if somebody dies or hurts badly every time cars encounter a "new" situation.

Do we not already have loss of life when someone encounters a sudden unpredictable situation? We may be able to learn on the spot from new data, but an automated vehicle can react orders of magnitude more quickly.

You have individual agents with actual intelligence that vary by skill.

The human regressions are pretty well defined and correctable: impairment, fatigue, distraction, vehicle condition.

“Very rare” doesn’t cut it for modern safety standards.

It does if it's a lot better than human drivers.

If your new standard is safer than the existing "standard", nobody says "well actually that's not enough of an improvement".

Sure it does. Planes crashing is very rare but it does happen.

'Very rare' doesn't mean anything.

Very rarely an asteroid falls on your head, how are you dealing with that?

The level 5 of autonomy is defined by the fact that the car can handle those new and extreme situations. I think that what you mention would match level 4 autonomy.

For example, definitions from https://www.techrepublic.com/article/autonomous-driving-leve...:

> Level 4: This is what is meant by "fully autonomous." Level 4 vehicles are "designed to perform all safety-critical driving functions and monitor roadway conditions for an entire trip." However, it's important to note that this is limited to the "operational design domain (ODD)" of the vehicle--meaning it does not cover every driving scenario.

> Level 5: This refers to a fully-autonomous system that expects the vehicle's performance to equal that of a human driver, in every driving scenario--including extreme environments like dirt roads that are unlikely to be navigated by driverless vehicles in the near future.

In my experience with machine learning, the last few percent of weird cases are exponentially varied and change over time so are always new. You hit a performance wall at a certain level. Researchers have been driven mad by this wall. See for example https://www.wired.com/2008/01/ff-aimystery/

For cars, behaving dangerously just 0.1% of the time is way too much.

A good AI should have really good 'one shot' deduction capability not just work by induction using tons of data. I don't think the current crop of algorithms are there yet.

Even humans have difficulty dealing with weird cases. Car accidents happen for many reasons, including the driver being unable to process the incoming information in the given time.

One of my pet peeves is the lack of raised markings (Botts' dots or other types) on highways. With the right combination of wet streets and lighting, all of the lines on a roadway can seem to disappear. A machine with different sensors may actually be able to detect the lines better than a human could in this particular situation.

Reply on news.ycombinator.com