Alphabet exec says self-driving cars 'have gone through a lot of hype'
cnbc.com
cnbc.com
In that case, I wonder what it says about our present reality. The need to make such gloriously optimistic projections about self driving cars essentially indicates that driving isn't fun to a lot of people (ergo, car makers are making boring cars), or that commuting is too unpleasant (ergo, city planners have been slacking)
VR is good for therapeutic immersive reasons, for example [1]. And 3D printing, well, shooting rockets into the stratosphere, for example [2].
IMO VR and 3D printing are real, albeit niche. Self-driving cars aren't fully there yet.
[1] https://www.geekwire.com/2018/snowworld-melts-away-pain-burn...
e.g., my experience with consumer-grade 3D printing is that it still kind of sucks. It makes home fabrication marginally more convenient than it was without 3D printing, and a lot more aggravating. Even with the rise of gig economy 3D printing services, which have made me a lot happier by removing the hassle of actually operating a 3D printer and giving me access to much better equipment than I'd ever buy for myself, I find that the best executed prints that I order still require significant manual finishing time. It's still just an extremely far cry from the "You'll never need stuff shipped to you because you can just print it at home" hype that surrounded the technology 5 or 10 years back.
VR, again, was posited as the end of "reality" itself, that we'd be living so much in our virtual worlds that the boundaries between the real and the virtual will cease.
While these might still happen, it's safe to say that their current applications are much narrower than the original scope
And now I remember why that is :P
[1] https://www.superdataresearch.com/superdata-xr-update/
[2] https://news.lockheedmartin.com/2018-07-11-Giant-Satellite-F...
I know that the topic at hand is about L5 fully autonomous, but wanted to share that though robotaxis are still some ways out, big benefits to personal driving is already here. From here things will rapidly evolve towards the ideal state.
An important concept pointed out in these conversations before is that human-in-the-loop driving is a moving target. That is, it's possible we may never get to the point of fully automated driving because we may discover that the safest method is humans doing some small fraction, and automated systems doing the rest. Since we have partial automation now, and will continue to in the future, we're not asking "When will automated systems replace human drivers?" but "When will fully automated systems replace humans using partially automated systems?" Because we're constantly improving the partially automated systems, we're constantly raising the bar that a fully automated system needs to meet.
If there is no car in front of you, stop and start at lights and intersections. Turn (while with FSD package it takes highway ramps automatically, it’s not reliable on any complex highway intersection). In heavy traffic, you have to change lanes manually, automatic system is not aggressive enough.
I'm personally most interested in the trucking example, where trucks can form a caravan with a single driver in the front.
In 2017 Google said the cars would be available in months according to the article.
Last year Elon Musk said there would be 1mm robo taxis by the end of 2020.
These corporations leverage the public's lack of understanding to spin a tale to investors and the market. Then, this tale got co-opted by politicians like Andrew Yang who say 4 million truckers will be out of work in 5-10 years, so we need to vote for him + UBI to protect us from the threat of AI in the future.
The view I take is that planes have pilots for a variety of reasons, many of which are not related directly to controlling a plane's path. We have 30 years of data on flight and flights have fewer inputs to analyze, yet we still keep pilots around because of the added sense of security and the utility pilots serve on edge cases.
I think self driving cars will be very similar and the future where no one drives probably will not exist for a long time.
Another thing I find interesting, the retort most bring to driverless cars is that they will prevent needless deaths, but say for example America has 40,000 deaths due to car crashes per year, and 20,000 are preventable, but autonomous vehicles malfunctioning causes 10,000 deaths. So a net result of 30,000 deaths with autonomous cars. In terms of death the result is positive, but now you are relinquishing control over your experience. Someone with excellent driving skills, who never drinks and drives, and who follows the speed limit, faces a way lower risk of death driving. However, self driving cars introduce new risks to that individual, and for that reason I do not I believe the optimal combination for drivers will be the same as for pilots: machinery + individual control.
So in that sense, they didn't lie, but the scale of the roll out has been much smaller and slower than what most people expected. I don't think Google specifically made comments about the scale. Elon on the otherhand did mention big numbers, which is a different story.
Which worse case takes out a minivan with 6 people in it? (Unlikely)
Why would we need truck drivers in a world of self driving cars?
Trucker drivers don't do these things and the payloads are not equivalent. However, truck drivers will still need to oversee the successful docking and unloading of goods, the refilling of gas, mechanical checks, special goods like diary product or gas, and the recipient of the goods.
My point was not simply that pilots are there as added security, but also that they serve secondary functions that provide utility from start to finish. In this sense, yes they are similar.
Faceless lives can be easy to judge, after all six people is less than 200, and given a median gdp per capita, 350k is less than 12mm. However, the moral hazard I am talking about is that you remove the responsibility from the individual to a faceless corporation. Btw you're not comparing apples to apples anyways, sure a truck may only kill 6 people, but if 100 trucks per day are doing that then that would be worse.
But I think you could hail a (mostly) conventional car (like a Tesla, but with Waymo's sensors) to most locations reliably. Even if you had to be behind the wheel to take control in some situations, but you could still rent the car by the minute/mile and take it wherever, that would still be 10-1000x better than renting a car conventionally, and it'd be 2-10x cheaper than Ubering.
I think the majority of people would pick driving themselves over Uber for most fo their trips if it meant half to 1/10th the cost. They could easily make a 20% profit on $0.55 a mile + $3.00 to start. Your average trip in NYC and SF is now ~$5 instead of ~$15.
I'm not sure why no one is pursuing this option. I know Google REALLY wanted fully autonomous cars. But this seems like it could just be another option in Uber/Lyft.
I would go as far as saying that we should be exploring this on the Municipal level. I wouldn't be surprised if 30%+ of car owners in major cities would get rid of their car if a reliable service at this price point existed.
What is that, a robotaxi for ants?
We’ve done the same to AI, what someone called AI a decade ago has been reinvented as AGI. We are doing the same to self-driving cars. “What we really meant by self-driving is self-driving in perfect conditions and normal intersections. You’re thinking of autonomous self-driving or Level X self driving.” All I’ve seen is driver assist tech improve.
That's because some people should think harder about their comparisons before they verbalize them. The quoted comment poo-pooed an existing product that you could buy with a mere trip to one's local Apple retailer.
Poo-pooing the hype of something that hasn't even been invented yet? To take take the opposite side of that argument with any fervor strikes me as arguing for the sake of it. I mean, until I can ride in one, we could go on all day about when they're likely to appear, couldn't we?
cmdrtaco was also criticizing the product as it existed at the time, not the product it would evolve into which, in retrospect, made his criticism seem ridiculous. He was criticizing the first revision which had poor connectivity (firewire only) and little storage (5GB on launch, 10GB later.) The fourth revision, the one that actually became a run-away success, had 20-60GB of storage and was the first to have full USB support.
https://en.wikipedia.org/wiki/IPod#/media/File:Ipod_sales_pe...
Suffice to say, I am almost certain that his daughter will be learning to drive very soon, and required to drive for years to come.
Everybody agrees we'll figure out autonomous cars sooner rather than later, but nobody thought a 2020 timetable was going to happen. It's undeniable that we've made huge advances since the first DARPA challenge, but there's a long way to go before we get them to be sufficiently robust.
Also true for other things such robot assistants etc.
My understanding is its not the sensor or control technology to blame for the pace of progress which makes me think AI isn't as capable of learning the extremely complicated task of driving as expected.
Anecdotal: I've noticed a lot of talk about ML/AI at my place of work is very hand wavey and talked about by business analysts who don't know what they're talking about. My friends at other places of work see this too.
I wouldn't be surprised if there was a large-scale loss of faith in AI simply because too many people tried to half-assedly leverage "using AI" for a promotion.
I feel that even the names ‘Artificial Intelligence’ and ‘Machine Learning’ were chosen as a marketing ploy.
My personal guess is that we can probably get to hands-off self-driving on limited access highways, possibly under a subset of conditions, sooner-ish rather than later. Which would actually be a huge win for those who do long highway drives or daily commutes on highways. But it's sort of a bucket of cold water over the head for those urbanites who bought the promise of never having to learn to drive, buy a car, etc.
I'm somewhat confident that there are still enough wins around ML that the whole thing won't peter out. But it may end up being less of a generalized revolution than some thought (or at least claimed they thought).
Self-driving cars are foundering because of the same problem that always sinks things that are sold as AI: We've still got zero idea how to program a computer to introspect and reason independently. Not even the foggiest. So we invariably fall back to an unmaintainable rat's nest of hand-coded rules when the time comes to implement the very core of the system. And then distract people by hyping up whatever fancy new technology (Stacks of NVidia GPUs let us fit hypersurfaces in lots more dimensions than we could on a single Xeon!. . . err, that's not marketable, let's call it "deep learning" instead.) is being plugged into that core.
The deeper issue is that people hear the term AI and invariably think "artificial general intelligence". You can try and walk that back all you want with disclaimer upon disclaimer, but it's kind of silly to even try. If you're describing something as Artificial Intelligence, of course people are going to think that you're claiming to have built something that is (1) Artificial, and (2) Intelligent. And, let's be honest, that's probably what they were trying to imply all along. If they wanted to describe the technology as "really really amped up parameter estimation", they'd have stuck with the term we already have for that: Machine Learning.
You are right, control is not the issue. Sensing is or isn't the issue depending on what you are talking about. On one hand, sensors are the best they've ever been and you could get a car to drive around a closed course in a robust enough manner. However, there's a lot going on in the world around a car in the wild. We can identify obstacles like other cars pretty robustly, but it is still a sticking point because small mistakes cost lives (see that early Tesla fatality that thought a truck was a billboard).
I think an AI winter is pretty inevitable at this point because AI has become so overhyped. It won't be as bad as previous winters though, because modern AI techniques have had a pretty major impact on other fields. Computer vision, for example, will never really be the same again. Moreover, for something like a self-driving car, ML will certainly play a big roll. Even though getting the computer vision models to work is incredibly tricky, I can't fathom how you would go about doing it with "conventional" algorithms.
The self-driving car (or "auto-auto") is a kind of sci-fi fetish.
What I want, and what could be built today I think, is a self-driving golf cart, covered in nerf and airbags, that doesn't go any faster than five or ten mph. It can do things like take seniors to and from the pharmacy or doctor appointments safely.
First, there's the self-driving car for the dense urban environment, which is basically a robo-taxi. You don't buy or own these cars, you hail them and they take you where you need to go. Basically mass transit with the option to go wherever you need instead of pre-defined stops, combining the best aspects of a taxi and a bus. There's no reason these need to be cars at all, aside from the fact that we have all the infrastructure for cars. They could just as well be small trains/"horizontal elevators", the tubes from Futurama, or whatever else your local sci-fi author can come up with. These types of vehicles could be centrally controlled and you could indeed replace complex AI with expensive infrastructure, but that is something we'd like to avoid (who knows if it's even politically possible), hence the push for using existing infrastructure. As an added bonus, these types of fully self-driving vehicles can take over tasks like delivery.
The second type of self-driving car is the car for every other environment, basically the same cars we have now with an autopilot add-on. These vehicle would be for people who live in small towns, suburbs, rural areas or in general travel outside of dense cities. It makes more sense to own vehicles like these, but they must cope with a larger variety of environments. Personally, I would never buy a car that does not have some form of manual override so I could drive it myself. If the AI is advanced enough this manual override doesn't have to be a steering wheel, but for a car I own, having the ability to violate the law, crash it or tell it to go someplace it doesn't want to (or, god-forbid, run someone over intentionally) is essential. I suspect that making an AI that can cope with all kinds of driving conditions in the country is a task so monumentally complex that cars (the kind you own) without manual override mode will basically never be a thing, both because it would be expensive and because I don't think the kind of people who would need to buy a car to own would accept not having full control. A couple of instances of the AI failing in some basic way (failure to drive into a small garage, for example) would be enough to convince people that a manual mode is essential.
What percentage of cars in the future will be of the first type and which will be of the second type is up for debate, but regardless I think the second type of car will be something you buy out of necessity or a luxury good, and overall car ownership will go down.
However, with technologies like self-driving electric delivery vans, telepresence, improved solar power and battery storage as well as better internet coverage, it's not obvious to me that the future lies in dense cities. I speculate that a certain tipping point will come in the near future where less and less people will need to live in a large city for work and migrate to small and medium sized cities for better living conditions and larger homes. As technology advances, it will become more and more comfortable (and possible) to live in a less dense environment and population will spread out. I'll admit that this involves a bit of wishful thinking since I hate big cities, but the alternative where population keeps concentrating in urban areas seems like a terrible dystopia to me.
That’s a big competitive advantage compared to Waymo’s simulations and paid drivers.
At least Waymo usually gets rearended by other cars as a result of being timid (except for the time they ran into a bus by being courageous and assuming the bus would move).
I actually thought for a moment that 'gore point' was a dark joke about places that Teslas tend to hit with their autopilots.
Waymo, as a privately-held company, has no strong incentive to deliver, and it doesn’t have strong delivery plan in the first place.
Waymo’s sensor package is an disadvantage too - apart from being expensive, in real world most likely it will be too finicky. You could have argued in 2000s that with current technology it’s not possible to do world modeling without Lidar, in 2020 you can do it with video cameras.
https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...