Hacker News Reader

Comment by minimaxir | Hacker News Reader

Funny in hindsight from 2016: "Lyft's president says 'majority' of rides will be in self-driving cars by 2021"

https://www.theverge.com/2016/9/18/12944506/lyft-self-drivin...

https://medium.com/@johnzimmer/the-third-transportation-revo...

I saw Levandowski do a presentation of self-driving at an Uber all-hands and it was then and there that I knew that self-driving is 20+ years away.

The example video showed so many small things that we take for granted as a human driver that would need to be built in or hand coded that AI or neural nets would never be able to get. There just isn’t a way that you can train every single little thing that could cause an accident. For some things you just need experience and fear which I imagine can’t be modeled via any current AI.

I think at a minimum all the cars need to be taking to each other over a protocol.

However that has to be secure that a hacker can't simply stop all traffic or cause speeding

Which is pretty much impossible if you let people have the physical cars. They will report their speed to other cars, they will measure their speed from pulses generated by something rotating on the vehicle, replace rotating thing with your function generator and then you can tell other cars you are going any speed you want. Turn up your reported speed and watch all the other cars get out of your way!

Well, you can take the input as a data point, but make the final decision based on sensor data. Thus the shared data can make the flow smoother.

But I'm also waiting for fake street signs to troll/manipulate self driving vehicles :)

It's easy to put up fake street signs to troll/manipulate human drivers, yet that hardly ever happens. I have read stories of people putting up fake speed cameras in their front yard to slow down traffic, but no one dares to fake speed limit signs.

Except Finland during elections. Where advertisement with nice white round circles with numbers inside are a standard. Single vote for candidate system where number of candidates can run to over hundred.

No human would think that number next to person is speed limit, but machines aren't trained for that...

Here in Gemrany I there are signs with "voluntarily 30" or similar combined with some surrounding images etc. and yes one can fake it, but most fakes can be identified realtively easily by humans (positioning, color, material (non-reflecting cardboard), ...) some high quality fake sign might however exist for a while.

Also there is the reverse: people stealing signs (https://news.ycombinator.com/item?id=25223633) or putting stickers on them (or snow, mud, ...) where most humans know what to expect (stop sign, yield, ... have specific shape for a purpose; empty frame where city starts; ...)

Going out of a somewhat controlled small area in SV/SFO or AZ is a huge task. Going to a different country requires complete re-learning of the models ... and then there are humans ...

I tried “tesla swatting” with a fake speed sign a couple of years ago. It worked then (autopilot went from 45 to 15 MPH) and I assume it still does.

Can I copyright the phrase “Tesla swatting”? :)

Sensor data is the input and can be faked!

I've always thought we should just start with a dedicated lane on the highway for self driving vehicles, raise the speed limit on it, and let the vehicles communicate with each other.

I don't disagree (and nor am I certain you are right), but if that is so, then there are some follow-on issues:

1) The system still has to react safely to pedestrians, cyclists and other entities outside of the network.

1.1) During the transition period, that is going to include vehicles driven by people (I do not think it will be feasible to ban the latter before self-driving vehicles are commonplace, except possibly in city centers and on some major highways.)

2) We do not want systems to learn how to take advantage of the protocol to drive aggressively and antisocially.

I wonder if the software industry has it in itself to produce a safe intra-vehicle network that wouldn't be taken advantage of, either by malicious actors or the manufacturers.

I am sure there is a technically sound protocol someone can come up with, but it would have to survive the corporate food chain, proprietary advances, and the three letter agencies.

Would we see the same dynamic we are seeing now with "net neutrality" and fast lanes for certain endpoints? Like if you pay for a premium service included in the BMW package, you get to travel at full speed, while the test of the vehicles are artificially slowed down. Or if you are on your way to Disney land with a paid ticket, all the stop lights will change in your favor.

Thankfully BMW doesn't own the roads.

They do this in LA already, you can choose to be toll'd more (wireless tolls) or you can stay in the normal lanes. I'm much more ok with the regional infrastructure charging it because you have at least a chance at knowing where the money goes.

The car manufacturers would love to have some sort of residual billing for crap like that I'm sure tho.

Plenty of places have privatized toll roads, in which case a private company owns the toll road and there's nothing (currently) stopping the owner from giving a special pass to particular car owners. BMW could partner with the toll road company to give free passage to BMW owners as an example.

The fact that it doesn't currently happen is a good sign, but as corporations grab hold of society in more ways I don't see why a premium car company wouldn't try to differentiate with those kinds of benefits. As cars become more like a service provided than an object you own, it makes sense that those kinds of things would spice up a particular "package deal" for your car service.

Can you share examples of these things? Isn't training data considered "experience"?

No, training data influences the values of weights and biases in a neural network. An over-fitted NN might encode the training data in its network, but in general, the training data can't be recalled in the same way that a human can recall, learn from, make judgments and adjust their future behavior and perceptions based on their experiences.

Example: One of the examples was a self driving car driving down a busy road. There was a car trying to cross that road perpendicularly. As a driver you could see that car inching its way into traffic but the self driving car didn’t recognize that as a threat. Suddenly the car drove into traffic and the self-driving car almost t-boned it but the human driver intervened before it hit. That’s something that as a driver I would have caught and slowed down because the car was acting weird.

AI/neural nets are good at detecting patterns, but only marginally so.

Humans have such incredible pattern recognition that we have built-in fear mechanics for when things are _not_ a pattern we recognize.

What percentage of the time do you experience fear when performing a task such as driving? Think about it - that's the margin for error that humans have in pattern recognition for that task.

You're raising an interesting point here. Humans pattern match everything (and integrate it into a world model). When something out of the ordinary appears, attention is brought to it immediately. By comparison, today's AIs just match what they can, leaving everything else out. I don't think that FSD can happen as long as that's the case.

I think a good rule to make is that to be fully selfreliant an AI would have to recognize random adjustments that are intentionally wrong or notably confusing for what they are or rather aren't.

I drove to work yesterday and a driver driving over the freshly painted lines to avoid a parked truck had basically caused a swerving lighter set of the lines in the middle of the road to appear. Similarly in my country at construction sites where the road is redirected for a while the new lines are often just in a different colour and the old ones remain. Sometimes they aren't and i've gotten confused at least once.

But what if it was intentional or more pronounced? Perhaps we really should paint ourselves a looney tunes-esque scenario. The kind where our hero paints some lines towards and a fake tunnel on a wall. Except instead of some by radar easily recognised wall it should be some other danger.

Also. If I put a cardboard cutout of a cartoon character next to or on a road will it cause these cars to slow down? Will it take a risk to swerve if the cutout appeared around a corner?

I think the recognition performance is irrelevant, the failure mode of a model not knowing when there's something it can't recognize is the issue. Most clssifiers will just fail silently and let the toy car drive into a trailer.

Your argument implies that "I don't know what is in front of me" can just be used as a signal to "stop the car", but that still isn't OK (slamming on the breaks on a freeway because a trailer has someone's face on the back of it, for example)

Recognition performance matters.

I don’t know much about the space, but always found it interesting that most of the conversation is about autonomous driving for individuals, which to me seems to be the last step after e.g cargo outside cities and public transportation has been automated. I reckon because the latter are less sexy, lead to job losses and maybe require more public acceptance?

I know there is a lot of work being done in this areas as well but what is the current state of it and what are the barriers?

A city bus carrying 20 people has 5% the drivers-per-person that an individual vehicle has. The driverless system only would save 5% the human effort that it would for a solo vehicle, so it needs to be 20x as good before it's worthwhile. Put another way, the bus with 20 riders is already equivalent to automating 19 cars.

Roughly the same idea with freight - a single truck driver serves the needs of many people.

Human drivers still do meaningfully cost money to employ, and they do annoying things like needing to sleep or take rest room breaks. Autonomous trucking would completely revolutionize the industry.

Autonomous trucks for urban trucking are a pipe dream, but maybe for long-haul routes.

That said, it's not at all clear to me that developing autonomous trucks for long interstate routes would be meaningfully cheaper than laying rail on literally every high-traffic route on the planet, which uses perfectly-understood technology and has entirely predictable returns.

Presumably you would be able to leverage a lot of that effort and expand into autonomous driving for other scenarios, right?

That's what's being attempted, essentially.

Autonomous driving in good conditions on highways is already working quite well, but once it's wet/foggy and there are animals wandering onto the roads and other cars are behaving unpredictably... those are not straightforward problems to solve.

We already have a lot of rail in the world, but still have lots and lots of trucks. Is it just not enough rail? Roads are also useful for people in a really dynamic way that rail isn’t (especially for big countries like the US).

There is some path dependency here, trains came earlier than cars and early industrialised nations build rail networks that were far more extensive than they are now. There were also freight trams in cities like Berlin, which delivered cargo to smaller industrial sites in the city.

You can have a driver in the truck for the city portions (at the beginning and the end of the trip), but sleeping on the highway rather than having to stop the truck for a full night.

Sure, but when you load 20 people onto a vehicle, if you don't have a driver you still need a security guard.

Subways the world over do without just fine.

The train system in Sydney, Australia has a guard cabin which it's recommended to ride near at night when there aren't many people about.

I don't know if it's because I've only been here four years, or because I'm from a city/country with a lot more crime, or it's my train line, but I've never felt unsafe on the late trains or the night buses.

Also, for the unfamiliar, 'guard' here does not mean an armed guard.

Subways in my country at least definitely do have guards (but they also have human drivers, so I don't think the two are really related).

Every train in London, even the fully automated ones, have at least one person onboard. Maybe at some point that will change, but there's a lot of pressure against it.

Profit margin. The driver just isn't an expensive or limiting part of cargo transport.

To be clear, the conversation also isn't about autonomous driving for individually-owned cars. Tesla's disasters aside, its about autonomous taxi services.

Are you in the freight industry? Because drivers are exactly what is expensive and limits growth right now. there is a humongous supply of trucking jobs, and they are paying really high rates right now (up to 70-80 cents per mile)

Fuel is still wayyyy bigger than the driver.

Both would be valuable - but a truck that is 30% more fuel efficient is probably going to save more money than one that doesn't need a driver (but still needs people more loading).

There are different estimates everywhere - but I usually see 20-30% of trucking costs are drivers. In other freight industries (cargo ships, trains, etc.) it would obviously be less.

https://www.thetruckersreport.com/infographics/cost-of-truck...

Thanks for that chart, but I read it differently. Just because something isn't the No 1. cost doesn't mean people aren't interested in lowering the cost. So let's say there are 3 million truck drivers in the US that earn an average of say 50K {1}. That's $150 Billion in costs per annum. Sure, there are other costs that are higher, but if you could lower that only by 50% -- say automate some (but not all routes), that technology would save $75 Billion per year, and that number would grow with inflation. Even at a real discount rate of 8% (10% with 2% inflation), that is worth 937 Billion dollars -- nearly a trillion. Now let's say the software vendor who solves this splits the savings with the trucking industry, so they get 37.5 B per year and the industry gets the other half. That's some nice recurring revenue, worth significant investment if it's technologically feasible.

So this is all about scale. Everyone is obsessing over Uber for cultural and political reasons, but there are 1/10 as many taxi drivers as truck drivers, and they earn half as much. The big money is in trucking, and even if the problem is only half solved (say just for some long-haul routes), it would be worth hundreds of billions of dollars, even if truck driver wages are not the single most expensive cost. Hell they could be the 10th or 20th most expensive, what matters is the billions that could be saved, not the rank ordering.

{1}https://www.bls.gov/opub/mlr/2019/article/is-the-us-labor-ma...

Another thing is that autonomous truck can drive 24h. In most countries there are regulation how often drivers have to take a break, e.g. in Poland after 4.5h of driving truck driver need 45min break, after 9h there is mandatory minimum 11h break.

In Australia on the major trucking routes, many trucks are company owned and are handed over to fresh drivers at rest stops. The trucks can proceed to destinations while the drivers sleep

Beyond experience and fear, I think an autonomous vehicle will need a reasonably accurate mental model of how other human drivers on the road are likely to react.

I feel this way too.

I've noticed over the years that the morning commute is a symphony of madness. Drivers going over the speed limit. Drivers doing illegial things in order to get to work, and keep the traffic moving.

Cops seem to let a lot of infractions go during the morning commute. (The rest of the time they let nothing go, but in the morning commute they just seem to know it's a symphony of driving?)

I picture a self driving car going 25 mph with 20 vechicles behind it.

I don't think you could legally program a car to break any law, even in emergencies. Well you could, but lawyers would have a fun day in court?

If the 25mph speed limit doesn't make sense on that road it needs to change. If it makes sense, then the self driving car going at speed limit is the right thing to do and human drivers should do the same.

It's as simple as that.

That's the kind of overly simplistic programmer thinking that doesn't work in the real world.

The technical term for this is 'situational awareness'

> Situational awareness or situation awareness (SA) is the perception of environmental elements and events with respect to time or space, the comprehension of their meaning, and the projection of their future status

There's an bearish case to be made - not only are the state-of-the-art algorithms nowhere near this - it takes kids quite some time to acquire this skill. Playing ball games with my kids, I've noticed even a 3yo still doesn't intuitively grasp Newtonian mechanics - they're trying to catch the ball where it is now, not where it's going to be in a few seconds. 8yo - notably better, though not fine tuned yet. Heck, I know some adults who're lacking in this respect.

Cats can catch a sparrow in mid-flight, though; it may take specialized hardware or software, but not human-level intelligence.

I chuckled! I spoke to Anthony at the Udacity office in Mountain View around 2018 (he’d already left Uber to found his trucking company). He seemed like he’d given up on city streets being autonomously drivable any time soon, but highways were still on the table in the next five years. One idea he floated was to employ drivers only from the exit ramp for the last mile journey. For trucking, this made sense.

Do you mean after he left Google? He left Google to form Otto which Uber bought. Frankly it’s the biggest mistake of TK’s career and caused the downfall in my opinion. He went against the advice of Emil and Salle, which was a huge mistake, and there were totally right.

Self-driving cars are perennially five years away.

Putting aside the question of whether the tech is there or not (and I would argue it isn’t there for commercial consumer use), the infrastructure and the laws will take years to catch up. This isn’t an overnight problem and many of the barriers are things that tech can’t just design around.

Not to mention there’s a huge voting block that’ll be against automation. I think the best first bet is just localized (robots in factories).

I’m personally of the opinion we are 10 years away or more technically. Main issue is weather and the risk threshold people can tolerate. All the tests thus far are in environments are high sun, no rain environments. If we are already localizing to southern cities, we may as well just add fiducial markers and guide wires.

In terms of risk threshold, look at COVID, risk of death under 45 is something like 0.5%. Or the black unarmed homicide count (<30). People have no way of measuring risk, just a few fatalities would have the public in a panic.

In the snow belt the reflective top layer of the lane markings gets scraped off by plows. When it rains the markings will completely disappear at night. Machine learning optimized for the ideal case is never going to work in many parts of the country.

The CDC's best estimate of COVID-19 infection fatality rate for people under 45 is more like 0.05%.

https://www.cdc.gov/coronavirus/2019-ncov/hcp/planning-scena...

Self-driving cars will always be x years away until purpose built and dedicated self-driving car roads are built to take out the variables that put safety at risk.

Back in late 2015, I was working at a media firm and one of our top directors called together a meeting with our engineering team to talk to us about the impact of our products/content. He was convinced that by 2018 most people would be using self-driving cars regularly and that this was going to hurt our core product.

I don't think I ever tried harder in my life not to LOL.

Ha! I hope he did not pivot the company or shutdown the core product

One of the curses of software projects is "premature convergence", where people to a ton of extra work for an overoptimistic launch date. I have to wonder how much wasted effort went in to trying to be seen as a leader in what turned out to be a hype bubble.

Steve McConnell also talks about a common estimation pattern: managers keep asking if something can be sooner, and are only stopped when some engineer proves that a date is 100% impossible. Normally that means that the launch date starts out with a 1% chance of success. Clearly it was lower than that here, but I wonder if that's because a) it's very hard to prove what's impossible here, or b) the Uber/Tesla hype bubble was so intense that it didn't matter what the nerds said.

I dunno, I remember that many of the commenters here were equally optimistic in 2015-16. It took a lot of arguments and discussion before people realised that it was much harder than it seemed.

Then that raises another interesting possibility: that absolutely nobody at the deranged-optimism companies knew what they were doing. Because there were definitely experts who knew those dates were impossible. Less 18 months after that Lyft prediction, roboticist and AI expert Rodney Brooks said that free-roaming auto-taxis wouldn't be running before 2035: https://rodneybrooks.com/my-dated-predictions/

It is difficult to get a man to understand something, when his stock price depends on not understanding it.

Wisdom for the ages.

Hard not to feel like we’re seeing a swoon in self driving like we saw in crypto back in 2017 and 2018.

They hype got ahead of the tech.

This is just the typical Gartner hype cycle. It applies to most new technologies, although the time scales differ.

https://www.gartner.com/en/documents/3887767

My kids are now 7 and 10 years old, respectively.

I wonder if they'll have to learn to drive, and make a driver's license?

It could be that by the time they come of age, there are self-driving cars, but they're much more expensive than making their license (which in Germany costs around 2k Euro, including lessons).

Or maybe there'll be no viable self-driving, fully autonomous cars by then? Who knows?

nothing new here, AI has been "5 years away" since the 1950s

It was 20 years away for a while. And may be again.

AI currently does lots of useful things in all sorts of industries. It just can't do everything that humans can do.

The world is already largely running on AI, just not the version of it that is in people's heads by watching sci-fi movies.

The world is running on good old software.

I don't know of anything important that's running on anything that can be reasonably called AI.

I think we can get by without Siri.

Of course, if you include the entirety of machine learning under the AI label, then whatever. But in my opinion, a principal component analysis does not an AI make. Statistical inference has been around since before computers were invented.

That's the rub, isn't it? Your definition of AI is basically "things we can't do by computer yet", which will always be unobtainable, by definition.

Meanwhile, you've got marketers that are also trying to use that definition, in order to sell you on some software solution, despite it being oxymoronic.

Two better definitions of AI that I'm a fan of are "Algorithms that model intelligence" (where "intelligence" is the ability to use knowedge in inference), or the broader definition of "Second-order algorithms" or algorithms that, rather than encoding a process that finds a solution, encode a process for finding a process that presents a solution.

Why, I'm perfectly willing to take either definition of AI.

My objection to the parent comment still stands; our world is not running on that.

For example, Wolfram Alpha, or OpenAI's algorithms that can learn how to play Super Mario will fit the definitions you provided, but our present world does not depend on systems like that.

People have wildly different definitions for what "reasonable AI" is, especially people working in the field. Colloquially the last decade or so of advancements in machine learning and data mining is all considered AI.

Slap AI on sufficiently many things, and sure, the world is running on AI!

For one, I think it's silly to apply that moniker to something that's not at least interactive.

Machine Learning / Data Mining pipelines that I've touched were running nightly, and furthermore, the systems they powered could continue functioning even if the pipelines crashed / weren't running.

Because with ML, both the input and output are data (information, at best). ML doesn't interact with commands.

The whole point if "A" in AI is that it's contrasted with Human Intelligence. Nightly batch processes simply ain't it.

Sure, people will still slap that label on anything, but it's boring to do so in the context of this discussion :)

I think there is some argument to be made that the Web is "running on" AI, since AI-driven ads businesses are the source of all money in the Web (essentially); and I do think that it's fair to call NNs "AI".

I for one am highly skeptical that the kind of automatic personalized targeting that AI enables is actually worth the price, but the industry overwhelmingly believes it is. So, removing AI from the picture would be considered to greatly diminish the value of advertising on the web per industry pricing, and with the ad money, much of the web would dry off.

The sci-fi version being as the term was defined by Alan Turing and the marketing version being as s/ML/AI

weak AI is here, AGI is a ways away..

I wonder if it’s possible that this is true in some restricted area where companies are heavily testing their self-driving cars, and COVID reduced the amount of driving people do.

From very brief observation a year ago, it seemed like Lombard Street [1] was 90% self-driving-car training. Definitely a very special case that just disproves the quote from Lyft's president.

[1] https://goo.gl/maps/oXrvVYhDXWVxaoQC8

Reply on news.ycombinator.com