Hacker News Reader

Comment by spamalot159 | Hacker News Reader

I've been acquainted with a number of engineers at various autonomous vehicle companies and they are some of the most reserved people when it comes to the reality of autonomous driving. Many of them going so far as to say they would never use one for themselves now or in the near future.

I think when domain experts speak up like this we should listen. The autonomous vehicle technology seems like it has a long way to go.

View on HN

I don’t know, I used to ride an autonomous vehicle every day in London. DLR is autonomous, the catch is - it has its own lane and it’s on rails. Also, someone sits in the car during rush hours, just in case.

Here’s how it looks to be in the front car of the train when there’s a supervisor: https://imgur.com/geFIidK

Outside of the rush hour, that lid is closed and it feels like the Half-Life train at beginning of the game.

I know it’s not possible to build trains to everywhere and USA is huge but I suspect that the answer for self driving would be somewhere in between of human+ level intelligence and dedicated infrastructure.

We could've vastly expanded a lot of train systems with the amount of money that's been burned trying to make cars drive themselves.

Unfortunately, most new transportation methods that are not trains are just attempts to build trains, while claiming they are not trains and will cost less than trains. But inevitably end up costing as much as trains.

Trains.

Is this really true? To be able to build a track you need to have government-level authority (e.g., in GP's example, to build something on its own lane in the road, or to displace people in homes in the way of the track). However, the entities funding SDCs are private.

Having dedicated tracks solve a lot of problems for autonomous vehicles.

Governments already claim a monopoly on roads in most places. Imagine if those roads had some sort of cheap embedded track that a number of vehicles could latch onto. Then you just need some sort of computer to coordinate the vehicles using the track, and understanding where the vehicles are on the track.

This isn’t exactly stopping the Boring Company or Loop from making its promises, is it?

Or they are attempts to reinvent the wheel because we can't possibly use a train built elsewhere. Or they are using trains as an excuse to push money to their political friends.

> We could've vastly expanded a lot of train systems with the amount of money that's been burned trying to make cars drive themselves.

I'd like to see the numbers on that. Given the cost for the CHSR is around $100B, I don't think that statement is true. The bigger issue here is the US's inability to build fast trains, affordably, and on a reasonable timeline.

Counting only published investment in companies like Cruise, Zoox, Argo, Nuro, Aurora, etc. gives investment of around 20B . But its likely that the money invested "privately" in Tesla's AV division, Uber ATC, lyft's AV division, Waymo, and Cruise post acquisition is much higher.

I think a very conservative guess is that there's been $40B invested in this space. That would, for example, cover the south bay BART expansion 4-5 times over.

> That would, for example, cover the south bay BART expansion 4-5 times over.

I've become more cynical about CA (and the US in general) lately. By reasonable costs it would, but I think the various corruption interests would find their way to increase the price tag until it is only half done, while they enjoy their $40 billion.

I'm pretty cynical myself, but the first phase of BART extension to the south bay is already completed and has been running for nearly a year, and cost $2.9B. The second phase is a bit more difficult because it requires substantially more underground work in downtown San Jose, but $40B would be quite a bit even by CA standards.

Madrid does all underground lines lines like your phase 2 for about 1 billion. Someone needs to learn about cost control. Think of how much more they could get with Madrid prices.

Not saying that CA prices aren't inflated, but there are also geological differences that probably matter. Madrid isn't near active fault lines (not to mention generic soil differences that can matter, but IDK if they do in this case).

I don't have Japanese costs, but they are lower, and deal with worse earthquakes than CA

There is basically no good place to build high speed rail in the U.S. Cross country routes aren't remotely competitive with air travel. Regional routes only make sense on the coast where they have unfavorable terrain like mountains, and some of the most populated/valuable land in the country. HSR really starts losing in time cost to flights in the 800-1000 mile range, and needs about 200 miles to make it better than driving. That a very narrow band for the U.S.

800 miles is quite far from the coast.

Also there are considerations other than time, it's much more convenient that driving or flying. I know in USA there's last mile problem going out of railway station but that can be solved by car trains.

Even if all that were true, it still does not explain the high costs. There are countless examples in other countries proving it's possible.

Imagine the investment wasted on self-driving cars being spent to solve that problem.

Another good example of a pretty neat, autonomous personal rail transport system is in Morgantown WV: https://en.wikipedia.org/wiki/Morgantown_Personal_Rapid_Tran...

A slight tangent, but the Victoria Line (London Underground system) was officially opened in 1969 to run trains automatically without the need for drivers. However, there is always a driver in the cab for emergencies etc.

On opening, the line was equipped with a fixed-block Automatic Train Operation system (ATO). The train operator closed the train doors and pressed a pair of "start" buttons and, if the way ahead was clear, the ATO drives the train at a safe speed to the next station. At any point, the driver could switch to manual control if the ATO failed.[27] The system, which operated until 2012, made the Victoria line the world's first full-scale automatic railway.

- from https://en.wikipedia.org/wiki/Victoria_line#Service_and_roll...

The Copenhagen metro has zero drivers, and not even room for one.

https://migogkbh.dk/wp-content/uploads/2019/09/Metro-1200x72...

No modern metro system should have anything new with drivers, as of about 20 years ago. Now most systems were build when automation wasn't an option, and it does take years to upgrade, so I'll give them another 10 year to do all the needed upgrades to get rid of the now obsolete drivers. I'll be shocked if any make them that haven't already, but the upgrade and remodel cycle should be 30 years long for all transport systems so I have to give them those 10 years.

FWIW, I work in AV eng/research and have never heard someone say this. Most of the people I know in the space are pretty clear-headed about how dangerous human driving is.

We are generally less optimistic about timing, but that was more true a couple years ago (before all the big cos' 2019 deadlines got pushed back) than it is now. Current timelines seem like an achievable challenge to me, and I don't think I'm unusually optimistic.

Musk's timeline pronouncements, as usual, just aren't taken seriously (and imo hurt the industry's credibility). I like the guy and respect what he's done, but for whatever reason he says really out-there things that are best ignored until he shows the receipts.

Two things can be true at the same time:

1. Human-driven cars are dangerous.

2. Autonomous driving (currently) is dangerous.

AD solves some human driving problems (like inattention) but introduces new ones (like driving under trucks).

Obviously so, and this doesn't contradict anything I said. But the types of people who work on AVs are quantitative enough to understand that these magnitudes can differ, and are able to process that difference instead of waving them away as "both dangerous" and sticking with the status quo out of blind habit.

Almost literally every decision trades off one danger for another: to pick an adjacent example, airbags introduce dangers that are different from the ones they solve. And yet, "airbags have dangers! Take them out of your car" is a woefully incomplete understanding of the risk landscape, driven more by our Neanderthal brain's superstition than our Homo Sapiens reasoning about risk.

That sort of magical thinking is a hurdle that every shift in the risk landscape needs to overcome ("what if vaccines have microchips in them??").

This is, of course, dependent on actual comparative safety stats, but I've yet to hear a safety argument against taking a Waymo ride over an Uber in Chandler[1]. If the same level of transparency and rigor is used in more complex environments, I'd take a cruise/waymo car in sf too (once they launch).

[1] https://www.forbes.com/sites/bradtempleton/2020/10/30/waymo-...

I don't find your analogy to airbags persuasive. The vector space of "weird situations where an airbag can kill a person who would not have been killed otherwise" is vastly smaller in dimension and size than the comparative space where autonomous driving can kill someone. That in and of itself does not make AD a bad idea, but it does make it a much harder engineering problem.

As engineers we can pretty much completely characterize the airbag space and make predictions about where airbags are dangerous, and then mitigate those. What makes AD a bad idea is that such characterization is impossible with the AD space because AD is currently done almost exclusively by feeding massive amounts of data into a neural net pattern matcher. In other words we have no idea why AD works, and thus we have no idea when AD won't work until somebody dies. This is no way to do engineering, especially when lives are at stake.

One of the biases in Hacker News is that comments like yours tend to get downvoted.

Some of my more reserved comments get downvoted, and then flooded with replies about 360 vision and cars never getting sleepy.

IMO: Most of the autonomous driving tech will evolve to be driver assistance and safety tech in the upcoming decade. I still believe we'll get autonomous driving someday, but not until we have a few more AI breakthroughs.

Until then, "Autonomous driving" will be like "Duke Nukem Forever."

... he comments under the top comment.

Skepticism of self-driving tech and "AI" in general is a very popular position here.

My most upvoted comments on HN[0] are the ones that pick on some hyped up technology. HN loves to be "contrarian" (in quotes because this contrarianism is actually the mainstream position).

[0]: I don't think this is good, it's just a statement of fact.

It has been increasing. 5 years ago we were more hopeful. However the needed parts have not delivered and so we are rightly looking harder.

5 years ago self-driving cars looked like technologists were letting their imagination run away.

There's always "something" that technologist think is right around the corner, until they realize that the hard problems are harder than they anticipated.

To be fair, Autopilot is about as good as Duke Nukem Forever.

This is because OP's comment is simply an appeal to authority without any direct evidence or reasoning. Listen to the experts. This is highly counter to the hacker ethos, and personally, I'm disappointed to see it as a top comment on HN.

This reminds me of the first bubble. Everyone on the inside were looking at each other as stock and salaries skyrocketed thinking this was crazy and unsustainable. But we all just shrugged because there was no stopping the hype train.

Here, anyone even just tangentially involved or know the some details of AI/ML are highly skeptical that Elon's vision of self driving cars will occur in our lifetime. Yet, Elon keeps pitching it, the press runs with it, and the public grabs it up, patiently waiting.

To be fair to Elon, there were a lot of naysayers coming at him about Tesla and SpaceX but he has accomplished more than they ever thought he would. Most of the time, however he has exaggerated the timeline in order to keep the hype around him and his companies. I think he would do better to under-promise and over-deliver. Especially with a technology that could potentially cost lives if delivered too early.

The way I like to think about it, Elon Musk over over over over promises, and "just" over delivers. Most CEOs could never dream of producing the kinds of real results that Elon Musk has, but also most CEOs aren't as comically obnoxious as he is.

with rockets and electric vehicles however science & engineering had an "angle of attack" based on fundamental discoveries in physics, and centuries of continuous progress , that's not the case with AI.

Deep learning is great for a few narrow cases but there is no path to general human-like intelligence required for a complex human-like activity like self-driving which has an infinitely long tail of edge cases you can't train for with these large and "dumb" ML models equivalent to curve fitting.

The mistake of Tesla is betting on covering more and more of these edge cases incrementally but what's needed is qualitative change rather than incremental improvement.

By qualitative change I mean actual model of human-like intelligence, including reasoning and "common sense", which is a prerequisite for self-driving. The solution to this problem requires more than the sum of its parts.

It's kind of like planning to build an airplane before figuring out Newton' basic laws of physics.

I don't want human like AI to drive cars. I want a deterministic system that we can prove always does the right thing.

I'm not sure that is possible on roads.

I tend to give him a break since he is one of the few modern entrepreneurs that I think actually produces things of real value.

The fact that it tends to be in businesses that are among the most mature and hardest to get into is really amazing.

> Most of the time, however he has exaggerated the timeline in order to keep the hype around him and his companies

All of Elon's companies are heavily subsidized by US tax payers (energy credits for Tesla and solar, rockets are paid for by NASA, military, etc...) So to use tax payer money you need to have the goodwill of the tax payers behind you. Its worth every penny to lie and bolster your image of doing cool shit so people don't mind their tax dollars continuing to go towards your cause. It's 100% PR. Without US tax payers none of these companies would still exist.

True, I am unfairly singling Elon out, but he's put himself at the front of this hype train. There are plenty of other "visionaries" that are in the engine with him.

Work at Waymo. Know lots of folks at Cruise. This is not reflective of the general outlook that most engineers have.

I have engineer friends at Waymo that I went to school with who tell me that real self driving is 30 years out.

What does "real" mean to you? L4 self driving in most parts of dense cities in the American sunbelt and similar constraints for trucking represents massive revenue and a real business in my eyes.

If you mean driving like a human does with no priors over the area and in nearly any circumstance, I would agree.

Shocking that the outlook at a self-driving car company doesn't match the general outlook of the entire automotive industry.

> Many of them going so far as to say they would never use one for themselves now or in the near future.

I'm less worried about the autonomous vehicle I might be driving and more about the ones others are driving. Here's an idea, how about being able to geofence a neighborhood as "<= Level 3 autonomy only". I'm only somewhat joking. I know it smells of NIMBY but it might be useful around schools, daycares, etc.

Based on how I've seen humans drive their vehicles for school dropoffs and pickups in San Jose, I would much prefer robots.

Not the kind of "robots" we have right now I think.

It's an incredibly hard problem to solve.

The Boring Company might be onto something. These tunnels are a very controlled environment, highly predictable. If they can be be made cheaper it might just be the initial "killer app" of self-driving cars.

See my comment here: https://news.ycombinator.com/item?id=27077320

The Boring company is essentially redesigning trains. With single-family train cars that are also very overpriced normal cars. That need to be substantially modified into very inefficient train cars to go down the tunnel.

if it's just fixed route stuff why not just build more advanced subways/underground trains? not that boring couldn't make that cheaper and faster to do the whole 2/4 door cars in a tunnel thing feels weird to me

I'm not trying to contrarian or anything but most of my life those have been my primary mode of transportation and it's worked out great especially if you throw a bike or now I guess a e-scooter into the mix

I think the Boring idea is that at the end of the trip you still have your car. So suburb to suburb works.

> why not just build more advanced subways/underground trains?

The train business is horrible to be in. Just look at Bombardier Transportation/Alstom. City A in State B and country C decides it needs a new train. The existing infrastructure was all custom-made in the 60's. Bombardier/Alstom has to design a mostly custom train for it. Everyone bike shed on the design. Then, because it's a government project, and State B financed some of it there's a clause in the contract that is must be manufactured in State B of Country C. Not only is a new production line (manufacturing hell) for the almost completely custom design required, they must also set shop in this new jurisdiction and work their supply chain around that. Ribbon cutting ceremony, positive polls for reelection.

Few years later, City D in State E of Country Z wants a new train system. The whole process starts all over again. Can't just retool the existing plant in State B of Country C, it has to be a new one too in State E of Country Z, or a local manufacturing partner (ribbon cutting ceremony obliges). Meanwhile the press in State B of Country C gets angry as workers are laid off since there's no more work to be done.

By using standard road vehicles, Musk eliminates one of the biggest cost. It's all off the shelf stuff that's user-supplied.

Reply on news.ycombinator.com