AI Podcast: Elon Musk on Tesla Autopilot [video] | Hacker News Reader

AI Podcast: Elon Musk on Tesla Autopilot [video]

youtube.com

51 pointsespeed56 comments

Full self driving by next year and yet accidents caused by autopilot are still common. Tesla says that you should keep your hands on the wheel at all times and stay vigilant, but Musk appears on TV demoing autopilot with his hands off the wheel and retweets videos of others doing the same. Three people have died using autopilot, 2 when their cars drove underneath a truck that the radar failed to pickup, and the Apple engineer who drove his car into a guardrail while under the influence of autopilot.

Telsa is going to continue pumping their technology until the investor day coming up on April 19 because they desperately need a capital raise. They announced closing stores, layoffs, price cuts, and the short-lived 35,000 model 3 because cash is tight and demand was way down in Q1. Q2 is looking like more of the same.

The worst part of this whole thing to me is Musk's blatant disregard for human life to protect the value of his company. 3 people have died because of autopilot and pumping this unfinished technology will lead to more deaths.

https://twitter.com/shabudibudi/status/1116514629268705280

https://www.youtube.com/watch?v=VWTMnnECBSo&feature=youtu.be

https://streamable.com/kpugc

https://twitter.com/NetflixAndLamp/status/111452700974343782...

In 2017, the USA had 37,133 motor vehicle fatalities, at a rate of 11.6 per billion vehicle miles driven.

I’m having difficulty finding a recent figure for how many miles Tesla autopilot has driven, but they reached over 1.2 billion by July last year.

3 deaths is 3 too many, of course, but it’s also 10.9 less than one should expect for 1.2 billion miles driven.

The number of Tesla-related deaths is much higher than Tesla reports.[1] At least 4 in California, at least 11 worldwide. That article points out how low the death rate for large luxury cars is, too. Tesla is at about 4x the death rate for the luxury car industry.

[1] https://medium.com/@MidwesternHedgi/teslas-driver-fatality-r...

That article is fairly damning.

I’m not convinced yet (and neither am I convinced of Tesla’s claims). You can’t make strong causal claims from observational studies. Run a real experiment and I’ll take these claims more seriously.

(For example, there are a ton of potential confounding variables that could be causing the difference other than the safety of the car, like demographic differences.)

The burden of proof is on the claimant, whether it’s Tesla claiming their vehicles are safe or skeptics claiming they’re not. Until then, neither adds much to the discourse.

Until the onus is on Tesla to prove their claims, rather than cherry picking statistics for marketing, it's better to be sceptical of what they say about safety of an unproven emerging technology that they are emphasizing to sell more cars. Basically, Tesla needs to prove that this reasonable analysis is wrong because they are the people making the claims about safety.

They really shouldn't get the benefit of the doubt. The burden of proof is on the company selling a system to control a two ton vehicle travelling over 70mph.

This is a biased article.

I do like that the article does some analysis to deduce that the NHTSA-FARS data is incorrect for Tesla and they make several well reasoned arguments why. Then using other data sources, they try to extrapolate what the correct data might be.

The problem is that the Tesla data in the NHTSA-FARS database is lower than expected. Yet they are comparing the "fixed" Tesla data with the rest of the data in the database that they take at face value.

It goes to reason if the NHTSA-FARS data for Tesla is easily proven to be underestimated (i.e. there are actually more fatal accidents than shown) that the data for other car manufacturers and models are also underestimated.

Their argument is the Tesla data is underestimated because, as a new and relatively unknown car manufacturer, Tesla is often miscoded.

It seems like these concerns are much less applicable to a BMW or other manufacturer of large luxury cars.

The author of the article admits that his fund is short TSLA. I don't know why I would believe his data.

Elon Musk is long TSLA so we shouldn't believe him either.

I would argue someone in a short position has more incentive to produce negative news than someone in a long position does to produce positive news.

Teslas are arguably sports cars more than luxury cars. You should really be looking at the driver statistics shouldn't you?

"Sports car" is just a marketing label now, says Road and Track.[1] Once a "sports car" was a little 2-seater with a roll bar and a 5-point harness. Not a 5-seater with a deluxe interior. Now, it's just a PR term.

[1] https://www.roadandtrack.com/car-culture/a28225/sports-car-d...

Perhaps, but they're marketing to different people and have clearly different driver populations and accident statistics even though the cars themselves are probably not much different.

Musk references OP's attitude in the interview. Tesla gets about 100x attention on it is other car companies. How many people gave died in a Lexus in the past 5 years? And how many of those deaths could have been been prevented by Lexus making safety a higher priority?

A better question though is what is the regulator's responsibility to approve a new technology like auto pilot when we know the general public will ignore the statistics and focus on the tangible deaths?

This is not an apples to apples comparison. Tesla autopilot only drives the easiest subset of what a human driver can do and has a human supervising it while it does it to catch mistakes. This paints a picture that is disingenuous.

It's like having a doctor that only accepts completely healthy 20 year old patients quoting their stats compared to the general population.

These cars are no where near being fully autonomous and the story that Elon is paining about this happening in a few years with the sensors in a Tesla and the current state of research is a fantasy. End of the day he's a car salesman so take it with a pinch of salt.

What is the general vehicular fatality rate on roads and under conditions where Autopilot is usable?

"For instance, in 2007 0.54 people were killed for every 100 million vehicle miles driven on urban interstates, compared with 0.92 for every 100 million vehicle miles driven on other urban highways and arterials, and 1.32 killed on local urban streets."

http://freakonomics.com/2010/01/29/the-irony-of-road-fear/

>Musk's blatant disregard for human life to protect the value of his company

He could also be developing autopilot to save human lives. There are a lot of road deaths each year which is not going to be fixed by doing nothing.

When he is working on making safe automated driving, then yes. When he is exaggerating the capabilities of what has been achieved so far, and thereby encouraging a dangerous over-reliance on the equipment, then definitely no.

Tesla’s autopilot has saved many lives already according to known statistics. Also, because Tesla’s have the highest safety rating, they have also saved lives in accidents.

Tesla compared itself to crash/fatality data that included motorcycles in order to look better than they are.

I'd accept the claim that autopilot is dangerous if it had brought a car into a state that an alert/situational-aware and normally skilled driver could not recover from. That's not what happened. These people have died because they thought they had got better things to do than monitoring the traffic.

The worst part of this whole thing to me is Musk's blatant disregard for human life to protect the value of his company. 3 people have died

It's all relative. Boeing CEO Dennis Muilenburg hears that and says: "hold my beer".

Boeing timeline:

   Oct 29, 2018 Lion Air crash, 189 dead

   Nov 10, 2018 pilots already talking about Boeing
   emergency airworthiness directive related to MCAS[1]

   Mar 10, 2019 Ethiopian Airlines crash kills 157

   Mar 11, 2019 Boeing CEO "confident in 737 MAX safety"[2]

Then the Boeing CEO drops his mic and says "Your move, Elan!"

Sorry to be so cynical, but the Boeing stuff really bothers me. Boeing management behavior was worse than Tesla's.

[1] https://www.pprune.org/rumours-news/614857-indonesian-aircra... [2] https://www.reuters.com/article/ethiopia-airplane-boeing-ceo...

3 people ain't that much. If I was Musk a good PR would be to visit their funeral. Since kills by autopilot are smaller than by human factor.

I'm skeptical of Musk's claim that Teslas are an "appreciating asset". That doesn't make sense in the automotive world, and it's claims like this that make him lose credibility.

It is pretty disingenuous on his part. The small set of cars whose selling price does appreciate satisfy a set of requirements which the Tesla doesn't.

1. Discontinued / Very few produced -> Tesla is ramping up production 2. Iconic design -> Teslas do not have stand out look and the same design language is being used for other Tesla vehicles too. 3. Movie / cult classic -> Tesla does not have its 'Back to the Future' or 'Fast and the Furious'. 4. Unique features not present in other cars-> Tesla doesn't necessarily have any

Lastly, tablet and screen based designs are ones that age the worst. So, Tesla is at a disadvantage there.

The car collector's market is completely at odds with what Tesla offers. Tesla lovers are high technology folk, who always want the latest and the greatest. They do not appreciate dated designs. Car collectors are the ones most averse to electric cars (as of now).

I think Teslas will hold their price better than other brands, but they certainly won't appreciate. (may be the original roadster, because they made so few of them)

Musk doesn’t have any “oracle” credibility to squander though. He’s known for many absurd utterances from his SEC violations to his pedophile accusations. The words coming out of his mouth are not given high credence by the investor class.

Well, it sounds pretty disingenuous, but you can still find some meaning out of that. An asset by definition is a useful/valuable thing or person. If Teslas become better just with OTA updates, then they become more useful and valuable for the person that owns them. Of course, financially speaking it doesn't make sense (they will depreciate like any other car that rack up miles), but I see some merit in that statement since no other vehicle in the market currently can get better just with the past of time (as you get new updates).

I don't know if his statement was about the financial depreciation of a Tesla or not, but if in 6 months, my car is better than it is today, with no custom parts or additions, you can undoubtedly claim that its value (the perception of what the car is doing for me) increased. Though, not its financial value.

>>That doesn't make sense in the automotive world

Well, there are plenty of cars which are appreciating from the moment you pick them up from the dealership. Rare limited editions of Lamborghinis, Ferraris and Porsches all come to mind. But those are rare exceptions from the rule, and Teslas are nowhere near that status.

Appreciating in the sense that they become more valuable, in utility, with time. Appreciating like software, not like a stock.

What I understood him to mean by that is that, because the software of the vehicle will be updated with more capabilities, the value of the car, perhaps not in monetary terms but in terms of utility, would increase over time.

Of course the use of the term 'asset', since it is typically used for financial value, is confusing if that was his intended meaning.

Autopilot drives far safer than I do, my mind wanders on a 1 hour commute. I know the gaps, on 101 where they just moved the lanes and used stickers for markers, autopilot doesn’t handle it well. But I feel WAY safer with it on. It’s pretty amazing, and it’s light-years better than other car manufacturers’ offerings. Try the Volvo self driving and then compare, that thing tried to kill me multiple times in a single weekend drive.

Volvo doesn't have self driving as a cunsumer offer. Or have I missed something?

You missed the Volvo S90

Point taken: the safety of such systems should be compared to how people actually drive, not as if everyone was driving with all due attention.

The most important concern is whether other people are safer when you are using autopilot than when you are not. Part of that calculation is how bad the accidents are when they happen.

There is also the question of whether safety overall would be better with a combination of lesser technologies, such as lane departure warning and automated emergency braking, than either full autopilot or unassisted manual driving.

Every manufacturer should be evaluated in the same way, and if Tesla is currently the best, then congratulations to them, but that does not mean autopilot is good enough yet, and what we need right now is more data objectively evaluated, not Musk's tendentious claims.

Driving with autopilot every day, I bet we are 5-10 years off the full self driving panacea that Musk is selling, but I don’t care. The car is blazing fast, safer than I am on my daily commute, and I don’t feel the pang of regret from hitting a gas pump.

I don’t buy into Musk’s bluster blindly, but he does accomplish most of what he says he will... just not on the time scale he promised. Seems fine to me though, I like what he promised and honestly I don’t care if my car appreciates, the automatic braking videos sold me.

so here is something interesting for you.

I work for a mining company that has autonomous and human driven trucks on the site (open pit mining).

after running the stats on efficiencies and wear and tear on the trucks we found that:

In the mornings (or start of shift) humans are far more efficient and cause less wear and tear on the trucks, towards the end of the shift the autonomous trucks are more efficient and wear the trucks out less.

At the beginning of the swing ( 8 days on 6 days off ) the human driven trucks are far more efficient and cause less wear and tear. Towards the end of a swing the Autonomous trucks are more efficient and less wearing on the trucks.

In wet weather humans are far more efficient and less wearing on the trucks, on hot dry days autonomous trucks are more efficient and wear less on the trucks.

At the end of the day, autonomous are more consistent but humans are always better but they wear out themselves or get board or get lazy and take shortcuts when its nearing end of swing or end of shift or when the conditions are hot and dusty.

I cant see why this wouldn't be replicated on open roads by every day humans and every day autonomous vehicles.

Softball interview.

Musk is still fixated on "What is a car" and "What is not a car". (At [28:12]). Teslas keep hitting things that are "not a car", yet clearly aren't driveable road. Musk doesn't have to evade that issue because the interviewer never brings it up.

Actual paper.[1] Which is actually an analysis of the data from [2].

[1] https://hcai.mit.edu/tesla-autopilot-human-side.pdf

[2] https://arxiv.org/pdf/1711.06976.pdf

I smell only fear here. I am not sure how many people have tried tesla autopilot, I have owned model 3 for a year and the rate of improvement, features on AP is crazy. I thought hackernews is a place where you can discuss about the technology but here it seems just blatant bias and no thought full conversation.

Anyone who's done anything sufficiently advanced with computer vision, AI, and cars knows that Musk is full of shit. The things he is promising are at least a decade away. Not 1 year. To put it another way, we're in "staging." We haven't hit production yet. If there's ever enough Tesla's on the road, you're going to start seeing a lot of problems with Tesla's "autopilot." Honestly the cars are fantastic enough without people overstating what autopilot can do.

Furthermore he's been funny with his companies' finances. Between Tesla, Solar City (remember when Tesla bought Solar City, because I do), and the fact that SpaceX needed NASA to bail them out with the falcon heavy should tell you all you need to know about how solid his companies' finances are. Tesla as a company, the forefront of Musk's empire (arguably), is showing signs of business turbulence. Layoffs, benefit slashing, etc.

Now, I don't think he's really tried to scam to anyone, sincerely. He's a businessman who's bit off a bit more than he can chew in the science department in my opinion.

> Anyone who's done anything sufficiently advanced with computer vision, AI, and cars knows that Musk is full of shit. The things he is promising are at least a decade away.

Citation needed.

I would like to know the name of the secret genius working at autopilot who knows something the rest of the self-driving industry doesn't.

If Tesla does well with HW3 they'll be able to get Autopilot into 'flashy demo' territory in the not-too-distant-future and that will have the Tesla fans thumping their chests for years to come before it dawns on them that there is actually is a massive gap between going a few miles in traffic and pulling off a couple fancy maneuvers along the way, and a car that can navigate a city safely without a human driver. Nobody knows how wide that gap actually is.

Progress in free-range robotics is inverse exponential, and by that I mean that to drive the error rate down by an order of magnitude, you need to do way more work than was needed to achieve the previous order of magnitude.

We can come back to this in 7 years or so and re-examine just how far off Tesla may be from 'The Tesla Network'. Although I suppose the Tesla Network can just be renting somebody else's Tesla for an afternoon, FSD not necessary, and in that case, any day, right?

Exactly. He's a car salesman. I work in CV and automotive and the stuff he's saying will be around the corner isn't solved in research with a $100k set of sensors and progress has plateaued. These are "two ton death machines", a direct quote for Elon in the interview, and overstating what these systems can do to customers who don't understand the state of computer vision is morally bankrupt.

Musk sounding confident about their self driving as usual.

>I'll be shocked if it's not next year at the latest that having having a human intervene will decrease safety (22:32)

Like how we got rid of elevator operators because automatic is safer.

>with a full self-driving car computer the rate of improvement is exponential (24:26)

>Tesla is vastly ahead of everyone (30:11)

I'll be shocked if it's not next year at the latest that having having a human intervene will decrease safety (22:32)

For certain interpretations of "intervene", that's true now. See this video of a Tesla hitting a construction barrier.[1] This is a dashcam view from a car behind the Tesla. At 00:21, it's too early for the vehicle to start turning, and at 00:22, the vehicle has already hit the barrier. The driver has to detect the failure and then take over. Much of the driver's allowed reaction time is taken up by having to wait until it's almost too late. Control handoff inherently implies a slower reaction time than manual driving.

(Not noticing the lane change warning and orange construction 45MPH speed limit sign back at 00:05 didn't help, either.)

This is the problem with shared control, and why Waymo rejected it early on.

[1] https://www.youtube.com/watch?v=-2ml6sjk_8c

That looked bad. It kept inside the lane right up to the moment when it hit the barrier and the way it stayed perfectly in the center of the lane really makes it look like it was on autopilot.

(By the way, the driver with the dashcam is driving terribly.)

"The rate of improvement is exponential" - Elon Musk on the state of improvement in self driving during this conversation.

This is literally the opposite of what is true.

The rate of improvement has slowed down as the low hanging problems have been addressed. Everyone has pushed their timelines out.

It's pretty bad that a research scientist at MIT didn't call him out on this. What's the point in someone who works in the field conducting interviews if he doesn't question the interviewee.

The thing that Elon is doing is called lying.

https://www.theverge.com/2016/10/19/13341100/tesla-self-driv...

https://twitter.com/elonmusk/status/823727035088416768?s=21

https://twitter.com/elonmusk/status/888053175155949572?s=21

And so on...

QOTD: "What's outside the simulation?"

If Elon would be the lucky ones along with Yudkowsky, Ruy etc.. to stand near the AI in the box then I guess his question would be answered

Reply on news.ycombinator.com