Hacker News Reader

Comment by pb | Hacker News Reader

This is a great quote: "Startups with no demanding customers can lose their urgency, with development cycles dragging on for years, says Rust, who worked at Google’s experimental research factory before joining Cruise. “We tried to accomplish what one of the most talented teams at Google did in two years and compress that into six months,” Vogt says."

This is one of the main reasons why "internal startups" at big companies fail. They lack urgency. The same thing often happens at startups that are over-funded too -- they think they can afford to take their time and that leads to compounding slowness.

View on HN

Yes, but in this particular case I would say that urgency is not the best thing to have. When my life is on the line, I would much rather have an autonomous driving system that was developed slowly and deliberately without time pressure instead of one that was rushed out the door to beat a competitor to market.

This is a fair point, but don't underestimate the value of getting something to market sooner rather than later, even for something like this.

Look at it this way: would you rather wait for five 9s against getting in a collision of any kind, in a car you won't be able to buy for another 25 years ... or settle for four 9s, and take delivery by christmas?

I don't know enough about the odds or the effect vs speed to comment with any confidence about whether Cruise's autonomous driving system fits this pattern or not, but I too often see a knee-jerk "slower and safer is always better when LIVES are on the line!" reaction, and I think it's worth being careful about whether it actually does make for a safer world.

To add perspective, the interim solution (driving yourself) is probably closer to three 9s. So it's a choice between a big improvement soon, or an even bigger improvement later.

This is false, and one of the reasons Tesla got a lot of deserved criticism over the death that occurred from using their autopilot system. So far, all of the data (which is limited, so it's far from conclusive) points towards self-driving technology being deadlier than manual driving.

This is not a condemnation of technology augmenting driving. ABS is clearly a win, airbags were once controversial, and even seat belts were once considered questionable. The technology will likely improve, but it isn't there yet and believing that it is can quite literally kill you.

I'll grant you that there are situations as you describe, where people lean too heavily on the current technology and die as a result. However, I believe that there are more situations where the technology works as intended, preventing mistakes and saving lives.

So, until there is clear data to the contrary, I'm sticking my original assertion that current-gen augmented driving technology improves safety.

With all the jockeying for position we've seen in the vehicle automation space the past couple years, one thing that's become clear is that both SV and Detroit have a lot to learn about the other guys home turf, and they've both humbled themselves with respects to what they can and can't do well.

This is a false dichotomy and just not representative of reality.

There can never be self-driving cars because the type of generalized intelligence that would be required is impossible to achieve. The fact remains that a computer program can only do what it is told, it cannot think. Code has no mind.

So what the choice is between is, instead, someone (Tesla) pushing out a falsely marketed support system that does none of the things it is believe to be able to, and causes unnecessary deaths and injuries. And, on the other hand, thorough testing to uncover the possibilities and limits of the technology, to produce a reliable, usable product.

This is a little simplistic. Computers can do things they aren't heuristically programmed to do - that's the basic idea behind machine learning. I've seen some incredibly impressive robots that ran entirely on neural nets and never failed, but because it was impossible to understand how they worked they could only be used for demonstration purposes as guaranteeing safety was impossible.

What you seem to be referring to is true AI. I agree we are nowhere close to anything resembling true AI, but I do believe something representative of a computer that can "think" will appear in the coming decades.

That said, you can program a car heuristically. Statistically speaking, the unknown unknowns are so rare that assuming it doesn't fail under expected conditions you can guarantee it to be safer than humans. The unknown unknowns are not the problem right now, it's just dealing with limited sensor input (cars have a hard time "seeing" what's around them) and making sense of the world around it.

There is a considerable ongoing research effort in formal verification of learned systems. Once an ML system has been trained, it becomes deterministic and can be tested.

That sounds neat. Can you link me to a paper?

"Computers can only do what they are programmed to do" is the slogan of a 1950s advertising/propaganda campaign by IBM to alleviate fears that the newly-birthed field of AI would create computers which would not only do what they were programmed to do.

Which is precisely what AI is about. Getting computers to effectively program themselves toward some specified goal. In which case, the problem becomes "be careful what you wish for". The paperclip maximiser AI is one such cautionary tale. HAL in 2001 was an exceptionally thinly veiled reference to IBM, no matter how loud Clarke and Kubrick's protestations to the contrary.

On the Media on IBM and AI:

http://www.wnyc.org/story/engineering-intelligence/

There are numerous IBM logos throughout 2001. One is noted in this Slate piece:

http://www.slate.com/blogs/browbeat/2013/01/07/hal_9000_ibm_...

And it points to a now-removed YouTube video pointing out numerous other instances:

http://www.youtube.com/watch?v=K1r5dOwUS6Y

The logos also appear on the Newspad, in various shipboard equipment, and elsewhere.

I would gladly take any number of 9s better than driving myself, but the problem is we're seeing products like Autopilot rushed to market that aren't there. I personally think Google is doing the best job of building a comprehensive system that can reliably replace drivers, and most other players are rushing to try and put something on the road. There's a happy medium where autonomous cars are significantly better than human drivers and still released sometime in the near future, and I'll wait the extra couple of years for a system that is proven safe instead of clamoring for one to be available tomorrow.

You're assuming that software developed "slowly and deliberately" is higher quality than the best engineers and a sense of urgency.

Pretty sure NASA would agree with that assumption.

NASA did get to the moon in under a decade though. I definitely suspect there was a heightened sense of urgency when they were working on that. Sub-orbital to lunar in 10 years is pretty quick.

https://www.fastcompany.com/28121/they-write-right-stuff

NASA puts lots of effort into avoiding any failure and then still encounters failures.

(Nothing wrong with failures if they're anticipated, of course.)

@pb did you watch comma.ai call out Cruise? It seems that comma.ai thinks Cruise lacks urgency, are sell outs. They clearly communicated that Cruise essentially had nothing.

It going to be interesting to see if a outsider like Hotz can outperform a well funded YC startup.

A person who hacked the PS3, iPhone versus whatever Vogt accomplished.

they aren't really competitors, comma.ai offers level 2 autonomous retrofitting for certain cars and Cruise pivoted from initially a similar offering to comma.ai at a higher price point to a level 4 system that uses lidar and is sold to car companies or in this case exclusive to GM with no direct to consumer offering

you think comma.ai is legit? seems kind of pointless ...

Hotz has described his product as 'fancy cruise control', and it isn't much more than that. I think he's about as legit as he says he is.

If he gets it out for under $1000 before 2017 he's a man of his word, and I think he'll do it.

After pulling off that impressive stunt, though, he's going to have to figure out how to turn comma.ai into a real company somehow. I'm not sure how he's going to do that.

Reply on news.ycombinator.com