Anthony Levandowski: Pronto AI | Hacker News Reader

Anthony Levandowski: Pronto AI

medium.com

77 pointsedward38 comments

"Over the past 18 months, as I’ve been painted into a villainous caricature"

You are right. The transfer of Google confidential files to Uber was a mere accident, they somehow ended up on a USB stick that you lost at Uber and someone at Uber found it and re-used it.

Edit:

I saw this gem: "We are not building technology that tells vehicles how to drive. Instead, our team of engineers is building tech that can learn how to drive the way people do."

For the love of $DEITY, don't build tech that learns how people drive. A lot of people drive like selfish jerks.

> Instead, our team of engineers is building tech that can learn how to drive the way people do."

I find this to be the 'faster horse' of the self-driving car world. It seems like a good idea for a first step automation but if you take a step back, most traffic problems are precisely because of how people drive.

> One thing I continue to stand by is my safety record, which is second to none in this industry. My teams have a demonstrably impeccable track record.

Please explain:

https://www.businessinsider.com/anthony-levandowski-google-s...

Over the past 18 months, as I’ve been painted into a villainous caricature...

Gratuitously unethical behavior will do that, my friend.

My willingness to believe anything this guy’s got to offer is, frankly, quite low. For some actual reasons that involve federal prosecutors, IP theft, and a host of other obviously bad ideas.

You can't fall as far as Levandowsky did without first climbing pretty high. He was instrumental in getting the Google self driving car project running, he is quite talented and driven when he sets his mind to it.

I’d agree, he’s quite insightful and a very creative engineer that is quite rare and few can match, sadly he also has a Gordon Gecko side too.

He's talented, let's pretend all the real shitty stuff he did never existed. It's settled guys...

> You can't fall as far as Levandowsky did without first climbing pretty high.

Yes, you can't fall as far as he did through discovery of theft and lies without first climbing very high by theft and lies.

I mean, we're kind of wandering off the main thread of things but in principle you can if you climb high by other means and are separately engaged (before or after your climb) in theft and lies.

You can climb and fall through completely different means, but in those cases, the fall usually doesn't invalidate the climb. People weren't dismissing ReiserFS because of its author's other actions.

Fair.

Can we apply this logic to Mo Zedong, so that he can be named the one of the greatest person of the 20st century?

      I’ve reconnected with a few old and very qualified friends and gotten to know the best young minds 
      in our industry that have fresh, truly innovative perspectives. 
      Together, we are building the best self-driving stack in the world.

      We are not building technology that tells vehicles how to drive. 
      Instead, our team of engineers is building tech that can learn how to drive the way people do.

Has Anthony Levandowski partnered with George Hotz?! This is exactly the same idea that Geo Hotz had for his self-driving startup Comma.ai!

"Levandowski clearly has a gift for self driving technology and is probably one of the few people in the world who could spend a year with 20 collaborators and come up with a fully self driving product. He clearly demonstrates a deep knowledge of the autonomous driving landscape and its future.

But he’s also clearly a flawed individual who not only ran off with a bunch of company secrets from Google after taking boatloads of its money, but also he’s run afoul of regulators on numerous occasions and locations. It is going to be hard for any individuals and companies to put their drivers behind Pronto.ai with this past in mind. His “religious” beliefs may also cause some concern."

https://electrek.co/2018/12/18/anthony-levandowskis-new-pron...

If I remember correctly his new 'Religion' is merely a tax shelter for the big payout he got from Uber purchasing Otto

The villain has a point: "The reason why nobody has achieved this level of functionality is because today’s software is not good enough to predict the future. It’s still nowhere close to matching the instincts of human drivers, which is the single most important factor in road safety."

For any model that makes predictions (be it an SVM, deep network, or 50ksloc of threshold functions written in C++), there are some set of parameters that one can use to tune the tradeoff between expected real-world true positives and false positives. SDC engineers have purposefully picked threshold values to favor conservative driving behavior and avoid crashes.

See for yourself:

In this video, the Waymo isn't avoiding a crash so much as potentially causing a rear-ender: https://youtu.be/vJovT6nsxB0?t=14 I am no Waymo engineer, but the car is clearly mis-predicting the behavior of other threats in the intersection.

Waymos tend to be "scared" because they're simply biased against being confident in their predictions about how clear the road is in front of them: https://www.youtube.com/watch?v=6itruDO57bk

Same with awkward lane changes. Waymos (and many other SDCs) have issues here because their engineers have convinced themselves via extensive manual testing and debugging that the cars would be less safe were they more "confident" in their predictions.

The "villains" of self-driving know what's going on. The best we can do is to try to out-compete them; let them make us stronger. Passion for technology and passion for society are largely orthogonal. We want the winner to be somebody who gets both right.

> In this video, the Waymo isn't avoiding a crash so much as potentially causing a rear-ender: https://youtu.be/vJovT6nsxB0?t=14

wtf. This is a shitty example. If you either (1) read the captions on the video or (2) watch their full video where they talk about this, you'd realize that the waymo car stopped because a car off-camera ran the red into the intersection and that braking actually was avoiding a crash (or rather, decreasing the severity of a crash).

https://youtu.be/spw176TZ7-8?t=41

I've watched this video several times. There is a red car clearly present in the intersection when the Waymo begins forward motion. That car partially occludes (and is a potential obstacle) for anything that's off to the left of the camera; it also is simply a dynamic obstacle in the intersection. The Waymo can certainly detect and predict if cross-traffic is going to roll through the intersection. But the Waymo started launching right when the light turned green... even though there's a car in the intersection at that time.

I agree this is a bad example due to the camera angle, but there's enough evidence to show that the Waymo failed to correctly predict traffic here. And it executed a safety stop (which could have caused a rear-ender) as a result.

Occluders are still really hard for SDCs. Humans can usually handle them well though. Handling occluders properly requires a very strong prediction system.

Lol and he touts his safety record like the shit he was directly working on didn’t kill somebody a few months after he left. This whole blog reads like somebody still in denial.

120 million dollar exit package and still a jerk.

If being a jerk gets you a $120 million exit package, why would your behavior change?

>We are building neural networks from the ground up that combine experience-based AI, end-to-end deep learning, and crowdsourced data with advanced computer vision to deliver a highly scalable and flexible driving stack. Nobody else is doing this.

I.E. precisely what Comma.ai has been working on for four years now. Except they actually have thousands of volunteers, millions of miles of crowdsourced data, and multiple hardware products for sale already.

And isn't Tesla also already doing this at a large scale (with owners being forced to give up their driving data)? And probably other car makers, but more quietly?

And what's the advantage over also using extensive simulation to add tons of experience quicky?

>Over the past 18 months, as I’ve been painted into a villainous caricature, I’ve had lots of time for introspection.

Anthony then proceeds to show literally no level of introspection. The designs he was working on at Google went on to mysteriously turn up at Uber. The designs he was working on at Uber went on to cause someone's death. Arguably his false claims leading to the acquisition of Otto went a long way to sinking Uber's self-driving cars. All either ignored or white-washed. This is clearly not a man capable of self-reflection, which seems like a pretty basic problem.

Any investor would have to ask themselves - if he succeeds will I get a pay off? Or will he abscond with the IP again. Any engineer would have to ask themselves - will I be on the team that finally pays the price for this guy's complete lack of concern for basic safety practices.

If at first you don't succeed, try try again?

This is kind of rediculous. He's pitching Otto + more magic deep learning juice. Is there sufficient evidence deep learning will solve these higher order planning tasks? Unless he has access to machine learning research not yet published, this is not a solved problem and better suited for a research focused environment than a startup promising engineered systems to investors.

Okay fine, maybe deep learning is a sufficient solution. It's patently false to claim no one else is doing end to end deep learning. The fact he is ignoring Wayve, one of the most interesting approaches in this space, is telling he has yet to get over his own ego and genuinely understand what is happening in the field. And if you're to go based on the excitement at NeurIPS, expect to see a bunch more deep RL papers and startups on this subject.

>"Is there sufficient evidence deep learning will solve these higher order planning tasks?"

What are those higher order planning tasks? I'm not familiar enough with the problem domain.

> Our approach? Much better software.

Oh. So that's the secret ingredient all the others have overlooked.

Here's my take on it:

He has lost all credibility as a trustworthy individual. Anything he builds will be seen as tainted by IP stolen from Google (whether true or not). Most companies will avoid doing business with him for the fear of being pulled in some lawsuit and/or bad PR. This will put a major drag on his company in attracting engineers, investment, etc. He is definitely an extremely talented engineer, but his problems may be beyond any technical ingenuity that he can come up with.

Ahh yes, when safety is your key concern, you absolutely want to make heavy trucks in excess of 15000 kg your test bed. What the actual fuck.

Maybe start with a self-driving trike.

Wasn't one of the reason Waymo (Google) figured out he took the files because one of the vendors accidentally contacted Waymo instead of Uber when they did an RFQ (since the designs were almost the same). I wonder how different things would be if that didn't happen.

> We are building neural networks from the ground up that combine experience-based AI, end-to-end deep learning, and crowdsourced data with advanced computer vision to deliver a highly scalable and flexible driving stack.

Oh yes, that's precisely what we need. An end-to-end black box NN driving our cars. So that when this completely alien mind causes an accident, and it will cause an accident, we'll have no idea what happened, or why, or what to do about it (except shove in more training data and keep our fingers crossed).

I have the distinct feeling that GPUs today aren't used to give intelligence to the machine, but to save developers from the need to think about the problem they're solving.

> (except shove in more training data and keep our fingers crossed).

You could also tune some parameters until the crash no longer occurs on that specific scenario, and then find some excuse why these parameters should have been chosen a priori.

>"I know what some of you might be thinking: “He’s back?” Yes, I’m back."

Exactly nobody is thinking that dude.

This person's sense of self importance is incredible. I guess the "hardships" he claims to have experienced the last 18 months failed to teach him any humility.

If this guy can still get some thing positive from this, like some vc findings, I'll be thoroughly disappointed at the society's judgement on good and bad people...

For people who downvoted my previous comments about why it’s better to fail big and spectacularly than small, this is a good example of why.

Why is this in the front page of HN?

Why does Silicon Valley celebrate thieves/theft to sociopathic behavior? Why is that accepted as its culture?

I think it's more schadenfreude than celebration.

This is newsworthy, although for atypical reasons.

Reply on news.ycombinator.com