Comment by halotrope | Hacker News Reader

Comment by halotrope | Hacker News Reader

This is indeed newly listed as a risk factor in connection with FSD features. Items from the previous 10k where about competition and regulation:

- "For example, Model 3 and Model Y face competition from existing and future automobile manufacturers in the extremely competitive entry-level premium sedan and compact SUV markets, including BMW, Ford, Lexus, Mercedes, Volkswagen Group and Volvo. A significant and growing number of established and new automobile manufacturers, as well as other companies, have entered or are reported to have plans to enter the alternative fuel vehicle market, including hybrid, plug-in hybrid and fully electric vehicles, as well as the market for self-driving technology and other vehicle applications and software platforms." [1]

- "In particular, we offer in our vehicles Autopilot and FSD features that today assist drivers with certain tedious and potentially dangerous aspects of road travel, but which currently require drivers to remain engaged. We are continuing to develop our FSD technology with the goal of achieving full self-driving capability in the future. There is a variety of international, federal and state regulations that may apply to self-driving vehicles, which include many existing vehicle standards that were not originally intended to apply to vehicles that may not have a driver. Such regulations continue to rapidly change, which increases the likelihood of a patchwork of complex or conflicting regulations, or may delay products or restrict self-driving features and availability, which could adversely affect our business. " [1]

1: https://www.sec.gov/Archives/edgar/data//1318605/00015645902...

While some of it was almost certainly hucksterism, I think it's also true that a lot of people saw the rapid pace of advance (mostly because of ML computer vision) and just couldn't imagine that the remaining kinks wouldn't be worked out sooner rather than later.

And young urbanites in particular so wanted to believe in a world where they wouldn't need to own a car and would effectively have a personal driver to whisk them around and run errands for them.

Go back a few years and many here were incredulous if anyone suggested we might be looking at decades to have full door-to-door self-driving.

I think in a recent CES talk Mobileye CEO mentioned that Googles models for speech recognition hit a wall after a certain level of expontial improvement. It is not clear that the required robustness and accuracy can be solved with Deep Learning and cameras alone (Not this year or next anyways). I cannot find the exact quote but the talk in itself is very interesting. [1] He is very bullish on vision but says Tesla is making a huge gamble by betting the house on Deep Learning alone. They seem to be far ahead of Tesla in terms of developing a safe and viable FSD solution with the necessary redundancy.

1: https://www.youtube.com/watch?v=h4BXxPvHi60

There's a comment in the last chapter of Russell and Norvig's AI textbook to effect of "We've made steady progress, right up to the top of the tree."

There is certainly a school of thought that we'll need to tap into interdisciplinary fields like cognitive science to continue making real advances. It certainly feels as a consumer that things like voice assistants, ML transcription, computer vision systems are not that different than they were a few years ago.

Yes. The advances in Deep Learning are very impressive and are already proving useful. Still there seem to be no real breakthroughs in actual intelligence, reasoning or something that resembles consciousness.

A lot of money and smarts has gone into studying cognitive processes and neurophysiology over <shudder> the past five decades. (I first took an Intro to Brain Science course in 1976.)

It's actually longer than that. Interesting bit of trivia is that the 1956 MIT IEEE conference that is usually dated as the origin of cognitive science as a field, though the term wasn't coined until later, was just a few months after McCarthy's Dartmouth workshop that's usually considered the origin of AI.

I dipped into a recent MIT class on video last year. As far as I can tell, some pretty basic things--like how do children learn--are as imperfectly understood now as they were in the late 70s.

> It is not clear that the required robustness and accuracy can be solved with Deep Learning and cameras alone

It sure is a good excuse for a bunch of rolling surveillance though.

mobileye previous CEO made a video a few years ago about the theoretical limits to level5, explaining exactly why the number of edge cases made a training approach like neural network impractical.

I hope that everybody will one day give this guy some credits, after the storm he must have faced when elon musk dropped mobileeye stating they didn't move fast enough, and claiming level5 was just around the corner.

As far as I understand Mobileye dropped Tesla because of their reckless marketing. Also they are heavily using cameras but additional sensors for redundancy.

> models for speech recognition hit a wall after a certain level of expontial improvement.

This is the logistic curve.

Solving the 50% case of self driving is something any talented hobbyist can do. Solving for 70% is something that a few companies with billions of dollars will be able to do. Solving for 90% is something that a few companies with billions of dollars of investment may be able to do. Solving for 100% is something that at present no entity on the planet can do. Claiming that you can is disqualifying, it turns you from a 'possible' into a 'must fail' because you lack the realistic constraints required to participate.

Tesla - unfortunately - falls squarely in the latter category, they have set themselves up for failure.

Solving for 100% is something that at present no entity on the planet can do.

Isn't this tautologically true right up until it isn't?

Tesla is claiming it is literally around the corner and lets people pay 10kUSD for the privilege. I think this is the difficult part as non-technical people might take the claims at face value.

Yeah, Tesla are acting like clowns. I was just wondering if the statement really meant anything as formulated, because it is qualified with 'at present'.

If some other company succeeds next week, it wasn't wrong, because the statement is about right now. And so on for any time in the future.

If it happens next week I'll be happy to retract it. If it happens in the next decade I'll be amazed. If it happens in the next century I'd still call it a very impressive accomplishment (only I won't be around to do so).

My prediction is we get to L4 (bounded limited access roads and weather conditions) soon-ish which is a huge win for both convenience and safety. Getting over the "uncanny valley" is a challenge but seems doable. And pretty much punt on dense urban driving as a commercial research project.

Imho, dense urban driving should be solved similar to limited access roads and parking lots with a a combination of pedestrian super blocks (such that Barcelona has). Deprioritize vehicles in urban settings and segregate them from other (more squishy) users. If you can’t AI solve for edge cases, build infrastructure to avoid them.

https://www.vox.com/energy-and-environment/2019/4/9/18300797...

Lots of things become easier when your answer is just redesign cities.

Assume spherical cows and whatnot.

And that is complicated by the fact that, all else being equal, people are generally willing to pay a lot more money for a house in pedestrian-friendly areas. I have friends in Houston, TX and it’s true even there.

If you had a free hand and could design a city with no-pedestrian auto sewer zones and no-FSD-vehicle pedestrian-friendly zones, you are basically going to create the stereotype that FSD is for poor people.

I'm not sure that is true in general. I live an hour outside a city and wouldn't trade a (certainly smaller) place in the city for the same amount of money.

American cities are build around cars- old places like Boston and NYC excluded. And there are a lot less pedestrians and cyclists to worry about.

Yes, but that is the point: Tesla claims that they are within striking distance of that goal, to the point that they ask people to pay for it today. Which is - not to put too fine a point on it - a fraudulent claim.

https://www.cnbc.com/2020/07/14/tesla-autopilot-self-driving...

I'm a little surprised the Michigan or Texas DAs haven't filed anything.

I doubt you ever see 100% if only because there will always be factors outside of a vehicle's control that maybe a newer version could mitigate better. But certainly the bar is very high (and should be).

I think solving 100% is possible but unlikely as I believe it would require two things that I doubt will happen in the next 50+ years:

All cars to universally communicate with each other.

The roads themselves to have sensors relaying information to the cars.

I’m skeptical car sensors alone will ever be able to accomplish the end goal they’re seeking, especially in the face of potential bad actors intentionally trying to break them. The only way I see that changing is with some unexpected monumental leap in technology. Like a 10000x increase in storage and compute power.

Also, any solution requiring the roads to be better and in tip-top condition is a nonstarter. It is really not that hard to find reports of a rich-world jurisdiction where the roads are potholed and the paint is barely visible, and we expect the authorities to change their lack of maintenance all of a sudden?

It is already technically feasible to have cheap mass produced RFID dust directly mixed into the tar/concrete pavement and programmed to supply navigation/road condition information. They don't even need sensors, the cars using the road could update the data instead.

The cynic in me could imagine how the road will inform the car where the potholes are, car will automatically avoid them and the road maintenance can be further postponed.

The amount of additional money in any road budget for this is $0. Are Tesla and Waymo et al supposed to fund this out of pocket? Because otherwise it's not happening.

Perhaps they would serve ads?

RFID is not capable of serving ads on its own. They're basically just wireless barcodes.

In any case, privately funded anything for roads only works on very heavily traveled roads where there's enough traffic to justify trying to reach out to potential customers that way. Your quiet suburban street, mountain road, or rural interstate is not going to generate enough traffic to pay for itself, so it won't happen. And billboards and roadside advertising are not even legal in all 50 states of the US, let alone other countries.

Nice. A road that serves you ads. The future sucks.

How exactly does the RFID road inform the car where the potholes are? More likely the dust moves and the car thinks the road isn't where it should be.

Sounds like there are so many ways this could be abused/vandalized/not work.

They are mixed in when laying the road, so they are firmly embedded in the material. Car that detects a pothole will save the information into remaining chips around it.

Sounds like a great way to pollute groundwater as it passes through the pavement.

So the car is detecting the pothole itself? What's the point then? Cars could potentially do this now without any rfid-in-the-road silliness. What problem is this solving?

Hacker News was full of people somehow assuming progress there would advance linearly, and plainly telling so ("look how much progress we've made in the last 5 years"). Or comparing the problem of autonomous self driving to much, much more solvable problems (plane autopilots: mostly flying straight lines in a featureless void).

Though to be fair, there were also always people in the comments who realized the immense complexity of the problem, suspecting that fully autonomous driving might need either massive infrastructure changes (bringing those cars closer to something akin of an autonomous train), or generalized AI.

> Go back a few years and many here were incredulous if anyone suggested we might be looking at decades to have full door-to-door self-driving.

As someone who has done this kind of push-back here, I can confirm - this place was highly incredulous of, if not downright hostile, to people with experience in AI who were explaining all these were false promises.

Anyone who actually did robotics knew just how much of a pipe dream this is, how slow progress is, and how solving the easy 50% is worth nothing because the real world gets superlinearly difficult. ML made a great jump, but it only solved parts of the problems we needed solving, and so far failed to introduce a bridge to those other parts. But you try telling people, and they won't listen, and now I understand the curse of Cassandra. :)

But I'm reserving special bile for those CEOs and Directors of Research who knew full well that we don't know how to make this stuff, and that most of it is an open research problem that's decades away from productizing - and yet they marketed it anyway, as if it was a done product that's ready to go, and took customers' and investors' money for something that experts knew was very, very, very clearly vaporware.

I'm not an expert in the area but I listen to experts like John Leonard, Rodney Brooks, and Missy Cummings who don't have a vested interest and I draw conclusions.

Trust me. I'd love to have FSD--although it's not without its problems--but I just don't see it in the near future whatever delusional software developers may think.

I know why companies like Uber presented themselves as they did. I'm genuinely curious whether Waymo thought they were closer than they were, drank too much of their own Kool-aid, or just lied.

Rod Brooks is a great example, he's someone who has done more in robotics, across a range of paradigms, than maybe anyone else in the world - and who has been very candid in his criticisms of promises around full self driving.

But it doesn't matter, people choose to believe marketing hucksters instead. :(

> and how solving the easy 50% is worth nothing

"The first 90 percent of the code accounts for the first 90 percent of the development time. The remaining 10 percent of the code accounts for the other 90 percent of the development time".[1][2] — Tom Cargill, Bell Labs

* https://en.wikipedia.org/wiki/Ninety–ninety_rule

... although in the case of self driving, it's not "the remaining 10 percent of the code", it's more like "the remaining 900 percent of the code".

The real world is fractal edge cases.

It sees difficult for anyone who was in the ML computer vision or self driving vehicle field to have made those predictions, which make it all the more problematic that Tesla still decided too, despite having all the information at hand.

Reply on news.ycombinator.com