8 ms·

Of course they did. There's only so much you can do with "feed more data in the machine learning pot, stir it, and see what comes out." We're doing a good job

by Syonyk 5y ago

Of course they did. There's only so much you can do with "feed more data in the machine learning pot, stir it, and see what comes out." We're doing a good job of proving what the limits of ML based self driving are... Computers aren't human, and as it turns out, human brains and visual processing systems are more than the "couple crappy cameras and a bit of a neural network" thing that a lot of SV types seem to believe it is. We're firmly into the "fusion" realm of self driving cars. A lot of money and a lot of time, effort, developers, etc, have been fed into the maw, and we've tried all the easy options - which it's clear don't work. I'll take the pessimal position of saying that the current approach Tesla is using (vision only) can't work, though I'm sure many will disagree with me. It comes down to a difference of opinion about the complexity and capabilities of the human brain. A sensor-heavy setup like Waymo is using has the greatest chance of working, though probably only in heavily-mapped areas where it can use the sort of peripheral data humans use to identify the lanes by building locations and such when the lanes are unclear. As long as not much changes, which probably requires very regular remapping of regions (or an awful lot of data upload from cars), that's the most feasible approach, though it's not a general purpose solution given the amount of data required to do high precision mapping of the entire country, much less the world. We'll see, but "a couple cameras and a car-scale computer" just doesn't seem a solution that's going to work.

tablespoon 5y ago
> Computers aren't human, and as it turns out, human brains and visual processing systems are more than the "couple crappy cameras and a bit of a neural network" thing that a lot of SV types seem to believe it is. Honestly, I think it's something approaching a general rule that most things are far harder than "SV types" believe they are. Maybe it has something to do with all the startups.
coz131 5y ago
I always find it odd that hacker news has such pessimism on technology advancements. It might not work now but give it another 5 to 10 years this may very well become reality. The pace of advancement in ML/AI is incredible.
- Syonyk 5y ago
  Maybe... but probably not. Care to set up a $10 wager for 10 years from now for agreed-on terms about self driving? I'm pretty pessimistic and jaded about computer technology. I've been in the field too long, and done too much computer security. (1) Whatever guarantees you think the hardware is supposed to be making, it isn't. The hardware is now complex enough that just about any security boundary can be violated with enough creativity, if you have enough fancy features to try and keep performance marching forward if you've hit a process wall. Rowhammer. Intel's... everything. The various branch predictor things. Etc. (2) Modern software, as a general rule, is dependency hell. You suck in hundreds of megabytes of code from God knows who, all of which is sucking in yet more, to do something. Nobody has any idea about all the code their project is sucking in, unless they've written it themselves. So, by that definition, nobody can actually understand their code, because some corner case, 15 layers down, can break something in absolutely unexpected and obscure ways. (3) Murphy is an ass, and that corner case will happen, when you least expect it and are least able to debug it. (4) Those piles of complexity, on top of piles of complexity, stacked on top of the previous complexity that exists to solve problems caused by the previous complexity, which was created to get around performance problems due to... etc, just go all the way down. It's all technical debt, and it leads to diminishing returns on investment, or, as I'm beginning to think, negative returns on investment. The more we try, the worse we make things. I'll point to Apple here, and the fact that despite the fact that they know messaging is the literal worst case for untrusted remote input, and that they've built explicit systems to sandbox it, firewall it, safely parse it, etc... it's still been, very recently, vulnerable to remote, zero-click, zero-user-awareness remote exploits that give the remote attacker root level privileges on the phone to do whatever they want and exfil whatever they want. Yay. (5) The "culture" around modern software development is absolutely terrifying when it comes to anything safety critical. Fine, ship whatever when it comes to the latest addictive social media app, and it if crashes, welp. But when it comes to things in the physical world, that's absolutely not acceptable. When Tesla OTA'd updates to the brake firmware of CR's Model 3 that utterly sucked in the braking tests and it got better, everyone was so excited about how, see, OTAs can fix issues! Very few people asked how on earth Tesla had shipped defective braking firmware in the first place, if it was a quick fix. Personally, I'd like my brake system firmware well tested before release, and absolutely "hard" in ROM, so that it can't be updated. Look at the papers from a while back where the security people were able to remotely compromise a Jeep's braking system through remote cellular connectivity and the radio. That's just not OK in the slightest! (6) As Tesla continues to demonstrate, the current way that ML/AI is done is absolutely rubbish for any sort of safety critical system, which automotive control systems are. "Lol, whoops, yeah... hey, thanks for reporting that your car can't take that turn that it used to take just fine!" is the sort of stuff that's funny, until it's tragic. Tesla's software practices are very clearly unsuited to anything remotely resembling the sort of safety critical systems they think they're developing. That's all before you get into the physical limits we're approaching, the economic cost of modern fabs, and the fact that something like 40% of the world's high end silicon production is in a country that China very much appears to be planning to take back. There's an awful lot of money chasing the hope, but I'm (obviously) very skeptical that much of anything useful will come out of it, other than making a bunch of people rich in the process of saying "Well, we're still 3-5 years out on this technology and I intend to retire before you can really call me out on the fact that in 3-5 years, it's still going to be 3-5 years out." I work in the low level weeds, and I see what an utter disaster modern tech is. I'm increasingly using less and less of the stuff in my personal life, and my life is better for it.
mettamage 5y ago
My opinion: How to solve travel for cars: build roads How to solve self-driving for cars: build digitaly aware roads You can’t make self-driving work with a pure black box model. Neural nets need to be able to explain why they work. Currently they can’t do that well enough.
- intended 5y ago
  It’s funny how unpopular this argument was. There is no way level 5 self driving works, unless you are aware of anomalies beyond what the models are prepared for.
- sumedh 5y ago
  > How to solve self-driving for cars: build digitaly aware roads Who is going to pay for it? Are you going to rebuild all the roads, if not then what is the point since your system needs to work on all kinds of roads.
  - mettamage 5y ago
    All good questions. IMO the government, cost benefit analysis needs to be conducted. My perspective is inherently Dutch. I think the roads here get replaced quite often anyway.
    - b112 5y ago
      While not sure of the value, I agree, a change at replacement time is essentially cost free. Paving while laying RFID tag strips would cost nothing per mile, compared to the labour of pulling up and resurfacing.
  - intended 5y ago
    You dont need to rebuild the roads. This is a dead simple idea. You need sensors every x distances, and then use WiFi/mobile data to update cars of the model it should follow or rules of the road. At this point though, you’ve started creating mass transit.
- 5y ago
pkulak 5y ago
> the current approach Tesla is using (vision only) can't work Of course it can. We do it. I'd phrase it myself as: The current approach Tesla is using (ML techniques available now, without major, new breakthroughs) can't work.
- Syonyk 5y ago
  It really comes down to what you think of the human brain. I don't think computers will ever come anywhere close to the human brain in terms of practical understanding and problem solving. Machine learning involves "Give a computer an awful lot of existing labeled data sets and try to get it to figure out, via un-debuggable magic, what they are." The human brain can figure out novel things without having to have had the huge dataset of previous examples, and I'll point to literally any small child as an example (I have two, and one solid rule of parenting is "They're more creative than you would ever expect"). The next question is then one of power efficiency, and there, again, even if we could do something similar to the human brain (which, see above, I don't think is at all likely), can we scale the required bits down to a car's power budget? A 10kW server rack in the trunk isn't really feasible. You seem to believe that anything humans can do, we can eventually design a computer to do. And I disagree entirely.
  - throwawayboise 5y ago
    And even with all of that, humans still get confused while driving. Think about the last time you came upon a construction zone with unclear lane markings, or a traffic pattern change in an area where you drive frequently, almost unconsciously.
    - Ekaros 5y ago
      A few days ago drove over intersection, but didn't turn as usual, was confused by old lane markings that were left there, from quite a time ago. Our road environment is quite far from clean.
    - intended 5y ago
      Absolutely, which is why it should be obvious that self driving is impossible. IF I stuck STOP signs every 5 meters, a human will realize that they can be ignored and its a prank. Unless autonomous cars are trained for precisely this kind of malice they will fail. IF someone decides to hunt autonomous cars, there really is nothing in their models which will save them. They aren’t designed for predator/prey dynamics. When you open up technology to the world, you open it up to the best and worst of humanity. There is no way fully autonomous cars survive.
SV_BubbleTime 5y ago
Automotive EE here: Agree, except you are missing / under selling something. It’s the car sensors that are great, and the human with the crappy cameras. Our eyes suck. Our attention spans suck. Our reaction times suck. To paraphrase my one of my favorite xkcds, “There is nothing in our evolution that could have prepared us for driving heavy boxes at high speeds”. However… somehow, we excel at it. We drive in all conditions despite the crappy inputs and processing. The IDEA that we could be able to make computers do it is sound. The execution is not! And I don’t think it will be for a long time. We had self driving vehicles 100 years ago, your horse would take you home despite how drunk you were. No one wants their car to get spooked of a shadow or floating bag though. The REAL solution to this is that cars need to be trains with figurative tracks. There needs to be an infrastructure in place that reports to the vehicle. V2I (Vehicle to Infrastructure) and V2V. Without that, this is all pointless and I’m surprised how much tome and power is being wasted on it. If the big brains behind this were truly serious, it would be in place now for semis that do the same A-B-A route over and over and can slow down to speed that would annoy a consumer car owner if it isn’t sure. There is just not going to be enough sensors and machine learning to handle no road paint, other drivers, new trolley problem, blinding whiteout, etc. Never will be.
- Syonyk 5y ago
  > Our eyes suck. Our attention spans suck. Our reaction times suck. To paraphrase my one of my favorite xkcds, “There is nothing in our evolution that could have prepared us for driving heavy boxes at high speeds”. I disagree. Our brains have massive plasticity, even into adulthood, and can quite literally rewire themselves for how they're being used. The first time you really get on the throttle on a sportbike, the effect is almost universally a "Woooooaoaaaaaaah!" sort of thing - the universe around you "goes plaid," and it feels like the hand of God has grabbed the horizon and yanked it to you. Hard. And then you get used to it, and people who ride powerful motorcycles or race them are perfectly capable of doing so for long periods of time, at high concentration. Go watch some clips of a motorcycle race, if you haven't. That's not born, that's learned. Same goes for driving cars, for flying planes at Mach 2, for... really everything humans do. I wasn't born knowing how to use a keyboard, but I've been using one so long that I don't even think about it anymore. The human brain is insanely adaptable, which is how it works well in all these cases. And I'll also disagree that our eyes suck. The human visual processing system is damned good at handling a 3D world. It's got a few quirks, but overall? It's properly impressive.
panick21 5y ago
> feed more data in the machine learning pot, stir it, and see what comes out. Very reductive and dismissive way of describing what they are doing.
- Syonyk 5y ago
  Do you have a better explanation for the fact that each release seems to fix a few things that used to not work, and break a few things that used to work fine (like "not deciding there's an object right in front of the car when there isn't")?
martindbp 5y ago
I'm more bullish on Tesla's approach, and it seems like many of you are misunderstanding the approach, or the technology. First of all, it's not a "pure ML approach" as in end-to-end. It's separated into perception and planning. As you can see in the UI now, the perception part is already quite good and gradually getting better. Adding lidar at this point wouldn't help much. What I do think will happen is that they will continue with the philosophy that the car should be able to drive without HD maps in most places (as if driving the route for the first time), but that they will augment the planner with low-res maps generated from other drivers in the hardest spots (e.g. weird intersections that humans also struggle with). They already have the capability to create these maps as seen during AI day, it's just a matter of admitting they are needed in highly complex areas. The second misconception is that the planner is somehow making instinctual decisions based on a simple neural network. Their approach is similar to AlphaZero, which includes Monte Carlo simulations with deep neural networks to predict paths of everyone on the road. I see no reason why AlphaZero cannot eventually beat humans at predicting traffic as well as it can predict moves in Go and chess. A third misconception I see is that there are so many disengagements, so it's clearly not safe. And that's true, it's not safe yet, but Tesla's strength is that they can have thousands of enthusiastic beta testers willing to drive around while paying careful attention, often intentionally seeking out difficult edge-cases. As the number of disengagements and the severity of them goes down over time they can roll out to even more users. Eventually they'll be safer than humans on average. I think that is likely to happen in a few years based on their rate of progress. What happens at that point is completely up to politics and human psychology. In summary, if you actually go through and watch a couple of videos, I challenge you to find examples of disengagements that are complete show stoppers that can't be solved within 2-3 years. Also, go check out videos of earlier version so you can get a sense of the rate of improvement, which really is the only thing that matters (and it's massive). You can argue that they will plateau somewhere below human level, and that's certainly a possibility, but I wouldn't bet on it.
- yellow_lead 5y ago
  I'm not bullish on Tesla's or Waymos approach. A Tesla may be able to predict other drivers, but how can the planner handle a situation like i.e a cop directing traffic in a particular way, or a complicated /non-standard sign that hasn't been seen before? I know these are edge cases but even one disengagement is a big fail if your car doesn't have a steering wheel (Elon's hope)

by Syonyk 5y ago

tablespoon 5y ago

> Computers aren't human, and as it turns out, human brains and visual processing systems are more than the "couple crappy cameras and a bit of a neural network" thing that a lot of SV types seem to believe it is. Honestly, I think it's something approaching a general rule that most things are far harder than "SV types" believe they are. Maybe it has something to do with all the startups.

coz131 5y ago

I always find it odd that hacker news has such pessimism on technology advancements. It might not work now but give it another 5 to 10 years this may very well become reality. The pace of advancement in ML/AI is incredible.

Syonyk 5y ago
Maybe... but probably not. Care to set up a $10 wager for 10 years from now for agreed-on terms about self driving? I'm pretty pessimistic and jaded about computer technology. I've been in the field too long, and done too much computer security. (1) Whatever guarantees you think the hardware is supposed to be making, it isn't. The hardware is now complex enough that just about any security boundary can be violated with enough creativity, if you have enough fancy features to try and keep performance marching forward if you've hit a process wall. Rowhammer. Intel's... everything. The various branch predictor things. Etc. (2) Modern software, as a general rule, is dependency hell. You suck in hundreds of megabytes of code from God knows who, all of which is sucking in yet more, to do something. Nobody has any idea about all the code their project is sucking in, unless they've written it themselves. So, by that definition, nobody can actually understand their code, because some corner case, 15 layers down, can break something in absolutely unexpected and obscure ways. (3) Murphy is an ass, and that corner case will happen, when you least expect it and are least able to debug it. (4) Those piles of complexity, on top of piles of complexity, stacked on top of the previous complexity that exists to solve problems caused by the previous complexity, which was created to get around performance problems due to... etc, just go all the way down. It's all technical debt, and it leads to diminishing returns on investment, or, as I'm beginning to think, negative returns on investment. The more we try, the worse we make things. I'll point to Apple here, and the fact that despite the fact that they know messaging is the literal worst case for untrusted remote input, and that they've built explicit systems to sandbox it, firewall it, safely parse it, etc... it's still been, very recently, vulnerable to remote, zero-click, zero-user-awareness remote exploits that give the remote attacker root level privileges on the phone to do whatever they want and exfil whatever they want. Yay. (5) The "culture" around modern software development is absolutely terrifying when it comes to anything safety critical. Fine, ship whatever when it comes to the latest addictive social media app, and it if crashes, welp. But when it comes to things in the physical world, that's absolutely not acceptable. When Tesla OTA'd updates to the brake firmware of CR's Model 3 that utterly sucked in the braking tests and it got better, everyone was so excited about how, see, OTAs can fix issues! Very few people asked how on earth Tesla had shipped defective braking firmware in the first place, if it was a quick fix. Personally, I'd like my brake system firmware well tested before release, and absolutely "hard" in ROM, so that it can't be updated. Look at the papers from a while back where the security people were able to remotely compromise a Jeep's braking system through remote cellular connectivity and the radio. That's just not OK in the slightest! (6) As Tesla continues to demonstrate, the current way that ML/AI is done is absolutely rubbish for any sort of safety critical system, which automotive control systems are. "Lol, whoops, yeah... hey, thanks for reporting that your car can't take that turn that it used to take just fine!" is the sort of stuff that's funny, until it's tragic. Tesla's software practices are very clearly unsuited to anything remotely resembling the sort of safety critical systems they think they're developing. That's all before you get into the physical limits we're approaching, the economic cost of modern fabs, and the fact that something like 40% of the world's high end silicon production is in a country that China very much appears to be planning to take back. There's an awful lot of money chasing the hope, but I'm (obviously) very skeptical that much of anything useful will come out of it, other than making a bunch of people rich in the process of saying "Well, we're still 3-5 years out on this technology and I intend to retire before you can really call me out on the fact that in 3-5 years, it's still going to be 3-5 years out." I work in the low level weeds, and I see what an utter disaster modern tech is. I'm increasingly using less and less of the stuff in my personal life, and my life is better for it.

mettamage 5y ago

My opinion: How to solve travel for cars: build roads How to solve self-driving for cars: build digitaly aware roads You can’t make self-driving work with a pure black box model. Neural nets need to be able to explain why they work. Currently they can’t do that well enough.

intended 5y ago
It’s funny how unpopular this argument was. There is no way level 5 self driving works, unless you are aware of anomalies beyond what the models are prepared for.
sumedh 5y ago
> How to solve self-driving for cars: build digitaly aware roads Who is going to pay for it? Are you going to rebuild all the roads, if not then what is the point since your system needs to work on all kinds of roads.
- mettamage 5y ago
  All good questions. IMO the government, cost benefit analysis needs to be conducted. My perspective is inherently Dutch. I think the roads here get replaced quite often anyway.
  - b112 5y ago
    While not sure of the value, I agree, a change at replacement time is essentially cost free. Paving while laying RFID tag strips would cost nothing per mile, compared to the labour of pulling up and resurfacing.
- intended 5y ago
  You dont need to rebuild the roads. This is a dead simple idea. You need sensors every x distances, and then use WiFi/mobile data to update cars of the model it should follow or rules of the road. At this point though, you’ve started creating mass transit.
5y ago

pkulak 5y ago

> the current approach Tesla is using (vision only) can't work Of course it can. We do it. I'd phrase it myself as: The current approach Tesla is using (ML techniques available now, without major, new breakthroughs) can't work.

Syonyk 5y ago
It really comes down to what you think of the human brain. I don't think computers will ever come anywhere close to the human brain in terms of practical understanding and problem solving. Machine learning involves "Give a computer an awful lot of existing labeled data sets and try to get it to figure out, via un-debuggable magic, what they are." The human brain can figure out novel things without having to have had the huge dataset of previous examples, and I'll point to literally any small child as an example (I have two, and one solid rule of parenting is "They're more creative than you would ever expect"). The next question is then one of power efficiency, and there, again, even if we could do something similar to the human brain (which, see above, I don't think is at all likely), can we scale the required bits down to a car's power budget? A 10kW server rack in the trunk isn't really feasible. You seem to believe that anything humans can do, we can eventually design a computer to do. And I disagree entirely.
- throwawayboise 5y ago
  And even with all of that, humans still get confused while driving. Think about the last time you came upon a construction zone with unclear lane markings, or a traffic pattern change in an area where you drive frequently, almost unconsciously.
  - Ekaros 5y ago
    A few days ago drove over intersection, but didn't turn as usual, was confused by old lane markings that were left there, from quite a time ago. Our road environment is quite far from clean.
  - intended 5y ago
    Absolutely, which is why it should be obvious that self driving is impossible. IF I stuck STOP signs every 5 meters, a human will realize that they can be ignored and its a prank. Unless autonomous cars are trained for precisely this kind of malice they will fail. IF someone decides to hunt autonomous cars, there really is nothing in their models which will save them. They aren’t designed for predator/prey dynamics. When you open up technology to the world, you open it up to the best and worst of humanity. There is no way fully autonomous cars survive.

SV_BubbleTime 5y ago

Automotive EE here: Agree, except you are missing / under selling something. It’s the car sensors that are great, and the human with the crappy cameras. Our eyes suck. Our attention spans suck. Our reaction times suck. To paraphrase my one of my favorite xkcds, “There is nothing in our evolution that could have prepared us for driving heavy boxes at high speeds”. However… somehow, we excel at it. We drive in all conditions despite the crappy inputs and processing. The IDEA that we could be able to make computers do it is sound. The execution is not! And I don’t think it will be for a long time. We had self driving vehicles 100 years ago, your horse would take you home despite how drunk you were. No one wants their car to get spooked of a shadow or floating bag though. The REAL solution to this is that cars need to be trains with figurative tracks. There needs to be an infrastructure in place that reports to the vehicle. V2I (Vehicle to Infrastructure) and V2V. Without that, this is all pointless and I’m surprised how much tome and power is being wasted on it. If the big brains behind this were truly serious, it would be in place now for semis that do the same A-B-A route over and over and can slow down to speed that would annoy a consumer car owner if it isn’t sure. There is just not going to be enough sensors and machine learning to handle no road paint, other drivers, new trolley problem, blinding whiteout, etc. Never will be.

Syonyk 5y ago
> Our eyes suck. Our attention spans suck. Our reaction times suck. To paraphrase my one of my favorite xkcds, “There is nothing in our evolution that could have prepared us for driving heavy boxes at high speeds”. I disagree. Our brains have massive plasticity, even into adulthood, and can quite literally rewire themselves for how they're being used. The first time you really get on the throttle on a sportbike, the effect is almost universally a "Woooooaoaaaaaaah!" sort of thing - the universe around you "goes plaid," and it feels like the hand of God has grabbed the horizon and yanked it to you. Hard. And then you get used to it, and people who ride powerful motorcycles or race them are perfectly capable of doing so for long periods of time, at high concentration. Go watch some clips of a motorcycle race, if you haven't. That's not born, that's learned. Same goes for driving cars, for flying planes at Mach 2, for... really everything humans do. I wasn't born knowing how to use a keyboard, but I've been using one so long that I don't even think about it anymore. The human brain is insanely adaptable, which is how it works well in all these cases. And I'll also disagree that our eyes suck. The human visual processing system is damned good at handling a 3D world. It's got a few quirks, but overall? It's properly impressive.

panick21 5y ago

> feed more data in the machine learning pot, stir it, and see what comes out. Very reductive and dismissive way of describing what they are doing.

Syonyk 5y ago
Do you have a better explanation for the fact that each release seems to fix a few things that used to not work, and break a few things that used to work fine (like "not deciding there's an object right in front of the car when there isn't")?

martindbp 5y ago

I'm more bullish on Tesla's approach, and it seems like many of you are misunderstanding the approach, or the technology. First of all, it's not a "pure ML approach" as in end-to-end. It's separated into perception and planning. As you can see in the UI now, the perception part is already quite good and gradually getting better. Adding lidar at this point wouldn't help much. What I do think will happen is that they will continue with the philosophy that the car should be able to drive without HD maps in most places (as if driving the route for the first time), but that they will augment the planner with low-res maps generated from other drivers in the hardest spots (e.g. weird intersections that humans also struggle with). They already have the capability to create these maps as seen during AI day, it's just a matter of admitting they are needed in highly complex areas. The second misconception is that the planner is somehow making instinctual decisions based on a simple neural network. Their approach is similar to AlphaZero, which includes Monte Carlo simulations with deep neural networks to predict paths of everyone on the road. I see no reason why AlphaZero cannot eventually beat humans at predicting traffic as well as it can predict moves in Go and chess. A third misconception I see is that there are so many disengagements, so it's clearly not safe. And that's true, it's not safe yet, but Tesla's strength is that they can have thousands of enthusiastic beta testers willing to drive around while paying careful attention, often intentionally seeking out difficult edge-cases. As the number of disengagements and the severity of them goes down over time they can roll out to even more users. Eventually they'll be safer than humans on average. I think that is likely to happen in a few years based on their rate of progress. What happens at that point is completely up to politics and human psychology. In summary, if you actually go through and watch a couple of videos, I challenge you to find examples of disengagements that are complete show stoppers that can't be solved within 2-3 years. Also, go check out videos of earlier version so you can get a sense of the rate of improvement, which really is the only thing that matters (and it's massive). You can argue that they will plateau somewhere below human level, and that's certainly a possibility, but I wouldn't bet on it.

yellow_lead 5y ago
I'm not bullish on Tesla's or Waymos approach. A Tesla may be able to predict other drivers, but how can the planner handle a situation like i.e a cop directing traffic in a particular way, or a complicated /non-standard sign that hasn't been seen before? I know these are edge cases but even one disengagement is a big fail if your car doesn't have a steering wheel (Elon's hope)