Just wait until they admit that they need LIDAR too... Cha Ching! Various Autopilot accidents strongly imply that whatever the computer “brain” is doing doesn’t matter if it doesn’t have sufficient “eyes” to actually see things like a fire truck, cop car, or bollard.
And yet squishy meatbags manage to sensor fuse a stereoscopic image (plus nascient sound based imaging) into a reliably rendered picture of the world...
Our eyes are incredibly advanced compared to any camera in existence, and hooked up to a real life AGI, and we still fuck up. No comparison. We meatbags are still far more sophisticated and battle tested than anything said meatbags managed to build, which is why so much of our building attempts to replicate our own capabilities.
Yeah, and they do that with dozens of years of accumulated knowledge. Do not have the impression that what you perceive with your eyes is the photons hitting your retina. All that input is heavily fused with decades of experience in dealing with people and the world.
If I trained a child to drive, I'd be comfortable letting them drive at 6. I believe I was around that age the first time I drove (handling steering, with assistance on gas / brake). My memory is that the only difficult part was a lack of power steering on hilly terrain.
Same age for walking around town unsupervised: the Japanese do and don't seem to lose many.
My original point was focused on the technical feasibility of driving with limited sensor input, which is physically possible.
Whether or not it takes a week, a month, or a million experience-years is a computational and algorithm problem.
All of which (based on previous technological progression) appear to be tractable.
And my point is that it's not just about experience-years, it's also about the breadth of knowledge. When driving, humans do not use just the experience acquired from looking at things from behind the wheel. They also use their general understanding of physics, of materials, of concepts related to visibility, traction, not to mention expected behaviour of other people and how to read face expressions.
For an average driver, I'd call BS on that. Humans are terrible at learning complicated things they rarely have to recall.
E.g. the disaster that is the first major snowfall, every year, or automatic transmissions, antilock brakes, and traction control systems becoming standard
The human can only reliably be trusted to keep the car between the lines, stop appropriately, and occasionally make turns. Which is something much simpler to compete with!
When your "existential crisis killer sentient ai" becomes as smart as this meatbag I'll listen to Elon musk. Till then he and you are out of touch with what computers are capable of
As someone who works in the field, I'd offer two points.
Our intuition about the future is linear. But the reality of information technology is exponential, and that makes a profound difference. If I take 30 steps linearly, I get to 30. If I take 30 steps exponentially, I get to a billion. - Ray Kurzweil
People once said the same thing about fire lances.
That squishing grey thing in your skull is pulling something in the vicinity of a petaflop of computation in a 10 watt thermal budget. Biological nanotech is a hull of a drug!
They seem to have a rather low number of mistakes per billion miles,* so I don't know if more sensors are actually required. I suspect the larger issue is "To Err is Human; To Really Foul Things Up Requires a Computer" https://quoteinvestigator.com/2010/12/07/foul-computer/
The current model makes assumptions that sometimes turn out to be false. So yea it's not going to get to perfection, but level five only needs better than human not perfection.
* Granted I am likely a bad driver, but I don't think I could do this well.
That number always comes up, and it always needs to be said that a billion miles with driver assist and a billion miles under autopilot are not the same, and it’s a bogus figure. As to better than a human, it needs to be better than the humans who buy luxury sedans, sober, and awake. Better than a random selection of humans in a random selection of cars is no good. Conflating all vehicle deaths is like conflating all gun and knife violence; the average person on the streets of Chicago has far less risk of being shot than someone in a gang, or engaged in the commission of a felony. Yet you can make all of Chicago seem like a war zone by conflating the risks of different populations and dumping the raw statistics.
As a matter of practical politics, that actually seems to be what the goalposts are. It might be better if things were different but in practice that's the standard a self driving car is going to need to meet to be accepted and allowed on the road.
This cars require the driver to pay 100% attention, by law and by Tesla agreements so isn't obvious that driver+drive assist+expensive car has better average numbers then driver + average (maybe old) car ?
I would like to see the numbers of how many times a driver had to intervene to save the situation
It's an option so insurance companies compare the exact same car with and without that option.
In terms of paying attention I really doubt most people are doing this very well. I am going to be stuck in the car either way so by default I am going to be watching the road, but as long as my relative stress level drops that's a huge win.
PS: You can make your own estimates. If people pay attention sufficient to avid accidents 99% of the time the car might be in 100x more accidents without drivers assuming those drivers where perfect. Drop that to IMO more likely 75% and it's closer to 4x at the absolute best case.
But if say me+ a computer program together defeat you at chess you can't conclude that the computer is better then you, just that the computer helped me maybe in some tricky situation , you can't conclude that the computer will win without me.
Same with Tesla, this is why I would like to see the number the drivers had to intervene and prevent crashes, without this number the only conclusion is about driver asist and not autopilot/self driving
IMO the difference between self driving and driver assist is who is usually in charge of the car. If the car is going to hit the breaks because it thinks you are going to hit someone otherwise that's driver assist. If you as a driver need chose to hit the breaks because you don't trust the car that's self driving.
Further, the difference between OK and an accident is normally fractions of a second. So, my suspicion is humans are really bad at this role and only catch the most obvious cases. AKA crap I am an a corn field, not hmm it's not slowing down enough I need to apply slightly more breaking power or I am going to hit that guy simply because you don't have time to react with the second situation.
Having said that, humans may be preventing a lot of issues well before they turn into accidents. The 'wanted to make a wrong turn down a one way street' is something a person can deal with easily, though it's not necessarily going to cause an accident.
Sure, that could still happen. But, they introduced it in 2014 and they are still on the road so they seem to be past the wildly unsafe stage I was expecting.
IMO, the largest risk is an over the air update killing a large number of people.
>PS: You can make your own estimates. If people pay attention sufficient to avid accidents 99% of the time the car might be in 100x more accidents without drivers assuming those drivers where perfect. Drop that to IMO more likely 75% and it's closer to 4x at the absolute best case.
You can't invent numbers and statistics.
Remember the Tesla that crashed in a side barrier and killed the driver, people could reproduce the incident in the same spot , you can see it on youtube, so without a driver and with say 1000 Teslas driving trough that section in the problematic interval you would get at least 1000x more deaths.
Only if you include the people that reproduced the incident mentioned above you will get the Tesla stats much below human drivers.
Computer failures are reproducible, but you don't get "1000 Teslas driving trough that section in the problematic interval" which shows up in the actual statistics.
Computers fail in different ways than humans do. But, that does not mean you can't reason about failure modes. If you ask someone to pay attention to something for an hour without doing anything most people are just really bad at this. Based on that I am rather shocked how well Tesla's systems work as I was expecting vastly more problems.
Could I have flipped to far in the other direction probably. But, I base things on my experience and the data I have available not just unchanging gut feelings.
Again with the LIDAR argument... Humans drive using only eyes (computer vision) and having a sense of location (localization/3D maps).
Early autonomous test vehicles used LIDARS because camera and compute tech are not at the same level they are today. It's a legacy system and more of a shortcut. In essence, Object recognition > LIDAR judging distances.
Yeah it's true that a vision based system is much harder to solve. However if Tesla cracks it, it will pay dividends.
Again with the LIDAR argument... Humans drive using only eyes (computer vision) and having a sense of location (localization/3D maps).
As has been discussed in some depth here before you commented, computer vision hooked up to some ML doesn’t come close to human eyes controlled by a human brain trained for many years.
Early autonomous test vehicles used LIDARS because camera and compute tech are not at the same level they are today. It's a legacy system and more of a shortcut. In essence, Object recognition > LIDAR judging distances.
Waymo is the cutting edge if you believe what people here say on thread after thread about them, and the raw stats. I’m sure they’ll be interested to hear your view of their “legacy” tech.
Yeah it's true that a vision based system is much harder to solve. However if Tesla cracks it, it will pay dividends.
Finally, reality. Yes, if they do something they’re yet to do, and their cars stop ramming stationary objects, then maybe something else will happen. No one is contesting that, there are just doubts given the state of Autopilot vs. leaders in the field, not to mention Tesla’s dodgy PR in general, and in the face of fatal incidents in particular.
LIDAR provides more than just distances. With full-360 LIDAR you also get object edges in 3d space, which means you can do object recognition on the LIDAR results.
In other words, you get everything that vision systems would offer except color, without having to do any processing to calculate depth, so you can go straight to the object-recognition and navigation processing.
Humans have general intelligence, that's why they recognize lane dividers and can tell where a truck ends, we kind of expect such things. Also why inebriated people can drive somehow (of course they shouldn't but I'm making a point).
Maybe driving on the roads we have now with computer vision alone needs strong AI, maybe simpler standardized roads and vehicles are needed, who knows. I think that strong statements about this should be backed by evidence, not gut feeling. And I'd like to see that evidence coming from the company that promises their customers their cars have all the hardware they need.
Does it matter if Tesla does the hardware swap on existing vehicles for free? Costs should not be high to do so if it’s their own silicon and the swap is done as part of regular maintenance at Tesla service centers.
The benefit of high margins and sophisticated engineering is you can get away with a lot of unknowns and promises.
Elon does not fall of anything. He will just admit that he was wrong and move on.
Its only Elon Musk haters that somehow want to proclaim every mistake of him as some sort of fall from grace where finally his auro of diseption will be lifted from his victims.
When you buy a Model 3 you can choose to pay up front for self driving. I'm assuming that this would be an included upgrade if you did so, so who cares.
This is pretty bad from a hardware/software verification perspective. People who care about correctness can just go home, I guess. It's amazing that we're gonna let these things on the streets. When you weigh almost 2 tons and can go >100mph, it doesn't matter if you have a gun mounted on you or not, you are a weapon.
I think you're right if you look at things from a statistical perspective but I don't think it's going to end up actually working out like that.
A person can explain "why" they did something when an accident happens. I've spent a lot of consulting hours helping unwind the "why" of ML/AI models, and we bill regardless of whether we can actually find that answer.
Putting aside by personal bias that I think Tesla is functionally incapable of doing anything well, lets pretend they deploy FSD - I think the fact that in that 1 accident to every 100+ human accidents you can't find "why" will override the fact that there are xx% less accidents.
Just from my experience consulting in healthcare and working with actuaries during the Obamacare debates and modeling out the impact of that -- people just honestly aren't open to statistical arguments when it comes to human life or emotional issues.
(Not saying they're wrong, me and my wife completely disagree about so many issues where it's emotions vs statistics, I think she has a very valid point on many of those - what I'm trying to say is this is going to be a philosophical debate more than a statistics debate so Tesla should gear up accordingly).
does correctness even work with deep neural networks to begin with? especially with the existence of adversarial examples, out of domain situations, etc
And we should be able to do better with our technology, instead of throwing shit ton of compute at black boxes no one understands, hoping it will all somehow work out in the end.
Yeah because all the verified neural networks made from meat travelling >100mph are legit ok to drive 100% of the time. There are no road accident deaths to speak of.
there is something fundamentally different between a machine and human. how the brain operates is not well known, so to claim our brain is a neural network made of meat is just a guess.
the truth is, people have a lot of experience with other people. we can usually predict behavior of other drivers on the road, and can even predict pathological behavior or know how to handle it. that is a major point. when machines and software fail, they can fail in big, unpredictable ways. humans fail in fairly predictable ways. none of us have this experience with neural networks or machines.
i barely trust my computer to work on a daily basis, much less a vehicle.
and lastly, when a human is driving, fault is usually clearly and easily assignable. how fault is assigned in the case of machines driving is not clear at all.
The fault hardly matters, the issue is one of safety. Humans are not safe already. Yet you feel entirely fine on the road. Humans are clearly not the same as ANNs but they are a network of neurones and the cognitive ability of a driver is not verified anywhere near frequently enough globally. In the UK post driving test about the only time anyone will check you for ability to drive is when you are declared blind from an eye test. But its not like you need to mandatorily have an eye test regularly, or any other kind of test. Unless you want to dispute a 70 year old human brain is not on average in the same kind of cognitive shape as a 30 year old human brain https://sci-hub.tw/10.1111/j.1532-5415.2012.04059.x Yet we have no policies saying over 70s cannot drive, or can only drive in good conditions etc. The verifiability of a driver is not something we have ever demanded before. Even when you could have done plenty for safety when the driver was a person.
i do not feel perfectly safe or fine with humans driving. i hate driving in the area i live because no matter how much i pay attention, it is simply an unsafe environment with the amount of cars, scooters, bicyclists, and pedestrians within the given infrastructure. however, my ability to predict both good and bad behavior in these environments helps a lot. there is no way i see automated vehicles handling city driving in dense urban areas any time soon.
of course safety is important. and i agree that verifying driving ability is well behind where it should be. elderly and inexperienced and inattentive and just plain bad drivers are major problems on the roads.
and fault certainly does matter. aside from legal responsibility, we are emotional beings. there is a difference between getting hit by a drunk driver versus a true accident. that affects people in real ways. as a motorcycle driver, i will be terrified of these automated vehicles. i am already terrified of human drivers, but i can often predict their idiocracy.
no one wants to be killed by a machine, especially one built by wide-eyed engineers.
That is only an argument for more assistive solutions. Just because humans have faults doesn't mean we should replace them with computers that are very likely to be even more dangerous. Augment, not replace should be the aim. And seems to be the aim for most manufacturers, but not Tesla. They have to claim full autonomy.
In the same sense that any object that can potentially kill someone (hint: everything) is a weapon. Which is technically correct but a useless observation.
If we use a more narrow definition of weapon, e.g. a tool optimized for the primary purpose of injuring or killing people then it certainly is not a weapon.
> In the same sense that any object that can potentially kill someone (hint: everything) is a weapon. Which is technically correct but a useless observation.
It's not useless; the damage that can be caused varies by degree between things. Then you have to look at two cases - suitability for object to be used as a weapon, and the damage it can cause on accident. Unlike knives or hammers, both those factors are very high for cars.
The fact how dangerous cars is is very much underappreciated by people in general, as evidenced by the number of morons on the road. We already lose hundreds of people daily in the US alone because of this; now we're trying to add another class of drivers into the mix - algorithms written by greedy optimizers caring primarily for short-term profit and being first to market. This should give us some pause.
I'm not saying this technology is not possible or not wonderful, but I think the current ecology of self-driving efforts is unhealthy. We have a for-profit race by companies, many of which can't be trusted with getting software right, and most (all?) of them pursuing self-driving capabilities by means of half-understood brute-force black boxes the neural networks are.
I was quibbling about semantics. I am aware that cars can be dangerous.
> and the damage it can cause on accident
You are conflating (un)safety of a tool used in the way it is intended (kitchen knife = cutting a steak) with accidents (cutting a finger) and with malicious use (stabbing people). Those three categories are not the same for object-that-may-act-as-weapon and object-designed-as-weapon.
Conflating them collapses the number useful things we can communicate.
So are you concerned about Tesla intentionally building killing instruments? Or potential for accidents? Or the potential for intentional misuse?
> The fact how dangerous cars is is very much underappreciated by people in general, as evidenced by the number of morons on the road.
Cars also provide immense utility. If all they did were providing the thrill of speeding then they would probably be banned as too dangerous. One of the tradeoffs is the overhead of enabling people to drive. We could drive down the number of morons by requiring astronaut training for vehicle operators but again, that tradeoff seems too harsh and it's more efficient to occasionally let people die in traffic accidents than letting them die because nobody qualified as ambulance driver.
> We have a for-profit race by companies, many of which can't be trusted with getting software right
In the short term this may cause more deaths than necessary. But on the other hand it might be the quickest way to find a winner and then hold the rest to the same standard. As long as the experimental fleets are small they are just a blip in the statistics. Right now they should be equated to the yearly batch of first-year drivers who have an inherently higher risk profile due to lack of experience. We still accept them on our roads in the expectation that they improve.
What is important is to make sure that they are as good as or better than humans once they roll out in large fleets.
> So are you concerned about Tesla intentionally building killing instruments? Or potential for accidents? Or the potential for intentional misuse?
The latter two.
> We could drive down the number of morons by requiring astronaut training for vehicle operators but again, that tradeoff seems too harsh and it's more efficient to occasionally let people die in traffic accidents than letting them die because nobody qualified as ambulance driver.
I don't think this is the real reason. You don't need astronaut-level training for vehicle operators, just more than the ridiculously low standard of today, and more importantly, much stronger and harsher enforcement of traffic laws. I doubt that this will reduce the number of qualified ambulance drivers.
I suspect the real reason we tolerate so many morons on the road is path dependence. When cars first appeared, they were rare, slow and safe. In the couple of decades it took to get to the present density and speed of cars, it became a social status symbol, and something politically impossible to rein in.
> What is important is to make sure that they are as good as or better than humans once they roll out in large fleets.
I'm afraid that with self-driving tech based on neural networks, with no ability to inspect and verify what's going on, we'll eventually have to eat the risk and roll them out in large numbers before we know they're as good as humans.
I do not agree that neural networks are a "black box" with "no ability to inspect and verify". Even putting aside the many methods to understand what a neural network is doing without running it, at core, neural networks are well tested instruments. That's how they learn-- by testing themselves.
Obviously it's possible for a neural network to have odd behavior in circumstances not accounted for but that was always going to be possible at the level of complexity we're talking about here.
We're talking about cutting edge technology here -- and I agree with your general sentiment. I just don't agree with pinning the blame on "... based on neural networks". The same factors would apply to any codebase of this complexity.
> Even putting aside the many methods to understand what a neural network is doing without running it
Name three :).
> neural networks are well tested instruments. That's how they learn-- by testing themselves.
Last I checked, neural networks are well-tested in a sense that if you throw a big database and a shit ton of compute at them, they'll learn to accurately work within that database. Step out of it, and all bets are off. We're better at this than we were 30 years ago - good enough to apply this technology to consumer-level products in which mistakes don't really matter. I'd be wary of applying even current neural networks to safety-critical tasks.
> Obviously it's possible for a neural network to have odd behavior in circumstances not accounted for but that was always going to be possible at the level of complexity we're talking about here.
The problem is that with NNs, the odd behavior is usually totally unexpected, and you can't really inspect the network beforehand to discover the possible ranges of error-generating inputs. Everything works fine but every now and then you get a patterned sofa classified as a zebra, or a car + little noise classified as a toaster. And then there's no obvious relation between multiple misclassifications, because the reasoning structure of the neural network is implicitly encoded in its weights.
> The same factors would apply to any codebase of this complexity.
I think there's a fundamental qualitative difference here. A codebase can be complex, but ultimately it has a structure, and usually (in case of ML) represents a well-understood mathematical structure. Neural networks have simple code, and the whole complexity is hidden in opaque matrices of numbers, where even single changes usually have global effects.
I'm not trying to dismiss NNs in general; I just don't trust them in applications where health and safety is at stake.
>Much stronger and harsher enforcement of traffic laws.
I'm sure that making the primary means of long (greater than walking) distance transportation for the majority of the population more expensive and higher stakes is going to work out great in the long term. I can see the parallels with healthcare. Creating yet another part of life where a single screw up that is capable of ruining the financial well being of someone living slightly better than paycheck to paycheck is not going to do positive things in the long run.
I'm not thinking about screwups, I'm thinking about reckless behaviour and endangering people. For example, treating speed limits as suggestions instead of hard constraints, overtaking in places where it's not allowed.
>I'm not thinking about screwups, I'm thinking about reckless behaviour and endangering people
But who determines which is which? The letter of the law really, really, really sucks when it comes to traffic law. I haven't collected data but I'd wager that pretty much nobody follows the letter of the law for an entire drive from A to B
>For example, treating speed limits as suggestions instead of hard constraints
That's more human than reckless. Outside of places with Orwellian enforcement (I'm looking at you Europe with all your cameras) they really are. The vast majority of people go at a speed they feel comfortable in the conditions. This is why (conditions permitting) traffic flows at 80 even when the sign might say 55. People only follow the speed limit when they feel it's a comfortable speed. This is why the general recommendation is to set speed limits for the 90th percentile speed. If you don't do this on highways you get people doing the (inappropriately low) speed limit in the wrong lane. Passing on the right, tailgating and all the other things caused by traffic friction which is more stuff for drivers to keep tabs on and that decreases safety for all. There have been studies on this (Google "traffic friction" and filter out everything that has to do with literal friction). If anything speed limits on multi-lane roads should be raised to reflect the speeds people actually drive. I hope we'll see more dynamic speed limits in the future since they'll help a lot.
>overtaking in places where it's not allowed.
If every 100th instance of an illegal pass (usually on the shoulder when waiting for someone who's stopped to take a left turn or on the right on the highway) in my state resulted in a ticket it would probably be about a year before half the state's drivers hit the three strikes cutoff and had their licensees revoked.
The picture I'm trying to paint here is that aggressive enforcement of existing laws would probably be bad for the population at large because people break traffic laws in inconsequential ways all the time and that stronger enforcement of them would just screw people over (unless of course you enforce them so well that important people get screwed which would result in the laws changing) and discretion isn't an answer because that just results in profiling.
I worry less about software / hardware crashes leading to death and more about cyberattack. I've written about it before[0]: We need regulations before a mass breach of autonomous systems happens. Not after. This whole bullshit where corporations claim all sorts of things about their security then get owned trivially has happened for every OS and device out there. The supposedly secure QNX was installed into countless tanks, nuclear power plants, and autonomous devices before getting owned in 2017[1]. I don't buy this "just trust us" attitude from automotive and other autonomous manufacturers. One of them is going to fuck up at some point.
YES. Security compromise is the huge achilles heel of self-driving cars - and anything that is algorithmically controlled.
When you have 100,000 self-driving cars on the road, and someone figures out how to hack them, and drive them simultaneously into crowds, that will be a BIG BIG problem.
And this problem isn't just about self-driving cars. It is a fundamental issue with all algorithms. It's a lot harder to hack humans to do bad things (although it can be done with sustained messaging and propaganda). But algorithms can be compromised and exploited.
> It's a lot harder to hack humans to do bad things
I'd say, it's probably easier to hack individual humans, but - like everything in computing - hacking algorithms scales well, while human factors generally don't.
Let us not also forget, we also want to prevent governments from having the ability to shut down them down as well. While there may be a need to disable a car in a chase or such I am against the idea of wholesale shutting down a highway full of cars or similar scenarios.
Has a flight ever been brought down through a cyberattack? I've heard a lot of concern over that over the years, but (thank goodness) I don't recall any news of an actual flight being brought down.
Pilots have the ability to manually disengage airplane autopilots, and more importantly the time to do so (on the order of minutes). It's very different in a car, where the difference between life and death can be a split-second reaction.
There are multiple safeguards for planes that don't work for self-driving cars.
First, they're not physically inspectable by every Tom, Dick, and Harry. This means it's easier to obscure interfaces and other potential areas of attack. Not state-actor proof, but probably ISIS-proof. Second, they only get software updates while parked at an airport, so it makes MITM attacks harder (though not impossible). Third, they're in the air where death isn't a sudden movement away. Pilots can override the system and fly it manually. Fourth, they're heavily monitored with errant flightpaths reported to militaries around the world (to stop another 9/11).
We have none of these safeguards for self-driving cars and there are going to be hundreds of millions of them.
We already have a number on embedded processors and controllers in cars today, including systems responsible for critical operational parts of the car.
How autonomous systems will make anything different? Why do you believe people aren't currently hacking cars?
I don't intend to be critical at all, I believe this is a very interesting point too that should be discussed. (I hope to check the links on my interval).
> We already have a number on embedded processors and controllers in cars today, including systems responsible for critical operational parts of the car.
> How autonomous systems will make anything different? Why do you believe people aren't currently hacking cars?
It might make always on data connections near-mandatory (to get maps data, etc). At least in my car, the insecure embedded computers are air-gapped from the internet.
> How autonomous systems will make anything different? Why do you believe people aren't currently hacking cars?
Part of the reason is that, in current ecosystem, companies will find ways in which self-driving requires being constantly on-line and connected to vendor's server. Off-line processing is not in fashion these days.
Off-line processing is absolutely in fashion with SDCs. Can't quickly react to a pedestrian jumping in front of your car when you have to bounce 12 1080p video streams to a AWS server.
Fashion or not, the speed of light is not going anywhere - latency can only be added, never removed. If the choice is "process locally or receive stale data", it's not actually a choice.
For example, chips have flaws. Hardware designers work incredibly hard to verify that their chip designs are functionally correct, because the cost of a correctness bug can be catastrophic. With the proliferation of custom hardware designs, verification of the whole stack is even more important. It's bad from the perspective that now instead of building on top and reusing a chip company's huge investment in a verified product, there is a custom product to which an unknown level of verification expertise has been applied. Doubly so because these chips are designed for approximate workloads, so they cut corners left and right on correctness for efficiency. People should be more skeptical.
I think his point is that you aren't backing up the statement that the chips are unverified. Nor does your statement that they are built for approximate workloads imply that the chips are unverified or cutting corners.
No it isn't. There are already standards in electronics for Functional safety and these chips will have to abide by them. Part of that is that there'll be fail-safe chips that are fully ASIL-D compliant, and the full system will be ASIL-D classified as well. The difficult part of this problem is the hardware, not the functional safety - we've got that pretty much well defined by now.
The benefit they point out is 200fps vs. 2000fps, but is that actually that useful? In the post of OpenAO's robot, they said that improving the response times gave no significant gains.
Only thing I can imagine would be allowances for wider and deeper neural nets.
More data points can hurt: it may just waste time processing data which doesn't improve results but it can actually make things harder if you're just adding more noise which has to be filtered out but there's no corresponding improvement in signal.
As a simple example in a different domain, years back I found that the best way to improve Tesseract's OCR accuracy was to ensure that I didn't feed it images at more than 150dpi because it would sometimes misrecognize dust, paper texture, etc. as characters.
My understanding is big problem is stationary objects which mostly get filtered out. With more processing power you might be able to do something about that.
My question is where the break-even point is between adding more resolution / frames-per-second versus performing more expensive analysis. Given that the human visual system is not especially high resolution or speed, I would bet that they're already at the point of diminishing returns on the sensor side and the big wins would come from more expensive processing techniques, which https://news.ycombinator.com/item?id=17671843 supports.
>2000fps will provide a lot more wiggle room to do higher level computation on each frame.
Yeah I think this is it. My best guess, it's 200fps vs 2000fps for the same workload. As Tesla's data model gets larger and its computations get more complex, that 200fps will start to sag.
2000fps will provide a lot more wiggle room to do higher level computation on each frame.
This is exactly right. They didn't mention it in the article, but I was on the call and Elon stated that one of the main benefits was that they could now process all of the camera images at full resolution and full framerate (implying they can't do that with the GPU).
Maybe that chip was not important in itself but was just a good opportunity to improve in house expertise in chip development. Having expertise gives a lot of power during negociations with providers.
> Only thing I can imagine would be allowances for wider and deeper neural nets.
...or more cameras, with better frame rates over a lot of cameras. My guess is that the fps rate is over all the cameras, so 10 cameras means going from 20fps per camera to 200fps per camera.
I think latency makes a lot of difference when an autonomous system is interacting with the world. Take a look at this video from 2009 where they use a high speed vision system (1000fps!). The things it could do are probably unequaled even today.
The higher the speed of the system, the less it has to predict in the future and better can adapt to unforeseen situations. It's basically using the world as a model, for free.
“Nvidia’s hardware was handling about 200 frames per second, its specialized chip is able to do crunch out 2000 frames per second “with full redundancy and failover”.”
What are the real-world benefits going from 200 to 2000 frames a second? 200 seems quite fast.
Nothing seems that fast at 50 m.p.h., however, excuse the mix in units, but each of those 200 frames is the equivalent of 10cm or so along the road. At 2000 frames per second then that is the equivalent of 1cm.
I don't make a habit of crashing cars but I did have one glancing head on crash, combined speeds around 125 m.p.h. and it was only centimetres that meant I did not wipe out that family and trash the car I was in.
As a consequence I would say that this 10x frame rate really is quite game changing, particularly for on-coming vehicles.
I think that comparison is just to illustrate the computational improvements. In other words, the new system is 10x faster. As Elon said on the call, that means they can now process all the cameras at full resolution and full framerate, which can improve accuracy and therefore ability and safety... which sounds like the real world benefits you're looking for.
The most interesting question about this is what their motivation was to justify building an in-house ASIC design team. That isn't a super hard thing to do for a company of Tesla's size, but it is nontrivial and seems like a significant investment. It is probable that they got massive performance improvements by specializing their hardware, but it isn't clear that they couldn't have gotten the same perf improvement by upgrading to the next gen nvidia system (drive Xavier or something like that, the one with Volta + NVDLA).
If you have your own ASIC it's a differentiator that's hard for other players to compete with. They'll take this to potential investors (or stockholders the "market") and say this is a big expensive thing we have that nobody else has. It's expensive and time-consuming for them to replicate and we also own a lot of IP that they would need.
You and I might know it's bullshit. That's not true of the investment market in general.
It's scary when a company with chronic implementation issues decides to take on a hard problem that is way outside its core competency.
If they succeed, it means slightly improved efficiency and cutting NVidia's profit margin out of their cost structure. If they fail, it would cost billions, and obscure chip bugs could put lives at risk. That risk/reward doesn't seem worth it unless you really know what you're doing and are confident you will succeed.
Of course, it fuels the hype machine in the short term.
"It's scary when a company with chronic implementation issues decides to take on a hard problem that is way outside its core competency."
This seems like the biography of Elon Musk's companies. He gets a crazy hard problem that people say he and his companies can not do, then he tries to implement it, often falling well behind and well short of the goalposts due to lack of foresight and overambitious scheduling. There were naysayers the whole way, including just as loud as there are now if not louder, and yet his companies' achievements and networth continue to climb.
What crazy hard problems have his companies solved?
Spaceflight was solved decades before Elon was born. Reusable rockets weren't a hard problem--they were simply a problem the incumbents were unwilling to address because it would have massively cut into their revenues and profits.
Electric cars actually predated ICE cars. The issue with EVs was always the charging infrastructure, which Tesla solved...by simply throwing a lot of money at it. (Not hard, just resource intensive.) Their batteries are built by Panasonic.
Boring Co literally is just a used tunneling machine. They have literally not done anything with it except test it out below the SpaceX parking lot (and the innovation in boring would come not from the tunneling but with the post-tunneling construction of the tunnel walls, stations, dirt removal, and ultimate extraction of the boring machine).
Even at Paypal, their biggest innovation was developed by Musk's biggest pre-Paypal competitor (Thiel's company) before they merged to form PayPal.
I'm willing to back up my statements with my existing login, and I don't need to be an engineer to read the history of resuable launch systems. They've been around since the 1960s, and space-capable launch systems were a thing in the 1980s. They were killed the first time around by the collapse of the USSR, and then again when the Iridium went bankrupt, eliminating the demand.
You have nothing to back up the statement that "reusable rockets are not a hard problem", and let me back up this claim of mine by pointing out that only partial reusability exists today. In addition I invite you to build reusable rocket systems, sir, since you have already claimed that the problem is "easy". You could make some good money solving that problem, and you would once and for all put this argument to bed.
Everybody including every other space CEO, people from all levels of NASA and pretty much everybody that is innvolved in Space is giving SpaceX top marks on innovation on almost every single level, production, operation, technology and so on.
But of course you know much better how simple all these things are. For you to disagree with that is just emberecing yourself.
The same applies in a lesser degree to your other comments. Do you seriously belive that the origonal EV in 1900 was the end of development and all Tesla did was 'throw money at it'? That is the hight of stupidity and you again wont find a single expert on the subject agreeing with you.
Also the battery chemstry is co-licenced by Tesla and Panasonic with the design of the cells and even part of the chemistry being inhouse at Tesla. Panasonic can not sell this technology.
So when you are spreading baseless FUD, at least get your information correct.
You seem to have lost all rationality in context to of this question.
>It's scary when a company with chronic implementation issues decides to take on a hard problem that is way outside its core competency.
Not quite... The chip was designed by Jim Keller, who was a key player in designing the highly successful AMD Zen architecture and Apple's A4 and A5 SoCs (pretty much a legend).
Jim is known in the industry as having an unquenchable thirst for really hard problems. Solves them, leave the company to recharge and goes on to the next company/project.
It has become easier to make ASICs in-house than it used to be, particularly if you're trying to do something relatively straightforward (versus trying to build a fully featured, backwards-compatible x86-based cpu architecture). The teams which developed Bitcoin mining ASICs were pretty small, for instance (although that's an extreme case of simplicity). You don't need your own foundry.
>By having its own silicone, Tesla can build for its own needs at its own pace. If they suddenly recognize something the hardware is lacking, they’re not waiting on someone else to build it. It’s by no means a trivial task — but if they can pull it off without breaking the bank (and Elon says it costs them “the same as the current hardware”), it could end up being a significant strength.
I outright snorted at this. Time to market is not why you build a custom chip instead of software. Is this article written by a complete bonehead?
Tesla has it's own hardware platform. Anyone with an ounce of sense will know that was always going to happen because of power consumption, cost and performance. The Nvidia Drive platform is designed to be a quick way to get to market with FuSa. The requirements for fully autonomous modes are estimated by some to be over 10x higher than the top CPU/GPU offerings.
But the materials are not even remotely similar. Silicon is a metallic element. Silicone is a polymer. Anyone who would make that mistake clearly has zero idea what they’re talking about, and it makes me question the entire article.
And both words have a one letter difference, which makes an easy typo, and could slip past an editor, I suppose. You’re centering on the disparate meaning of the words, leading to incredulity, when you might do better to focus on the question of how such an obvious mistake slipped through the cracks.
IOW, yeah, it’s a crap article, but not for this reason.
I don't know; it's not really surprising that a polymer has a name reflective of the metalloid that is the key element along with oxygen in its backbone chain.
They also published the quote about iPhone leading in performance and go-to-market because of their custom SoC efforts, and they mentioned Tesla's SoC program manager Pete Bannon, but they didn't even bother to connect the two ideas: Pete Bannon was the guy who led Apple's SoC efforts to develop the first 64 Bit ARM chip for mobile.
They also didn't mention what Elon said about now being able to run all the cameras at full resolution and full framerate.
">By having its own silicone, Tesla can build for its own needs at its own pace ... "
This is not what Musk said on the conference call. Surely some bonehead reporter is writing this.
Musk said the latter, precisely to go from current 200 fps speed to something like 2000 fps (I might have the numbers wrong, but it's a 10x improvement) and that they have been doing this in stealth for the last 2-3 years. They expect the cost of the new chips to be the same as that of the current nVidia lineup. That got me thinking though. What are they currently using - the Titans ?
The board that was shipping since Oct 2016 is a drive px2 custom. Basically, a pascal gp106(gtx 1060 equiv) and a denver 2+cortex a57 cpu. I think hardware 2.5 added another parker chip. Figure $400-500 max per board.
>Tesla has it's own hardware platform. Anyone with an ounce of sense will know that was always going to happen because of power consumption, cost and performance
... so that they can create performance-intensive features that nobody else in the market yet has, because the off-the-shelf hardware won't allow it. Due to power consumption, cost and performance issues. And this custom hardware is being developed to solve them.
You're basically saying the same thing in different words. I mean, what else do you do with better power consumption, cost and performance other than new, better-working features? You think Tesla will sell cars to the mass market with impressive computer specs?
Apple's own custom hardware has enabled it to experiment with performance-intensive features unlike other smartphone manufacturers, Tesla is going for the same strategy.
Other people are going to be using inferencing chips soon too. Tesla might even just be basing it all on NVDLA (Nvidia's open source inferencing hardware targeted at self driving) and calling it their own.
We made it easy to switch out the computer, and that’s all that needs to be done. You take out one computer, and plug in the next.
Will existing Tesla owners be able to swap the computer? Or is the swap happening only inside the factory, to the next generation of the Teslas they're building?
This may be one way in which Tesla's approach differs from Apple's.
Can future gen GPUs bridge the gap? The NVIDIA CEO seemed really confident when asked previously about the challenge that specialized AI chips posed. Is his confidence misplaced/just bluster or what?
The hardware platform for self-driving cars isn't just Sensor -> AI Chip -> Steering wheel. There will be custom chips for accelerating some of the hardest parts of the algorithm, but there's no reason not to think there'll be a significant Nvidia chip in there too.
Probably current gen GPUs can bridge the gap. The nvidia platform after Drive PX 2 (which is what Tesla ships) uses Volta (with "tensor cores" to speed up fp16) and has a deep learning accelerator (http://nvdla.org) for inference (i think both int8 and fp16).
what does he mean with regards to GPUs and CPUs "emulating"? they are computation machines doing computations. an algorithm running on one or the other isn't emulating anything in this context. running on "bare metal" is just an optimization along a certain axis. it isn't any more real of a computation.
and talking about this now seems like another desperate attempt at generating hype.
I think he is saying that CPUs and GPUs are generalized computational mechanisms. They have increasingly complex overhead that attempt to optimize general problems.
By limiting the context of use, specialized hardware can perform better than CPU/GPU. I think his thinking relates strongly to ASIC crypto mining hardware https://en.bitcoin.it/wiki/ASIC .
"bare metal" is just a shorthand for the level of access to hardware that does not require a kernel/OS.
Ultimately its down to dimensionality CPUs (at least originally) were designed for scalar calculation, you loaded something a piece at a time from RAM and did a single scalar calculation before putting it back into RAM, rinse repeat. A GPU works in a more vectorised/matrix level way with essentially a pipeline of linear transforms on a set of data values all at once. The issue with both of these two is they still have a linear memory model, because that works for more serial lower dimension operations, usually three dimensions for a GPU maybe a couple more. Networks or graphs as objects aren't neatly linear in memory or nicely shaped into regular dimensions for doing pipelined vectorised calculations. Though the GPU model is a little closer than the CPU one. Graph algorithms its not random access or serial access either its /structured/ access based on the graphs connections. So ideally you want memory shaped and optimised for this sort of access, as well as computation well linked to utilise this rearchitected RAM. Many companies have been exploring newer architectures for dealing with ANN calculations that inherently benefit from changing the model of memory access and computation. So you can "emulate" as in reformulate the mathematical representation of a graph into a list or a matrix, an adjacency list based algorithm would work well on a CPU and an adjacency matrix based one would work better on a GPU. But neither of those fit perfectly to either model they are just dimensionally the nearest with many tradeoffs you have to make for each.
As a naive example, if in a CPU you are doing some specific 3-step algorithm which would normally require 3 clock cycles, you can instead bake that algorithm into a circuit along with the memory it needs that will run in 1 cycle. In Elon's terminology, the CPU has to "emulate" what the baked-in circuit simply is.
But from the sounds of it this chip goes beyond that. It sounds like some kind of RAM/GPU hybrid with matrix-specific instructions specifically targeting NN inference for Tesla's specific style of NNs, which as a first guess I take to mean "lots and lots of layers."
In the call they mentioned increasing their investment in the chip team and technology, which is a sign that they're confident it's going to pay off.
what i was getting at that the term "emulation" is rarely used, at least from my experience, in this context except in cases where something is truly faking it like game emulation, OS emulation, hardware emulation, etc. i personally feel like musk is using this term as a marketing and hype technique, that all of the sudden CPUs and GPUs are "faking it" for these applications.
Right now, with HW2 Tesla is in kind of an awkward no man's land where their hardware is overkill for basic ADAS, but nowhere near enough for any kind of autonomy that allows the driver to turn their attention away from the driving task.
By going to a custom ASIC with inferencing. I have little doubt about Musk's claim that it'll offer a 10x improvement over the current stripped down PX Drive2 variant for deep learning tasks. If I'm to decrypt Musk's branded terminology and take him at his word, then they'll be going from about 8 TOPS to 80 TOPs with that they'll be able to get even deeper into the no man's land that's even more overkill for basic ADAS but still not nearly enough to handle full self driving safely in all conditions, though maybe they'll get L3 highway out of it.
Nvidia's upcoming Pegasus board will supposedly do 320 TOPS, and their competitor, Intel/Mobileye's EyeQ5 will likely have comparable specs. These motherboards are designed to replace Robotaxi trunks brimming with ~$150k in liquid cooled compute that draw 3-4000 watts.
It's hard to say for sure they are no where near full autonomy in their research although their current vehicles crashing into barriers and fire trucks obviously don't cut it.
I'm only commenting on what little there is to glean about the compute, but there are so many other red flags about their autonomous development strategy. Crippling per-unit cost constraints, No Lidar, SLAM-lite, ANN heavy, insufficient redundancy and they've got a major data collection problem because they're relying on what they can get over the air. Plus Musk scared away most of their their top engineers.
Serious question- why is LIDAR so important? Obviously it’s possible to drive without it (humans do it), so it’s conceivable that eventually machines will be able to do it as well... So I’m guessing your point is, that won’t happen anytime soon and LIDAR is the fastest way to market; can you confirm/expand/clarify?
The standard of safety for autonomous drivers will have to be far beyond parity with humans. Humans cause car accidents all the time. LIDAR gives unambiguous depth and probably much easier object segmentation, no complexity or reasoning required like computing from stereo images.
Good point, plus if you have to choose between having LIDAR and acceptable self-driving capability in 3 years, or no LIDAR and acceptable self-driving capability in 10 years, I think most people would prefer the first option.
Estimating distances robustly is very important for navigation. LIDAR is accurate at relevant distances and most importantly very robust against environmental effects. Without LIDAR you can use stereoscopic vision (like human with two eyes) but that is very demanding, not that accurate, and very error prone. Sure, human can do it somewhat well, but eye and brain are extremely complex things to implement (and it still takes years to learn to understand what you see).
Personally, I’d refuse to implement self-driving vehicle without at least a ”backup LIDAR” to check vision system results. Otherwise you are forced to assume stuff like ”things at stand still are either above the road, beside it, or just shadows”, causing crashes when there is suddenly a stopped car in front of you. (If you didn’t do that assumption, you would be dodging shadows and other clutter..)
[source: I’ve been researching vision algorithms in a related field.]
>> Without LIDAR you can use stereoscopic vision (like human with two eyes) but that is very demanding, not that accurate, and very error prone.
Because that isn't how people judge distances, at least not distances beyond a few feet in front of their faces. Plenty of people with only one eye do very well. It is a difficult problem because we use a variety of techniques and 'hardware' when estimating distances and speeds. Car companies are trying to do with one tool (ie lidar) something we do with many.
Yep, we do huge amount of assumptions to derive ”model of a world” from quite limited amount of data. And we do lots of mistakes without ever realising it. Fortunately, most of those mistakes are irrelevant, and safety margins let us correct most of the relevant mistakes. Rest become accidents.
I guess the same logic applies directly to self-driving cars as well..
> Without LIDAR you can use stereoscopic vision (like human with two eyes)
At any but very close distance, don't humans mostly use a combination of lighting cues, a priori knowledge of actual size vs. apparent size, motion parallax, and other flat-image cues instead of stereoscopy?
You can use stereoscopic using the full width of the car as the baseline instead of eye distance. Humans have a pupillary distance of about 60mm which is good for about 10 meters (possibly much more: https://jov.arvojournals.org/article.aspx?articleid=2191614). A Model 3 is over 6 feet in width, so the pupillary distance is thus 30 times that of humans, and so should be good to about 300 meters, which is comparable to high end LIDARs (although the stereoscopic approach won't be as precise at those distances).
Anyway, if LIDAR becomes small and cheap, Tesla can just strap it on.
That is until it rains and LIDAR falls flat on its face. LIDAR Is Great for training in perfect rainless conditions. For everything else we’ll have to use other tech, very likely camera based. Which is what tesla is doing.
I find the "humans can do it with eyes" argument pretty weak. Humans also have brains, which are doing most of the heavy lifting. With something like LIDAR, you're shifting more of that heavy lifting into the sensors, so that you don't have to go as deep into trying emulate the human brain in processing. It's such an obvious point.
Yeah, a direct a distance measurement is much more straightforward than the integration of all the monocular and binocular depth cues human perception relies on.
Musk commented about LIDAR in a prior earnings call[1]. Tl;dr he believes that LIDAR is a short-term crutch that distracts from the real long-term challenge, which is the machine learning, not the sensors.
Elon mentioned on the Q2 conference call yesterday that specific demo is easily doable.
He also added that it would be sort of cheating, because that's a very defined route and he prefers it to be able to be much more dynamic. i.e. pick any two US cites and the car will drive you there.
I'm not the one claiming that Autopilot is able to make a cross-country trip but delaying the test because it would be "too easy" to prove anything. If it were actually that easy, Tesla would have done it, because Musk has never passed up the chance for that sort of publicity.
Ergo, logically, if Tesla has not performed such a test, it is strong evidence that Autopilot is not currently capable of passing it.
Elon said the exact same thing in the Q4 2017 earnings call. And then said they would still do the demo in the next 3-6 months. It's been 6 months and this "easily doable" demo has still not been done. Tesla has shown zero evidence of their self driving program making any meaningful progress.
>Tesla has shown zero evidence of their self driving program making any meaningful progress.
Google's autonomous efforts far precedes Tesla, yet they have not out in the market yet. I would say on the contrary, they made significant strides in a short amount of time even after a complete overhaul Mobileye (AP1) to an in-house solution (AP2/2.5) https://vimeo.com/192179727
Has Tesla demonstrated any progress whatsoever since that 3 minute demo video they posted over a year and a half ago? If they were really making amazing progress they would have shown it off. Tesla LOVES showing off their tech when it's actually working, and it is often very impressive.
Waymo has been operating a real actual passenger service for over a year now. They drive tens of thousands of miles a day. They're "not out in the market yet" only in the sense that they're not charging money for it yet, which they seem to plan on doing soon.
That's navigation, and that's actually a solved problem. Actually following the roads...oh look, that's a nice set of traffic cones. What could that possibly mean? IDK, just continue through it.
Waymo pretty top-secret about their hardware, so I'm not sure. I would venture to guess that if they don't have a custom ASIC yet they're working on one. I know they had Intel FPGAs and it's a safe bet they've integrated Google's tensor core architecture.
A bunch of details about their lidar came out in the depositions, and lots of juicy gossip about their internal politics, power struggles and backroom dealings. I don't recall reading anything about their compute.
Waymo's advantages include LIDAR and extremely detailed maps, IIRC. If I had to guess, those maps probably help them distinguish between roadside objects that happen to be static right now and ones that are really not going anywhere.
I ordered a Tesla Model 3 yesterday. When I see the "Full Self-Driving Capability" checkbox at checkout time (it's an option you can pay $3,000 for right now) I just have zero confidence that they can deliver on this. Almost every other aspect of the car is fantastic. But Tesla's delusions about achieving full self-driving with their current approach just make me lose some respect for them. It makes me worried that they might be deluded about other things too.
I use the word "deluded" above realizing that I might be completely wrong. If they pull it off I'll eat crow and give them props. But I just don't see it happening.
Are the model 3 hardware modular? For example, is it possible they’ll have you bring it to the dealer in a few years and get outfitted with these new chips?
Yes, Musk said that on the earnings call. But the hardware upgrade is not the hard part of self driving. At least they have stopped claiming that their current hardware is powerful enough for full self driving already, which was a ridiculous and worrying position for them to take.
I think “fraudulent” would be more accurate; it was factually false, it was at best entirely unfounded if not actually knowingly false, it was material to decisions of people to purchase the self-driving add-on.
> At least they have stopped claiming that their current hardware is powerful enough for full self driving already
The Autopilot page [1] still states the following:
"All Tesla vehicles produced in our factory, including Model 3, have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver."
If you consider “hardware” as sensor suite only (quite a stretch of an interpretation, I know), then it’s kind of true, especially if they will replace the brains at no cost once they have an actual use for them
This is a commonly missed point. You can buy the car today and be very happy with a fantastic car as it is. I did and I'm quite happy, you'll probably be very satisfied as well. I didn't even buy the EAP, as I don't do much highway driving and I think almost no one actually buys the FSD.
Having said that, I find it hilarious that people on here have such immensely strong opinions about something that doesn't exist. They'll go at length about how LIDAR is a hard requirement, about how many FLOPs you need, etc etc etc. Fact: FSD doesn't exist and no one has it. Nobody knows which approach will work and if more than one approach works no one knows the timelines or economics of the different approaches and very certainly no one knows the market impact of these. But people start thinking they know the future and then become very emotionally vested in their make-believe stories.
It's actually quite simple: companies sell you stuff, buy it if you want it/need it. Today Waymo sells me nothing, Cruise/GM sells me nothing, Tesla sells me a great BEV with Level 2 driving assistance. I got the car, but not the level 2 because it doesn't fit my driving needs. If they start selling level 3-4-5, I'll buy it if the price is right and I need it.
>>It makes me worried that they might be deluded about other things too.
Tesla will fail. But they will fail at 5% of their goals, and achieve 95% of that goals. It will still be called 'Failure'.
To avoid that 'Failure' tag, other companies won't even try.
Do this process enough number of times. And it will appear like Tesla is failing all the time, and yet achieving almost 100% more than other companies.
The problem here is in defining success in terms of binary outcomes of success and failure.
Oh, it's capable all right. Unless you actually try to use it anywhere that's not a perfectly spherical road in a vacuum, then it's also capable of killing you.
The interesting thing about the $3000 "upgrade" is that if you actually want to get it, I think you're better off taking the $3000 and investing it in Tesla.
Assuming autonomy is make or break for Tesla in the long term: if Tesla fails at it your stock will still be worth something (versus a useless upgrade). If Tesla succeeds you'll make more than enough to cover a retrofit.
Still not going to work for autonomy. Expect better ADAS and nothing else. The computer vision problems they need to solve to make autonomy work reliably are not solved yet, and I would be willing to bet good money that they won’t be able to get anywhere near L5 with the sensors they install today and without a LIDAR.
Are they hardware limited? It seems like the limiting factor is the convnet rather than the hardware...do they think it would be more safe with less latency?
Curiously there already is (was?) a probably eastern European chip manufacturer named Tesla (search images for "tesla ttl chips") but I wasn't able to find info on the manufacturer and if they're still in business.
Czechoslovak Tesla was shared brand of various state-owned electronics companies in the communist era (the companies in fact didn't have common parent company). Several of such companies are still in business.
In particular the semiconductor manufacturer Tesla Roznov and later TESLA SEZAM is today a subsidiary of ON Semiconductor (anecdotally they manufacture the power transistors used in Tesla Motors drive units, but I have no way to verify that claim).
We knew this almost a year ago [1]. Tesla held a panel discussion during NIPS 2017 in Long Beach where Elon Musk, Andrej Karpathy, and Jim Keller talked about AI. (I was at the event.) Jim and Elon talked briefly about how their hardware will overcome the limitations of current GPUs by making the bus massively parallel, along with other interesting tid-bits.
As for how they’ll get the chips into existing Teslas, Elon says: “We made it easy to switch out the computer, and that’s all that needs to be done. You take out one computer, and plug in the next. All the connectors are compatible.”
I'll believe it when they successfully install it to the previous gen car. More often than not it turns out that some other hardware is incompatible with software improvements or has bugs and full system is not functional anyway without full upgrade.
Really working or not, I'm just happy that an upgrade path crossed their minds. My car is 6 years old, and the console/nav looks so dated I wish they'd just put in dials.
237 comments
[ 2.7 ms ] story [ 105 ms ] threadSame age for walking around town unsupervised: the Japanese do and don't seem to lose many.
My original point was focused on the technical feasibility of driving with limited sensor input, which is physically possible.
Whether or not it takes a week, a month, or a million experience-years is a computational and algorithm problem.
All of which (based on previous technological progression) appear to be tractable.
E.g. the disaster that is the first major snowfall, every year, or automatic transmissions, antilock brakes, and traction control systems becoming standard
The human can only reliably be trusted to keep the car between the lines, stop appropriately, and occasionally make turns. Which is something much simpler to compete with!
Our intuition about the future is linear. But the reality of information technology is exponential, and that makes a profound difference. If I take 30 steps linearly, I get to 30. If I take 30 steps exponentially, I get to a billion. - Ray Kurzweil
People once said the same thing about fire lances.
https://hips.seas.harvard.edu/blog/2013/01/30/what-is-the-co...
The current model makes assumptions that sometimes turn out to be false. So yea it's not going to get to perfection, but level five only needs better than human not perfection.
* Granted I am likely a bad driver, but I don't think I could do this well.
See you can try to move the goal post, but as soon as they would be saving lives by taking tired or drunk drivers off the road that's very useful.
I am curious on what numbers are you basing this on, the Tesla PR numbers are not valid statistics.
I trust insurance companies to have good data, they think it improves safety when you compare rates.
This cars require the driver to pay 100% attention, by law and by Tesla agreements so isn't obvious that driver+drive assist+expensive car has better average numbers then driver + average (maybe old) car ?
I would like to see the numbers of how many times a driver had to intervene to save the situation
In terms of paying attention I really doubt most people are doing this very well. I am going to be stuck in the car either way so by default I am going to be watching the road, but as long as my relative stress level drops that's a huge win.
PS: You can make your own estimates. If people pay attention sufficient to avid accidents 99% of the time the car might be in 100x more accidents without drivers assuming those drivers where perfect. Drop that to IMO more likely 75% and it's closer to 4x at the absolute best case.
Same with Tesla, this is why I would like to see the number the drivers had to intervene and prevent crashes, without this number the only conclusion is about driver asist and not autopilot/self driving
Further, the difference between OK and an accident is normally fractions of a second. So, my suspicion is humans are really bad at this role and only catch the most obvious cases. AKA crap I am an a corn field, not hmm it's not slowing down enough I need to apply slightly more breaking power or I am going to hit that guy simply because you don't have time to react with the second situation.
Having said that, humans may be preventing a lot of issues well before they turn into accidents. The 'wanted to make a wrong turn down a one way street' is something a person can deal with easily, though it's not necessarily going to cause an accident.
Anyway I am waiting for some government institution or maybe the lawsuit to get more data then we can use that to make informed conclusions.
IMO, the largest risk is an over the air update killing a large number of people.
You can't invent numbers and statistics.
Remember the Tesla that crashed in a side barrier and killed the driver, people could reproduce the incident in the same spot , you can see it on youtube, so without a driver and with say 1000 Teslas driving trough that section in the problematic interval you would get at least 1000x more deaths.
Only if you include the people that reproduced the incident mentioned above you will get the Tesla stats much below human drivers.
Computers fail in different ways than humans do. But, that does not mean you can't reason about failure modes. If you ask someone to pay attention to something for an hour without doing anything most people are just really bad at this. Based on that I am rather shocked how well Tesla's systems work as I was expecting vastly more problems.
Could I have flipped to far in the other direction probably. But, I base things on my experience and the data I have available not just unchanging gut feelings.
Early autonomous test vehicles used LIDARS because camera and compute tech are not at the same level they are today. It's a legacy system and more of a shortcut. In essence, Object recognition > LIDAR judging distances.
Yeah it's true that a vision based system is much harder to solve. However if Tesla cracks it, it will pay dividends.
As has been discussed in some depth here before you commented, computer vision hooked up to some ML doesn’t come close to human eyes controlled by a human brain trained for many years.
Early autonomous test vehicles used LIDARS because camera and compute tech are not at the same level they are today. It's a legacy system and more of a shortcut. In essence, Object recognition > LIDAR judging distances.
Waymo is the cutting edge if you believe what people here say on thread after thread about them, and the raw stats. I’m sure they’ll be interested to hear your view of their “legacy” tech.
Yeah it's true that a vision based system is much harder to solve. However if Tesla cracks it, it will pay dividends.
Finally, reality. Yes, if they do something they’re yet to do, and their cars stop ramming stationary objects, then maybe something else will happen. No one is contesting that, there are just doubts given the state of Autopilot vs. leaders in the field, not to mention Tesla’s dodgy PR in general, and in the face of fatal incidents in particular.
In other words, you get everything that vision systems would offer except color, without having to do any processing to calculate depth, so you can go straight to the object-recognition and navigation processing.
LIDAR > vision systems.
Maybe driving on the roads we have now with computer vision alone needs strong AI, maybe simpler standardized roads and vehicles are needed, who knows. I think that strong statements about this should be backed by evidence, not gut feeling. And I'd like to see that evidence coming from the company that promises their customers their cars have all the hardware they need.
Even if they "crack" it, they will just be spending much more money to get the same benefits as a $100 mm wave system.
In the future, vision systems will only be seen as a gimmick of this era.
Radars beat both ladars and vision systems at phase, speed, and distance estimation.
The benefit of high margins and sophisticated engineering is you can get away with a lot of unknowns and promises.
Its only Elon Musk haters that somehow want to proclaim every mistake of him as some sort of fall from grace where finally his auro of diseption will be lifted from his victims.
If you pay for self-drive you will get an update.
Elon has always said it should be enough and if not you get an upgrade.
But some nuonce is lost in all his statments ofter it got threw the hype and couter hype cycle.
Also, while I agree that you can "program" a human, you can't program them to jump up and down and touch their toes (so to speak)
A person can explain "why" they did something when an accident happens. I've spent a lot of consulting hours helping unwind the "why" of ML/AI models, and we bill regardless of whether we can actually find that answer.
Putting aside by personal bias that I think Tesla is functionally incapable of doing anything well, lets pretend they deploy FSD - I think the fact that in that 1 accident to every 100+ human accidents you can't find "why" will override the fact that there are xx% less accidents.
Just from my experience consulting in healthcare and working with actuaries during the Obamacare debates and modeling out the impact of that -- people just honestly aren't open to statistical arguments when it comes to human life or emotional issues.
(Not saying they're wrong, me and my wife completely disagree about so many issues where it's emotions vs statistics, I think she has a very valid point on many of those - what I'm trying to say is this is going to be a philosophical debate more than a statistics debate so Tesla should gear up accordingly).
the truth is, people have a lot of experience with other people. we can usually predict behavior of other drivers on the road, and can even predict pathological behavior or know how to handle it. that is a major point. when machines and software fail, they can fail in big, unpredictable ways. humans fail in fairly predictable ways. none of us have this experience with neural networks or machines.
i barely trust my computer to work on a daily basis, much less a vehicle.
and lastly, when a human is driving, fault is usually clearly and easily assignable. how fault is assigned in the case of machines driving is not clear at all.
of course safety is important. and i agree that verifying driving ability is well behind where it should be. elderly and inexperienced and inattentive and just plain bad drivers are major problems on the roads.
and fault certainly does matter. aside from legal responsibility, we are emotional beings. there is a difference between getting hit by a drunk driver versus a true accident. that affects people in real ways. as a motorcycle driver, i will be terrified of these automated vehicles. i am already terrified of human drivers, but i can often predict their idiocracy.
no one wants to be killed by a machine, especially one built by wide-eyed engineers.
If we use a more narrow definition of weapon, e.g. a tool optimized for the primary purpose of injuring or killing people then it certainly is not a weapon.
Any object that can be remotely programmed to drive itself at high velocity into a target, yes.
It's not useless; the damage that can be caused varies by degree between things. Then you have to look at two cases - suitability for object to be used as a weapon, and the damage it can cause on accident. Unlike knives or hammers, both those factors are very high for cars.
The fact how dangerous cars is is very much underappreciated by people in general, as evidenced by the number of morons on the road. We already lose hundreds of people daily in the US alone because of this; now we're trying to add another class of drivers into the mix - algorithms written by greedy optimizers caring primarily for short-term profit and being first to market. This should give us some pause.
I'm not saying this technology is not possible or not wonderful, but I think the current ecology of self-driving efforts is unhealthy. We have a for-profit race by companies, many of which can't be trusted with getting software right, and most (all?) of them pursuing self-driving capabilities by means of half-understood brute-force black boxes the neural networks are.
> and the damage it can cause on accident
You are conflating (un)safety of a tool used in the way it is intended (kitchen knife = cutting a steak) with accidents (cutting a finger) and with malicious use (stabbing people). Those three categories are not the same for object-that-may-act-as-weapon and object-designed-as-weapon.
Conflating them collapses the number useful things we can communicate.
So are you concerned about Tesla intentionally building killing instruments? Or potential for accidents? Or the potential for intentional misuse?
> The fact how dangerous cars is is very much underappreciated by people in general, as evidenced by the number of morons on the road.
Cars also provide immense utility. If all they did were providing the thrill of speeding then they would probably be banned as too dangerous. One of the tradeoffs is the overhead of enabling people to drive. We could drive down the number of morons by requiring astronaut training for vehicle operators but again, that tradeoff seems too harsh and it's more efficient to occasionally let people die in traffic accidents than letting them die because nobody qualified as ambulance driver.
> We have a for-profit race by companies, many of which can't be trusted with getting software right
In the short term this may cause more deaths than necessary. But on the other hand it might be the quickest way to find a winner and then hold the rest to the same standard. As long as the experimental fleets are small they are just a blip in the statistics. Right now they should be equated to the yearly batch of first-year drivers who have an inherently higher risk profile due to lack of experience. We still accept them on our roads in the expectation that they improve.
What is important is to make sure that they are as good as or better than humans once they roll out in large fleets.
The latter two.
> We could drive down the number of morons by requiring astronaut training for vehicle operators but again, that tradeoff seems too harsh and it's more efficient to occasionally let people die in traffic accidents than letting them die because nobody qualified as ambulance driver.
I don't think this is the real reason. You don't need astronaut-level training for vehicle operators, just more than the ridiculously low standard of today, and more importantly, much stronger and harsher enforcement of traffic laws. I doubt that this will reduce the number of qualified ambulance drivers.
I suspect the real reason we tolerate so many morons on the road is path dependence. When cars first appeared, they were rare, slow and safe. In the couple of decades it took to get to the present density and speed of cars, it became a social status symbol, and something politically impossible to rein in.
> What is important is to make sure that they are as good as or better than humans once they roll out in large fleets.
I'm afraid that with self-driving tech based on neural networks, with no ability to inspect and verify what's going on, we'll eventually have to eat the risk and roll them out in large numbers before we know they're as good as humans.
I do not agree that neural networks are a "black box" with "no ability to inspect and verify". Even putting aside the many methods to understand what a neural network is doing without running it, at core, neural networks are well tested instruments. That's how they learn-- by testing themselves.
Obviously it's possible for a neural network to have odd behavior in circumstances not accounted for but that was always going to be possible at the level of complexity we're talking about here.
We're talking about cutting edge technology here -- and I agree with your general sentiment. I just don't agree with pinning the blame on "... based on neural networks". The same factors would apply to any codebase of this complexity.
Name three :).
> neural networks are well tested instruments. That's how they learn-- by testing themselves.
Last I checked, neural networks are well-tested in a sense that if you throw a big database and a shit ton of compute at them, they'll learn to accurately work within that database. Step out of it, and all bets are off. We're better at this than we were 30 years ago - good enough to apply this technology to consumer-level products in which mistakes don't really matter. I'd be wary of applying even current neural networks to safety-critical tasks.
> Obviously it's possible for a neural network to have odd behavior in circumstances not accounted for but that was always going to be possible at the level of complexity we're talking about here.
The problem is that with NNs, the odd behavior is usually totally unexpected, and you can't really inspect the network beforehand to discover the possible ranges of error-generating inputs. Everything works fine but every now and then you get a patterned sofa classified as a zebra, or a car + little noise classified as a toaster. And then there's no obvious relation between multiple misclassifications, because the reasoning structure of the neural network is implicitly encoded in its weights.
> The same factors would apply to any codebase of this complexity.
I think there's a fundamental qualitative difference here. A codebase can be complex, but ultimately it has a structure, and usually (in case of ML) represents a well-understood mathematical structure. Neural networks have simple code, and the whole complexity is hidden in opaque matrices of numbers, where even single changes usually have global effects.
I'm not trying to dismiss NNs in general; I just don't trust them in applications where health and safety is at stake.
America banned alcohol by an amendment, pretty much the hardest political barrier we have
It's just an excuse IMO
I'm sure that making the primary means of long (greater than walking) distance transportation for the majority of the population more expensive and higher stakes is going to work out great in the long term. I can see the parallels with healthcare. Creating yet another part of life where a single screw up that is capable of ruining the financial well being of someone living slightly better than paycheck to paycheck is not going to do positive things in the long run.
But who determines which is which? The letter of the law really, really, really sucks when it comes to traffic law. I haven't collected data but I'd wager that pretty much nobody follows the letter of the law for an entire drive from A to B
>For example, treating speed limits as suggestions instead of hard constraints
That's more human than reckless. Outside of places with Orwellian enforcement (I'm looking at you Europe with all your cameras) they really are. The vast majority of people go at a speed they feel comfortable in the conditions. This is why (conditions permitting) traffic flows at 80 even when the sign might say 55. People only follow the speed limit when they feel it's a comfortable speed. This is why the general recommendation is to set speed limits for the 90th percentile speed. If you don't do this on highways you get people doing the (inappropriately low) speed limit in the wrong lane. Passing on the right, tailgating and all the other things caused by traffic friction which is more stuff for drivers to keep tabs on and that decreases safety for all. There have been studies on this (Google "traffic friction" and filter out everything that has to do with literal friction). If anything speed limits on multi-lane roads should be raised to reflect the speeds people actually drive. I hope we'll see more dynamic speed limits in the future since they'll help a lot.
>overtaking in places where it's not allowed.
If every 100th instance of an illegal pass (usually on the shoulder when waiting for someone who's stopped to take a left turn or on the right on the highway) in my state resulted in a ticket it would probably be about a year before half the state's drivers hit the three strikes cutoff and had their licensees revoked.
The picture I'm trying to paint here is that aggressive enforcement of existing laws would probably be bad for the population at large because people break traffic laws in inconsequential ways all the time and that stronger enforcement of them would just screw people over (unless of course you enforce them so well that important people get screwed which would result in the laws changing) and discretion isn't an answer because that just results in profiling.
[0] https://www.zachaysan.com/cars and apologies for the length; it was a very stressful time in my life.
[1] http://support.blackberry.com/kb/articleDetail?articleNumber...
When you have 100,000 self-driving cars on the road, and someone figures out how to hack them, and drive them simultaneously into crowds, that will be a BIG BIG problem.
And this problem isn't just about self-driving cars. It is a fundamental issue with all algorithms. It's a lot harder to hack humans to do bad things (although it can be done with sustained messaging and propaganda). But algorithms can be compromised and exploited.
I'd say, it's probably easier to hack individual humans, but - like everything in computing - hacking algorithms scales well, while human factors generally don't.
https://www.bbc.com/news/technology-33650491
First, they're not physically inspectable by every Tom, Dick, and Harry. This means it's easier to obscure interfaces and other potential areas of attack. Not state-actor proof, but probably ISIS-proof. Second, they only get software updates while parked at an airport, so it makes MITM attacks harder (though not impossible). Third, they're in the air where death isn't a sudden movement away. Pilots can override the system and fly it manually. Fourth, they're heavily monitored with errant flightpaths reported to militaries around the world (to stop another 9/11).
We have none of these safeguards for self-driving cars and there are going to be hundreds of millions of them.
https://www.computing.co.uk/ctg/news/3020901/dhs-team-manage...
How autonomous systems will make anything different? Why do you believe people aren't currently hacking cars?
I don't intend to be critical at all, I believe this is a very interesting point too that should be discussed. (I hope to check the links on my interval).
> How autonomous systems will make anything different? Why do you believe people aren't currently hacking cars?
It might make always on data connections near-mandatory (to get maps data, etc). At least in my car, the insecure embedded computers are air-gapped from the internet.
Part of the reason is that, in current ecosystem, companies will find ways in which self-driving requires being constantly on-line and connected to vendor's server. Off-line processing is not in fashion these days.
Will this mistake never go away?
Only thing I can imagine would be allowances for wider and deeper neural nets.
As a simple example in a different domain, years back I found that the best way to improve Tesseract's OCR accuracy was to ensure that I didn't feed it images at more than 150dpi because it would sometimes misrecognize dust, paper texture, etc. as characters.
Yeah I think this is it. My best guess, it's 200fps vs 2000fps for the same workload. As Tesla's data model gets larger and its computations get more complex, that 200fps will start to sag.
2000fps will provide a lot more wiggle room to do higher level computation on each frame.
36 / 200 frames = 18cm travel distance between each calculation.
So you either can calculate much more each 18cm or calculate shorter distances.
I think a shorter distance is not very useful but calculating more might be.
I think n = 8
...or more cameras, with better frame rates over a lot of cameras. My guess is that the fps rate is over all the cameras, so 10 cameras means going from 20fps per camera to 200fps per camera.
https://www.youtube.com/watch?v=-KxjVlaLBmk
The higher the speed of the system, the less it has to predict in the future and better can adapt to unforeseen situations. It's basically using the world as a model, for free.
What are the real-world benefits going from 200 to 2000 frames a second? 200 seems quite fast.
I don't make a habit of crashing cars but I did have one glancing head on crash, combined speeds around 125 m.p.h. and it was only centimetres that meant I did not wipe out that family and trash the car I was in.
As a consequence I would say that this 10x frame rate really is quite game changing, particularly for on-coming vehicles.
If you have your own ASIC it's a differentiator that's hard for other players to compete with. They'll take this to potential investors (or stockholders the "market") and say this is a big expensive thing we have that nobody else has. It's expensive and time-consuming for them to replicate and we also own a lot of IP that they would need.
You and I might know it's bullshit. That's not true of the investment market in general.
If they succeed, it means slightly improved efficiency and cutting NVidia's profit margin out of their cost structure. If they fail, it would cost billions, and obscure chip bugs could put lives at risk. That risk/reward doesn't seem worth it unless you really know what you're doing and are confident you will succeed.
Of course, it fuels the hype machine in the short term.
This seems like the biography of Elon Musk's companies. He gets a crazy hard problem that people say he and his companies can not do, then he tries to implement it, often falling well behind and well short of the goalposts due to lack of foresight and overambitious scheduling. There were naysayers the whole way, including just as loud as there are now if not louder, and yet his companies' achievements and networth continue to climb.
Spaceflight was solved decades before Elon was born. Reusable rockets weren't a hard problem--they were simply a problem the incumbents were unwilling to address because it would have massively cut into their revenues and profits.
Electric cars actually predated ICE cars. The issue with EVs was always the charging infrastructure, which Tesla solved...by simply throwing a lot of money at it. (Not hard, just resource intensive.) Their batteries are built by Panasonic.
Boring Co literally is just a used tunneling machine. They have literally not done anything with it except test it out below the SpaceX parking lot (and the innovation in boring would come not from the tunneling but with the post-tunneling construction of the tunnel walls, stations, dirt removal, and ultimate extraction of the boring machine).
Even at Paypal, their biggest innovation was developed by Musk's biggest pre-Paypal competitor (Thiel's company) before they merged to form PayPal.
But of course you know much better how simple all these things are. For you to disagree with that is just emberecing yourself.
The same applies in a lesser degree to your other comments. Do you seriously belive that the origonal EV in 1900 was the end of development and all Tesla did was 'throw money at it'? That is the hight of stupidity and you again wont find a single expert on the subject agreeing with you.
Also the battery chemstry is co-licenced by Tesla and Panasonic with the design of the cells and even part of the chemistry being inhouse at Tesla. Panasonic can not sell this technology.
So when you are spreading baseless FUD, at least get your information correct.
You seem to have lost all rationality in context to of this question.
Not quite... The chip was designed by Jim Keller, who was a key player in designing the highly successful AMD Zen architecture and Apple's A4 and A5 SoCs (pretty much a legend).
Jim is known in the industry as having an unquenchable thirst for really hard problems. Solves them, leave the company to recharge and goes on to the next company/project.
I outright snorted at this. Time to market is not why you build a custom chip instead of software. Is this article written by a complete bonehead?
Tesla has it's own hardware platform. Anyone with an ounce of sense will know that was always going to happen because of power consumption, cost and performance. The Nvidia Drive platform is designed to be a quick way to get to market with FuSa. The requirements for fully autonomous modes are estimated by some to be over 10x higher than the top CPU/GPU offerings.
IOW, yeah, it’s a crap article, but not for this reason.
They also published the quote about iPhone leading in performance and go-to-market because of their custom SoC efforts, and they mentioned Tesla's SoC program manager Pete Bannon, but they didn't even bother to connect the two ideas: Pete Bannon was the guy who led Apple's SoC efforts to develop the first 64 Bit ARM chip for mobile.
They also didn't mention what Elon said about now being able to run all the cameras at full resolution and full framerate.
This article could have been so much better.
This is not what Musk said on the conference call. Surely some bonehead reporter is writing this. Musk said the latter, precisely to go from current 200 fps speed to something like 2000 fps (I might have the numbers wrong, but it's a 10x improvement) and that they have been doing this in stealth for the last 2-3 years. They expect the cost of the new chips to be the same as that of the current nVidia lineup. That got me thinking though. What are they currently using - the Titans ?
... so that they can create performance-intensive features that nobody else in the market yet has, because the off-the-shelf hardware won't allow it. Due to power consumption, cost and performance issues. And this custom hardware is being developed to solve them.
You're basically saying the same thing in different words. I mean, what else do you do with better power consumption, cost and performance other than new, better-working features? You think Tesla will sell cars to the mass market with impressive computer specs?
Apple's own custom hardware has enabled it to experiment with performance-intensive features unlike other smartphone manufacturers, Tesla is going for the same strategy.
Will existing Tesla owners be able to swap the computer? Or is the swap happening only inside the factory, to the next generation of the Teslas they're building?
This may be one way in which Tesla's approach differs from Apple's.
and talking about this now seems like another desperate attempt at generating hype.
By limiting the context of use, specialized hardware can perform better than CPU/GPU. I think his thinking relates strongly to ASIC crypto mining hardware https://en.bitcoin.it/wiki/ASIC .
"bare metal" is just a shorthand for the level of access to hardware that does not require a kernel/OS.
But from the sounds of it this chip goes beyond that. It sounds like some kind of RAM/GPU hybrid with matrix-specific instructions specifically targeting NN inference for Tesla's specific style of NNs, which as a first guess I take to mean "lots and lots of layers."
In the call they mentioned increasing their investment in the chip team and technology, which is a sign that they're confident it's going to pay off.
By going to a custom ASIC with inferencing. I have little doubt about Musk's claim that it'll offer a 10x improvement over the current stripped down PX Drive2 variant for deep learning tasks. If I'm to decrypt Musk's branded terminology and take him at his word, then they'll be going from about 8 TOPS to 80 TOPs with that they'll be able to get even deeper into the no man's land that's even more overkill for basic ADAS but still not nearly enough to handle full self driving safely in all conditions, though maybe they'll get L3 highway out of it.
Nvidia's upcoming Pegasus board will supposedly do 320 TOPS, and their competitor, Intel/Mobileye's EyeQ5 will likely have comparable specs. These motherboards are designed to replace Robotaxi trunks brimming with ~$150k in liquid cooled compute that draw 3-4000 watts.
Personally, I’d refuse to implement self-driving vehicle without at least a ”backup LIDAR” to check vision system results. Otherwise you are forced to assume stuff like ”things at stand still are either above the road, beside it, or just shadows”, causing crashes when there is suddenly a stopped car in front of you. (If you didn’t do that assumption, you would be dodging shadows and other clutter..)
[source: I’ve been researching vision algorithms in a related field.]
Because that isn't how people judge distances, at least not distances beyond a few feet in front of their faces. Plenty of people with only one eye do very well. It is a difficult problem because we use a variety of techniques and 'hardware' when estimating distances and speeds. Car companies are trying to do with one tool (ie lidar) something we do with many.
I guess the same logic applies directly to self-driving cars as well..
At any but very close distance, don't humans mostly use a combination of lighting cues, a priori knowledge of actual size vs. apparent size, motion parallax, and other flat-image cues instead of stereoscopy?
But, yeah, LIDAR cuts through all that, too.
Anyway, if LIDAR becomes small and cheap, Tesla can just strap it on.
Whether this will succeed is another question.
[1] https://www.youtube.com/watch?v=FaW85fIos64&t=21m40s
Supposed to be fully autonomous since, what, 18 months ago? How about that SF->NY demo?
He also added that it would be sort of cheating, because that's a very defined route and he prefers it to be able to be much more dynamic. i.e. pick any two US cites and the car will drive you there.
I can only assume then that you have some kind of insider info to support your claim?
Ergo, logically, if Tesla has not performed such a test, it is strong evidence that Autopilot is not currently capable of passing it.
http://www.autonews.com/article/20180208/MOBILITY/180209770/...
Google's autonomous efforts far precedes Tesla, yet they have not out in the market yet. I would say on the contrary, they made significant strides in a short amount of time even after a complete overhaul Mobileye (AP1) to an in-house solution (AP2/2.5) https://vimeo.com/192179727
Waymo has been operating a real actual passenger service for over a year now. They drive tens of thousands of miles a day. They're "not out in the market yet" only in the sense that they're not charging money for it yet, which they seem to plan on doing soon.
https://www.bloomberg.com/news/features/2018-07-31/inside-th...
I use the word "deluded" above realizing that I might be completely wrong. If they pull it off I'll eat crow and give them props. But I just don't see it happening.
I think “fraudulent” would be more accurate; it was factually false, it was at best entirely unfounded if not actually knowingly false, it was material to decisions of people to purchase the self-driving add-on.
The Autopilot page [1] still states the following:
"All Tesla vehicles produced in our factory, including Model 3, have the hardware needed for full self-driving capability at a safety level substantially greater than that of a human driver."
[1] https://www.tesla.com/autopilot
Having said that, I find it hilarious that people on here have such immensely strong opinions about something that doesn't exist. They'll go at length about how LIDAR is a hard requirement, about how many FLOPs you need, etc etc etc. Fact: FSD doesn't exist and no one has it. Nobody knows which approach will work and if more than one approach works no one knows the timelines or economics of the different approaches and very certainly no one knows the market impact of these. But people start thinking they know the future and then become very emotionally vested in their make-believe stories.
It's actually quite simple: companies sell you stuff, buy it if you want it/need it. Today Waymo sells me nothing, Cruise/GM sells me nothing, Tesla sells me a great BEV with Level 2 driving assistance. I got the car, but not the level 2 because it doesn't fit my driving needs. If they start selling level 3-4-5, I'll buy it if the price is right and I need it.
It's certainly not nothing
But there's no need to pick on GM, they all have self-driving tech capable of killing you.
Tesla will fail. But they will fail at 5% of their goals, and achieve 95% of that goals. It will still be called 'Failure'.
To avoid that 'Failure' tag, other companies won't even try.
Do this process enough number of times. And it will appear like Tesla is failing all the time, and yet achieving almost 100% more than other companies.
The problem here is in defining success in terms of binary outcomes of success and failure.
Assuming autonomy is make or break for Tesla in the long term: if Tesla fails at it your stock will still be worth something (versus a useless upgrade). If Tesla succeeds you'll make more than enough to cover a retrofit.
In particular the semiconductor manufacturer Tesla Roznov and later TESLA SEZAM is today a subsidiary of ON Semiconductor (anecdotally they manufacture the power transistors used in Tesla Motors drive units, but I have no way to verify that claim).
It's not a unique name in any shape or form. There's a smartphone "manufacturer" from Serbia that uses Tesla as its brand name: https://tesla.info/en/
Here's the direct link to my notes from the event: https://www.reddit.com/r/teslamotors/comments/7iczp7/elon_mu...
[1]: https://www.theregister.co.uk/2017/12/08/elon_musk_finally_a...
I'll believe it when they successfully install it to the previous gen car. More often than not it turns out that some other hardware is incompatible with software improvements or has bugs and full system is not functional anyway without full upgrade.