-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: FSD Beta tweets
- Thread starter Terminator857
- Start date
DanCar
Active Member
Interactive Highway Safety Design Model: Accident Predictive Module | FHWA
The Summer 1994 issue of Public Roads introduced the Interactive Highway Safety Design Model (IHSDM).
highways.dot.gov
From the above diagram we see that 57% of accidents are solely attributed to driver issues, back in 1995. Roadway and driver issues accounted for 27% of accidents.
Elon sure loves his pseudoscientific metaphors and dumbed down marketing that his followers lap up without scrutiny and repeat endlessly. I think he’s implying FSD+human is safer than human. This is the same messaging as in the Tesla marketing material called ”safety report”. 40 years of research show that human+system isn’t much (or at all) safer than system due to automation complacency, so this line reasoning isn’t proved to be relevant if the design goal is to make an ADS.
If FSDb would magically become 100x safer it’s still only going to be a fraction of human level safety. And that would likely, somewhat counterintuitively, cause accidents to go up since we humans suck at monitoring.
There is simply no way that FSDb on current cars will get anywhere close to human level performance of 2-3 accidents per human lifetime on current hardware in a large ODDbefore we have computers that can reason about the world.
To suggest that will likely happen by the end of this year is at the very least misleading and simply not likely at all. FSDb isn’t even designed for DDT fallback handling at this point. I’m guessing it’s still an L2 in 3 years based on the current rate of reliability improvement. and the general state of where computer vision safety is at.
Last edited:
Yeah, but as Koopman says, computers will make different mistakes. Currently CV has a very hard time understanding that a cyclist or a stop sign on a billboard isn’t an actual cyclist or a stop sign. A human isn’t prone to that sort of failure because we have reasoning and actual intelligence, so this “a human drives with two eyes, so then computers will be able to” snakeoil marketing doesn’t hold up for even basic critical thinking. Perhaps in the next decade cameras will be enough, but not likely in the coming years.View attachment 958767Interactive Highway Safety Design Model: Accident Predictive Module | FHWA
The Summer 1994 issue of Public Roads introduced the Interactive Highway Safety Design Model (IHSDM).
From the above diagram we see that 57% of accidents are solely attributed to driver issues, back in 1995. Roadway and driver issues accounted for 27% of accidents.
Last edited:
DanCar
Active Member
In recent earnings call Elon said he is going to spend $1 billion on dojo. It is obvious to me if you are spending $1 billion on dojo, then it is not just for FSD. This is a springboard to general A.I.
www.fool.com
www.theregister.com
Separately Elon said Dojo not as good as nVidia, but have to proceed with dojo because can't get enough nVidia product.
In another tweet, Elon not sure that dojo will succeed.
Tesla Will Spend Over $1 Billion on Dojo Supercomputer -- Is This the Top EV Stock to Buy Now? | The Motley Fool
Tesla's bet on high-performance computing could turn into a huge profit driver. Is it time to buy the stock?
www.fool.com
Tesla spending over $1B on Dojo for self-driving AI training
More data means better neural net training, but it also means more cores
www.theregister.com
Separately Elon said Dojo not as good as nVidia, but have to proceed with dojo because can't get enough nVidia product.
In another tweet, Elon not sure that dojo will succeed.
At the end of the day it's just a faster training computer. It solves nothing but faster round trips. If everything that was needed was more compute and more data, self driving on computer vision alone would have been "solved" years ago.
It's not that autonomy is an unsolved problem, only autonomy on budget hardware is unsolved. And will likely remain so for a long time regardless of training dataset and compute. If Elon want to make FSD driverless this decade he would have more luck adding more and better sensors instead of removing them. Even it it's only for the safety case, it's incredibly stupid not to add a $500 Lidar at this point for reliability.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
AVs have accidents because of a variety of failures: hardware and software bugs, perception errors, bad decision-making and actuation errors. So in theory, if you can reduce all those types of failures enough, you could eventually reach 10x safer than humans. I think Elon believes that with enough ML training and data, FSD will improve, make fewer and fewer errors that lead to accidents, until it is eventually 10x safer than humans. But I think one problem with this is that the accident rate needed to be 10x safer is really hard to achieve with just vision-only ML. According to US highway statistics, human drivers have a fatal crash on the highway every 3.75M hours of driving. So 10x safety would be 37.5M hours of driving per fatal accident. So to achieve 10x safer than humans on highways, Tesla would need to train vision-only FSD to only make one safety critical error on the highway every 37.5M hours of driving. That's a very high bar IMO. In fact, Mobileye argues that vision-only cannot achieve that high of a MTBF since vision-only right now has a MTBF of less than 1000 hours of driving. That is why basically everybody argue for adding radar and lidar to do "eyes off". They argue it is necessary to get to to the needed MTBF.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
I would point out that the hardware and software that Tesla is using is vastly inferior to the human brain and eyes. Human eyes have a resolution of about 576 MP while Tesla uses 1.2 MP cameras. So Tesla's cameras have a resolution about 480x worse than human eyes. The human brain has an estimated 86B neurons and performs at around 18-640 trillion TEPS (traversed edges per second, a measure of fast information is communicated internally). The fastest supercomputers operate about 23 TEPS so the low end of the human brain. The FSD computer in a Tesla is nowhere near the power of a human brain.
Brain performance in TEPS
Traversed Edges Per Second (TEPS) is a benchmark for measuring a computer's ability to communicate information internally. Given several assumptions, we can also estimate the human brain's communication performance in terms of TEPS, and use this to meaningfully compare brains to computers. We...
aiimpacts.org
Another big difference is that humans have generalized intelligence. At an early age, the human brain learns to understand the world around it and reason. By the time a human learns to drive, they already have about 15 years of generalized intelligence. For humans, learning to drive is more about applying that generalized intelligence to the specific task of driving. Vision-only FSD has to train perception from scratch to understand the world in the context of driving. And no matter what Elon says, it is actually not generalized intelligence since FSD is being trained with specialized intelligence for the purpose of driving. So yes, there is generalization in the perception (ex: it can recognize all cars from being trained on some cars etc) but it is not generalized intelligence.
So I would say that "humans only need 2 eyes and a brain to drive" is a nice analogy and it works to prove capability since vision-only can drive a car. But FSD needs to do achieve very high reliability in order to be "eyes off". I think the analogy fails to prove that vision-only can achieve "eyes off" reliability since Tesla hardware and software is nowhere near the same level as humans. It's that big difference between capability and reliability that many Tesla fans seem to miss. Maybe when we have cameras with similar resolution to the human eye and we have computers comparable to the human brain with generalized intelligence, we can achieve the needed reliability with vision-only. But we are very far from that IMO.
And if we look at airplanes, birds did inspire humans that flying was possible. And yes, we did take some lessons from watching birds fly. But when we tried to mimic how animals fly by flapping their wings, we failed. But we did find another way with fixed wings and engines to achieve the same end result. I think airplanes are proof that we don't need to always mimic humans or animals exactly in order to achieve the same result. Same with AVs. We can learn some lessons from how humans drive, like the fact that the AV does need vision to see the world and does need a "brain" to process information and make driving decisions. But we can find a different way for AVs to drive like using cameras, radar and lidar and HD maps to help AVs drive more reliably, since AVs lack the hardware and software of humans.
Last edited:
This 576 MP number is often quoted but is highly misleading. This number is derived by taking the receptor density in the very narrow-angle central fovea region of the retina, being very roughly 200,000 color-sensitive cones, and then pretending that the entire very wide field of human vision is filled up with receptors of that density. If you had something like 2 minutes to scan your foveal vision across your whole visual field, l and nothing moved during those minutes, you could then take in the equivalent detail of something like 576 megapixels.
By the way, your suspicions and skepticism are should be aroused as soon as you read such an obviously stretched equivalence calculation, that results in a widely repeated number having 3 significant digits!
Please understand that I'm not denigrating human vision (except in my own case which could use a lot of help
), but it's really a highly fraught comparison with digital camera sensors. Human vision works in a substantially different way. The non-foveal peripheral vision has very poor resolution and color sensitivity, but specializes in edge-movement detection - evolved exactly to trigger the visual cortex to redirect and refocus those ~200,000 central receptors onto the latest and most important threat/opportunity.The considerations of what today's digital cameras need to for driving, what they can see and how they are is in some ways inferior and in other ways superior to human visual perception, it's beyond the scope of a post that anyone's going to read. If you've followed my own arguments for a while, you know that I tentatively disagree with some of Tesla's choices of camera positioning, and I have various engineering suggestions of how it could be done differently and I think better. (I say tentative, because for some reason, no one has ever invited me to review the topic with the Tesla Autopilot team
)A huge benefit of the multi-camera system (anyone's) is that it monitors pretty good detail, simultaneously and tirelessly all around the car - setting aside said debate over the exact implementation. It also has mostly better night vision, though again I have some suggestions for improvement. The real point here is that it's substantially different from human vision.
To me, a theme of ADAS engineering should be to leverage available and cost-effective hardware for its "superhuman" capabilities (and there are plenty of those in even the now-dated HW3 cameras) to offset aspects that are inferior or less adapted to the driving environment.
I'm not going to get into the computer vs. brain "specs" debate; it's harder and even less well understood. My take is that the camera vision is more generally capable but less specifically capable then human. By contrast the present silicon computer is more specifically capable (e.g. in precision, repeatability and reaction time) but less generally capable and adaptable (i.e. intelligent) then human. This is essentially what you said also. TLDR: the computer and brain are still very different, despite the significant point that huge progress is being made by trying to understand and mimic processes of the brain.
diplomat33
Average guy who loves autonomous vehicles
This 576 MP number is often quoted but is highly misleading. This number is derived by taking the receptor density in the very narrow-angle central fovea region of the retina, being very roughly 200,000 color-sensitive cones, and then pretending that the entire very wide field of human vision is filled up with receptors of that density. If you had something like 2 minutes to scan your foveal vision across your whole visual field, l and nothing moved during those minutes, you could then take in the equivalent detail of something like 576 megapixels.
By the way, your suspicions and skepticism are should be aroused as soon as you read such an obviously stretched equivalence calculation, that results in a widely repeated number having 3 significant digits!
Please understand that I'm not denigrating human vision (except in my own case which could use a lot of help
The considerations of what today's digital cameras need to for driving, what they can see and how they are is in some ways inferior and in other ways superior to human visual perception, it's beyond the scope of a post that anyone's going to read. If you've followed my own arguments for a while, you know that I tentatively disagree with some of Tesla's choices of camera positioning, and I have various engineering suggestions of how it could be done differently and I think better. (I say tentative, because for some reason, no one has ever invited me to review the topic with the Tesla Autopilot team
A huge benefit of the multi-camera system (anyone's) is that it monitors pretty good detail, simultaneously and tirelessly all around the car - setting aside said debate over the exact implementation. It also has mostly better night vision, though again I have some suggestions for improvement. The real point here is that it's substantially different from human vision.
To me, a theme of ADAS engineering should be to leverage available and cost-effective hardware for its "superhuman" capabilities (and there are plenty of those in even the now-dated HW3 cameras) to offset aspects that are inferior or less adapted to the driving environment.
I'm not going to get into the computer vs. brain "specs" debate; it's harder and even less well understood. My take is that the camera vision is more generally capable but less specifically capable then human. By contrast the present silicon computer is more specifically capable (e.g. in precision, repeatability and reaction time) but less generally capable and adaptable (i.e. intelligent) then human. This is essentially what you said also. TLDR: the computer and brain are still very different, despite the significant point that huge progress is being made by trying to understand and mimic processes of the brain.
I appreciate the thoughtful post. Thank you for providing more detail. I think the general point still stands that the "humans can drive with 2 eyes and a brain" argument to defend FSD is flawed since as you pointed out eyes and cameras work differently and the human brain and computers work differently.
I do think that making computers more generally intelligent is the main obstacle for autonomous driving. That's because even if cameras were as good as the human eye today, AVs would likely still struggle in certain situations due to lack of intelligence. But if computers were more generally intelligence, AVs would be able to handle those edge cases even if the cameras were less good. And AVs can use radar and lidar to compensate for weaknesses in cameras. So cameras being less good as the human eye in certain ways can be addressed now. But it is harder to compensate for the lack of intelligence. I feel that we see that today with the current robotaxis. They don't have perception issues, they have intelligence issues.
Last edited:
I would agree that the argument "humans drive with only two eyes" is not helpful. It's certainly incomplete, it's not really true in most cases, and it focuses the argument on a justification for sensor hardware when, as you say, it's much more than that.
By the same token however, pointing to a particular difficulty or failure of vision-based FSD as proof that lidar or radar would have solved the problem, is also an incomplete argument and often a probably false conclusion. Spin serves the narrative on both sides.
I found both your posts insightful and educational beyond the typical Google search spin most posts offer. Thanks for sharing.
Point is: When one does disengage L2 driver assist (as they should), total miles driven without an accident is completely meaningless metric. It does not measure whether the driver assist is good or not. Instead, it measures whether the driver (who is always in control regardless of the L2 assist) is a good driver or not.
willow_hiller
Well-Known Member
It's only meaningless when comparing very different cohorts of vehicles and drivers. But when there are 0.31 accidents per million miles for Tesla drivers with FSD Beta, and 0.68 accidents per million miles for Tesla drivers without FSD Beta, it doesn't make sense to attribute that difference to driver performance alone.
Unless you're saying that Tesla drivers that opted for FSD just so happen to be twice as safe as Tesla drivers that didn't.
We really do not know how Tesla calculates the numbers, so I would take the numbers with a salt of grain. The history is that Tesla has not been exactly trustworthy with statements related to autonomy.
Furthermore, Tesla do push people to drive carefully with FSD: first they only let in people who got a sufficient safety rating and nowadays they disable FSD for people who fail to drive safely. This surely impacts the numbers however they are calculated.
There could also be demographics problems: Maybe the ones who are willing to spend to lease/buy FSD feature are by nature safer drivers. Could be that they are wealthier and thus more likely older and more experienced. Just speculating here, but as long as the demographics are not addressed by Tesla in statistics they release, we just do not know.
All that said: People seem to interpret the mentioned safety numbers as something to do with autonomous driving. They have little to do with that. It would be great if a L2 driver assist might help with safety and that should plausible (my own first hand experience surely supports that), but lets not confuse them with autonomous driving (L3-L5).
willow_hiller
Well-Known Member
I don't think anyone in this thread was doing that. The quote that started this conversation was "Interesting. Pretty sure I have averaged 10k annually on FSD or really close to that. No accident yet." in response to someone else saying that using it anywhere other than the highway was inviting accidents.
Meanwhile, there are people that like to imply that FSD Beta is too unsafe to warrant public testing. I just don't think the statistics support that idea. Even with all the demographic issues you raised, the direction and magnitude of the difference between FSD Beta and all other Tesla drivers is too great. It's implausible for those with FSD Beta to be something like naturally 4 times safer than all other Tesla drivers, and be lowered to only 2 times safer by FSD Beta.
This is probably the wrong place for this debate, but I'd like to stick my oar in.
Over on the 11.x.x thread, there's a contingent of users who really, really, don't like the FSD-b. And they report some really scary behavior: Trying to cross over three lanes in a short, unmanageable distance; running red lights; running stop signs; very poor merging onto a freeway or incorrect movement off of a freeway, and so on. I would guess that five or six frequent posters on that thread, who have previously loudly expressed their discontent and disgust with the FSD-b, are currently stating that they're sending off the appropriate emails to Tesla so they can get off the program. Some of these loudly bewail the day that they ever put any money down on the software.
And then.. there's, I guess, about a 3X number of posters going, "Are you guys for real?". I'll admit that I'm one of those. Yes, the beta's got its bugs, no question, and it's not ready for Prime Time. But the seriously scary stuff being batted around simply hasn't been happening, or has been happening a lot less frequently over time. On a given 20-mile run, I usually get somewhere between zero and two interventions, which is loads better than, say, five months ago. Admittedly, I've had the car attempt to run a stop sign with the current 11.4.4 release, but that was once. And I did say, "Not ready for prime time." Still jerky; doesn't handle unprotected left turns all that well; and so on.
A recent 11.4.5 release went out to employees, but not to others. There's an 11.4.6 release that's getting some employee goodness, but we'll see.
Main point: If all the complainers do what they claim they're going to do and dump FSD-b, the rest of us will go back to the regularly scheduled, "Well, this works. That doesn't. Comments?" business, like we used to.
Disparity in expectations for what "Prime Time" means seems to be the main source of heated debate. For some it seems to be an advanced L2 driver assist, for some an L3 with and some expect L4+. When everyone debates readiness of FSD without specifying which of these vastly different product they speak of it confuses everyone involved. It does not help that Tesla has marketed the feature as L5 and L2 (at different times) and thus people have purchased a very different product (promise) under the confusing "Full Self Drive" product name. Furthermore, CEO statements keep making the situation even more confusing. To top all this off, there are a large groups of uninformed fans and investors whose interest are not in seeking truth, but defending Tesla in any online debate.
Knightshade
Well-Known Member
They haven't though.
They used to sell a thing they said would eventually but definitely is not today at least L4.
Then they changed it to be clear they'd deliver L2 today and didn't promise more than that anymore.
But in both cases it was very explicit what you were getting at purchase was only operating at L2.
And the system reminds you of that when you first activate it.
And every single time you activate it after.
So nobody should be confused it's L2 regardless of if you bought the very that promises more later
Similar threads
