We should measure accidents while testing compared to average, before having knee jerk reactions before implementing unnecessary regulations.
When I've seen YouTube videos of customers testing, they seem very responsible and to be honest I think it's a great idea in general and I'm usually very impressed with Tesla's progress seeing those unfiltered videos.
It's obviously not as simple as that. Tell that to all the families whose members die each year to any other vehicle accidents. Should we ban vehicles in general?
Presumably HN has many intelligent people on it. I am continually perplexed, given that, by all the comments assuming that autonomous driving will be better than human driving any time soon. Can you name one open problem domain that automation does better than a human at present? Does AI make better art? Does it make better written arguments? Does AI wire houses? Does AI deliver mail? Does AI cook food in a kitchen? Does AI repair machines? No to all these and many more. The only domain that "AI" has been demonstrated to work better than or close to as good as a human is in closed problems like board games and to an extent language translation and similar things, perhaps image analysis, and things like welding a seam in the same spot on an assembly line. So given that we have at present 0 examples of competent open domain AI, and given that driving on a road is an open domain problem, why the heck is it that presumably intelligent people think that general autonomous driving technology is likely to be around the corner (and better than human)?
Maybe we should try making a robotic chef first before we attempt a driver.
To date all AI is pattern matching in closed domains. It is not capable of solving open ended problems like how do I safely pilot a car given an indefinite number of "edge conditions."
I personally didn't say it was around the corner. What I'm asking is that is it dangerous to allow Tesla to do supervised training? For this FSD doesn't have to drive as good or better than a human driver. They just have to be safe in combination with the supervising human. We should look at miles driven and accidents/deaths happened per mile while someone was supervising a training Tesla. We also allow people to learn with a supervisor on the side.
If they are as safe or safer than your average human in a similar situation why disallow it? It will help us to get to FSD faster, whether it is in 2 years, 15 or 30 years, doesn't matter.
We don't know when it happens, but we can try our best to make the progress faster if it's safe and sustainable.
There are many, many confounding variables. Not all miles driven are qualitatively the same. I.e. driving 1000 miles on winding, narrow New England roads, is qualitatively very different from driving 1000 miles on a grid in sunny Phoenix.
The issue I have is the assumption that FSD will ever be an economically attainable goal. Basically FSD requires AGI applied to driving domain. Nobody is making investments based on an assumption that AGI is 2, 15, or 30 years away, so why should they be making investments on the assumption that FSD is closer?
Even if at some point FSD is technically feasible it remains to be seen if it will be economically competitive with human drivers and if it will be actually safer. Many people enjoy driving, myself included, and many people are very safe drivers.
Speaking of New England how does FSD react to a deer jumping out of dense brush 30 feet ahead of the car at night. First it has to recognize the deer as an obstruction. Then it needs to decide whether to brake and maintain line of travel or brake and try to evade the deer while maintaining control of the car.
How does FSD identify black ice at night vs. wet pavement. Experienced human drivers can identify ice by the "feel."
How does FSD know what branches in the road after a wind storm are safe to drive over and which require braking and navigating around?
What about a limb hanging down into the road way?
How does FSD know where to drive on a back road with no lines on it?
How does FSD react to a plastic bag in the road? What about a rock that is similar colour and shape?
Etc.
Nothing about FSD is simple or close to being solved.
I am suggesting that yeah we should not attempt FSD until we get traction in more constrained "AI" problems. FSD is probably one of the hardest AI problems out there. Even if it becomes technically possible it remains to be since if it will be economically and perfomance-wise competitive to a human. In the world of present day facts, as opposed to media hype, there is zero evidence that AI is a better open ended decision maker than a human. The hard part of driving is not rapid sensing but rather rapidly making sense of what is seen, i.e. making decisions based on incomplete or dynamic information based on an internal model of the world and past experience.
Right now, all numbers from Teslas Autopilot look better than average. There are a bunch of potential asterisks to the numbers (mostly freeway miles, Teslas are safer than average, etc), but overall the numbers suggest humans + Tesla Autopilot are better than humans alone.
Regulation to require reporting is fine and I'd definitely like to see that happen. I'm just hoping there won't be unnecessary regulations due to emotional or similar reasons.
14 comments
[ 4.5 ms ] story [ 40.8 ms ] threadWe should measure accidents while testing compared to average, before having knee jerk reactions before implementing unnecessary regulations.
When I've seen YouTube videos of customers testing, they seem very responsible and to be honest I think it's a great idea in general and I'm usually very impressed with Tesla's progress seeing those unfiltered videos.
Maybe we should try making a robotic chef first before we attempt a driver.
To date all AI is pattern matching in closed domains. It is not capable of solving open ended problems like how do I safely pilot a car given an indefinite number of "edge conditions."
If they are as safe or safer than your average human in a similar situation why disallow it? It will help us to get to FSD faster, whether it is in 2 years, 15 or 30 years, doesn't matter.
We don't know when it happens, but we can try our best to make the progress faster if it's safe and sustainable.
The issue I have is the assumption that FSD will ever be an economically attainable goal. Basically FSD requires AGI applied to driving domain. Nobody is making investments based on an assumption that AGI is 2, 15, or 30 years away, so why should they be making investments on the assumption that FSD is closer?
Even if at some point FSD is technically feasible it remains to be seen if it will be economically competitive with human drivers and if it will be actually safer. Many people enjoy driving, myself included, and many people are very safe drivers.
Speaking of New England how does FSD react to a deer jumping out of dense brush 30 feet ahead of the car at night. First it has to recognize the deer as an obstruction. Then it needs to decide whether to brake and maintain line of travel or brake and try to evade the deer while maintaining control of the car.
How does FSD identify black ice at night vs. wet pavement. Experienced human drivers can identify ice by the "feel."
How does FSD know what branches in the road after a wind storm are safe to drive over and which require braking and navigating around?
What about a limb hanging down into the road way?
How does FSD know where to drive on a back road with no lines on it?
How does FSD react to a plastic bag in the road? What about a rock that is similar colour and shape?
Etc.
Nothing about FSD is simple or close to being solved.
AGI is of course a wider domain than just driving, so it seems logical that we'll solve self driving before we reach AGI.
There will be problems and challenges that we'll be facing, but I'm not sure what the eventual point of your comment is?
How do we get truthful data on that without regulation? Just trust Tesla to report data fudging?
Even if Tesla is truthful, we need regulations to compare data in a consistent manner across car manufacturers.