Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Feasibility of FSD (incl. AI and Dojo)

This site may earn commission on affiliate links.

Singuy

Well-Known Member
Jun 28, 2018
9,005
90,563
US
Don't know if posted before, but David Lee just made a video about it.

View attachment 654057
This can mean a lot of things.
It can be a plausible explanation for why the FSD download button is not there yet - as well as an 'excuse' for the lateness of FSD v9.

My interpretation is that it is stating that the 'purely' car-version of FSD is not enough. Elon does not say AGI directly, or limited subset AGI. But ... he seems to be inching in that direction. I have a sense that he is re-framing FSD as not only car FSD, but a kind of general 'being-in-the-world' awareness/navigation.
Perhaps not FSD, but FSM (Full Self Moving)

This may be short term bad news for people expecting FSD to be right around the corner. Middle or long term it is good news: As I see it, when FSD(FSM) arrives, it will be surprisingly good. And very generic. (Hint: Robots)

TLDR:
FSD is not around the corner, it takes a lot of time. But ... it will work well when (if) it arrives. And perhaps solve a large part of robotics as well.

UPDATE: Just saw all of Daves new video. One of his best - highly recommended.
Dave thinks that Tesla will go all in on various AI-applications. I think that is likely. But it is even more likely, to me, that Tesla will actually go into robotics, specifically humanoid robots. Tesla's core strengths are not AI in isolation, but products using AI (at least, that is what we think is the likely outcome for the FSD-project)
What do humanoid robots need?
  • Vision based AI, (FSM)
  • Very good manufacturing making robust, high quality products in high volume.
  • Batteries: cheap and good.
  • Systems integration.
Who can do all of above?
Who would you buy a robot from? Goggle or Facebook - no seriously, with that privacy baggage? And no fabrication expertise?
Also: What is the most challenging task? Making a flying drone to deliver pizzas? Or a golf cart with limited autonomy for delivering packages?
No.
The biggest challenge is to develop a humanoid robot platform, for performing a wide range of tasks in everyday real life. Tesla robots.

Also: Robots are very useful, or perhaps necessary for Mars colonization
Also+: The latest Neuralink mind-pong video is interesting in many ways. Specifically relevant here is the very lean physical envelope of the coin-sized compute unit in the monkey's head. Having a large human-sized robot is going to present problems re. cooling, and energy conservation in a space much more limited than a car. Additionally, in a related tweet, Musk discussed the importance of good cooling when designing Dojo. Energy conservation, tight envelopes, cooling - there seems to be some very common trends amongst the various Musk Enterprises.

TLDR+:
Given Elon-time, the AI Day will be sometime in the autumn. Tesla will present Dojo. And talk about FSD and more general form of FSD (FSM). I think Musk will also bring up robotics. It is the next logical step.
You usually don't say "hey turns out in hindsight FSD requires us to solve real world ai problems first" without having solved FSD. Did Tesla freaken solved FSD? You can't claim you need x, y and z for the answer with NO OTHER WAY if you haven't gotten the answer already. How can he be so sure that this is the ONLY way unless FSD was not only solved but he tried to solve it in other ways and all failed?
 
You usually don't say "hey turns out in hindsight FSD requires us to solve real world ai problems first" without having solved FSD. Did Tesla freaken solved FSD? You can't claim you need x, y and z for the answer with NO OTHER WAY if you haven't gotten the answer already. How can he be so sure that this is the ONLY way unless FSD was not only solved but he tried to solve it in other ways and all failed?


The same way they were sure HW2.0 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW2.5 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW3.0 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW3.1 had "all" the HW needed for full self driving-until it didn't.

The same way they're sure NOW that HW3.2 has "all" the HW needed for full self driving.

The same way they were sure using radar as a primary sensor was a key to full self driving- until it wasn't.

The same way they were sure using way more robots was the solution for building Model 3s- until it wasn't.

The same way Elon has been varying degrees of "sure" FSD would be ready to do amazing stuff it still hasn't done annually since at least 2017.



Depending on your perspective it's either a towering strength or a glaring weakness (or more realistically some of both) that they go ahead SUPREMELY confident in the approach they're sure must be right even if they haven't actually figured out if it is yet.... but once they do realize they're wrong they're at least willing to go down a different path... (even if they now sound exactly as sure as last time THIS is the right one...)

Seems better than, say, pounding away with LIDAR for 10+ years and still being stuck in one tiny geofenced suburb in Arizona.



DWS (Dojo Web Services) as a cash machine - it is already 98% built*

It's not even expected to exist in version 1.0 until this fall, let alone the version that's ready to sell as a service- so I think your 98% thing is...optimistic.

I don't at all disagree it's potentially a massive revenue stream.

But let's maybe actually get the first version working and see how it is.

Could be mind boggling right out of the gate. Could be another HWx situation where we're several years and versions away from it being that way.
 
Glad to see that others are coming around to the idea that if Tesla can solve car autonomy, that means their solution is easily transferable to any other AI navigation puzzle. I know I've pointed this out many times and I'm sure many others have as well. Home navigation? Easy. Tanks or other large vehicles? Easy. Airplanes/drones? Just a matter of adding a Z axis. Boats? A breeze. Robots moving around factories?

It's bullish AF. Not willing to hazard a guess on the available market but could be similar to the autonomous car market.

The same way they were sure HW2.0 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW2.5 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW3.0 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW3.1 had "all" the HW needed for full self driving-until it didn't.

The same way they're sure NOW that HW3.2 has "all" the HW needed for full self driving.

For real robotaxi we will at least need hardware to clean camera lenses. It's not a huge deal for personal use, but we can't have robotaxis disabled on the side of the road every time roads get salted.
 
Last edited:
You usually don't say "hey turns out in hindsight FSD requires us to solve real world ai problems first" without having solved FSD. Did Tesla freaken solved FSD? You can't claim you need x, y and z for the answer with NO OTHER WAY if you haven't gotten the answer already. How can he be so sure that this is the ONLY way unless FSD was not only solved but he tried to solve it in other ways and all failed?

Sorry, but that's bad logic. You can try all the other methods, fail to solve FSD, and still come to the conclusion that solving general AI must be a requirement (read that as pre-requisite) to solving FSD.

So no, FSD isn't solved yet. The pre-requisites were simply identified and confirmed.
 
Glad to see that others are coming around to the idea that if Tesla can solve car autonomy, that means their solution is easily transferable to any other AI navigation puzzle. I know I've pointed this out many times and I'm sure many others have as well. Home navigation? Easy. Tanks or other large vehicles? Easy. Airplanes/drones? Just a matter of adding a Z axis. Boats? A breeze. Robots moving around factories?

It's bullish AF. Not willing to hazard a guess on the available market but could be similar to the autonomous car market.



For real robotaxi we will at least need hardware to clean camera lenses. It's not a huge deal for personal use, but we can't have robotaxis disabled on the side of the road every time roads get salted.
I believe the front facing and back facing cameras has a lot of redundancy. So the chance of all or multiple cameras blocked is very rare. This redundancy is how they built a birds eye view.
 
I have an alternate viewpoint on Elon's timeframes. I think it's possible that he gives these overoptimistic timeframes on purpose, to put pressure on his people to perform beyond reasonable expectations. After all, he does seem to elicit more out of his employees than anyone else I've ever seen. It may just be his way to keep the screws tight to ensure Tesla stays far ahead.
Can you imagine the FSD team coming to work the morning after his first “button” tweet? “Fuuuhhhhhh.....”
The same way they were sure HW2.0 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW2.5 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW3.0 had "all" the HW needed for full self driving-until it didn't.

The same way they were sure HW3.1 had "all" the HW needed for full self driving-until it didn't.

The same way they're sure NOW that HW3.2 has "all" the HW needed for full self driving.

The same way they were sure using radar as a primary sensor was a key to full self driving- until it wasn't.

The same way they were sure using way more robots was the solution for building Model 3s- until it wasn't.

The same way Elon has been varying degrees of "sure" FSD would be ready to do amazing stuff it still hasn't done annually since at least 2017.



Depending on your perspective it's either a towering strength or a glaring weakness (or more realistically some of both) that they go ahead SUPREMELY confident in the approach they're sure must be right even if they haven't actually figured out if it is yet.... but once they do realize they're wrong they're at least willing to go down a different path... (even if they now sound exactly as sure as last time THIS is the right one...)

Seems better than, say, pounding away with LIDAR for 10+ years and still being stuck in one tiny geofenced suburb in Arizona.





It's not even expected to exist in version 1.0 until this fall, let alone the version that's ready to sell as a service- so I think your 98% thing is...optimistic.

I don't at all disagree it's potentially a massive revenue stream.

But let's maybe actually get the first version working and see how it is.

Could be mind boggling right out of the gate. Could be another HWx situation where we're several years and versions away from it being that way.
Willingness to fail, learn, recover quickly and try again is one of Elon’s superpowers. Doing it quickly with an almost $1T company? Unheard of.
Glad to be along for the ride!
 
I believe the front facing and back facing cameras has a lot of redundancy. So the chance of all or multiple cameras blocked is very rare. This redundancy is how they built a birds eye view.

So apologies for my terrible art- but there are tons of spots all around the car with no redundancy either at all, or beyond a given distance.

See all the spots (roughly) circled in red only with 1 camera coverage.

redundancy.png
 
  • Like
Reactions: ThisStockGood
Do you consider instantaneous vision in this case (a single frame, interpreted one at a time) or do you take into account the "4D" strategy of vision over time, where something that was in a frame is in shallow memory for FSD?

I was discussing in the context of the discussion that L5 will require some way to clear dirty cameras.

Because otherwise lacking redundancy you have zero frames from one or more of those spots in red if a single camera gets blocked/dirty.


Functionality of basic AP will be fine either way- the front cameras besides there being 3 of em, are behind the windshield that can clean itself...

But anything that needs to know if an adjacent lane is safe to change into, or if it's clear to make a turn at an intersection from cross-traffic, could potentially be impossible to do safely with a single side camera (either front/side or rear/side facing) out of service.

If you're stopping at an stop sign wanting to turn right into cross traffic that has no sign, and say the front/left side camera is blocked, once the last car the front camera saw as you approached the sign has crossed, there's nothing for the car to "remember" to know if it's safe to go or not because there might have been other cars behind that one- just to cite one easy example of where memory won't fix the hole in the cars vision with just 1 blocked camera.
 
Last edited:
I believe the front facing and back facing cameras has a lot of redundancy. So the chance of all or multiple cameras blocked is very rare. This redundancy is how they built a birds eye view.
All I know is that Autopilot shuts down quite often when I drive in salt slush. Perhaps the front camera is enough for the car to pull over safely but that's not enough. But of course this is a trivial problem compared to creating a friggin AI that can drive. They could probably just design a new camera assembly with self contained cleaning hardware.

Worst case our older cars just can't be robotaxis in those weather conditions.
 
All I know is that Autopilot shuts down quite often when I drive in salt slush. Perhaps the front camera is enough for the car to pull over safely but that's not enough. But of course this is a trivial problem compared to creating a friggin AI that can drive. They could probably just design a new camera assembly with self contained cleaning hardware.

Worst case our older cars just can't be robotaxis in those weather conditions.
I'll put this video here for reference. Tesla cameras require a lot of blockage to stop functioning. So a slight occlusion will most likely not stop the system.

 
You can't come to that conclusion. There are infinite amount of methods, most haven't been developed. So you can't say anything is certain unless you already reached your answer. Also who to say you can't solve FSD without solving general AI first? You can have more confidence in that stance only if you solved FSD with general AI and it has been a many years later no other people have solved it without solving general AI first.

Now I'm in the camp that Tesla haven't solved FSD. However it's weird for Elon to be this certain, like the angels of future time travel visited him in his sleep and told him out of the infinite possibilities of solving FSD, none will get him there unless solving general AI first.

We're talking about interpretation of the signals from the cameras here, aka computer vision. There aren't infinite methods for that, since lidar has been explicitly removed as an option.

The driving part is pretty deterministic, as that's what the rules in the driver's handbook are for, so we're not talking about navigation algorithms.

Determining what the cameras see, that's what the AI is for, and that's why general AI is a pre-requisite, because you need to be able to identify what most of the objects are, in order to determine whether or not they're a potential threat. So from Musk's tweet, we can assume that they tried to limit the scope of what the AI needed to solve and found that there were too many edge cases where this failed. That's why the AI needs to be able to identify and understand its environment more generally, rather just as vehicles, pedestrians, walls, roads, lights, and signs.
 
I'll put this video here for reference. Tesla cameras require a lot of blockage to stop functioning. So a slight occlusion will most likely not stop the system.



That's not a terribly useful video, given it's an early beta of an L2 system.

The system inherently is meant that if it's about to do something stupid because a camera is blind (or any other reason), the human will stop it.

L5, not so much.

AI is very powerful. It's not magic.

It can't somehow fill in a literal black hole with no data of what is, for example, coming in cross traffic through the intersection while you're stopped with no visibility to that intersection.

For that matter, it keeps "working" even after he blocks -all- the cameras, presumably just using map and radar data (which is going away if you remember)- clearly the system in beta like that is not disabling itself for safety when degraded as it would have to for an L5 system.
 
Last edited:
That's not a terribly useful video, given it's an early beta of an L2 system.

The system inherently is meant that if it's about to do something stupid because a camera is blind (or any other reason), the human will stop it.

L5, not so much.

AI is very powerful. It's not magic.

It can't somehow fill in a literal black hole with no data of what is, for example, coming in cross traffic through the intersection while you're stopped with no visibility to that intersection.
I love that this post is so perfectly aligned with your TMC username “knightshade”
 
As I was watching all those M3 and MY in Shanghai, waiting to get loaded for transport, it occurred to me that Tesla will need a different way to transport the CyberTruck. I don’t think it will fit in a regular car hauler. Anyone knows if it does and any ideas how they will ship them?
I bet it will FSD itself to the buyer's home address.
 
We're talking about interpretation of the signals from the cameras here, aka computer vision. There aren't infinite methods for that, since lidar has been explicitly removed as an option.

The driving part is pretty deterministic, as that's what the rules in the driver's handbook are for, so we're not talking about navigation algorithms.

Determining what the cameras see, that's what the AI is for, and that's why general AI is a pre-requisite, because you need to be able to identify what most of the objects are, in order to determine whether or not they're a potential threat. So from Musk's tweet, we can assume that they tried to limit the scope of what the AI needed to solve and found that there were too many edge cases where this failed. That's why the AI needs to be able to identify and understand its environment more generally, rather just as vehicles, pedestrians, walls, roads, lights, and signs.
Are you saying it will take AGI (Artificial General Intelligence) to solve autonomy?
 
I have an alternate viewpoint on Elon's timeframes. I think it's possible that he gives these overoptimistic timeframes on purpose, to put pressure on his people to perform beyond reasonable expectations. After all, he does seem to elicit more out of his employees than anyone else I've ever seen. It may just be his way to keep the screws tight to ensure Tesla stays far ahead.

That's what I thought at one time, and it probably caused the team to scramble.

Now I bet the AI team has a hundred jokes about Elon's timelines and ultimatums. They know from experience that rushing things and pulling out all the stops to meet a deadline is wasteful and causes them to make mistakes. Now they probably just joke about it and keep doing what they know how to do best while making rational decisions regarding the best way to proceed.

They know they are the strongest FSD team in the world, on the most productive path, no one else is even breathing down their necks, they are the closest to the big prize and there's not a damn thing Elon can do about missed deadlines. They are working hard and not in a wasteful manner. But the process is such that it responds most favorably to a methodical nature and good decision making than to rushing and taking short-cuts. Elon knows this too but he's the kind of guy that wants the progress to happen yesterday and it must be a little frustrating for him. He knows the team thinks very highly of him, he probably goes in there and tells them how much he wants it to happen. Karpathy knows Elon tends to want to send out half-baked stuff so he convinces Elon to wait another few iterations and tells him how rapidly it's improving and how much better it will blow everyone away if it's more developed when the public gets their hands on it. I bet Karpathy knows there is a threshold that needs to be reached to ensure the public releases don't get shut down by the regulators. Karpathy seems very much like the kind of person that doesn't let the pressure get to him, he just calls it how he sees it and convinces Elon to wait a little longer. Elon respects Karpathy and knows how much it would set back the program if he left so he leaves him in the driver's seat (so to speak, lol!).

It will come when Karpathy wants to release it. And we will be better off for it. Meanwhile, Elon has to keep stringing us along. When it finally comes it's going to be awesome (but still not fully level 5 yet, obviously).
 
Are you saying it will take AGI (Artificial General Intelligence) to solve autonomy?

My reading of it is that AGI means something different from General AI. I read AGI as being artificial human intelligence (like being able to play billiards, poker, drive, AND troll the forums - aka pass a turing test), while general AI is still an expert-systems level intelligence, but without a defined scope like how watson, deep blue, and alpha go were topic specific AI. Like how Data couldn't tell a joke, despite possessing more knowledge (not the same as being smarter) than the average trekkie.
 
My reading of it is that AGI means something different from General AI. I read AGI as being artificial human intelligence (like being able to play billiards, poker, drive, AND troll the forums - aka pass a turing test), while general AI is still an expert-systems level intelligence, but without a defined scope like how watson, deep blue, and alpha go were topic specific AI. Like how Data couldn't tell a joke, despite possessing more knowledge (not the same as being smarter) than the average trekkie.
So the term "General AI" seems to be something you just made up and defined all by yourself.

I don't want to derail the thread so I'll leave at that. (Mods, you would probably do everyone a favor by deleting the entire discussion).
 
  • Like
Reactions: friit and mblakele