Musk drives FSD Beta in Palo Alta

[CoachTerry]

https://apple.news/AXQFXy-R7R9-oiUEhBmIJEA

Sponsored

 

[Raebrek]

I watched that entire video and now I know who owns the Potato phone we have seen all the bad pictures from. It’s Elon.

 

[JBee]

I watched the whole thing too, was a bit boring except for the fact that FSD 12 is all taught by neural nets and nothing is explicitly coded...not even stop signs, lights, traffic or lanes. ?

They also went past Stanford and Suckerburgs place...

 

[charliemagpie]

In the event of something going wrong, TV networks would only have a potato video to go on.

 

[Arctic_White]

This is massive. I'm shocked that the stock didn't move, which tells me that FSD isn't priced in.

Once Tesla solves this, it'll be the largest asset value increase in modern history (not financial advice).

For those that didn't watch the video, it basically means that FSD is solved. We just need more data and computing power.

 

[FutureBoy]

For those that didn't watch the video, it basically means that FSD is solved. We just need more data and computing power.

I saw the video.
To make a small quibble..... FSD isn't really "solved". But it now has a very clear demonstration and proof of concept. There is still work to do getting it to work everywhere. Plus getting regulatory acceptance. As much as the tech is now to the point that we know the approach works, there are other things that have to "work" before FSD can be fully put to work.

 

[Arctic_White]

I saw the video.
To make a small quibble..... FSD isn't really "solved". But it now has a very clear demonstration and proof of concept. There is still work to do getting it to work everywhere. Plus getting regulatory acceptance. As much as the tech is now to the point that we know the approach works, there are other things that have to "work" before FSD can be fully put to work.

Yup. I meant solved in the sense that the path is very clear.

Prior to version 12, there was always the question of "what-if" and if we can engineer and solve all of the edge cases. However with version 12, when the AI writes its own code, this is no longer the case. This is huge, massive news yet no one is talking about it!!!

 

[TyPope]

I hope they let me check some box in FSD that says "Stop at stop signs." I know not all of you, heck, most of you, probably don't stop at stop signs and yes, yield signs are underutilized... But, around here, most people stop at stop signs and I don't want some California roll being required for me to comfortably use FSD.

 

[JBee]

This is massive. I'm shocked that the stock didn't move, which tells me that FSD isn't priced in.

Once Tesla solves this, it'll be the largest asset value increase in modern history (not financial advice).

For those that didn't watch the video, it basically means that FSD is solved. We just need more data and computing power.

Well "solved" in the sense that you now have the mechanism to teach it to get the results you want. There will still be a lot of teaching, but essentially, taking out the fixed coding meant that it could solve and find viable behavioural patterns without constraints, meaning it could solve from the base up.

The key here is that the learning is done by feeding a supercomputer in a building the correct behaviour (EM comments on that only 0.5% of people stop at a stop sign, so that data is rare) and that it then produces the correct algorithms that can be run onboard the vehicle itself (without being connected to the supercomputer) to create that behaviour in the FSD car.

If they can really pull this off, it has far reaching consequences, not only for FSD and robotaxi, but also anything that needs to navigate and behave in a certain way in "our reality". Be that Optimus for household or aged care, or manufacturing robots, military or interplanetary exploration. Technology that can mimic human behaviour with unnatural precision.

 

[FutureBoy]

Yup. I meant solved in the sense that the path is very clear.

Prior to version 12, there was always the question of "what-if" and if we can engineer and solve all of the edge cases. However with version 12, when the AI writes its own code, this is no longer the case. This is huge, massive news yet no one is talking about it!!!

At the moment I'm happy that the stock is not reacting. I need more time to buy more stock. The longer the price stays down the better for me. At some point it is going to actually pop. I definitely want to push that out as far as possible at this point.

 

[Arctic_White]

Well "solved" in the sense that you now have the mechanism to teach it to get the results you want. There will still be a lot of teaching, but essentially, taking out the fixed coding meant that it could solve and find viable behavioural patterns without constraints, meaning it could solve from the base up.

The key here is that the learning is done by feeding a supercomputer in a building the correct behaviour (EM comments on that only 0.5% of people stop at a stop sign, so that data is rare) and that it then produces the correct algorithms that can be run onboard the vehicle itself (without being connected to the supercomputer) to create that behaviour in the FSD car.

If they can really pull this off, it has far reaching consequences, not only for FSD and robotaxi, but also anything that needs to navigate and behave in a certain way in "our reality". Be that Optimus for household or aged care, or manufacturing robots, military or interplanetary exploration. Technology that can mimic human behaviour with unnatural precision.

100% this.

The impact of this is very far-reaching.

As you suggested, it'll take time. FSD won't be solved tomorrow and may not even be solved by 2024. But end of 2025, or 2026 looks more promising.

My model assumes Tesla solves FSD by 2030, so I've got 7 more years to go. With version 12, however, I think it'll be sooner.

 

[Arctic_White]

At the moment I'm happy that the stock is not reacting. I need more time to buy more stock. The longer the price stays down the better for me. At some point it is going to actually pop. I definitely want to push that out as far as possible at this point.

Makes sense, if you're buying each month.

I already have enough and given that I'm still working, I don't care once the pop actually happens. I think it'll happen when people realize what Tesla is able to do and what the implications will be for everything.

This is both scary and exciting.

 

[JBee]

I hope they let me check some box in FSD that says "Stop at stop signs." I know not all of you, heck, most of you, probably don't stop at stop signs and yes, yield signs are underutilized... But, around here, most people stop at stop signs and I don't want some California roll being required for me to comfortably use FSD.

lol. Technically, if everyone else stops at stop signs, you can just always drive through without ever getting hit.

BTW the US 4way stop crossings are more fist come first serve, and are terrible for traffic flow. Better off having a roundabout...aka traffic circle. (not sure what the difference is in the US, the satnav used both terms)

Here in Australia, they don't do crossings very much, and try to avoid them by intentionally offsetting the more minor road so it forms two t-intersections.New sub-divisions are intentionally designed without any crossings at all.

 

[TyPope]

I get that people don't want FSD to stop for them. I just want the choice of actually stopping.

 

[Jhodgesatmb]

I watched the whole thing too, was a bit boring except for the fact that FSD 12 is all taught by neural nets and nothing is explicitly coded...not even stop signs, lights, traffic or lanes. ?

They also went past Stanford and Suckerburgs place...

They have not 'explicitly coded' anything for image recognition in a long time, but they have extensively trained (i.e., labeled and then done reinforcement learning) all aspects related to video for a long time. What they are doing now is training the navigation parts of the puzzle, so someone still has to do the reinforcement part (what is right behavior vs. wrong behavior) rather than using heuristics (i.e., rules, presumably in code). Rules allow the system to tell us, for example, that the vehicle is changing lanes to maintain speed, so all of that interaction will be gone forever. If FSD does something, whether or not you like it, you will have no explanation (and will never get one). On the other hand, instead of having to teach it to recognize, and respond to, warning signs (a pet peeve of mine), it will be trained to slow down, for example when on windy mountain roads, or around schools, which is the right thing to do. There are pluses and minuses. It would be a good idea to watch @aidrivr's YouTube video as he did a good job of articulating many of these things, though he got the reinforcement learning thing wrong (imho).

Sponsored