Feasibility of FSD (incl. AI and Dojo)

[Discoducky]

At what extent the FSD possibility is currently priced in TSLA stock price? When there is a drop in SP because of FSD FUD from MSM, is it because of a panic sell off or the lower probability of FSD occurring because of negative political legislation from FUD?

Has anyone every evaluated the probability of L5 FSD occurring, is it above 90%?
I’m halfway through Intelligence from Jeff Hawkins.
is FSD based on self learning neural network DOJO or programmed based on new input that has to be coded through the software?
just getting anxious that a machine can only be intelligent if it can make predictions. Can FSD make actual predictions from the future that is going to happen on the road from what the neural network learned from the past or from new input programmed in the software? At what extent L5 FSD has to make intelligent new predictions based on events of the last for it to be possible.

TL;DR - Yes, great read, soft supervised is my guess, HELLs YES, need to pass through L3 and L4 filters first but will be fast for humans.

Long boring version....

I personally do not think FSD has been priced in. It is such a huge deal that **IF** it were priced in, we'd see such a huge influx of buyers it would drain all other momentum stocks. Just the thought of never having to drive in traffic again makes someone dream of how much they would pay for that. Anyway...as it hasn't happened yet and it is effectively **magic** on 4/20/21, I think that when it is enabled as a BUTTON, people will lose their *sugar*. And yes, I think it is right around the corner, a very hard unprotected left turn, but nonetheless, Tesla's path is technically sound (yes I'm highly biased; some may know how much...tee hee hee).

Everyday, everyday, everyday that goes by is another day for training the most advanced AI in the functional world. You could argue that GPT-3 is more advanced, but the company that made it is an Elon company.....hmmmmm....

Based on self learning? No, I believe it be soft-supervised. In other words, it is supervised at an arms length. I think they are trying to figure out intent along with much better path planning and controls. But until they understand intent, it will always be hesitant, tentative, jittery, passive and above all hard to trust. My 2 cents...

L5 is a distant future, but not in calendar days, we have to pass the great L4 filter first and prior to that we need to pass the L3 filter, before that we need at least a few million turns that we can trust at L5. A better way to say this is your path to the grocery store has many turns (most likely) and by the beauty of math, a fraction of them are hard for AI. Some folks have a path to the food that is easier, these will go full L3 first and holy moly L4 will follow quickly in calendar days. This is what I think we are on the cusp of.

IT CAN TOTALLY MAKE PREDICTIONS!!! That is what it is doing! It is effectively predicting the very near, ms (millisecond by millisecond) future. That is only what it does, nothing else. Anyone that tells you different is trying to sell you something that is similar to the Shamwow!

And again, true L5 is off in the distant future. We will STILL be AMAZED at true L3. Realize that this means that you have like ~10 seconds to respond to the warnings. If the car is only doing L5 or L4 turns, you'll never have to pay attention. You won't know that, but the car will, the AI will, the engineers will and when they review the logs of the turn that returned a crappy confidence or a horrible intervention they will scratch their heads, actually they won't do that, the soft supervised AI should hopefully do the right thing on the next iteration of training. If it doesn't then it will need to be human analyzed.

Ok, I think I've said enough.

 

[JohnnyEnglish]

is FSD based on self learning neural network DOJO or programmed based on new input that has to be coded through the software?
just getting anxious that a machine can only be intelligent if it can make predictions. Can FSD make actual predictions from the future that is going to happen on the road from what the neural network learned from the past or from new input programmed in the software? At what extent L5 FSD has to make intelligent new predictions based on events of the last for it to be possible.

Probably the best way to answer these questions is to watch the Dave Lee Investing videos on youtube where Dave talks to James Douma. The first two explain how FSD uses neural networks for perception. There are further videos on FSD hardware (comparison with nVidia etc), Dojo etc.

Rob Maurer has just started a similar discussion with James. It is covering similar ground but the series is not complete yet.

 

[LiveLong&Profit]

Dave Lee made a new video:

It is just a teaser for a upcoming video, but interesting nevertheless.

In the upcoming interview the ARK analyst focuses on data advantages.
So, in FSD, data is the 'holy' trifecta

• Very valuable
• Hard to get (scaling is super hard, expensive - or both)
• Does not become less valuable over time (Always valuable to continue to collect new edge cases used to improve FSD thus saving future lives)

There is a strong network or actual forcing function re. FSD data: The more data, the better FSD gets. The better FSD gets, the more users. The more users, the better data. Repeat the loop.

So, it follows that FSD data is digital gold.

Another metaphor (my own) would be to compare FSD with bitcoin mining:

The car drives and mines FSD data - worth gold.
In contrast to bitcoin, Tesla pays nothing for the 'compute' (the car) - in fact, it generates income!

Also, bitcoin mining does require power - sometimes a lot of it. Again, payed by the Tesla owners.
Lastly, looking after a bitcoin farm can be automated, but you have to follow up and follow through: You need some admins, who will shop around for cheap power, monitor that all the compute is running and earning money. Sometimes, but rarely, the bitcoin mining admins have to take action: Take mining equipment out of commission, or buy more compute in the cloud world, or many such tasks.
A Tesla' driver fulfills the same function: Much of driving is banal and of little or no value.
But, sometimes the Tesla driver (Bitcoin miner admin) provides tremendous value, by getting the automated setup out of trouble: Take action when FSD makes a mistake or hesitates.

Finally, you can argue that FSD data is more worth than bitcoin per unit of compute. Now, I haven't done the heavy lifting to fully prove this, so I will just outline the argument. Bitcoin scales sorta linearly: It gets progressively more expensive to mine, but what you get is also progressively worth more (I know, wild fluctuations in bitcoin value , but still directionally true that even though the cost of mining has exploded, so has the value.

Contrast to FSD data 'mining': It scales sub-linearly - it gets cheaper per data unit the more data is gathered. And the data continues to be worth a lot for the foreseeable future.
Buying huge amounts of GPUs and FPGA is not cheap, neither is electricity, nor do admins for monitoring/automating huge compute arrays of equipment work for free. In contrast, the primary use case of a Tesla car is transportation. The FSD 'mining' piggy-backs on this. The initial and running cost of sensors and custom-built chips in all new Teslas and the cost of having a great FSD team is not cheap, for sure. The great thing is that it scales very well: The total costs for harvesting FSD data for 50 million robotaxis is not 100X of harvesting data from ½ million cars but the data you get is 100X. And because the data is so widely distributed (cause of the huge amount) you are sure to get hold of the nuggets of not gold, but 'diamond': The edge cases and related training which makes the real difference.
The real difference between an 'FSD sorta works most of the time FSD', to 'OK robotaxi, but sometimes you have to take the wheel as a customer', to the robotaxi service you would, quite literally, trust with the lives of your spouse, or young kids.

TLDR:
Tesla FSD is like bitcoin mining: It creates digital gold.
But, not only is compute free for Tesla, power and admin workforce is free also.
And not only free - Tesla makes money on creating equipment that will mine them digital gold in the years to come. FSD data might become, over the years to come, more worth that bitcoins, per unit of compute.

It follows, that we should heartily thank all Tesla car owners driving with FSD who are not also investors:
They literally pay money and work for free to gift us future fortunes!

 

[lascavarian]

So, it follows that FSD data is digital gold.

I would say that Driving Data is foundational to NN success in autonomous driving. Digital Gold is a made up marketing term IMO.

 

[LiveLong&Profit]

I would say that Driving Data is foundational to NN success in autonomous driving. Digital Gold is a made up marketing term IMO.

True. Just trying to make the metaphor re. Bitcoin stick, but perhaps overdid it.

 

[Xepa777]

In my podcast, "Driving Green", we recently interviewed Sven Beiker, a Stanford Professor and Managing Director of Silicon Valley Mobility. There's a lot of great insight on city design considerations, most economical models of robotaxis, and a history of the field here. Though he's not a huge Tesla fan, his general insights are really compelling in terms of the economical model of FSD.

Will robotaxis actually make things cheaper for us? That's a concept he challenges.

There's a fantastic Elon Musk cameo in the beginning

00:00 - Show Open
00:30 - Introduction, Sven Beiker
05:30 - History of Autonomy and ADAS
10:15 - Autonomous Hardware Trends
18:30 - City vs Suburban Environments
25:00 - What is Mobility?
35:15 - What will Autonomy Unlock?
44:00 - City Design
49:00 - Ride Sharing
52:55 - Billy Riggs Zoom-Bombing
56:25 - Tesla's "FSD" Approach
01:09:53 - Carbon Footprint

 

[ZeApelido]

In my podcast, "Driving Green", we recently interviewed Sven Beiker, a Stanford Professor and Managing Director of Silicon Valley Mobility. There's a lot of great insight on city design considerations, most economical models of robotaxis, and a history of the field here. Though he's not a huge Tesla fan, his general insights are really compelling in terms of the economical model of FSD.

Will robotaxis actually make things cheaper for us? That's a concept he challenges.

There's a fantastic Elon Musk cameo in the beginning

00:00 - Show Open
00:30 - Introduction, Sven Beiker
05:30 - History of Autonomy and ADAS
10:15 - Autonomous Hardware Trends
18:30 - City vs Suburban Environments
25:00 - What is Mobility?
35:15 - What will Autonomy Unlock?
44:00 - City Design
49:00 - Ride Sharing
52:55 - Billy Riggs Zoom-Bombing
56:25 - Tesla's "FSD" Approach
01:09:53 - Carbon Footprint

Just watched Tesla portion. He talks about Tesla is going to hit a wall in progress soon...

Except I don't know how this guy has the background to even competently talk about that?

 

[Xepa777]

Just watched Tesla portion. He talks about Tesla is going to hit a wall in progress soon...

Except I don't know how this guy has the background to even competently talk about that?

Because you literally skipped the whole podcast lol. I wasn't there to debate him, especially since he's not an AI engineer, just get his POV. And he does provide a lot of great problems and considerations that are industry relevant that are not talked about often now.

 

[ZeApelido]

Because you literally skipped the whole podcast lol. I wasn't there to debate him, especially since he's not an AI engineer, just get his POV. And he does provide a lot of great problems and considerations that are industry relevant that are not talked about often now.

I will watch the rest I promise!

Certainly I understood he's not an AI engineer. That's generally my frustration with "experts" in autonomy. They're not actually experts in the core competency that will drive the success of autonomous cars - signal processing and machine learning.

Thus it would be better if added more of the "I don't know I'm not an expert in this area" preface before saying Tesla's improvements are about to hit a wall

I would actually implore more podcasts to try to interview machine learning and signal processing experts, as they have the closest to "real" insight into this area. Kinda like James Douma

*yes I am one of those people, but would be better to find some of the folks who've dealt with more video and lidar algorithms.

 

[Xepa777]

I will watch the rest I promise!

Certainly I understood he's not an AI engineer. That's generally my frustration with "experts" in autonomy. They're not actually experts in the core competency that will drive the success of autonomous cars - signal processing and machine learning.

Thus it would be better if added more of the "I don't know I'm not an expert in this area" preface before saying Tesla's improvements are about to hit a wall

I would actually implore more podcasts to try to interview machine learning and signal processing experts, as they have the closest to "real" insight into this area. Kinda like James Douma

*yes I am one of those people, but would be better to find some of the folks who've dealt with more video and lidar algorithms.

the 30 second Elon musk intro and 52:55 zoom bombing parts are great comedic relief

 

[SOULPEDL]

OT:

https://twitter.com/x/status/1387893690354515970

View attachment 658307

What does "Probability Distribution" look like visually?
Is that like seeing double and have to decide which line to follow? Oh... carry on.

 

[heltok]

What does "Probability Distribution" look like visually?
Is that like seeing double and have to decide which line to follow? Oh... carry on.

First I am not sure if Elon is correct that it is a probability distribution, more likely it is a score of 0 to 1 where 1 is that the algorithm is very certain it is something and 0 it is very certain it is not. Maybe the score can be normalized to be a probability but I think that is too much work. As for the distribution my guess is that they will show a 2D image where higher scores(or as Musk sees it probabilities) have higher opacity. So for example a car would be very solid if the camera can see it, but fade out as the camera cannot see it anymore. Something like this:

Where more white means that they are certain of what is driveable area and more red indicates intersection and dark blue indicates non-driveable area.

I think Musk is using the term probability distribution a bit different from how most other roboticists are using it. Not sure how much statistics/sensor fusion/probabilistic robotics he has studied. I could be wrong, we will see in a few weeks, but that is my take.

This is how a probability distribution can be visualized:

But I don’t think that is how Tesla will do it...

 

[22522]

First I am not sure if Elon is correct that it is a probability distribution, more likely it is a score of 0 to 1 where 1 is that the algorithm is very certain it is something and 0 it is very certain it is not. Maybe the score can be normalized to be a probability but I think that is too much work. As for the distribution my guess is that they will show a 2D image where higher scores(or as Musk sees it probabilities) have higher opacity. So for example a car would be very solid if the camera can see it, but fade out as the camera cannot see it anymore. Something like this:
View attachment 658374

Where more white means that they are certain of what is driveable area and more red indicates intersection and dark blue indicates non-driveable area.

I think Musk is using the term probability distribution a bit different from how most other roboticists are using it. Not sure how much statistics/sensor fusion/probabilistic robotics he has studied. I could be wrong, we will see in a few weeks, but that is my take.

This is how a probability distribution can be visualized:
View attachment 658376

But I don’t think that is how Tesla will do it...

Hi,

Thank you for making conversation happen. I expect opacity to be used as 1-y axis as drawn in the second row in Figure 2.
I don't think an abstraction as described as "more red indicates intersection" will be used. That is more of a time history probability of crossing traffic - information not known to the vehicle/cameras.

So probability of a solid object in a location (now) is what can be shown. And it can most intuitively be shown as opacity/size of an image on the screen.

 

[LiveLong&Profit]

For sure: Dojo as a service is more imminent than (for now an extremely speculative future division of Tesla namely) Tesla Robotics.
My personal opinion is that even though Elon says 'any ML' I would assume that he implicitly means 'any ML with a strong vision component'.
For Tesla to be competitive in the solution space of Vision-based ML is already extremely bullish.
To think that Tesla is also very competitive in generic ML without a strong vision component is more bullish that even I am prepared to be. It would mean that Tesla has leapfrogged Google on all fronts. Strategically, that is not necessary. Strong-vision based AI/ML combined with manufacturing expertise and systems integration, and good batteries is more that enough for making excellent humanoid robots.

Even being very optimistic, I don't see a Tesla Robot before 2025. (Perhaps before for strictly internal use, and held close to the vest)

Not even sure that Elon will mention robots at the AI Day sometime in the coming autumn - it might be 'too much' in the sense that the current product catalogue Tesla already has teased or mentioned is far from being fulfilled. (Roadster2, Semi, CT, Dojo, FSD, Tesla future 25K car, etc)

Even though Dojo as a service is more immanent, consider, however the TAM of humanoid robots:
Self driving cars, and Dojo might be very big. But what is the upper limit of humanoid robotics market? 1 robot per employee in all service and production sectors - a 'robot buddy'?
A couple of robots per wealthy household? One robot for most households?

Looking futher ahead, AFIK that there are around 3 very hard constraints on wealth for any age:

• Regulation/peace/law/justice
• Labor
• Energy

Peace/law etc is very hard to 'solve' - how society should be structured varies greatly over time. No company can solve that.
But abundant labor and abundant energy is very important to the wealth of any society. Tesla can solve both!
Combine that with UBI (which Elon is a strong proponent of) I could se the TAM of humanoid robotics being virtually without limits in a 30-50 year time frame.

Dave Lee has a new video out describing Tesla Vision/FSD as just the first baby step in a future world of real world robots doing, in the limit, basically all tasks done by humans today - and better than humans:

 

[Missile Toad]

Ph'nglui mglw'nafh lawyers R'lyeh wgah'nagl fhtagn

 

[DrDabbles]

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?

He isn't. He's holding a hammer, so everything he sees is a nail. Computer vision rain sensors should pretty much give you the red flag that they're incapable of even replicating analog device accuracy with their neural nets. And as far as solving "AI" in a general sense? We're centuries away from that, probably. If the last >50 years of work is any prediction for the future (it is), then we aren't even at a phase where we're standing, let alone walking or running.

 

[mickificki]

Interesting look at pillars during Galileo Russell’s fsd drive thru Seattle.

https://twitter.com/x/status/1415033967749697538

 

[DrDabbles]

The fact it doesn't identify the pillars as an obstacle is crazy. IMO, this is the clearest sign of just how far they have to go still, and how very far away anything beyond L2 assistance is.