Feasibility of FSD (incl. AI and Dojo)

[Discoducky]

I'll keep this brief, but we are on the cusp of release.

Tesla can solve (9 9's success rate) FSD for 99% of miles/turns/intersections driven (knowing relative environmental conditions through a basic API call) as they will then also know the 1% where they haven't (simple GPS and maps) and then set whatever price they want.

Today is 10k and you just have to ask yourself what you'd pay to not have to drive for 99% of the time.

Yes, this will happen in phases, but it is going to happen soon with a wide release.

My guess is 20k by July and $199/mo subscription.

 

[LiveLong&Profit]

These are all great applications of a best in class AI supercomputer/ NN training machine - but in my opinion, from a revenues POV I think the more immediate and profitable outcome to look forward to is the fact that Tesla will open their Dojo computer to anyone, aka Dojo Web Service (DWS) - in case you forgot, Amazon didn't have much profits UNTIL AWS (Amazon Web Services) launched, and is now powering the infrastructure of a lot of website engines. And has been the top target of Microsoft/ IBM/ Oracle etc ever since.

View attachment 654078

DWS (Dojo Web Services) as a cash machine - it is already 98% built*, as Elon mentioned all it needs is working out the bugs, then of course setting up the infrastructure for letting others rent time on Dojo. Just imagine the military applications (sorry, a bit out of line with Tesla's mission), precision targeting of enemy aircraft/ tanks, pharma drug research - the Nurburgring equivalent would be pitching Dojo vs AlphaGo

*the fact that FSD hasn't been solved to the 99.999999% degree is not relevant to the DWS launch, BTW


View attachment 654079

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.

 

[LiveLong&Profit]

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 may be right. Yes, that is a very logical way to interpret Elons statement.
If so, that is even more bullish than I have previously thought.'

However, there is another possibility.
Remember a few year back when Elon said in a tweet that FSD was basically solved? At that time, before FSD-beta, an other hints, it seemed a bit premature, bragging/hyping.
That doesn't mean he was wrong, but it may mean that even though it is solved in Elons mind, it takes time for the solution to be ready to release.

You may be right that FSD it is fully or almost solved now, but consider that even though FSD may be solved, in principle, it can take time to manifest. Meaning that Tesla still think they need to accumulate more data, more 'evidence', more perfection before they dare to declare FSD 'done'. It is not a matter of being timid, but thinking the next 5-10 steps ahead.

Note, that when (if) FSD is really done, and done well, Tesla will be even more subject to strong hostile attacks than before, because Tesla having the only working FSD immediately puts a traditional automotive and big oil in a very weak position.

So, it is not only a matter of is FSD ready or not, it is also a matter of having a fortified position before releasing if because Tesla will come under very hard attack from industries facing a mortal threat.

 

[WarpedOne]

So, it is not only a matter of is FSD ready or not, it is also a matter of having a fortified position before releasing if because Tesla will come under very hard attack from industries facing a mortal threat.

Having 100k, 1M or 5M cars on the road at that moment will make a significant difference in winning public support.
It is easier to kill something almost no one ever saw than something many did see and experienced first hand.

 

[heltok]

Real World AI = AI that is acting in the physical world, interacting with humans and nature. In contrast with games, simulation etc where the world is very controlled, less complex and it costs less to make mistakes.

Tesla has solved how to solve real world AI problems. To do this you need
1. A method of growing a dataset with useful edge cases to converge on high performance (software 2.0 stack, data engine, labelers etc)
2. Hardware to train the neural network (Dojo)
3. Hardware to run the neural network( HW3/HW4 etc)

What Tesla has done is set up a complete system for this, developing the entire software 2.0 stack.

Why are Tesla not ready with FSD yet? Because it turned out that solving FSD was a lot harder than we intially thought.

Google had driven 300 000 autonomous miles around San Francisco in 2012. They have had a lot of skilled engineer, led by the ”Probabilistic Robotics” author Sebastian Thrun, having had plenty of hardware infrastructure and grown the team a lot since then. Yet for 9 years of intense work by skilled engineers they are still not ready. Because it was a lot harder than they thought.

In comes Tesla. They had the best autopilot on the streets, but in 2016 they split from Mobileye. The split was really messy, they had to reimplement entire Mobileye intellectual property from scratch, which set them back, but they did this very quickly. But already in 2016 they had a FSD car up and driving for the demo ”Paint it black”. Clearly they had a pretty great team to be able to do this. Since then it has been mostly delays and minor releases, because it was harder than they thought.

Here comes Elon. Like Google and like Tesla he is really frustrated that progress is not happening as fast as he wants. He tries to understand the problem. He kicks out Chris Lattner, the top lead from Apple who made LLVM and Swift, clearly a really good project leader, because he realizes that it’s not enough to have a team of great developers, like Google and Tesla had. There needs to be a totally different approach, it’s needed to solve ”Real World AI”. What is needed is three components, the entire data engine software stack, see Karpathy software 2.0, dedicated hardware, see Jim Killer in Lex Fridman both for training(Dojo) and for inference(HW4) see Pete Bannon in Autonomy day.

With this approach, there is a clear path with constant improvements towards the end goal. The only way that we know that progress can be made on what the elite teams at Waymo and Tesla has struggled so much with over the last five years.

It might seem that Tesla are constantly missing goals and that no progress is happening. But finally all the pieces are almost into place and we will start seeing the march towards 99.9999%. And consider, if the elite teams at Tesla and Waymo(who managed to do so much from 2009-2016 and have since then gotten so much more resources) have struggled with making progress, how hard will it be for VW, Toyota etc to make progress?

Relevant videos:

 

[Buckminster]

Elon's deleted tweet (biggest company) timeline lines up perfectly with AI day in July. Just saying...

 

[LiveLong&Profit]

I'll keep this brief, but we are on the cusp of release.

Tesla can solve (9 9's success rate) FSD for 99% of miles/turns/intersections driven (knowing relative environmental conditions through a basic API call) as they will then also know the 1% where they haven't (simple GPS and maps) and then set whatever price they want.

Today is 10k and you just have to ask yourself what you'd pay to not have to drive for 99% of the time.

Yes, this will happen in phases, but it is going to happen soon with a wide release.

My guess is 20k by July and $199/mo subscription.

Don't have any problem with 20K (actually think it is too low). But, 199/m seems low in comparison...

 

[LiveLong&Profit]

Having 100k, 1M or 5M cars on the road at that moment will make a significant difference in winning public support.
It is easier to kill something almost no one ever saw than something many did see and experienced first hand.

Yes - agree. The more cars, the merrier.

I vaguely recall the (in)famous Tesla uber bull (and criminal defense lawyer) Warren Redlich detailing that 80-90% of all traffic violations, including very dangerous ones are committed by *very* few people. So FSD doesn't have to be very good for it to be vastly better than alcoholics, drug-users, severely depressed and maybe suicidal drivers, or very inexperienced teenagers with boiling hormones. Besides from all the emotional torment of the suffering people involved in accidents and their families, having working, adult people killed all the time in car crashes is very bad for the economy!

Benevolent and educated politicians know this. And they will side with Tesla, knowing that FSD is a great boon for society.

On the other hand, FSD has to be good enough to withstand future foul play, opposition and skullduggery from traditional auto and big oil, using all the tools of their disposal: FUD, captured politicians sabotaging progress, pay-for-play media bamboozling the public - and other methods too foul to be mentioned.

 

[jeewee3000]

Real World AI = AI that is acting in the physical world, interacting with humans and nature. In contrast with games, simulation etc where the world is very controlled, less complex and it costs less to make mistakes.

Tesla has solved how to solve real world AI problems. To do this you need
1. A method of growing a dataset with useful edge cases to converge on high performance (software 2.0 stack, data engine, labelers etc)
2. Hardware to train the neural network (Dojo)
3. Hardware to run the neural network( HW3/HW4 etc)

What Tesla has done is set up a complete system for this, developing the entire software 2.0 stack.

Why are Tesla not ready with FSD yet? Because it turned out that solving FSD was a lot harder than we intially thought.

Google had driven 300 000 autonomous miles around San Francisco in 2012. They have had a lot of skilled engineer, led by the ”Probabilistic Robotics” author Sebastian Thrun, having had plenty of hardware infrastructure and grown the team a lot since then. Yet for 9 years of intense work by skilled engineers they are still not ready. Because it was a lot harder than they thought.

In comes Tesla. They had the best autopilot on the streets, but in 2016 they split from Mobileye. The split was really messy, they had to reimplement entire Mobileye intellectual property from scratch, which set them back, but they did this very quickly. But already in 2016 they had a FSD car up and driving for the demo ”Paint it black”. Clearly they had a pretty great team to be able to do this. Since then it has been mostly delays and minor releases, because it was harder than they thought.

Here comes Elon. Like Google and like Tesla he is really frustrated that progress is not happening as fast as he wants. He tries to understand the problem. He kicks out Chris Lattner, the top lead from Apple who made LLVM and Swift, clearly a really good project leader, because he realizes that it’s not enough to have a team of great developers, like Google and Tesla had. There needs to be a totally different approach, it’s needed to solve ”Real World AI”. What is needed is three components, the entire data engine software stack, see Karpathy software 2.0, dedicated hardware, see Jim Killer in Lex Fridman both for training(Dojo) and for inference(HW4) see Pete Bannon in Autonomy day.

With this approach, there is a clear path with constant improvements towards the end goal. The only way that we know that progress can be made on what the elite teams at Waymo and Tesla has struggled so much with over the last five years.

It might seem that Tesla are constantly missing goals and that no progress is happening. But finally all the pieces are almost into place and we will start seeing the march towards 99.9999%. And consider, if the elite teams at Tesla and Waymo(who managed to do so much from 2009-2016 and have since then gotten so much more resources) have struggled with making progress, how hard will it be for VW, Toyota etc to make progress?

Relevant videos:

You know you're spending (too?) much time on a subject when you've seen all these videos before they were posted.

("No honey, I'm not going down a youtube rabbithole, I'm researching for investment purposes!")

But great post either way. If you haven't seen these videos, do so.

 

[WarpedOne]

I vaguely recall the (in)famous Tesla uber bull (and criminal defense lawyer) Warren Redlich detailing that 80-90% of all traffic violations, including very dangerous ones are committed by *very* few people. So FSD doesn't have to be very good for it to be vastly better than alcoholics, drug-users, severely depressed and maybe suicidal drivers, or very inexperienced teenagers with boiling hormones

Yes, but this also means it has to be better than your average common "non-violator".
Though I suspect this has been so for some time already...

 

[WarpedOne]

("No honey, I'm not going down a youtube rabbithole, I'm researching for investment purposes!")

This is *not* a hyperbole - my best investment of time was invested in TMC and YT.
Deal with it.

 

[Discoducky]

Real World AI = AI that is acting in the physical world, interacting with humans and nature. In contrast with games, simulation etc where the world is very controlled, less complex and it costs less to make mistakes.

Tesla has solved how to solve real world AI problems. To do this you need
1. A method of growing a dataset with useful edge cases to converge on high performance (software 2.0 stack, data engine, labelers etc)
2. Hardware to train the neural network (Dojo)
3. Hardware to run the neural network( HW3/HW4 etc)

What Tesla has done is set up a complete system for this, developing the entire software 2.0 stack.

Why are Tesla not ready with FSD yet? Because it turned out that solving FSD was a lot harder than we intially thought.

Google had driven 300 000 autonomous miles around San Francisco in 2012. They have had a lot of skilled engineer, led by the ”Probabilistic Robotics” author Sebastian Thrun, having had plenty of hardware infrastructure and grown the team a lot since then. Yet for 9 years of intense work by skilled engineers they are still not ready. Because it was a lot harder than they thought.

In comes Tesla. They had the best autopilot on the streets, but in 2016 they split from Mobileye. The split was really messy, they had to reimplement entire Mobileye intellectual property from scratch, which set them back, but they did this very quickly. But already in 2016 they had a FSD car up and driving for the demo ”Paint it black”. Clearly they had a pretty great team to be able to do this. Since then it has been mostly delays and minor releases, because it was harder than they thought.

Here comes Elon. Like Google and like Tesla he is really frustrated that progress is not happening as fast as he wants. He tries to understand the problem. He kicks out Chris Lattner, the top lead from Apple who made LLVM and Swift, clearly a really good project leader, because he realizes that it’s not enough to have a team of great developers, like Google and Tesla had. There needs to be a totally different approach, it’s needed to solve ”Real World AI”. What is needed is three components, the entire data engine software stack, see Karpathy software 2.0, dedicated hardware, see Jim Killer in Lex Fridman both for training(Dojo) and for inference(HW4) see Pete Bannon in Autonomy day.

With this approach, there is a clear path with constant improvements towards the end goal. The only way that we know that progress can be made on what the elite teams at Waymo and Tesla has struggled so much with over the last five years.

It might seem that Tesla are constantly missing goals and that no progress is happening. But finally all the pieces are almost into place and we will start seeing the march towards 99.9999%. And consider, if the elite teams at Tesla and Waymo(who managed to do so much from 2009-2016 and have since then gotten so much more resources) have struggled with making progress, how hard will it be for VW, Toyota etc to make progress?

Relevant videos:

Wow, really great summary! I'd vote this for the merit thread. And I was a part of that path! Well done!

 

[Discoducky]

Don't have any problem with 20K (actually think it is too low). But, 199/m seems low in comparison...

The $199 comes from the value associated to a monthly spend of expendable income as well as a gain in value due to a loss of productivity during commute times. I know too many people who would give anything to leave their house an hour later each day to send email or texts while commuting in their car. The percentage of people willing to pay more than $199 is way less due to phycological reasons. It just won't compute to put it bluntly.

 

[insaneoctane]

Weekend OT re: Tesla Autopilot

So, I don’t have the fancy new beta FSD, but I have AP and use it often. Last night, while driving on the freeway (6 lanes wide on my side BTW) at about 80 MPH, my car began to slow down told me “stopping for traffic signal“. In the middle of the freeway! Luckily it was late at night with very little traffic. I over-rode it before it slowed too much. But, really? I know that the team has been working on city streets FSD, but this “edge case” seems like a significant miss in a product (freeway AP) that Tesla considers fairly mature. I’m sure the vision system saw a light on a neighboring street, but There’s no real easy way to transmit this data to Tesla. I personally think freeway AP needs better situational awareness like city streets Is showing... Just some perspective.

It happened here while traveling N or W bound....

Ronald Reagan Fwy · Los Angeles, CA

Los Angeles, CA

goo.gl

 

[Xd85]

Weekend OT re: Tesla Autopilot

So, I don’t have the fancy new beta FSD, but I have AP and use it often. Last night, while driving on the freeway (6 lanes wide on my side BTW) at about 80 MPH, my car began to slow down told me “stopping for traffic signal“. In the middle of the freeway! Luckily it was late at night with very little traffic. I over-rode it before it slowed too much. But, really? I know that the team has been working on city streets FSD, but this “edge case” seems like a significant miss in a product (freeway AP) that Tesla considers fairly mature. I’m sure the vision system saw a light on a neighboring street, but There’s no real easy way to transmit this data to Tesla. I personally think freeway AP needs better situational awareness like city streets Is showing... Just some perspective.

It happened here while traveling N or W bound....

Ronald Reagan Fwy · Los Angeles, CA

Los Angeles, CA

goo.gl

happened to me too. My daily commute has me going up a bend on an overpass that briefly brought traffic light below into view. After a few weeks of it stopping for it and needing overriding, it ignored the erroneous light. Not sure how, but proves to me NNs are working, and this type of edge case has been solved.

still tho, it thinks flashing yellows attached to "intersection ahead" signs are traffic signals.

Edit: I do not have FSD beta, just your run-of-the-mill HW3.0 on '17 S75D

 

[Singuy]

happened to me too. My daily commute has me going up a bend on an overpass that briefly brought traffic light below into view. After a few weeks of it stopping for it and needing overriding, it ignored the erroneous light. Not sure how, but proves to me NNs are working, and this type of edge case has been solved.

still tho, it thinks flashing yellows attached to "intersection ahead" signs are traffic signals.

Flashing yellows are still an issue for FSD beta testers. Hopefully that issue is gone in v9

 

[Featsbeyond50]

Benevolent and educated politicians know this. And they will side with Tesla, knowing that FSD is a great boon for society.

Hey that's a good one. What's the acronym for laughing my ass off, slapping my knee and holding my belly as my laughter rolls into a full howl?

Mark my words, there will be so much resistance to FSD that the only way to get it through regulators will be to have overwhelming evidence that even the most stubborn AND UNEDUCATED will not be able to deny it's worthiness. And, conveniently, there will also have to be competition. There's no way the world will let Tesla have a monopoly on this...

Also, "benevolent politician" is an oxymoron.

 

[LiveLong&Profit]

Hey that's a good one. What's the acronym for laughing my ass off, slapping my knee and holding my belly as my laughter rolls into a full howl?

Mark my words, there will be so much resistance to FSD that the only way to get it through regulators will be to have overwhelming evidence that even the most stubborn AND UNEDUCATED will not be able to deny it's worthiness. And, conveniently, there will also have to be competition. There's no way the world will let Tesla have a monopoly on this...

Also, "benevolent politician" is an oxymoron.

I cannot fault your cynicism re. politicians, there are plenty bad examples so maybe I am naive. I still choose hope.

If the bar is as high as you argue here, that is sad, because it will take much longer to collect the mountain of evidence needed - lots of people will die from traffic accidents in the meantime. Actually, it also reframes the issue: it is not factual evidence as is a court of law that is needed, but something convincing to a large populace.
Hm... I wonder what it would take?
Maybe showing people that you can now have a cheap cab-fare like experience instead of drunk-driving and thus avoid facing fines or criminal charges?
Showing that you can now send a car for your annoying mother-in-law, so you don't have to drive 1 hour back and forth to get her to join a family celebration?
Showing that you can just send your date right home instead of them spending the night?

So, a kind of public information campaingn, disguised as ads?

 

[OrthoSurg]

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.

 

[MC3OZ]

Obviously any ranking like this is inherently controversial, and many on this forum feel Tim Lee has an anti-tesla bias.

But the fact remains that almost every other company is using lidar and HD maps because it believes they are helpful.

This is where I disagree, if Elon and Tesla decided lidar was necessary, they would have invented their own cheap lidar.

We can gauge the progress of the FSD Beta so far, and the executive summary is, it is an improving solution, that will continue to improve.
It is pure speculation to imagine it would hit a problem that required lidar and HD maps, Elon and the team seem confident that will not be required.

All the lidar based solutions ultimately need to solve vision or be very dependent on accurate HD maps, a model which can break as the environment changes.

So what is the cost of building and maintaining accurate HD maps of a complex city like New York, LA, or London? I suggest that is a never ending task, some portion of the maps will always be out-of-date. In other locations, the city might be smaller, and change more slowly, some places you go back there 10-20 years later and nothing has changed, but in major cities, change is the only constant.

Why people assume they know more about the problem than Elon and Karpathy, is beyond me.
My impression is Karpathy seems confident, I trust his knowledge and judgement.