Tesla glory/failure

[Sigmetnow]

Elon Musk
 
This weekend, the @xai team brought our Colossus 100k H100 training cluster online. From start to finish, it was done in 122 days.
 
Colossus is the most powerful AI training system in the world. Moreover, it will double in size to 200k (50k H200s) in a few months.
 
Excellent work by the team, Nvidia and our many partners/suppliers.

9/2/24, 12:53 PM https://x.com/elonmusk/status/1830650370336473253

[nadir]

Powerful! More GPUs! More! More!

Ok. I bet the AI revolution, and in particular the FSD revolution, lies not in more power but in better ideas and knowledge.

[KiwiGriff]

Ok. I bet the AI revolution, and in particular the FSD revolution, lies not in more power but in better ideas and knowledge.

Me thinks you don't grok what the AI revolution is.

[nadir]

Ok. I bet the AI revolution, and in particular the FSD revolution, lies not in more power but in better ideas and knowledge.

Me thinks you don't grok what the AI revolution is.

Probably, but watching the strong disagreements among the experts, seems I am not alone. Anyway, I don’t think Tesla can solve the FSD problem just by increasing their models size. That also creates problems of its own.

[oren]

Despite the high certainty displayed in this thread by some posters for the longest time, I have lingering doubts about Tesla's actual ability to solve the unsupervised automated driving problem (aka Robotaxi) satisfactorily, rather than an unveil, a demo, or a never-ending string of revisions each with lots of pats on the back. And as Musk bet the whole company on this issue, this could yet lead to Tesla failure even after achieving earlier glory. Only time will tell.

[NeilT]

Well its AI5. If you buy HW4 you’re almost outdated as much as HW3 and you’ll see the end of FSD updates very soon, unless you upgrade HW.

See, the only thing Tesla has applied is scaling up the FSD neural network, because apparently there’s a range of improvement just by increasing the number of nodes. That’s the arithmetic Musk used in his head several months ago to go all in on AI and FSD.

Unfortunately other companies ahead on the AI game are observing some sort of saturation in the learning, and additional problems such as, too much training may even make the machine lose (“forget”) elements of the basic initial training that need re-training at a much higher cost.

But Musk didn’t factor in any of this, like he never does because he knows little, he just pushes ahead as a ram and sometimes his customers win, often they lose.

Btw, customers needing faster bigger hardware on current strategy. HW3 is almost history.

False dawns and glass ceilings.

They believe with enough training grunt they can get around all of these issues. Remembering that the more training power they have the more times they can go round the loop, tune it, refine it and then come to a streamlined AI that can drive.

They know it is brute force right now and HW4 has the scope to deliver with that brute forced approach.

They will need to refine afterwards and prune the dross in order to get a good performance on HW3 and HW4.

Time will tell if this approach is valid or not.  Nobody else is even close let alone ahead.

[NeilT]

Despite the high certainty displayed in this thread by some posters for the longest time, I have lingering doubts about Tesla's actual ability to solve the unsupervised automated driving problem (aka Robotaxi) satisfactorily, rather than an unveil, a demo, or a never-ending string of revisions each with lots of pats on the back. And as Musk bet the whole company on this issue, this could yet lead to Tesla failure even after achieving earlier glory. Only time will tell.

Failure is always an option for every company.

The absolute worst case, nastiest solution, to resolve this would be to go with Lidar and mm Lidar mapping.

The Tesla solution would blow away every single other solution on the planet if they did that, hands down, no arguments.

Think about it.  Tesla isn't going for some crippleware.  It is going for the whole solution and if they achieve it they will walk away with everything.

In order to even try to do that you Have to believe you can do it and you have to go for it.  Half measures will never do.

In the end driving is nothing more than a set of circumstances and a set of actions and the recognition of the circumstances and mapping it to actions.

If you think about it that way, the more circumstances that are understood and the full set of actions on meeting those circumstances, the more capable a driver it will be.

Let me give you an example from my own life.  I was driving on a relatively small highland back road in Scotland.  It is usually totally deserted and I was going a long way over the 60mph speed limit.  It was also raining (well it is Scotland you know).

I came over a rise and faced a vehicle coming the other way and also I noted in that split second a large rut running well down the road under my left wheel (UK so kerb side).  That rut was full with about 8 inches of water.

What did I do?  I didn't think about it, I just put my foot down on the accelerator hard.  Just to understand my car did not have traction control or limited spin diff.

Do you know what happened?  Or even why I did that?  I didn't even think about it until it was over it was pure cerebral decision time and my brain stem used my knowledge to get me out of trouble.

These are the kinds of things an autonomuos vehicle has to face every single day and deal with.  Millions or even billions of edge cases in every conceivable manner.

There is only one company trying to do that.  Which is why it is so hard and so slow.  But compute with few limits will be the single most likely avenue to success.  Because the problem is recognition, actions and decision mapping.  These are standard compute problems.  Anyone who doesn't understand that doesn't understand the challenge.  Elon does understand.

[oren]

The absolute worst case, nastiest solution, to resolve this would be to go with Lidar and mm Lidar mapping.

The Tesla solution would blow away every single other solution on the planet if they did that, hands down, no arguments.

Think about it.  Tesla isn't going for some crippleware.  It is going for the whole solution and if they achieve it they will walk away with everything.

I don't need and am not interested in detailed examples.
My position is that there is another option. The worst case is not to go with LIDAR and mapping, it is to go without autonommous driving at all.
 Tesla could go the route of the best/cheapest/most advanced electric car company, and do without FSD altogether. This would have gone very well with the Tesla mission and master plan. Considering the huge advantage they had in EVs (maybe still have, really not sure but it's surely eroding), they could continue trailblazing that path that brought them to glory, and could have controlled a huge share of a fast growing market for the next 20 years. However, Tesla/Musk chose otherwise and decided to bet the whole company on FSD. Hopefully it succeeds, but I am quite uncertain on that.
I am sure I don't need to remind Elon's promises (which he did believe at the time, by my assessment) in 2016 and in 2020 of self driving nearly complete and all the hardware sufficient. Both were gross errors of judgement. The next time the prediction turns out to be wrong (if it does), Tesla is gone and the opportunity to push the EV market hard and fast goes with it.

[Sigmetnow]

Elon Musk
 
This weekend, the @xai team brought our Colossus 100k H100 training cluster online. From start to finish, it was done in 122 days.
 
Colossus is the most powerful AI training system in the world. Moreover, it will double in size to 200k (50k H200s) in a few months.
 
Excellent work by the team, Nvidia and our many partners/suppliers.

9/2/24, 12:53 PM https://x.com/elonmusk/status/1830650370336473253

< xAI’s $4 billion (estimate) supercomputer
9/2/24, https://x.com/sawyermerritt/status/1830651283155656713

EDIT:

Brett Winton
@wintonARK
Probably a 4 year build if relying on power/datacenter vendors, contractors, and consultants
@xai gets it done in 4 months
Given the dramatic performance improvement rate in AI, velocity/urgency is the key determinant of success

Elon Musk
The most optimistic (ie unrealistic) quotes we received were 12 to 18 months

9/2/24, https://x.com/elonmusk/status/1830816418813604163

 
=====
 
Elon Musk
 
Video of the inside of Cortex today, the giant new AI training supercluster being built at Tesla HQ in Austin to solve real-world AI
8/26/24, 4:08 AM
➡️  https://x.com/elonmusk/status/1827981493924155796
20 sec. Row after row of servers. With audio.
 
——-
 
xAI Colossus versus Tesla Cortex computer clusters.
 
The xAI Colossus and Tesla Cortex are two distinct AI supercomputer clusters, each with its own purpose and specifications, primarily driven by Elon Musk's companies, xAI and Tesla respectively.
 
For the AI-response tolerant, Grok summarizes the differences here:  https://x.com/i/grok/share/xaf4xLYq3sl9OWy5LCEOVkaAF

 
=====
 
Teslascope
ACTUALLY SMART SUMMON HAS ARRIVED FOR THE FIRST TIME.
This is not a drill.
   —-
Buckle up for the ride of your life, except, surprise! You're not in the car. ASS (Actually Smart Summon) allows your vehicle to come to you, or head to a spot that you choose, all on its own.
9/3/24, 12:27 AM https://x.com/teslascope/status/1830825004474535995
Details at the link.

 
ΛI DRIVR
ASS (Actually Smart Summon) test #1
 
9/3/24, 3:27 AM
➡️  https://x.com/aidrivr/status/1830870120677417470
3 min. Two examples. Around a hefty curb divider in a parking lot. And banishing it to a point on the far side of the lot.
The car moves fast! — this is not the old Smart Summon.

Actually Smart Summon test #2 (Roundabout)
I'm officially impressed
 
9/3/24, 4:26 AM
➡️ https://x.com/aidrivr/status/1830885165838877178?
2 min. Will it take an illegal short cut, or go all the way around?

The distance involved and the speeds being used are now much closer to human, or impressive robotaxi, level.

[NeilT]

My position is that there is another option. The worst case is not to go with LIDAR and mapping, it is to go without autonommous driving at all.

That's the point.  As far as Tesla is concerned that is the cross they will die on.  It is absolutely and completely not an option.

Absolute worst case they will even entertain is something like Lidar pre mapping and today that is also a cross they will die on.  They can back away from that one, they are completely unable to back away from FSD totally.

I thought Everyone got that?

[Rascal Dog]

These are the kinds of things an autonomuos vehicle has to face every single day and deal with.  Millions or even billions of edge cases in every conceivable manner.

There is only one company trying to do that.  Which is why it is so hard and so slow.  But compute with few limits will be the single most likely avenue to success.  Because the problem is recognition, actions and decision mapping.  These are standard compute problems.  Anyone who doesn't understand that doesn't understand the challenge.  Elon does understand.

Elon does not understand.

If he did, he would not have promised "full self driving" back in 2019 at "Autonomy Day". And a promised robotaxi rollout in 2020.

I'm not sure anyone understands the full depth of the problem.

[NeilT]

I'm not sure anyone understands the full depth of the problem.

I think understand and know are two different things.  I believe they understand the depth of the problem.  What they do not know, or at least cannot calculate, is how much effort it will take to resolve it.

It is calculating the scope of the problem which evades them.  As they solve one set of problems, they find that their solution should resolve another set but does not.

This spawns another complete set of problems to be resolved.

This is where the "knowing" comes in.  You can understand what you need to resolve but knowing the full scale of the specific use cases and edge cases is an entirely different metric.

As they brute force more and more edge cases they will spawn hundreds or even millions of new one's.  As the compute continues to ramp they will find more and more resolutions faster and faster.

Eventually there has to be convergence.  Because if there were not a human would never be able to drive.  It is a finite problem.  We just don't know the full size of the picture yet.

[Sigmetnow]

Elon has said for years that getting approval for FSD involves a “march of 9’s” — how good does the system need to be for regulators to approve it?  99.99% safe?   99.99999% safe? 

The process is not direct, or obvious, or hardware-specific.  It involves solving new problems as they appear.  Tesla tried HD maps and ultra-sonics, but those were not the answer.  Although parts of the system have been using neural nets for a while, the entire stack could not be integrated under AI until there was sufficient data, the means to train it, and the hardware to use it.  That capability wasn’t there a few years ago — like the iPhone, it simply wasn’t possible until all the ingredients came together.  Now, with billions invested in the new supercomputer clusters, it becomes possible.  (And yes, they can make FSD backwards compatible to HW3; it will simply be not quite as good as AI4 or AI5.)

Other companies’ autonomous systems won’t have this level of capability for many more years, if ever — until they spend billions of dollars on their own, or they license FSD for themselves.  Many have already given up, and decided it is better financially for them to simply use a human tele-operator to deal with the edge cases, or severely limit allowable conditions; or geofence with HD maps as their way to reach enough 9’s to be acceptable.

[Rascal Dog]

I'm not sure anyone understands the full depth of the problem.

As they brute force more and more edge cases they will spawn hundreds or even millions of new one's.  As the compute continues to ramp they will find more and more resolutions faster and faster.

Eventually there has to be convergence.  Because if there were not a human would never be able to drive.  It is a finite problem.  We just don't know the full size of the picture yet.

Humans are the result of billions of years of evolution. What if it takes millions of iterations to get to a solution? Sure, it converges... but not in our lifetime.

Yes, it must be a finite problem. But might not be a problem of just throwing enough computes at the problem. Might be just a missing idea or three on how to organize the computes efficiently. Brute force doesn't always find solutions in finite time.

I've been pondering, with the innocence of a little knowledge, on the difference between human and machine learning as a neural network. Training a neural network is usually a huge separate task from doing the actual task. Humans both learn and do the task at the same time. I suspect that getting a neural network to learn while doing is going to be part of the needed tools for a solution.

Humans can sometimes learn much faster than neural networks do. Why? Understanding why again might be a needed tool for a solution.

[nadir]

I think you are onto something fundamental, Rascal Dog. Right now machines are very stupid, but who will have the balls to let machines learn and perform at the same time? Not only it would require a hardware heavier than today’s cars to do that (that can change). It would also escape the control of the programmers and the users.

[nadir]

The Actually Smart Summon is another scam. So, the ultimate responsible of whatever might happen between the parking spot and the user is… the user! Even when the car may be wherever hidden in a huge parking lot, which is the main use case of the Summon.

I really wonder what Musk has done to paralyze US regulators in order to allow this thing along with FSD “supervised”.

[Rascal Dog]

I think you are onto something fundamental, Rascal Dog. Right now machines are very stupid, but who will have the balls to let machines learn and perform at the same time? Not only it would require a hardware heavier than today’s cars to do that (that can change). It would also escape the control of the programmers and the users.

Thanks, but I'm not even close to the first to point out that machine learning for NN is slow, expensive and offline. Fixing any of that was well above my pay grade back before I retired.

The hardware requirements are mostly unknown as better learning methods are unknown. A fixed trained NN might need to be larger than a NN that can learn.

Lastly, many things are outside the real control of the users, but work well anyways. How many user of old cars really understand how a carburetor works, or even could fix a carburetor? Their control is based on if "I push this thing the car goes faster" experience, which isn't based on real understanding.

Outside the control of programmers/designers is perhaps a more serious issue. But is it? The designers want an intelligent car that will find joy in getting a person to work and back, and perhaps gossiping with the other cars at the charging station. If the programmers/designers can control that, isn't that enough?

[oren]

My position is that there is another option. The worst case is not to go with LIDAR and mapping, it is to go without autonommous driving at all.

That's the point.  As far as Tesla is concerned that is the cross they will die on.  It is absolutely and completely not an option.

Absolute worst case they will even entertain is something like Lidar pre mapping and today that is also a cross they will die on.  They can back away from that one, they are completely unable to back away from FSD totally.

I thought Everyone got that?

I guess I should have said there "was" another option, Tesla "could have gone".
I fully agree with you, Tesla intends to get FSD or die trying.
I think it may still turn out to be a huge mistake, as the jury is still out on whether they actually ever achieve true unsupervised autonomy, despite the frequent assurances on this thread. And I believe it is not the path most logically aligned with Tesla's original/previous mission. It's the main reason why I stopped investing in Tesla and why I became much less interested in Tesla.
Be that as it may, the dice is cast already, just need to wait and see the result. My main point was to try and poke what I believe is some people's (and Elon's) overconfidence in Tesla's ability to achieve this. It's not a given and foregone conclusion.

[nadir]

Even if they achieve it, the question is, will users adopt it? Will it be so revolutionary economically? Long term, there is a case for all cars go autonomous, but that’s a 2050+ vision, will require major legislation, with tremendous opposition from milions.

In the meantime, will the FSD business cases be really so profitable? Or will there be competition where it does matter, that is, urban/suburban areas? Waymo and others slowly but surely growing their reach.

What about the robot? Who will pay $20,000 (in the optimistic opinion of Musk) to let a human-size mechanical apparatus enter their home if it’s not with all the assurances that it will not confuse the toddler sleeping in their cradle with a piece of furniture that needs to be removed?

[SteveMDFP]

I'm not sure anyone understands the full depth of the problem.

As they brute force more and more edge cases they will spawn hundreds or even millions of new one's.  As the compute continues to ramp they will find more and more resolutions faster and faster.

Eventually there has to be convergence.  Because if there were not a human would never be able to drive.  It is a finite problem.  We just don't know the full size of the picture yet.

Humans are the result of billions of years of evolution. What if it takes millions of iterations to get to a solution? Sure, it converges... but not in our lifetime.

Yes, humans are the result of billions of years of evolution.  But we didn't evolve to drive cars (and the skills of many of my fellow H.saps shows this!)

Autonomous driving is a challenging task.  But AI is evolving far faster than any biological organism.  My expectation is that driving with high level of skill and safety will probably be soon. 

[nadir]

Wow big regression on latest udstes of FSD 12.5.
Where is the 5x to 10x improvement?

Miles between critical disengagement remain at levels of ~ 2 years ago. And this is the future of Tesla.

https://x.com/realdanodowd/status/1831114541838860590

[NeilT]

Autonomous driving is a challenging task.  But AI is evolving far faster than any biological organism.  My expectation is that driving with high level of skill and safety will probably be soon. 

I read a story about the Google game playing AI they created.  It took them 3 years to build the AI, during which time they gathered 3 years worth of  human game playing.  It took 3 days to train it with the data.

That game playing AI exceeded average human ability after training.

Machines learn at machine speed.  It is the gathering of the relevant training data and correctly labelling it which causes the main problem.

Driving is a skillset many orders of magnitude harder than game playing.  But the resolution, once you have end to end AI, is the same.  But the data is also many orders of magnitude larger and harder to gather as cars with technology to gather it are not in billions of homes.

What most detractors are missing is the broader picture.  Since Tesla went from a combination of code and AI, the argument has no longer been about whether FSD is competent to drive but in critique of how well it is driving and how many mistakes it makes.

It should be obvious that Tesla has moved from IF to when.  But many people simply don't see it.

[NeilT]

Wow big regression on latest udstes of FSD 12.5.
Where is the 5x to 10x improvement?

Miles between critical disengagement remain at levels of ~ 2 years ago. And this is the future of Tesla.

https://x.com/realdanodowd/status/1831114541838860590

"Critial" disengagement is a combination of driver confidence and poor actions.  My wife, for instance, would critically disconnect every 100yds for daring to drive any way other than the way she thinks is perfect.  Absolutely Zero tolerance for any other kind of driving.  She has to suffer me but she won't suffer AI.

These figures are extremely subjective and also subject to 1 month free trial issues where people who don't know the vehicle or capabilities are in control of deciding what the FSD should do or not.

Only Tesla knows which are confidence disengagements and which are error based.

[nadir]

Wow big regression on latest udstes of FSD 12.5.
Where is the 5x to 10x improvement?

Miles between critical disengagement remain at levels of ~ 2 years ago. And this is the future of Tesla.

https://x.com/realdanodowd/status/1831114541838860590

"Critial" disengagement is a combination of driver confidence and poor actions.  My wife, for instance, would critically disconnect every 100yds for daring to drive any way other than the way she thinks is perfect.  Absolutely Zero tolerance for any other kind of driving.  She has to suffer me but she won't suffer AI.

These figures are extremely subjective and also subject to 1 month free trial issues where people who don't know the vehicle or capabilities are in control of deciding what the FSD should do or not.

Only Tesla knows which are confidence disengagements and which are error based.

At the very least that website is comparing apples with apples, and you can’t see the 5x to 10x improvements. Regressions are expected, but a real “wow” moment has not happened at all.

I can’t believe they are allowing the Summon move the car driverless. What are regulators in the US thinking or doing?

[nadir]

Autonomous driving is a challenging task.  But AI is evolving far faster than any biological organism.  My expectation is that driving with high level of skill and safety will probably be soon. 

I read a story about the Google game playing AI they created.  It took them 3 years to build the AI, during which time they gathered 3 years worth of  human game playing.  It took 3 days to train it with the data.

That game playing AI exceeded average human ability after training.

Machines learn at machine speed.  It is the gathering of the relevant training data and correctly labelling it which causes the main problem.

Driving is a skillset many orders of magnitude harder than game playing.  But the resolution, once you have end to end AI, is the same.  But the data is also many orders of magnitude larger and harder to gather as cars with technology to gather it are not in billions of homes.

What most detractors are missing is the broader picture.  Since Tesla went from a combination of code and AI, the argument has no longer been about whether FSD is competent to drive but in critique of how well it is driving and how many mistakes it makes.

It should be obvious that Tesla has moved from IF to when.  But many people simply don't see it.

The “when” can be a problem if it is delayed beyond expectations, even reasonable expectations. It can bankrupt the company. I don’t doubt FSD would eventually work and be approved, but the “when” is essential. Especially for a publicly traded company that, absent of breakthroughs that would provide 10x and 100x profits, is in a super-inflated condition ready to burst.

[gerontocrat]

The “when” can be a problem if it is delayed beyond expectations, even reasonable expectations. It can bankrupt the company. I don’t doubt FSD would eventually work and be approved, but the “when” is essential. Especially for a publicly traded company that, absent of breakthroughs that would provide 10x and 100x profits, is in a super-inflated condition ready to burst.

One order of magnitude, 2 orders of magnitude additional profits required? Why?

Given the current share price, the current market "values" Tesla at USD 670 billion, with a p/e ratio of 61.

To what extent are future profits totally reliant on Full FSD/Robotaxi and Optimus AI etc etc etc?

Cybertruck sales and production are on the ramping up path, and the USA has vast numbers of MAGA Would Be Alpaha males desperately wanting/needing a 100% Macho Truck to shore up their egos battered by competent successful women taking centre stage. I reckon plans for 100,000 annual sales by end 2025 is achievable, but mostly for the USA (and likeminded idiots elsewhere, e.g. Chechen Warlords).

The semi factory is being built. Mass production starting slowly and ramping in in 2025-2026. It's a big market and very international.

The US and Europe are building fortresses against China's Auto industry. How vulnerable is Tesla-China to retaliation by China against foreign automakers?
Will we see a USD 25K EV in mass production from sometime next year? If yes, a big market.
This should also help maintain sales volumes of Tesla's existing range of vehicles.

The mass storage market is growing by leaps and bounds, and Tesla is doing very well in it.

A 2 fold increase in profits by sometime in 2027 looks relatively easy. It would at the current share price bring down the p/e ratio to around 30, sort of OK for an extremely cash-rich auto company with a proven ability to use HighTech in very productive ways.

As much as I would not like / hate to be on Mars with Musk in charge, I do hope Tesla does not push Robotaxi and Optimus so hard that somewhat more traditional business development is neglected. After all transition from ICEVs to BEVs is a significant and eminently achievable part of the energy transition. Methinks the next 12 months will tell us the path Tesla has chosen.

[Rascal Dog]

Machines learn at machine speed.  It is the gathering of the relevant training data and correctly labelling it which causes the main problem.

It should be obvious that Tesla has moved from IF to when.  But many people simply don't see it.

Machine learning is much slower than human learning. A human driver gets about 40 hours of instruction and can pass a driving test.

Can FSD pass a driving test? That's with many thousands of hours of training data. A million miles of training data?

[Sigmetnow]

Machines learn at machine speed.  It is the gathering of the relevant training data and correctly labelling it which causes the main problem.

It should be obvious that Tesla has moved from IF to when.  But many people simply don't see it.

Machine learning is much slower than human learning. A human driver gets about 40 hours of instruction and can pass a driving test.

Can FSD pass a driving test? That's with many thousands of hours of training data. A million miles of training data?

Is every human who passes a driving test a safe driver?  Do humans always follow the rules of the road?  Are they courteous to pedestrians, cyclists and other vehicles?  Do they remain completely alert and focussed on driving, with 360° vision?  Do they not drive when they are not feeling well?

The goal, or rather the requirement, for a general autonomous driving system, is that it must be much better than even the best human driver.

And yes, FSD worked well enough years ago to pass a human driving test.

Please note the Y axis units in the graphs below.  Click to enhance.

[NeilT]

Machine learning is much slower than human learning. A human driver gets about 40 hours of instruction and can pass a driving test.

I've said this before but will say it again.

Humans are not allowed to learn to drive until they have matured for a minimum of 16 years.  Learning the skills used for driving.

Tell me in about 13 years time which was faster.  I say 13 because it's only in the last 3 years that the AI has really begun to evolve.

Also that human has the whole world to gain experience in whereas the driving AI needs specific vehicle related training material which must be produced.
But once the material is there AI can learn in hours what it takes humans decades to learn.  This is not in doubt, it has been proven.

[nadir]

The “when” can be a problem if it is delayed beyond expectations, even reasonable expectations. It can bankrupt the company. I don’t doubt FSD would eventually work and be approved, but the “when” is essential. Especially for a publicly traded company that, absent of breakthroughs that would provide 10x and 100x profits, is in a super-inflated condition ready to burst.

One order of magnitude, 2 orders of magnitude additional profits required? Why?

Given the current share price, the current market "values" Tesla at USD 670 billion, with a p/e ratio of 61.

To what extent are future profits totally reliant on Full FSD/Robotaxi and Optimus AI etc etc etc?

Cybertruck sales and production are on the ramping up path, and the USA has vast numbers of MAGA Would Be Alpaha males desperately wanting/needing a 100% Macho Truck to shore up their egos battered by competent successful women taking centre stage. I reckon plans for 100,000 annual sales by end 2025 is achievable, but mostly for the USA (and likeminded idiots elsewhere, e.g. Chechen Warlords).

The semi factory is being built. Mass production starting slowly and ramping in in 2025-2026. It's a big market and very international.

The US and Europe are building fortresses against China's Auto industry. How vulnerable is Tesla-China to retaliation by China against foreign automakers?
Will we see a USD 25K EV in mass production from sometime next year? If yes, a big market.
This should also help maintain sales volumes of Tesla's existing range of vehicles.

The mass storage market is growing by leaps and bounds, and Tesla is doing very well in it.

A 2 fold increase in profits by sometime in 2027 looks relatively easy. It would at the current share price bring down the p/e ratio to around 30, sort of OK for an extremely cash-rich auto company with a proven ability to use HighTech in very productive ways.

As much as I would not like / hate to be on Mars with Musk in charge, I do hope Tesla does not push Robotaxi and Optimus so hard that somewhat more traditional business development is neglected. After all transition from ICEVs to BEVs is a significant and eminently achievable part of the energy transition. Methinks the next 12 months will tell us the path Tesla has chosen.

Yeah the 100x might be a bit exaggerated, but not the 10x. My point is that there have to be substantial sources of revenue while they spend billions in R&D for the highly speculative Tesla AI revolution, which may take many years more than implied by Tesla.

You think they can do that with Cybertruck, Semi, model refreshes, 25K vehicle? Well, they better start to focus (or refocus) and apply themselves to it. In any case there are some question marks about the Semi practicality/price ratio, and the Mexico factory for the 25K may not be completed especially if Trump wins.

[Sigmetnow]

My point is that there have to be substantial sources of revenue while they spend billions in R&D for the highly speculative Tesla AI revolution, which may take many years more than implied by Tesla.

Tesla’s $30 Billion cash on hand should help nicely.

[kassy]

But is it Tesla AI since it is another company that runs it or is that owned by Tesla?

[nadir]

It’s part of Tesla.

Btw they released an interesting roadmap on Twitter…

Glad that it’s not Elon Musk giving away his guesstimates but the company itself comes up with something for the public.

https://x.com/tesla_ai/status/1831565197108023493

They mention some expected improvements on intervention metrics for FSD, not sure if they have ever released or will release those, I guess internal, metrics.

[Sigmetnow]

Due to popular demand, Tesla AI team release roadmap:
 
September 2024
- v12.5.2 with ~3x improved miles between necessary interventions
- v12.5.2 on AI3 computer (unified models for AI3 and AI4)
- Actually Smart Summon
- Cybertruck Autopark 📐
- Eye-tracking with sunglasses 🕶️
- End-to-End network on highway 🛣️
- Cybertruck FSD 📐
 
October 2024
- Unpark, Park and Reverse in FSD
- v13 with ~6x improved miles between necessary interventions
 
Q1 2025
- FSD in Europe (pending regulatory approval) 🌍
- FSD in China (pending regulatory approval) 🇨🇳
 
9/5/24, https://x.com/tesla_ai/status/1831565197108023493
 
< Tesla FSD could launch in the UK and Australia in the first half of 2025.
 
Elon Musk
Hopefully, RHD markets in late Q1, early Q2, pending regulatory approval
9/5/24, https://x.com/elonmusk/status/1831601961965080933

Note: Tesla has made the FSD versions designed for AI4 run on AI3 (HW3).  It’s the same version, not a crippled one.

Note: FSD is not coded specifically for the US.  A Swiss regulator had an intervention-free drive in Germany not long ago.  This will make it simpler to expand to other countries.

Fun fact:  if you consider all the parking lot surface that can be used with Actually Smart Summon, Tesla can now operate without a person in the driver’s seat in a larger area than Waymo can.

 
===
 
Musk causing major Tesla Brand Damage? Not So Says Major Automotive Intelligence Firm
 
Tesla is the leader in brand loyalty, according to a new report from S&P Global Mobility.
Sep 2, 2024
“Tesla continues its run as the leader in brand loyalty with a rate of 67.8% for the first half of 2024. Customer loyalty has remained relatively constant."

“We can only report on what we see in the data. In this instance, there is some decline in Tesla’s loyalty for the first half of 2024 vs. 2023; however, it is below 1 percentage point,” he said, adding that Tesla is clearly beating other brands.
 
“The brand still remains the industry leader in brand loyalty by a healthy margin. For comparison’s sake, the industry brand loyalty average stands at 52.5% for H1 2024 and no other brand has a loyalty rate above 60%,” he said. …

https://www.forbes.com/sites/brookecrothers/2024/09/01/tesla-brand-damage-not-so-says-major-automotive-intelligence-firm/

 
—-
Vehicle internal crash safety
Tesla among car brands least likely to be involved in fatal accidents: study
By Simon Alvarez August 31, 2024

A study from vehicle history report firm EpicVIN has revealed a ranking of car brands with the lowest percentage of fatal injuries in reported accidents. Among the carmakers outlined in the study, electric vehicle maker Tesla was dubbed as one of the ten least likely car brands to be involved in fatal accidents.
 
The firm noted that its analysis used the National Highway Traffic Safety Administration’s (NHTSA) Fatality and Injury Reporting System Tool to gather data on occupants involved in fatal vehicle crashes between 2017 and 2022. Injury types were then analyzed to show the percentage of those involved in fatal accidents that suffered fatal injuries. …

https://www.teslarati.com/tesla-least-likely-involved-fatal-accidents-study/


—-
 
The ‘glory days’ for global automakers in China are over
Sep 4, 2024
“Tesla’s manufacturing of the Model 3 in Shanghai transformed consumers’ perspective of electric cars.” They became “the new cool.”
https://finance.yahoo.com/news/glory-days-global-automakers-china-220020627.html

 
—-
NEWS: After a trial run, the Tesla app’s Charge on Solar feature is now available to all Australian owners of a Powerwall and compatible Tesla vehicle.
  The feature enables owners to charge only when surplus electricity is being produced by their solar power system, without dipping into mains grid supply (or the Powerwall).
 
Source: https://www.solarquotes.com.au/blog/tesla-charge-solar-mb2990/
 
8/20/24, https://x.com/sawyermerritt/status/1826058548507222234
 
 
—-
The Powerwall 3 is flood resistant in over 2 ft (70cm) of water.
Here’s a clue how:
Tesla Powerwall 3 internals. ⬇️
8/20/24, https://x.com/niccruzpatane/status/1825950993957212308

[Sigmetnow]

But is it Tesla AI since it is another company that runs it or is that owned by Tesla?

“Cortex” is Tesla’s AI super computer, located in the Austin Gigafactory. This cluster is part of an expansion to support Tesla's ongoing AI and autonomous driving ambitions.
 
It is owned and operated by Tesla.

xAI’s “Colossus” supercomputer is in Memphis, Tennessee.

Which is not to say that no Tesla or xAI employees have worked on both.

[Rascal Dog]

...because it's only in the last 3 years that the AI has really begun to evolve.

I would say more like the past 30 years. DARPA grand challenge, for example? Compare 2004 and 2005 results.

AI is overhyped. That doesn't mean it hasn't advanced, or that it might make major advances in the future...

Elon is just doing the Silicon Valley game of "fake it until you make it".

[NeilT]

I have mentioned this before in a different way, much more detailed, but you need to actually read what they say.

3x improved miles between necessary interventions

Not all interventions are necessary and Tesla knows the difference.  They have all the intervention data, the video clips, the decision process of the intervention and the final outcome.  They can then map the final outcome to the predicted path of the AI and then they know which interventions were down to people who have a different driving style and no trust and when the AI goes totally off piste and heads for an accident.

If this company direction is not met what are you going to say then?  Elon told them what to say?

[NeilT]

...because it's only in the last 3 years that the AI has really begun to evolve.

I would say more like the past 30 years. DARPA grand challenge, for example? Compare 2004 and 2005 results.

AI is overhyped. That doesn't mean it hasn't advanced, or that it might make major advances in the future...

Elon is just doing the Silicon Valley game of "fake it until you make it".

I'm talking specifically about autonomous driving AI.  Which is orders of magnitude harder than anything which has been tried with AI before.

[Rascal Dog]

...because it's only in the last 3 years that the AI has really begun to evolve.

I would say more like the past 30 years. DARPA grand challenge, for example? Compare 2004 and 2005 results.

AI is overhyped. That doesn't mean it hasn't advanced, or that it might make major advances in the future...

Elon is just doing the Silicon Valley game of "fake it until you make it".

I'm talking specifically about autonomous driving AI.  Which is orders of magnitude harder than anything which has been tried with AI before.

https://en.wikipedia.org/wiki/DARPA_Grand_Challenge

Autonomous driving AI.

[Sigmetnow]

My point is that there have to be substantial sources of revenue while they spend billions in R&D for the highly speculative Tesla AI revolution, which may take many years more than implied by Tesla.

Tesla’s $30 Billion cash on hand should help nicely.

From the Q2 2024 slide deck:

Cash
 
We have sufficient liquidity to fund our product roadmap, long-term capacity expansion plans and other expenses. Furthermore, we will manage the business such that we maintain a strong balance sheet during this uncertain period.

[NeilT]

...because it's only in the last 3 years that the AI has really begun to evolve.

I would say more like the past 30 years. DARPA grand challenge, for example? Compare 2004 and 2005 results.

AI is overhyped. That doesn't mean it hasn't advanced, or that it might make major advances in the future...

Elon is just doing the Silicon Valley game of "fake it until you make it".

I'm talking specifically about autonomous driving AI.  Which is orders of magnitude harder than anything which has been tried with AI before.

https://en.wikipedia.org/wiki/DARPA_Grand_Challenge

Autonomous driving AI.

Not quite.  From the wiki.

The 2012 DARPA Robotics Challenge, focused on autonomous emergency-maintenance robots, and new Challenges are still being conceived.

[Rascal Dog]

...because it's only in the last 3 years that the AI has really begun to evolve.

I would say more like the past 30 years. DARPA grand challenge, for example? Compare 2004 and 2005 results.

AI is overhyped. That doesn't mean it hasn't advanced, or that it might make major advances in the future...

Elon is just doing the Silicon Valley game of "fake it until you make it".

I'm talking specifically about autonomous driving AI.  Which is orders of magnitude harder than anything which has been tried with AI before.

https://en.wikipedia.org/wiki/DARPA_Grand_Challenge

Autonomous driving AI.

Not quite.  From the wiki.

The initial DARPA Grand Challenge in 2004 was created to spur the development of technologies needed to create the first fully autonomous ground vehicles capable of completing a substantial off-road course within a limited time. The third event, the DARPA Urban Challenge in 2007, extended the initial Challenge to autonomous operation in a mock urban environment. The 2012 DARPA Robotics Challenge, focused on autonomous emergency-maintenance robots, and new Challenges are still being conceived. The DARPA Subterranean Challenge was tasked with building robotic teams to autonomously map, navigate, and search subterranean environments. Such teams could be useful in exploring hazardous areas and in search and rescue.

Fixed that for you.

[NeilT]

It is absolutely Zero relation to FSD.  We are talking simple navigation by robots, like bipedal.  Not AI driving cars among all the humans.  Totally different challenge and what is stated here does not really need AI in the way FSD does.

[Rascal Dog]

It is absolutely Zero relation to FSD.  We are talking simple navigation by robots, like bipedal.  Not AI driving cars among all the humans.  Totally different challenge and what is stated here does not really need AI in the way FSD does.

Before AI drives cars among humans, getting AI to drive cars down an empty road sounds like a good first step. Don't you agree?

[NeilT]

It is absolutely Zero relation to FSD.  We are talking simple navigation by robots, like bipedal.  Not AI driving cars among all the humans.  Totally different challenge and what is stated here does not really need AI in the way FSD does.

Before AI drives cars among humans, getting AI to drive cars down an empty road sounds like a good first step. Don't you agree?

Already completed. Long ago. But autonomy is about interacting with existing traffic to drive a vehicle. Not a
 hypothetical  no vehicle case.  So baby steps bear no relation to w
The challenges of FSD. You can do that without AI.

[Sigmetnow]

“The Wall Street Journal reported on Saturday that Tesla would license xAI's artificial-intelligence models to help power its driver-assistance software, full self-driving technology and share some of that revenue with the startup, according to the proposed arrangement as described to investors.”

Elon Musk
 
Haven’t read the article, but the above is not accurate.
 
Tesla has learned a lot from discussions with engineers at xAI that have helped accelerate achieving unsupervised FSD, but there is no need to license anything from xAI.
 
The xAI models are gigantic, containing, in compressed form, most of human knowledge, and couldn’t possibly run on the Tesla vehicle inference computer, nor would we want them to.
 
The Tesla AI models have incredibly “dense” (in a good way lol) intelligence, as they compress video of reality into driving commands, but must operate on a ~300W computer with memory size and bandwidth far lower than say an H100 GPU.
 
Tesla real-world AI also has a vastly larger context size than an LLM, as the combined video history from all cameras is several gigabytes in size.

9/7/24, 11:13 PM https://x.com/elonmusk/status/1832618127189774374
 Emphasis mine.

ALEX: Thanks for the clarifications. This is the 5th time WSJ has published inaccurate info about xAI

[nadir]

The dangers of having a “smart” car

[Sigmetnow]

BREAKING: CAMOUFLAGED TESLA ROBOTAXI SPOTTED TESTING IN LA 👀
 
9/13/24, https://x.com/thesonofwalkley/status/1834453975422161128
 
9/13/24, https://x.com/theevuniverse/status/1834514369968828899

< Did you notice the Robotaxi front and rear wheels are in different size?

   —-
Speculation is it might have a lower front end, a large windshield and significant storage space in the back.
 
   —-
 
The Simpsons did it again …
⬇️ 9/13/24, https://x.com/sp_myp/status/1834506935577595927

[Sigmetnow]

—- CYBERTRUCK -—
 
 NEWS: S&P Global Mobility says Tesla delivered 5,175 Cybertrucks in July (up 61% MoM), while all other EV pickups combined sold just 5,546 units during the same period.
 
< That's run rate of 62100 per year or nearly 50 percent of the intended line capacity of 125k. Impressive!

Jesse D. Jenkins
Welp. I could not have been more wrong. Not only has the Lightning or R1T not outsold the Cybertruck, as I predicted last year, but in July, sales of the Cybertruck Foundation Series (priced at $100k!) nearly matched ALL other EV trucks combined. 🤯 🤦‍♂️
9/15/24, https://x.com/jessejenkins/status/1835301269340062177

 
The Tesla Cybertruck, which starts at $100,000, is currently outselling the combined sales of all these EV pickup trucks:
 
• Ford F-150 Lightning
• Rivian R1T
• Chevy Silverado EV
• GMC Sierra EV
• Hummer EV
 
9/16/24, https://x.com/sawyermerritt/status/1835800585532002767

EV registrations surged 18% in July, boosted by Tesla Cybertruck and new models
https://www.cbtnews.com/ev-registrations-surged-18-in-july-boosted-by-tesla-cybertruck-and-new-models/

Pretty insane how many Foundation series Cybertrucks Tesla has sold.
Quite possibly may become the largest volume EV truck on the road this year. Period. Of all of them. Ever.
Last year, Ford sold nearly 24k F150 Lightnings while Rivian sold 19,410 R1T trucks.
Tesla will likely sell more than both of them combined, in their first year, all while being sold at a $20k premium.

9/13/24, https://x.com/greggertruck/status/1834451070946607141
 
 < Pretty remarkable that they are still selling Foundation Series 9 months and 30,000+ units into this.
<< Weird how other e-trucks can’t seem to move at any reasonable pace.

——
Sawyer Merritt
A positive Tesla headline from Jalopnik?!
9/16/24, https://x.com/sawyermerritt/status/1835810896964145287
< Whoa… What’s the catch?
SM: Surprisingly there isn't one. It's a pretty fair article
 
Tesla Cybertruck Is Kicking Every EV Truck’s Ass In Sales
https://jalopnik.com/tesla-cybertruck-is-kicking-every-ev-truck-s-ass-in-sal-1851649308

——
 
< instead of comparing against other EV truck, let's compare against ICE pick up trucks
<< Fine. It’s outsold the Nissan Titan and will outsell the Gladiator too.
Next year it will likely exceed the Toyota Tundra.
9/13/24, https://x.com/greggertruck/status/1834581690574078278

https://www.goodcarbadcar.net/2024-u-s-pickup-truck-sales-figures-by-model-with-rankings/

 
—-—-
About $85 million worth of Cybertrucks spotted at Tesla's Giga Texas factory on Sept 12.
https://x.com/sawyermerritt/status/1834345916054048978
 
Video via Brad Sloan: https://www.youtube.com/watch?v=wfSVpaOtXW0&feature=youtu.be
 

Joe Tegtmeyer
Lots and lots of @cybertruck at the Giga Texas outbound lot this morning, with a large group being moved into covered carriers. Production seems very robust too as I saw many more exiting the factory along with a big group of Model Ys. …

9/11/24, https://x.com/joetegtmeyer/status/1833876336475054096
⬇️ pic below; more at the link.
 
< I wonder if CyberTruck production in Austin is > 1400 per week now. That was the volume that Elon mentioned at the Tesla shareholder meeting back in July.
JT: Closer to 2K per week

——
 
James May finally drives the Tesla Cybertruck
https://www.youtube.com/watch?v=CQzYhMDNLPA&feature=youtu.be
14 min. Sept 11, 2024
He likes it.  But wishes such styling were available in a smaller vehicle.
 
Render of a Model-Y-size Cybertruck/car
➡️ https://x.com/cybermikeog/status/1833917924433940569

—-
Cybertruck has been approved for rides in The Boring Company’s LVCC (Vegas) Loop.
People waited out in the 108°F heat to ride in it.
 
➡️ https://x.com/serobinsonjr/status/1834670042702921838
3 pics. 

—-
Kids LOSE it for Cybertruck! 😂
9/12/24, ➡️  https://x.com/greggertruck/status/1834293371357241809

[John Batteen]

In absolute terms, none of these electric truck sales are large numbers, so I don't think the Cybertruck can take a victory lap yet.  It will be interesting to see how this continues to play out with the Cybertruck's well-documented quality issues.