Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

[Usain]

FWIW I assume he was telling the truth when he said that about HW2. Then again when he said it about HW2.5. Then again when he said it about HW3.

But until they actually achieve it, nobody including Elon actually knows how much in car compute is "enough" to do the job.

True. It's just that Elon knows more than we do about the specifics so his opinion is a little better than ours.

Over time, Elon will be less wrong than those who are less informed.

 

[Linkmeup]

Look. as an investor I am always trying to deepen my understanding, I'm not here to argue (there's Twitter for that..). So since, if I understand you correctly, you seem to object to me not placing a great value on Dojo, would you care to elaborate on why you feel it is worth a lot? Is it the point made by others that it decreases the dependency on NVIDIA, which I understand but still don't see as a considerable value driver, or is it something else?

I'm not an expert on LLM's, but I did have an interest in going into supercomputers rather than Internet engineering back in undergrad.

Each computing device has a single processor to do jobs. These supercomputers have many processors running in parallel which compound to allow for really complex calculations (which lead to complex simulations). The simulations, which have barely been around, are accurate for various tasks like asking questions, navigation, etc. when fed great data.

Tesla has the best data store for 3d world maps (outside Google), using their cameras from the vehicle fleet, running in real-time...which can lead to all sorts of useful functions if tied to supercomputers that allow for people to run requests for information/summaries from that data outside of just navigation that FSD (a feature) provides. Hope that provides some context for TAM.

 

[SOULPEDL]

Not sure if serious.

Lower huh?

 

[jerry33]

Folks are completely capable of filling up pages of nonsense without my input.
I simply asked if TSLA was a good buy, to test the crowd's fear level or appetite.
Thanks for your response. Of course HODL.

Isn't TSLA similar to most stocks, either a good buy or good-bye. HODL is of course correct.

 

[Tony73]

Sounds like made up numbers when Elon said compute is no longer a constraint, but you are here saying they need 90% more chips.

Training technique and software optimization trumps number of chips. Give Tesla 300k H100 in 2016 and they wouldn't know what to do with it as lots of the end to end training techniques were not even developed until 2020.

Compute may not be a constraint for the amount of data they currently have but data still needs to increase by an order of magnitude for full autonomy. Current compute gets them to reliable FSD which increases take rate which brings jn more data.

So yes to get to the point where disengagements are measured in the order of thousands of hours, they need a lot more compute.

 

[M|S|M MYP]

Lower huh?

Lower, higher, sideways...take your pick. I will go with higher if given enough runway.

 

[ItsNotAboutTheMoney]

True. It's just that Elon knows more than we do about the specifics so his opinion is a little better than ours.

Over time, Elon will be less wrong than those who are less informed.

Although you have to take into account that he's a saleman for his companies, which means that what he says can simply not be true.

 

[2daMoon]

Compute may not be a constraint for the amount of data they currently have but data still needs to increase by an order of magnitude for full autonomy. Current compute gets them to reliable FSD which increases take rate which brings jn more data.

So yes to get to the point where disengagements are measured in the order of thousands of hours, they need a lot more compute.

Which is what led to Elon's instruction to expose Tesla customers to FSD at every opportunity when picking up their car from a service center or showroom.

Though, another approach to demonstrate it could be to offer some incentive for Tesla owners to try out FSD, either with a store rep or on their own.

Tesla could have a drawing for a new vehicle of the winner's choice that coincides with the free FSD trial cutoff date. The only way to enter for the drawing is to be verified for having taken an FSD demo ride. (only new FSD users qualify)

The drawing could include multiple winners, as many as Tesla feels like giving away to promote more FSD sales/subscriptions from the demos in order to add to the compute.

 

[SOULPEDL]

Isn't TSLA similar to most stocks, either a good buy or good-bye. HODL is of course correct.

Maybe you're right. I don't want to antagonize, but I also might lose interest.

100% HODL is freaking boring and it's precisely why dead people do best, on average, so they say. I'm not dead yet.

Thumbs up if I should tone down the day-trading stuff, betting. However, do not confuse this with wanting to blame others for an unfair stock price on days like these. I really thought I was adding to the fun.

 

[Unpilot]

Although you have to take into account that he's a saleman for his companies, which means that what he says can simply not be true.

More explanation needed.

 

[2daMoon]

Maybe you're right. I don't want to antagonize, but I also might lose interest.

100% HODL is freaking boring and it's precisely why dead people do best, on average, so they say. I'm not dead yet.

Thumbs up if I should tone down the day-trading stuff, betting. However, do not confuse this with wanting to blame others for an unfair stock price on days like these. I really thought I was adding to the fun.

Best to keep in mind how your enthusiasm might lead someone to try doing the same without understanding the strategy, and the potential consequences.

I guess this is why "The Wheel" thread was created, so people are at least offered an opportunity to read all the information to first form a strategy. That opportunity for learning rarely happens when individuals post in this thread about short-term trading.

 

[mongo]

Compute may not be a constraint for the amount of data they currently have but data still needs to increase by an order of magnitude for full autonomy. Current compute gets them to reliable FSD which increases take rate which brings jn more data.

So yes to get to the point where disengagements are measured in the order of thousands of hours, they need a lot more compute.

As the march of 9s progresses, the amount of driving needed to collect the samples increases, but training size only depends on the number of situations, not their rarity. 10x the raw data is filtered down to some number of units of training data.
Needing an order of magnitude compute increase implies there is an order of magnitude increase in training cases (or NN size, or blend thereof). This could be, however that may indicate an overfit/ overtrained NN versus a more generalized one that can handle both sampled and abstracted driving conditions.

 

[scaesare]

I havent seen this posted?

Ford (NYSE:F) Cuts Jobs at F-150 Lightning Plant - TipRanks.com

Legacy automaker Ford (NYSE:F) announced that it’s cutting jobs at the Dearborn, Michigan factory responsible for the F-150 Lightning electric truck due to a dip in...

www.tipranks.com

F150 lightning is in real trouble.

I'm sure the mainstream media would rather do cartwheels than suggest that the cybertruck is now taking EV truck buyers attention. I reckon we will have to see 4x the CTs sold per quarter than the lightning before any mainstream media grudgingly accepts it. Will likely happen this year.

I wonder if folks waiting for a native NACS port, rather than having to futz with an adapter, is playing in to this at all.

I'm rather disappointed, as the F-150 Light(e)ning felt like it light be a vehicle to helpe segu the average "truck guy" in to EV's. That and Farley seemed to be willing to put (some) company politics aside and cooperate with Tesla on things....

I had hoped it would be a runaway success that would force Ford to pursue EV's more significantly....

 

[Motor Mouth]

As the march of 9s progresses, the amount of driving needed to collect the samples increases, but training size only depends on the number of situations, not their rarity. 10x the raw data is filtered down to some number of units of training data.
Needing an order of magnitude compute increase implies there is an order of magnitude increase in training cases (or NN size, or blend thereof). This could be, however that may indicate an overfit/ overtrained NN versus a more generalized one that can handle both sampled and abstracted driving conditions.

I had a growing fear it was happening. I am officially too dumb for this MB.

 

[growler23]

I wonder if folks waiting for a native NACS port, rather than having to futz with an adapter, is playing in to this at all.

I'm rather disappointed, as the F-150 Light(e)ning felt like it light be a vehicle to helpe segue the average "truck guy" in to EV's. That and Farley seemed to be willing to put (some) company politics aside and cooperate with Tesla on things....

I had hoped it would be a runaway success that would force Ford to pursue EV's more significantly....

I think the F-150 Lightning, specification-wise, looks (or at least looked) like a real contender. But when you see that Ford is subsidizing it to the point of $40k (reported as the overall average loss on each of their EV's - IMO they are likely subsidizing the Lightning even more than this), it must be considered a loser.
Why the sales are dropping, given the massive ICE subsidy Ford is putting into the purchase price - I do wonder this as well.
Could the dealerships be getting tiny profits for moving these, since they sell overall at a loss, so dealers have little incentive to learn and sell them?
Did the Cybertruck "Osborne" the Lightning (I doubt this given the reported lack of knowledge of the CT out there in general public land)?
Did the lack of decent charging options until VERY recently put off folks? Combination of multiple of these? I am curious if anyone has insights on this as well. I expect GM is a lost cause, but I really want Ford to pull through.

 

[2daMoon]

As the march of 9s progresses, the amount of driving needed to collect the samples increases, but training size only depends on the number of situations, not their rarity. 10x the raw data is filtered down to some number of units of training data.
Needing an order of magnitude compute increase implies there is an order of magnitude increase in training cases (or NN size, or blend thereof). This could be, however that may indicate an overfit/ overtrained NN versus a more generalized one that can handle both sampled and abstracted driving conditions.

Could a Filter NN be trained to extract good examples for review before they are added to the data used for training?
Would that help manage an increasing data well being drawn from?

 

[DarkandStormy]

I would be curious what the per day delivery rate of the F150 Lightning compared to the cybertruck. Are they getting close?

I believe Ford exited 2023 with a ~1,200/week run rate of the Lightnings. They've made a bunch of different scaling back announcements, but I think they're anticipating ~1,600/week this year? Could be less as I can't keep track of what they're scaling back and when.

 

[mongo]

Could a Filter NN be trained to extract good examples for review before they are added to the data used for training?
Would that help manage an increasing data well being drawn from?

That's what has run in the vehicles in the past to trigger collecting samples. Not sure if E2E breaks that functionality.
The autolabler could potentially further classify scenes (whether from triggers or disengagements) and flag things it doesn't understand for human interpretation.

 

[Singuy]

As the march of 9s progresses, the amount of driving needed to collect the samples increases, but training size only depends on the number of situations, not their rarity. 10x the raw data is filtered down to some number of units of training data.
Needing an order of magnitude compute increase implies there is an order of magnitude increase in training cases (or NN size, or blend thereof). This could be, however that may indicate an overfit/ overtrained NN versus a more generalized one that can handle both sampled and abstracted driving conditions.

Yup, Tesla throws away 99.999% of all data collected and only focus on the unique situations that requires training. I don't see why they need 10x more compute to fix the 0.0001% of the problem. I argue they needed more compute when they had to fix over 50% of the problems when V1 started as that would require lots of training passes to reduce error rate.

Right now Tesla's limited step is not compute but combing the gigantic dataset, where most of it are useless, to find the right data to feed the GPU cluster. I believe they spent a lot of time figuring out techniques trying to appropriately filter the useless data away as that's more important right now than H100 clusters.

 

[mongo]

Yup, Tesla throws away 99.999% of all data collected and only focus on the unique situations that requires training. I don't see why they need 10x more compute to fix the 0.0001% of the problem. I argue they needed more compute when they had to fix over 50% of the problems when V1 started as that would require lots of training passes to reduce error rate.

Right now Tesla's limited step is not compute but combing the gigantic dataset, where most of it are useless, to find the right data to feed the GPU cluster.

Speaking purely theoretically:
If each 9 has 10x the situations which each occur at 1/100 the frequency (1/10 overall occurance), then it would need 10x the compute at each step.
However, it seems unlikely that so many unique cases would be needed.
1. Don't hit things
2. Don't get hit by things
3. Don't break laws (in general) and get to destination