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

[Remus]

You don’t know the solution, but you do know which direction it’s in(with a small amount of noise from sampling bias). SGD works by finding that direction, moving a small amount in that direction, and then recalculating that direction. “Stochastic Gradient Descent” is literally exactly what it says.

In a multi dimensional curve SDG is essentially a greedy algorithm when all you can "see" is your nose tip. I argue the way FSD pick edge cases to work on likely follows the same principal, randomly identify many scenarios to gather data at the same time. The scenario that gathers the most number of cases in a fixed time frame is the "gradient" points to the next direction you should go.

 

[Alphacrux]

This post shows the first example of the diagnostics reporting an issue before it causes a problem and has the part automatically shipped to your service center that I have seen.

Hopefully they can do more predictive failure detection like this and that it actually streamlines service.

“Dave, I am predicting a failure in the AE-35 unit”

 

[jhm]

Prepare For Battery Metal Shortage: Tesla | OilPrice.com
This is pretty cool in a humble brag sort of way. Tesla is expressing concern about the longterm supply of copper. EVs use about twice as much copper as ICE vehicles. So this is quite bullish. Who has been worried that EVs would scale so fast as to encounter the limits of the massive copper supply?

 

[Carl Raymond]

I do think at some point they could simply put every other unaccounted scenario as "other", and activate "survival mode" after anomaly detection where the generic neural network recognize it's a situation that it cannot handle (possible detection by huge spikes of conflicting decisions made in a short period of time). In survival mode, the only rules are:

1) Avoid physical collision with all other objects.

2) If collision is not avoidable, minimize relative velocity of impact

The behavior in survival mode could be trained in a simulation - just keep generating random objects (cars, rocks, humans) and throw it at the car with optimization goal being minimizing collision score.

Pretty much what humans do, very badly and with slow reaction times, in extreme circumstances. I hope the AI will be forgiven for occasionally failing in impossible situations. Doubt it however - there will always be somebody willing to claim “I could have avoided/foreseen that”, with perfect hindsight.

 

[Singer3000]

Prepare For Battery Metal Shortage: Tesla | OilPrice.com
This is pretty cool in a humble brag sort of way. Tesla is expressing concern about the longterm supply of copper. EVs use about twice as much copper as ICE vehicles. So this is quite bullish. Who has been worried that EVs would scale so fast as to encounter the limits of the massive copper supply?

I hope this is a misquote. Because there’s a problem if a senior Tesla executive thinks EVs could cause a problem with the supply of copper and the problem is not with the supply of copper.

 

[Prunesquallor]

“Dave, I am predicting a failure in the AE-35 unit”

Ummm. That ended badly.

 

[tinm]

This post shows the first example of the diagnostics reporting an issue before it causes a problem and has the part automatically shipped to your service center that I have seen.

Hopefully they can do more predictive failure detection like this and that it actually streamlines service.

Ha, I’m old enough to remember when HAL detected a malfunction on the AE-35 unit in Discovery’s antenna then we all know what happened to Dr. Frank Poole and Dr. David “Open the pod bay door, HAL” Bowman. Ha!

Oh yeah, I’ve seen this movie before. Ohhhhh yeah. Don’t like it onnnne bit, I tell ya. History gonna repeat itself, yesiree.


(EDIT: just noticed someone else caught the same drift before me. Chalk it up to great minds thinking alike.)

 

[EVNow]

I do think at some point they could simply put every other unaccounted scenario as "other"

The problem is the edge cases would usually get classified as a known scenario incorrectly, rather than as unknown. If it is unknown it would be so much easier to just disengage instead of crashing.

 

[izemize]

Most people here have roughly the same outlook, incentive, and eschatology -- Tesla will be an astounding megacap 20x return. I find it interesting that most of the conversations here miss the mark on what would be most useful because they tend to be either 'who can out-optimist who' or 'who can detect infiltrating shorts better than who' (kruggerand is getting hungry)

The real optimal strategy should be to outline the moderately pessimistic outcome and demonstrate why even that is sufficient for a bull case. It's an inequality. We can handle >= 'a fairly bad case'. That's the path to confidence. I'll avoid bring Taleb into the discussion but you get the gist. Otherwise you get the reek of insecurity that most people coming here probably experience. i.e. a band of cultists underperforming a market for 5 years yet still following their lord of light.

Here is a premise. Tesla will probably post negative GAAP earnings every quarter for the rest of the year. Why is that ok?

Truth is hardware is the easy part

I was thinking about the fact that Apple launched Siri 7 years ago. State of the art AI. Elon promised 1 million robotaxis on our streets next year. Please forgive me to question his timeline. Sorry for being such a non-believer, you can now return to talk about NN, FSD, wiggle worms and the zip zorps.

 

[neroden]

I don’t read much that wasn’t already known. And if there was some new information, it wasn’t very material.

Panasonic STILL not up to full production capacity is news. In fact, every week they're not up to full production capacity, it's news that they're still not up to it.

It's extremely material news too.

 

[PANN]

My take on the lease price adjustment is Tesla want to maximize both production and leases, in order to build up the Robotaxi Fleet as fast as they can in next 3 years in semi-stealth mode, with lease customer pay for the upfront cost. If one don't see Elon as a fraud with his FSD/Robotaxi plan, one will understand this is the logical and smart move.

 

[neroden]

I think we are saying Tesla can solve x number of scenarios per year, on flat budget. May be more than x (say 2x) - but not 10x.

So, what does that mean in terms of march of 9s ? It depends on the distribution of edge cases by their probability of occuring. In one case this leads to exponentially better FSD, in others not.

Assuming Tesla can solve 100 scenarios in a year, in this ideal case - it takes 100 scenarios to go from 90% quality to 99%, as each of them have 0.09% of occurring. Next year, again Tesla solves 100 scenarios which have 0.009% probability of (i.e. 1/10 the 1st year scenarios). So FSD goes from 99% to 99.9%.

View attachment 403804

But if the edge cases are such that the second year ones have 0.005% probability of occurring, FSD only goes from 99% to 99.5%. So, as I said whether solving a certain number of edge cases a year results in exponentially better quality or not depends on the probability distribution.

View attachment 403822

BUT, since we are talking about the long tail and the total of all probabilities need to be 100, it is somewhat reasonable to assume that the probability goes down exponentially i.e. asymptotically approaches zero. So, Tesla FSD can get exponentially better if
- Tesla figures out a way to solve the most probable edge cases first i.e. prioritization is very important
- Tesla doesn't start running out of enough training data / NN nodes

Thanks for this clear analysis. It's pretty clear to me that as you go out the long tail of the probability distribution you have more individual corner cases each of which is less likely, so progress slows down as you get further out.

 

[EVNow]

In a multi dimensional curve SDG is essentially a greedy algorithm when all you can "see" is your nose tip. I argue the way FSD pick edge cases to work on likely follows the same principal, randomly identify many scenarios to gather data at the same time. The scenario that gathers the most number of cases in a fixed time frame is the "gradient" points to the next direction you should go.

Yes - a truly random selection of scenarios over a week of sufficient numbers and manually classifying to figure out highest probability cases would work.

If we look at all the disengagements and shadow mode errors of 400k cars - I'd guess they will be getting a few million cases a day. For eg. I personally disengage - more than 10 times a day (because I'm using on city roads). So, over a week, we are looking at over 10 million cases. If there are 1000 scenarios of various probabilities, how many random samples do you need to catch the highest probability scenarios with sufficiently good accuracy ? My stats is a bit rusty … anyone ?

We can figure out how realistic the random approach is.

 

[neroden]

The real optimal strategy should be to outline the moderately pessimistic outcome and demonstrate why even that is sufficient for a bull case.

This is what I do. Here's the scenario: No robotaxi income. Sell every car they can produce. Can't ramp up production as fast as they want. Continue to have a bad service reputation.

Here is a premise. Tesla will probably post negative GAAP earnings every quarter for the rest of the year. Why is that ok?

Well, this is a dumb premise. But since Tesla will post positive cash flow every quarter, who cares?

 

[neroden]

As long as they can solve certain # of prioritized edge cases per year, they can ensure the time doesn't increase exponentially.

No, they can't. As you work out the far end of the curve, it's just more and more edge cases of lower and lower frequency.

 

[EVNow]

Truth is hardware is the easy part

Easy as in more well known with definite timelines.

BTW, I've to call out Tesla on the whole power usage stuff. There is nothing sacrosanct about either 100 watts or why 500 is too much. Tesla had a design limitation of 100 watts because they want to be backward compatible. 500 watts isn't a big deal in a car that may use 10 kWh or more per hour even in urban areas.

So, if we ignore that power usage, I'm not sure Tesla has a 4 year lead over NVDA. Ofcourse, for an OEM o build a car with that NVDA chip and needed sensors is a different story.

 

[EVNow]

These "investor calls" where Tesla management discloses material information to a select few investors and doesn't make it public are totally illegal. They violate Reg FD.

Where's the SEC?

(I know, in the pocket of Wall Street crooks.)


------
There were definitely material statements, including that Model Y location decision is next week; that if it's at Fremont, they'll expand the south paint shop; that Model S/X orders in Q1 were weak because news of the refresh leaked; that Panasonic has not reached full capacity yet; that they're lining up other cell suppliers in China.

And that's what we got from the notes from the short-seller. We have no idea if they made other material statements.

This is illegal.

Yes. Infact, everytime an investor meets or tours a company facility (even with NDA) there is good possibility of getting more information than available in public. For a long time I had stopped individual stock trading because of this - only started on Tesla recently because of unique circumstances.

 

[dqd88]

Buy TSLA

FYI I do have some TSLA, about 8% of my net worth. What would a bullish investor put in as a % of net worth? 10%? 50%?
10% seems like a % that is material and can move the needle, yet still allows for decent diversification, yeah?
Is there a rule of thumb (for "casual" investors)?

 

[Unpilot]

Interesting question...I'm at about %65 but that might be a bit over done. Damn fire sales.

 

[Remus]

No, they can't. As you work out the far end of the curve, it's just more and more edge cases of lower and lower frequency.

If a million cars can't see such cases in enough frequency to train the model, how likely these situations be encountered by a human driver in a lifetime?