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

[Singuy]

Disagree. As long as operating cash flow is positive, serious investors won't worry about capital expansion, which should eat cash.

I was referring to cash flow negative from operations. That would be devastating.

 

[capster]

@capster, what do you believe Elon means by feature complete? What percentage of “tricks” of the human brain should the NN be able to master for full autonomy... or even, for the enhanced summon (since that’s also unsupervised)?

Off the cuff, my guess (hope?) is that Feature Complete FSD is this:
* A driving control system that is vastly better than human drivers for almost all common (in the US/developed world at the mo) driving situations.
* A driving control system that may not quite reach the the flexibility and range of human drivers such that in certain locales/situations the driving performance is below that of a human.
* A driving control system system that is 'performance fenced' in that it can recognize that certain situations and locales have driving conditions outside of its abilities and can avoid them or request assistance.
* A driving control system with an overall performance so good that on average passengers are much safer than with human drivers even if it were to drive in the few conditions in which humans are better drivers.

The second question is interesting though not simple to answer. I’ll take a shot though my knowledge of the driving problem and Tesla’s work on it is almost purely conjecture.

It would range from none to a small but select percentage.

None because as I mentioned you could in principle ’evolve’ something completely new. I’m impressed by how much they appear to have achieved with inference alone.

If you were to take inspiration from the brain, the percentage would be small for a couple of reasons. Biological brains have a lot of requirements — developing through growth, regulating hormones, maintaining circadian rhythms and sleep, … — that built systems don’t. Much of what the human brain does — speech, adaptation to changing and compliant effectors, abstract reasoning, … — isn’t required to drive well.

I would reframe the the question as: What would you cherry pick from biological brains to guide and accelerate the evolution of an artificial driver?

Models from vision, sensory integration, associative learning/behavioral conditioning, and, maybe less likely, motor control would be relevant. For Tesla, the remaining missing pieces at the moment I’m guessing have more to do with behavioral control than sensors.

I previously outlined, in a tongue and cheek manner, a quite serious team competition development plan in which one team would try to leverage biological models:

If you are betting on the known to be decisive Team Don’t Rule out Rules (DRoR) in the fantasy leagues of the Tournament of Cybernetic Frankenstein Drivers being hosted (hypothetically for the mo’) at Tesla, you might want to play along with this book (ignoring the parts about clutches, etc).

BookBub shows it on special: How to Drive by Ben Collins - BookBub

I just picked it up because I’m guessing my fantasy team, Team Hey Nice Conditioning Baby (HNCB), isn’t a pure play. Why not embed rules if you know how to train them in or adjoin them you if can adaptively parameterize them? Hey, it’s a no holds barred Tournament — survival of the fittest and all.

In HNCB’s defense, it’s likely the late, dark horse anyway. Also, Meta-Team All Your Nodes are Belong to Us (AYNBU) is just going to be poaching, distilling, and integrating all the other teams’ best work anyway. Bunch of frickin’ monster makers over at AYNBU, let me tell you!!

Of course, Team Infernal Inference Engine (IIE) may have the lead going out of the gate, what with all their current glory. Showboaters, sheez.

Should be some exciting developments! You never know, it’s possible Meta-Team R. Fast Arbitrator (R.FA) will be as good (well close enough to ship along with some teams) as AYNBU.

Going to be an exciting Tournament for sure! What kind of mutant cybernetic driver will emerge superhuman??

We can surmise that Tesla is gonna dim the lights in the Intertubes when those matches hit the clouds! (Not to worry folks, shouldn’t be more than a total of 30 match days with chances of traffic delays)

Go out there and step right up and place your bets!! Watch your step around those fantasy shares, please.

Not any advice here either.

Tl:dr; Can Elon reach FSD before or within epsilon delta of his forecast (sorry, had to do it fans of “order of magnitude”)? He has a sporting chance.

 

[neroden]

I'm thinking more and more that lower priced EVs like the model 3 will hold their value much better than similar ICE cars.

AS cars get older fuel and maintenance costs take up more and more of the total cost of ownership pie... but 'fuel' and maintenance are so much lower on EVs that a used $15k EV is going to have the same TCOE of an ICE have that has a significantly lower price.

I'm thinking this effect will make ICEs depreciate much lower than ICe cars.

This -- EVs depreciating less quickly than ICE cars -- will happen eventually. It isn't happening yet because everyone's just mindlessly using the standard depreciation curves based on miles driven and years.

I think it's going to change once people notice what good shape the Model Ses are in after 10 years. But that won't become clear to the general public until 2023 or 2024. Which is around the time that the value of ICE cars will drop off a cliff and they'll be practically unsaleable, so I guess we won't notice the more subtle depreciation effect immediately.

 

[SO16]

Look at auto dealer donations to the state legislature. Bigger in some states -- which pass taxes on EVs -- smaller in other states -- which have incentives for EVS.


Paid off.

In Michigan, many democrats are union members and therefore don’t like Tesla.

 

[neroden]

I was referring to cash flow negative from operations. That would be devastating.

Yeah. (Unless it was due to a big inventory-in-transit buildup. But I don't see how they can deliver 92K cars AND do inventory buildup -- they aren't producing fast enough, surely?)

When Musk says cash flow, though, he generally means cash flow -- he says something else when he means cash flow from operations.

 

[neroden]

In Michigan, many democrats are union members and therefore don’t like Tesla.

It would be more accurate to say that many are *employees of Ford, GM, and Chrysler* and therefore don't like Tesla.

 

[EVNow]

In Michigan, many democrats are union members and therefore don’t like Tesla.

Unions apparently have more FUD thrown at them than Tesla can ever imagine.

ps : Its no secret that GM actively campaigned against Tesla !

Rick Snyder Signs GM-Backed Michigan Tesla Ban

 

[Carl Raymond]

I'm pretty sure it'll be lithium, partly because there has been a bizarre failure of financial funding to miners to ramp up lithium production. I don't see this for other mining situations. Some of the other raw materials bottlenecks would be classified as processing, not mining. (There's not that much synthetic graphite manufactured, but there's plenty of petcoke to make it from; there may not be that much nickel sulfide, but there's plenty of nickel to make it from).

Australia has lithium. Thinking from Australian perspective, I’m wondering if coal and lithium are seen as competing interests. If coal lobby has the power to shutdown climate action in this country, perhaps they also have the power to put a go slow on lithium.

It’s consumable v asset. Corporations would rather sell consumables than materials for assets. Coal gets dumped into a furnace, gone. Lithium goes into a battery that may soon last a million miles.

Same logic as why toothbrushes are designed to wear out and printers are sold as loss leaders to proprietary ink cartridges (which cost 60 cents and sell for 60 dollars).

 

[neroden]

We ride along in Tesla Mobile Service's customized Model S - Video

Tesla mobile service

Thanks for this! This is a very useful informational video. It's also marketing. BTW, basic Tesla marketing would include readily providing links to things like this on a consolidated "PR information" page -- so reporters looking for information would find this immediately.

Interesting thing to note: these are mostly pre-Facelift. This iwhat Tesla is doing with the older lease returns and trade-ins. In case anyone was wondering why lease returns and trade-ins aren't showing up in used inventory.

It is mentioned that some Model Xes are also being converted.

 

[Garlands]

Interesting reddit thread on service improvements:

Our shop became Tesla collision certified recently and it has been a very welcome change. (Part delays are also getting a lot better for those worried about repair turn around time!) : teslamotors

 

[Dutchie]

OK, I went on Twitter this afternoon and saw that I had like 30 notifications! Well that never happens to me. Looking at it it was all about a reply I gave to some bloke who wanted to accelerate Tesla's demise.

My reply was just a simple question: "Why?" Tesla has about 30 thousand employees who will all lose their jobs with its demise and many more people who get in trouble because they rely on the pay cheques. Why do you want Tesla's demise?

Reading trough the many reactions people stated that they did not want Tesla's demise but that it should file for chapter 11 immediately and restructure. I mean what are these guys thinking. So Elon should file Chapeter 11 with the courts now and they really think the courts will approve? So they think Tesla cannot pay the bills now and Tesla has not a s..t load of money coming each day due to all deliveries? And if you see how many "likes" these posts get.....
Have these guys no sense of reality?

Then there was another one who said that Elon wasn't allowed to talk because of some SEC rule. I replied, since when aren't people allowed to talk? Yes, but he lies all the time.. My reply, yes he exaggerates and has his timing off but lies all the time is a bit of an exaggeration..

Then again another one. Tesla is already in demise but Wall Street doesn't know it yet. I mean, on which planet are these guys living. I just showed them all the sales from 2011 onwards and asked them what is demise in that stat. No reaction

Then there was another one who was complaining about Tesla as being a fraud employer... My reaction was then why does everyone want to work there and why does it get overall positive reviews? No answer

Oh and the biggest one was that Elon was a fraud. I asked in what way? Again he exaggerates and his timelines are off but in the end he always delivers. Yes, he also made mistakes but to call him a fraud?

What I have learned is that these TSLAQ people are really "foam mouthed" fanatics who really cannot keep it straight anymore. I think a lot of them need to seek help because some of them aren't really well at all.

 

[neroden]

Even if we say that Tesla scale data sets are necessary but not sufficient, who says that Tesla is solely using inference a la multi-multi-layer backprop?

Biological brains are a bag of tricks. You can find mis-matched based learning for categorical memory, match based learning in motor control, and opponent processing based reinforcement learning for timed behavioral selection. Some of the tricks involve learned mappings from one coordinate system to another, e.g. retinal coordinates to spatial to motor codes in movement. Some of the tricks transform input, e.g. visual inputs are spit into form and boundary channels that reconverge for further processing. And so on.

You can pick and choose from this bag of tricks. Oftimes you can leverage some of these insights in data preprocessing and input representations. Sometimes you might need to do more. You can even use a hammer: You can add 'rules,' algorithmic pre-, inter-, or postprocessing.

Just because Tesla isn't publicizing every trick they use doesn't mean they aren't using them. Don't be blinded by your own strawperson.

Perhaps my favorite thing from the Autonomy Day was the information that they are making a large number of different specialized neural nets to do different specific things, and then passing information between them. I think they'll end up needing to have an expert system or two too, to enforce the rule-based traffic laws and the logic-based parts of driving reasoning but they can do that.

It is important to use a lot of different specialized modules that talk to each other. That is something Tesla is doing. I wish Musk wouldn't underestimate the number of modules he's going to need (far far more than he thinks), but it's the right approach.

FYI this is another way of saying why it's going to take a long time to get full self driving. Each time you realize you need a new module, you're basically starting the development process from scratch. Tesla may develop each module faster than anyone else can -- and they will. But the only way to accelerate the full development is to accelerate the speed at which you identify the *need* for the new modules, and so far Tesla's doing that fairly slowly -- some paranoid expert drivers might be able to speed up the process of listing what they're going to need to develop.

 

[Remus]

There can't be much of an inventory buildup if orders > production (he misspoke saying sales > production). There is also no evidence of the wave unwinding - and with another "beat the record at any cost" quarter, they will try to deliver as much as possible. This means wave good bye to wave ending plans

While the other 2 consume cash - it would be $600M+ big.

I think its ok for EM to be somewhat circumspect, after Q1 may be he learnt a lesson.

We really don't know how much they are unwinding. Q1 they had those "deliver them regardless of cost" things. If they stopped doing that would be somewhat unwinding.

 

[brian45011]

This is great. Does that mean I can go into a crowded theatre and yell "Fire"? Because freedom of speech.

Yes; theaters are mostly empty these days. Just don't do it in a crowded night club--unless the pyrotechnics for the grand finale were poorly designed.

 

[Mars ☰mperor]

You know what actually pisses me off? That Tesla doesn’t sell merch at the damn showrooms/service centers. Why? Seriously wtf? Merch is an impulse purchase.

Yes, as I mentioned before there is no online merch store in many regions around the world.

And Tesla should really include merch store and design studio in the damn Tesla mobile app, so current non-owners can also interact with the app, thus leading to even more demand.

 

[neroden]

Semi OT

It will be years before the tri-motor architecture appears in the S/X. I predict sub-10 sec quarter miles, a 300+ kph top speed, and a 500 mile range on a 125 KWh Maxcell pack to be released around 2023. Prices should be comparable to current levels.

Three motors is still clearly an interim step. As Rimac knew, there will eventually be one motor per wheel. It makes the suspension more complicated, and at the moment that means more expensive. I wonder when Tesla will develop a four-motor design with low unsprung weight which they consider cheap enough to mass produce -- I'm sure every engineer working on the drivetrain wants to do that.

 

[neroden]

Good Point - make it whatever the lowest price available on the website is. $39,900?

Or, don't even put the starting price. Only reason I mentioned that is because another thing most people ask me is, "Don't those things cost about a hundred grand?".

Fascinating. Uninformed people are still surprised when I tell them the Model 3 starts around $40K.

But it's not new!

People were actually overestimating the price of Teslas back when I bought my $80K 2013 Model S -- people were guessing it was $200K!

And people thought the Roadster was a million-dollar car!

So the pattern of people being shocked by how low Tesla prices are is actually consistent.

There's got to be a way to leverage this phenomenon in marketing.

All are shocked when I say they start around $37,000 (or whatever).

 

[EVNow]

Off the cuff, my guess (hope?) is that Feature Complete FSD is this:
* A driving control system that is vastly better than human drivers for almost all common (in the US/developed world at the mo) driving situations.
* A driving control system that may not quite reach the the flexibility and range of human drivers such that in certain locales/situations the driving performance is below that of a human.
* A driving control system system that is 'performance fenced' in that it can recognize that certain situations and locales have driving conditions outside of its abilities and can avoid them or request assistance.
* A driving control system with an overall performance so good that on average passengers are much safer than with human drivers even if it were to drive in the few conditions in which humans are better drivers.

Interesting "cross platform" features / quality levels they might be targeting.

I've listed what I think are the functional requirements they might be targeting for FC.

Tracking FSD Feature Complete

 

[dakh]

Off the cuff, my guess (hope?) is that Feature Complete FSD is this:
* A driving control system that is vastly better than human drivers for almost all common (in the US/developed world at the mo) driving situations.
* A driving control system that may not quite reach the the flexibility and range of human drivers such that in certain locales/situations the driving performance is below that of a human.
* A driving control system system that is 'performance fenced' in that it can recognize that certain situations and locales have driving conditions outside of its abilities and can avoid them or request assistance.
* A driving control system with an overall performance so good that on average passengers are much safer than with human drivers even if it were to drive in the few conditions in which humans are better drivers.

Elon is probably optimistic in what he thinks he can deliver but in my opinion your guess/hope is wildly optimistic. How about this: all the NNs are identified and reasonable training sets/pipelines are established, but accuracy isn't necessarily good enough to declare victory and will have to be worked on to get the required accident per mile stats? Tweak hyperparams, better training sets including more corner cases, maybe some tweaks to pipelines, things like that.

Being able to identify stuff the car sees is a "feature". You can call it feature complete when you have a system that does that and knows what to do with the stuff it sees. But the fact that the feature is there doesn't mean the implementation is good enough for L4 driving. I don't see any other reasonable interpretation of why Elon would call it "feature complete".

Industry standard for "feature complete" is that you have all the right stuff in the right places, but it might crash on corner cases, not able to deal with more than trivial input size, not be able to function in all target environments etc.

 

[neroden]

We really don't know how much they are unwinding. Q1 they had those "deliver them regardless of cost" things. If they stopped doing that would be somewhat unwinding.

Even if they stopped doing that for *overseas sales* and kept doing it for US sales (to help out people who need to get their cars before the tax credit drops again), it would be *somewhat* unwinding.