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

[scaesare]

They didn't mention it at autonomy day, therefore they're not doing it yet.

Incidentally, I think it's this tenet that underlies many of your assertions about Tesla's autonomy approach or framework: If they haven't publicly announced/demoed something, or you can deduce it from a "visible source" (i.e. job postings), that initiative doesn't exist.

Whereas my opinion is that all you conclude from a non-announcement is that it wasn't announced.

 

[dmckinstry]

It's a standard engineering concept...how engineers are taught to think. Until you rid the world of engineers, this phrase isn't going anywhere.

It's been a standard for scientists and mathematicians for centuries and has nothing to do with Elon and Tesla.

 

[dc_h]

I guess it depends on what you mean by "custom hardware and software".

This is not a completely apropos counterexample, but I was involved in the Supercomputer industry back in the 1980s. There was the "big iron", dominated by companies like Cray, CDC, IBM, Fujitsu, Amdahl. Then there were a number of "mini-supercomputers", startups using different architectures and specialized hardware. Most of them you'll never have heard of: Elxsi, CHOPP, Pyramid, nCUBE, and a few others. What killed them? "The attack of the killer micros". Basically, machines with lots of commodity microprocessors. (I was involved in a project at IBM Research that replaced a vector mainframe with a rack full of Riscsystem 6000s. Got into a lot of political trouble. Deep Blue was a rack full of RS6000s too, which is how I got to meet Gary Kasparov. Long story.)

Companies like Intel (especially), Motorola, National Semiconductor, at the time could throw more money and effort into making better microprocessors than these other companies could improve their specialized hardware. At any given instant, their lead was only 18 months over a machine with a bunch of micros at 1/10th the price (Sequent, Sun, others.)

Fast forward to a decade ago, same thing happened with general purpose micros versus dedicated GPUs, we still need a CPU but it controls an array of GPUs. Fast forward to now, and we see dedicated (but soon to be commodity) NN machines taking part of the market that was dominated by GPUs and vector instruction sets.

I see us at a fork in the road. Who will be the big player in NN processor chips? Who will be the Intel, or NVidea, Bitmain, for NNs? There are a bunch of incumbents who have recognized the space being important (Intel, NVidea). There are a bunch of new startups (LMGTFY). Then there's Tesla. If it was any other company/entrepreneur I'd say that Tesla was not going to succeed in this area, because they would be the supercomputer manufacturers facing the killer micros, and one of the startups would win. But they're not, they are (in this area anyway) a well-funded startup with some of the best talent and a captive market. I think they will lead, not get crushed. I certainly hope so. BUT, to lead, they have to make the chips available to others too; their captive market isn't big enough, and I don't just mean other car manufacturers. They have to open to other entire industries. In other words, they succeed if their hardware is no longer "custom", rather "commodity".

Corollary: Intel and NVidea will not win this race. Incumbents never do, too much baggage. If it isn't Tesla it'll be a different startup. Hmmm, must start looking around and following some of those startups.

I think the analogy is the tesla chip is like an ARM vs an Intel chip. Instead of a phone chip using a fraction of the power at a fraction of the cost, it’s a car.

 

[Zhelko Dimic]

Firstly, are you really disputing that ABS has much longer braking distance on ice?? It's simple physics, and I cited the numbers: ABS has more than 50% longer braking distance on ice. Are you calling braking ice on roads a 'limited and rare set of circumstances'?

Secondly, many racing series ban ABS altogether, and even when they allow it it's not your everyday street car ABS, it still allows the wheels to slip:

https://www.quora.com/In-the-real-world-Do-race-car-drivers-use-assists-ABS-TCS-STM

"That said, ABS and TC in a race setting are much different than in a street car. We have multiple maps and are changing them on the fly in response to conditions. They also do not ensure that a car doesn't get out of shape, they just allow us to commit slightly early to things, or protect the tires."​

I.e. racing cars typically have weaker ABS that still allows the wheels to slip when cornering, because not having an ABS nanny is beneficial to a professional driver's driving performance. Which was exactly the whole point I made.



If you read my whole comment and don't just grab out of context quotes then you'll see that I'm not arguing against ABS - to the contrary, in aggregate it has saved countless lives.

Also, you launch personal attacks, then you refuse to reply:



Classy.

This unfortunately requires answer as you're playing with half-truths that may resonate with people.

Breaking on ice: there is no way you can treshold brake on ice better than decent ABS. After all, you don't have interface to do it. Brake pedal engages all 4 brakes, and you are not even in control of the brake bias front to rear, let alone left to rear. Which means you as a driver have a primitive tool (one pedal) that doesn't allow you granularity of control as wheels travers over varying surfaces; even if you were a superhuman. ABS computer can release and brake each wheel individually.

Or are you thinking about mashing the brake and locking wheels? Yeah, depending on condition of tires and surface, drag that's created that way may stop car in shorter distance than ABS, but at what price? Complete loss of steering. If that's acceptable, you may as well carry an anchor in the trunk that you can drop, or maybe parachute, or no. I know! you can jump out of car and tumble, you're likely to outperform ABS!

This is what I'm talking about you pulling some obscure links, edge cases and half-truths.
I wish you stopped weighting and arguing about stuff you know little about. I'm really done now.

 

[capster]

Keep in mind that perception differs from reaction which differs from tuning (of dynamics for compliant effectors) though they are interwoven in many ways.

Most animals use reinforcement learning to survive. They use it to choose and time reactions in real time to threats and opportunities in the world and in light of the state of internal drives (hunger vs satiety, fear vs relief, ...). I suspect it might already be deployed in systems somewhere for which that capability might be helpful or even necessary.

There are reasons to think cars might benefit, for example for timing jockeying during merges while avoiding crashing. The compute power required could even be quite modest relative to perception. So you would not necessarily need to do it on board this new chip.

Personally, I’d start with a brew of some of the more biologically plausible conditioning models, Tesla’s driving simulations on lots of servers, and some genetic algorithms. I’d call it, I dunno, Project Dojo or something.

Tl;dr: I’m pretty sure Tesla isn’t sharing every trick up their sleeves, though I am not here asserting they do use it (yet).

The work I'm thinking of is probably pretty obscure to most everyone here, so here are a couple of relevant citations that will give you a taste.

For rigorous mathematical modeling of behavioral conditioning using biologically plausible neural networks with actual brain correlates, look at the first three chapters of The Adaptive Brain I, Stephen Grossberg editor, North-Holland, 1988. (The Adaptive Brain II has interesting stuff on vision, though Grossberg and his colleagues did quite a bit more vision later)

The article Associative Learning and Selective Forgetting in a Neural Network Regulated by Reinforcement and Attentive Feedback by Grossberg, Levine, and Schmajuk in the the book Motivation, Emotion, and Goal Direction in Neural Networks edited by Levine and Levin, Lawrence Erlbaum Associates, 1992 is also nice.

For spectral timing during learning, check out the article Neural dynamics of adaptive timing and temporal discrimination during associative learning by Grossberg and Schmajuk in Volume 2, Issue 2 of Neural Networks, 1989.

All of this stuff is heavy sledding. Many engineers may lack the background in Psychology (non-extinction of conditioned fear, anyone?) and Neurophysiology. The math (dynamical systems) is no picnic either which makes this entire vein of work challenging for many.

Still, those of you speculating about what can and can't be done technologically might care to be aware that there is powerful and relevant modeling work that's been around for decades available to those who know about it and persevere to understand it. The folks at Tesla seem to me to be the kind of people who would be aware and would persevere.

 

[EVNow]

OT :

At our present level of understanding, the primary contributor to most heart attacks is bad luck. Otherwise known as "we haven't a clue" or "no known risk factors".

Bad luck as in DNA ?

Yes, you can get into an accident because of bad luck. But you know the risk increases if you live dangerously.

 

[printf42]

James Locke on Twitter

@elonmusk have you given thought to allowing owners who didn't pre-buy EAP and/or FSD but want to allow car into robotaxi fleet doing so, but with Tesla taking more of the sales to compensate for not paying for the FSD license? Up until the value is paid off with margin?

Elon Musk on Twitter replies:

Will consider

This has more implications than it sounds on the surface:

• Now FSD-as-subscription is officially on the table, and it's requested by customer, you can not blame Tesla to be greedy if they offer that.
• FSD is the main value creator in RoboTaxi service, if customer's car doesn't have FSD, Tesla has every right to take lion's share of the revenue, much larger than the expected 25%-30% when car already has FSD. I would expect owner only get reasonable return for the car's wear plus electricity, which is at ~$0.2.
• Considering they calculated gross profit on RoboTaxi at ~$0.65/mile. This means Tesla could charge >>$0.2/mile for FSD alone. With (vastly underestimated)400k car life span, it would sum up to 80k, substantially higher than what they're charging for FSD now.

If this calculation is not off by order of magnitude, there are two logical strategy for Tesla on FSD pricing:

1. Significantly raise FSD price, by at least an order of magnitude, or separate RoboTaxi FSD from personal use FSD, and charge a hefty price for former.
2. Turn FSD into subscription only service, charge per mile for RoboTaxi, and maybe per month for personal use.

My interpretation when Elon said Tesla will substantially raise FSD price, he does not mean a mere 3k(or 1k), he means 10x-20x. It's software after all, pricing for Software is relative to how much value it creates, not how much it costs to develop.

But they can not do that before they open Model 3 order to all existing Tesla markets, and give all reservation holders a chance to order FSD at current price level.

Now we would see a slow and steady FSD price hike until they get RoboTaxi approval.

And they could turn FSD into subscription only any moment from now, which I think would make more sense, since HW4 etc in the pipeline.

IMHO, this is the path for TSLA to 4k, for real.

 

[EVNow]

No. I thought after they added most of worldwide and introduced SR and leasing that they’d go to more than 20k / month.

If anyone thought differently 6 months ago, I sure don’t recall it.

But why do you think it is below 20k/month ? Even in Q1 with all kinds of headwinds they delivered 17k per month (with 10k or so not delivered).

ps : 6 months back i.e. in Oct '18 Tesla delivered ~ 17,750 cars according to Insideevs.

 

[Doit]

This is not the relevant number when comparing to Autopilot's crash rate of one accident every 2.87 million miles with Autopilot and every 1.76m miles without AP. There may be a crash every 436k miles, but with 1.8 cars per crash, the average US car has an accident every 246k miles. So Tesla is seriously underselling its accident rate improvement vs the US average. Presumably they just forgot to adjust for number of vehicles per crash, but this is pretty stupid when the average human crash rate is one of the key benchmarks in their FSD progress.

You are mistaken. With the assumption that there are approximately the same 1.8 cars per crash involving any Tesla, the numbers (436k, 2.87m, and 1.76m) represent the same metric, i.e. vehicle miles traveled (VMT) per crash, a jargon from your linked doc.

On second thought, I might be mistaken too: it should be the other way around, i.e. Tesla's numbers should be divided by 1.8. The reasoning: as there are only tiny fraction of Teslas driving around, the 0.8 vehicle of the averaging 1.8 vehicles of each crash with any Tesla is another Tesla is also tiny. Therefore, the miles traveled by the 0.8 vehicles are unaccounted for. Unless, of course, Tesla already factored this into its calculation.

Of course, this factor of 1.8 needs to be adjusted downwards in areas with much higher percentage of Tesla vehicles.

In context, using the miles per car per crash metric, rather than the miles per crash numbers, then using 3x safety improvement, 1 million Robotaxis and 100k miles per car per year, Tesla would see:

• 135k Tesla Robotaxis in accidents per year >>> 741k VMT vs. currently: 2.87m VMT with AP, 1.76m without AP
• 42k accidents involving an injury per year >>> 2.38m VMT
• And 548 accidents involving a fatality per year (this may be the passenger, a pedestrian or another car driver. The odds of the Tesla passenger dying are lower due to the crash safety design of the Tesla cars).

It makes no sense that Tesla would push out an FSD robotaxis that significantly worse than accident rates, with or without AP.

BTW, according to Tesla's numbers, Teslas already 3x (without AP), 5.6x (with AP) better than the average vehicles. Nobody cares?

 

[EVNow]

And they could turn FSD into subscription only any moment from now, which I think would make more sense, since HW4 etc in the pipeline.

They already plan to take a cut of TN - so subscription for FSD just means a larger cut (for people who haven't purchased it already) - like the tweet says.

 

[EVNow]

BTW, according to Tesla's numbers, Teslas already 3x (without AP), 5.6x (with AP) better than the average vehicles. Nobody cares?

With how many disengagements / overrides ? Once the disengagements/overrides become negligible (as low as crash rates), then people will start caring (from FSD POV).

 

[EVNow]

Incidentally, I think it's this tenet that underlies many of your assertions about Tesla's autonomy approach or framework: If they haven't publicly announced/demoed something, or you can deduce it from a "visible source" (i.e. job postings), that initiative doesn't exist.

Whereas my opinion is that all you conclude from a non-announcement is that it wasn't announced.

I think they are busy just trying to get to FC. They will get to things like this once they finish FC and knock down a 9 or two. If they implement this kind of thing now - it would be unusable. It will stop all the time.

 

[kbM3]

But why do you think it is below 20k/month ? Even in Q1 with all kinds of headwinds they delivered 17k per month (with 10k or so not delivered).

I hope it’s higher than 20k / month by now.

I started being concerned about demand because of the rate at which Tesla has pulled demand levers this year.

My concern grew when Gene Munster stated that orders have dropped off in the last 6 months. I don’t know what data has led him to believe that, but it would line up with the pulling of demand levers so quickly. Hopefully, if he’s right, it was from earlier and orders have now picked back up.

 

[Carl Raymond]

And they could turn FSD into subscription only any moment from now, which I think would make more sense, since HW4 etc in the pipeline.

Down the track when driverless cars are proven income earners, subscription model seems inevitable. People who cannot afford the then huge upfront cost will subscribe and pay for it out of their Tesla network revenue stream. Without the subscription model, the car goes without FSD and neither the owner nor Tesla (shareholders) profit. Bad all round.

Until that time, x people will take the plunge at y dollars, taking a punt that the value will jump on ‘teslanet day’, and enjoying the features in the interim. Tesla will know what this x y curve looks like and position y to max the product of x times y.

Because people like to gamble, fixed price seems the best option for now.

For people who want the features but don’t believe driverless will happen, there is the lease the whole car option. Tesla can price FSD quite reasonably here, because they literally get it back later.

 

[AudubonB]

Also, could Tesla cars collect data about the cars around them.
Speed, tag number, all worth a lot of money to the other insurance companies.

IE That car going 50 mph faster than the speed limit that passed you?
How much would that be worth to Allstate, Progressive, or any other insurance company ?

Just an FYI for those otherwise inclined to pishtush the above:

Reasonably recently I was forced for the first time in my insured life to push outside the cozy USAA bubble, as the vehicle I had just purchased was one they will not carry.
So when I shopped around other carriers, I was shocked (first time) and stunned (thereafter) to learn that every other insurer knew instantly and blithely stated to me words along the lines of "Ah. I see you have USAA as insuring the Tesla Model S and Model X and the _A_ and _B_ and _C_and...... (I've a lot of vehicles...).

The important point is that that confidentiality you - if you're like me - think you have with your carrier?.....It doesn't exist. In another era, another industry, it would have been the affirmative answer to the once-famous question "Does Macy's tell Gimbel's?".

 

[Maitri982]

Not a huge deal, but at autonomy day Elon outlined future items he will do through end of 2020 and the new roadster was nowhere in sight

 

[EVNow]

My concern grew when Gene Munster stated that orders have dropped off in the last 6 months. I don’t know what data has led him to believe that, but it would line up with the pulling of demand levers so quickly. Hopefully, if he’s right, it was from earlier and orders have now picked back up.

I think Muster is confusing demand, orders, delivery (sales). Tesla doesn't have a long waiting list as it did six months back. So, in that sense "demand"/"order" has dropped.

 

[Zhelko Dimic]

@ggr : Ok, this thread by all accounts has gone off the rails and I admit I have been a part of it, infrequently, but still a part.

Your job as 'mod' is thankless but I have to question moving my link to Trump imposing an increase tariff tweet that I think most would agree will effect the SP of TSLA off'Tesla and the Investment World' (this) thread.

From what I can tell it did not send us spinning off into some political debate.

C'mon @AIMc, you've been here long enough that discussing stuff that can affect SP isn't allowed in this thread: no TA and no crazy market moving tariff talks!
Everything else is fine...

 

[woof]

The important point is that that confidentiality you - if you're like me - think you have with your carrier?.....It doesn't exist. In another era, another industry, it would have been the affirmative answer to the once-famous question "Does Macy's tell Gimbel's?".

Not sure if Macy's ever did tell Gimbel's, (aren't they both out of business?) but in my neck of the Commonwealth it's the State that knows and tattles about insurance carriers. I get hand crafted snail mail all the time telling me that company Z can beat the rate I'm currently getting from my current insurer, with my current coverage clearly shown line by line.

 

[AudubonB]

A Google study estimated 4.2 crashes per million miles. This is 99.9996%. Tesla needs to be at 6 nines to be better than humans. 7 nines would make it better by an order of magnitude. I'd look at this as the point when most people would stop driving.

Crash rates for self-driving cars less than conventional car: study - Reuters

Number of crashes reported to nhtsa should be taken with a pinch of salt because of under-reporting. Insurance providers have the best data here. All state says on average someone gets into a crash every ten years. That makes it about one every 150k miles or so.​

The above poster is not the only one who has mangled the important data; I have sat on my hands long enough and so only now am going to respond....in the hopes that this confusion stops now. Am not picking specifically on EVNow.

But first, some further falsehoods, the better to clarify to all why I've been gnashing my teeth:

Let's take the data above. In the prior post we learned that the Google study estimated 4.2 crashes per (1E6 mi * 5280 feet/mi * 12"/ft) = 7 billion inches! 99.999997%!!!! Level 5 here we come!!!

Whoops. Wrong denominator. In fact, there actually are 4.2 crashes per each gizzlewiwicq*. Oh no. 0.0001% perfection!!! We yumans have a long way to go.....

*(Martian SI. Equivalent to 4.2 million earth miles).


'Nuff said, everyone? Careful out there: it's a statisticasticocal jungle.