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

[kbM3]

Havent seen this posted yet:

About halfway through, Munster says he was at the factory recently and noted the production bottlenecks have been released

Interesting that he said very matter-of-factly that the order flow has diminished.

I wonder on what data is the conclusion based?

On the good side, the number their model predicts is only 20K below Tesla’s guidance range.

 

[Boomer19]

Havent seen this posted yet:

About halfway through, Munster says he was at the factory recently and noted the production bottlenecks have been released

cnbc is basically “soap opera” disgusting.

AR sorkin acts soooo confident in his implications that tesla has a demand problem, and even if not, they have a supply problem (which they do to an extent), and if not, they won’t be able to stay solvent and/or the competition will catch up. the cap raise was a kick the can, munster pretty much alluded to this as well. gene, stay with apple. that’s who made you famous, and about all you could handle. you haven’t said anything new.

this bias is why mindless viewers stay uninformed. then melissa brings the weed back into the equation at the end of the clip

i mean jesus effing christ they are shameless trolls

what a hack organization
if sorkin and co were so damn smart, why are they crappy tv hosts? sorkin co-produces billions on showtime btw, so it’s no wonder he became, let’s use the phrase that’s become so annoyingly overused by everyone now, an “order of magnitude” douchier because he’s been in the HF circles and breaking bread with them more and more. clearly.

if you were to believe their self assurance portrayed when the camera was on was an indicator of their mastery, they’d be billionaires and not need to be on tv banging stocks around for their real masters...a55holes

how do people not see right through this BS? embarrassing, really.

 

[Doggydogworld]

I think Tesla may be using or may be planning to use reinforcement learning in its simulations (but simulations aren't currently its key training method).
This would be like AlphaGo (and not like AlphaZero). 1) Begin with supervised learning based on high graded human driving actions collected from the fleet. 2) Use reinforcement learning with a reward system to train the driving policy. 3) Train a value network to predict the value of each decision.

The difficulty is that the real world is far more complicated than any computer game. The movements of other people, cars and objects are extremely difficult to model when building your simulation. The simulation training also doesn't help you solve cases the software engineers couldn't imagine and didn't design for. Reinforcement learning in simulation is only really going to work if you already know the case you are trying to solve. This is why Waymo's simulation first approach isn't likely to work, and why in fact Tesla also has the advantage in simulation.

Tesla's fleet will be able to identify and collect data on far more rare problem scenarios. For these scenarios it can use the data collected to feed good driving examples into a supervised learning model in a simulation and improve the driving policy through reinforcement learning. It can also use the real world data to feed a neural network to much better predict and simulate the decisions of other cars and pedestrians. Waymo is using this approach to try and train the other moving objects in its simulations, but I don't see how they can make much progress with this without the volume of data Tesla has access to.

Karpathy and Musk comments on the (lack of) value of simulation indicate they aren't even thinking about RL. Maybe someday. People here vastly overestimate NN involvement in Tesla's driving policy. Based on Autonomy Day I see their NN giving path prediction hints, but heuristics constructing a 3D scene and making all trajectory decisions.

Even Waymo's advanced R&D ChaufferNet project didn't use RL, and they've put much more into simulation than Tesla.

 

[Maitri982]

Tesla has a history of actuarial incompetence, mispricing extended warranties and service plans, miscalculating warranty reserves, so unless they have hired some new talent, they could lose everything by writing insurance. The plan revealed in the California filing looks pretty conservative though so maybe they have some actuaries...

I've noticed Tesla hires a lot people who I would classify as smart, but young (i.e. minimal proper experience) and unqualified for the position they are put in...and this is at the exec level too. If they indeed are *smart*, they will get what they do right eventually, but there is that cutting teeth period where they do stupid things too because they lack the proper background. There's a lot of learning on the job there.

That's not all bad. Sometimes you do something novel when you come in with zero a priori knowledge.

But most of the time, especially for relatively rote things, experience matters.

 

[SW2Fiddler]

This (terrible) scenario is so obvious, that Tesla surely will train or already have trained their FSD computer to handle it.

It’s arguably a meme, even. It (Ball>STOP>Child>“Phew!”) is in half of commercials about any kind of car autonomy. It’ll be the bloody Second Law of Car Robotics.

It’s the basis of a Jack Rickard quip that everyone likely skipped past:
“Every loose ball will be accompanied by a child. About half of all children are accompanied by balls.”

 

[Zhelko Dimic]

• ABS-brakes are actually braking worse than the best human driver: in slippery conditions a professional driver can just stay at the edge of wheel slip, without waiting for the wheel to actually slip. To make matters worse ABS typically cannot even be turned off by the human driver, it's an all-or-nothing feature included by manufacturers. This means that braking distance with ABS is actually longer than with the "best human" driver, hence by your logic ABS technologies would have been "shut down" already.

Bunch of BS. I go to advanced driving days on racetrack all the time, have over 100 track days, hang with very skilled drivers all the time, and you would be hard pressed to find one to support your viewpoint. All advanced drivers organizations and instructors I've been with teach opposite of what you're saying. At some events we even do parking lot exercises to prove the point to newbies and have them internalize it, as well as feel effects of weight transfer.
Of course, with you, I have to immediately point to a very limited and rare set of circumstances where opposite may be true, or you'll dig out some obscure link(s) that proves it, but that's just spinning. Yeah, on dry track with warm, sticky tires, careful weight transfer cognizant breaking beats ham-fisted ABS engagement or ham-fisted tires lockup (for a skilled driver). I won't respond any further on this subject.

What's next after ABS: arguing against benefits of vaccination?

 

[Doggydogworld]

It's also good that Tesla tries to smooth the delivery wave with the newly raised capital:

Smooth delivery requires less capital than the wave. Smooth delivery would vastly improve intra-quarter cash balances.

The EOQ numbers don't look as good with smooth delivery, however, which is why Tesla keeps backsliding into the wave despite promising to unwind it for over a year.

 

[Maitri982]

Interesting that he said very matter-of-factly that the order flow has diminished.

I wonder on what data is the conclusion based?

On the good side, the number their model predicts is only 20K below Tesla’s guidance range.

the whole demand issue as far as I can tell was a red herring put out by shorts. I have seen zero evidence anywhere that this is true other than for Model S/X refresh. The idea that Model 3 demand has all been met now and they need something new is ridiculous. We are not even at the first pitch in a nine inning game that plays out over decades. TSLA is peerless right now and unless someone does something incredible and amazing, that won't change anytime soon.

And this is with FSD not ever working. If FSD actually works we have a 10-bagger, if it doesn't and they execute the business its still a 3-5 bagger IMO. And BTW if their self driving gets to something like 99.99% perfect, which is not good enough for Level 5...this is still huge demand driver as it makes drivers life so much easier and insurance companies (<cough> also Tesla) risk profile much better.

So some level of autonomy from Level 3-5 is a big winner no matter what. I am personally hoping for the home run of FSD...and i have coded and even created novel neural nets and started a machine learning company...so I know a little about this field...but I confess to not having enough deep info on where that line is and what all the edge cases are and when they will handle most of them, etc etc to say "for sure it will work" or "for sure it won't work" or give any timeline.

they have a great FSD team and Musk is going to look like a total ass if he is wildly wrong...he has their latest stuff in his own car too. BUT...he drives in a very constrained area...so his experience with it is nowhere near adequate in sunny LA to say how good it is. But I note that he said by end of 2020 he feels very confident that "somewhere" it will have regulatory approval. That could be metro LA or some other relatively easy scenario.

 

[Bladerskb]

This should be an internal deadline, not one for the media and investors. All it does is setup a (99% likely) "fail" situation for Elon and perpetuate the "forever over-promising" narrative.

I overstate it a bit, bit that's exactly what the MSM, bears and shorts will do, they love these forward-looking statements, easy wins for them.

Elon uses these statements to sell sales and upsell features and spread hype and marketing. You cant have it both ways.

 

[AlMc]

Time to clear another poster that I don't need to hear comments from anymore.

 

[ZeApelido]

cnbc is basically “soap opera” disgusting.

AR sorkin acts soooo confident in his implications that tesla has a demand problem, and even if not, they have a supply problem (which they do to an extent), and if not, they won’t be able to stay solvent and/or the competition will catch up. the cap raise was a kick the can, munster pretty much alluded to this as well. gene, stay with apple. that’s who made you famous, and about all you could handle. you haven’t said anything new.

this bias is why mindless viewers stay uninformed. then melissa brings the weed back into the equation at the end of the clip

i mean jesus effing christ they are shameless trolls

what a hack organization
if sorkin and co were so damn smart, why are they crappy tv hosts? sorkin co-produces billions on showtime btw, so it’s no wonder he became, let’s use the phrase that’s become so annoyingly overused by everyone now, an “order of magnitude” douchier because he’s been in the HF circles and breaking bread with them more and more. clearly.

if you were to believe their self assurance portrayed when the camera was on was an indicator of their mastery, they’d be billionaires and not need to be on tv banging stocks around for their real masters...a55holes

how do people not see right through this BS? embarrassing, really.

I'm mean, realize that Sorkin writes in "Billions" 2 anti-Tesla bits.

1) Assistant AG is pouring over financials of some shady company, says something like "these finances don't make any sense, like Tesla"

2) Has the charismatic founder of a rocket company (!!!) blown up on his own rocket ship.

How the eff is this guy allowed to act impartial to Tesla?

 

[winfield100]

More better batteries!!
-----------snip---------------
Electric Vertical take off & landing drones/ambulances to cut response times

Drones Could Save 20,000 of Lives Lost to Out-of-Hospital Cardiac Arrests Annually

Each year in the US, roughly 350,000 cardiac arrests occur outside of hospitals but only 12% of the victims survive. One of the reasons the survival rate is so low is that ambulances arrive an average of eleven minutes after a call to 911. While eleven minutes is impressive, as shown below even shorter response times could save many lives.

Electric vertical take-off and landing drones (eVTOLs) could be life-saving time savers. Opener’s Blackfly, for example, could get a paramedic to the scene first, in just five minutes according to ARK’s research, saving as many as 20,000 lives annually, as shown below.
-----------------snip-----------------------

Source: ARK Investment Management LLC;
https://www.ahajournals.org/doi/10.1161/CIRCULATIONAHA.116.024400;

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6156551/

 

[ReflexFunds]

Karpathy and Musk comments on the (lack of) value of simulation indicate they aren't even thinking about RL. Maybe someday. People here vastly overestimate NN involvement in Tesla's driving policy. Based on Autonomy Day I see their NN giving path prediction hints, but heuristics constructing a 3D scene and making all trajectory decisions.

Even Waymo's advanced R&D ChaufferNet project didn't use RL, and they've put much more into simulation than Tesla.

Their comments were not denying the value of simulation. They simply said simulation only works if you know the problem and already have the data - It cannot be a substitute for real world driving experience, but it can help with improving your performance on known scenarios. I agree its definitely not necessary for their simulation to use RL, but it seems potentially well suited to it in the future.

I think it is very clear Tesla plans to move as much of its prediction and driving policy to NNs as possible. They have their current driving policy heuristics in place as a base infrastructure and will gradually replace parts of the code base as they work on improving performance. Even without comments like the following from Elon and the team, Tesla’s entire FSD strategy seems clearly geared towards this:

“The neural net is eating into the software base more and more. There are some things that are very simple for a heuristic and very difficult for a NN, so it makes sense to maintain some level of heuristics in the system because they are computationally easier. But over time It moves to just training against video. Video in, steering and acceleration out, almost entirely. That’s what we are going to use the Dojo system for.”​

I would guess at the moment Tesla’s 3D scene is almost entirely populated by Vision and radar neural nets, including NN distance and velocity labels. They are increasingly moving object future path prediction to NNs too. This includes predicting the future curve of the road and predicting future “cut ins” from NNs which have been trained to identify them before they happen. Heuristics still take care of driving policy, but if a NN tells Tesla there will be an object in its current planned path in 3 seconds, the heuristics to avoid this become much simpler.

I think the future project Dojo Elon mentions is where imitation learning plus RL would be well suited. I can see Tesla collecting data on the 50k/500k most common scenarios – likely each only 1 – 10 seconds long, and then use their data to train a NN to deal with each of these scenarios. One option is to use imitation learning to pre train a driver in simulation for each of these scenarios. They also feed their prediction NNs into the simulation to model the other cars/pedestrians. They then use reinforcement learning to the point where they are highly reliable with “video in, steering and acceleration out”. Now, in the real world, whenever the car identifies a scenario in its capability set, it temporarily hands over to the NN for that scenario. As they add more and more scenarios to the set, more and more of the driving policy gets done by NN. Obviously they are not doing this currently, and in the end it might not work, but I think this is broadly their current plan (whether or not it uses simulation & RL or other ML techniques).

 

[EVNow]

More better batteries!!
-----------snip---------------
Electric Vertical take off & landing drones/ambulances to cut response times

Drones Could Save 20,000 of Lives Lost to Out-of-Hospital Cardiac Arrests Annually

Each year in the US, roughly 350,000 cardiac arrests occur outside of hospitals but only 12% of the victims survive. One of the reasons the survival rate is so low is that ambulances arrive an average of eleven minutes after a call to 911. While eleven minutes is impressive, as shown below even shorter response times could save many lives.

Electric vertical take-off and landing drones (eVTOLs) could be life-saving time savers. Opener’s Blackfly, for example, could get a paramedic to the scene first, in just five minutes according to ARK’s research, saving as many as 20,000 lives annually, as shown below.
-----------------snip-----------------------

Source: ARK Investment Management LLC;
https://www.ahajournals.org/doi/10.1161/CIRCULATIONAHA.116.024400;

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6156551/

If people were to eat better, they won't have a cardiac arrest to start with.

 

[kbM3]

the whole demand issue as far as I can tell was a red herring put out by shorts. I have seen zero evidence anywhere that this is true other than for Model S/X refresh. The idea that Model 3 demand has all been met now and they need something new is ridiculous. We are not even at the first pitch in a nine inning game that plays out over decades. TSLA is peerless right now and unless someone does something incredible and amazing, that won't change anytime soon.

And this is with FSD not ever working. If FSD actually works we have a 10-bagger, if it doesn't and they execute the business its still a 3-5 bagger IMO. And BTW if their self driving gets to something like 99.99% perfect, which is not good enough for Level 5...this is still huge demand driver as it makes drivers life so much easier and insurance companies (<cough> also Tesla) risk profile much better.

So some level of autonomy from Level 3-5 is a big winner no matter what. I am personally hoping for the home run of FSD...and i have coded and even created novel neural nets and started a machine learning company...so I know a little about this field...but I confess to not having enough deep info on where that line is and what all the edge cases are and when they will handle most of them, etc etc to say "for sure it will work" or "for sure it won't work" or give any timeline.

they have a great FSD team and Musk is going to look like a total ass if he is wildly wrong...he has their latest stuff in his own car too. BUT...he drives in a very constrained area...so his experience with it is nowhere near adequate in sunny LA to say how good it is. But I note that he said by end of 2020 he feels very confident that "somewhere" it will have regulatory approval. That could be metro LA or some other relatively easy scenario.

1) Gene Munster, a long time Tesla bull, stated the order flow is dropping.
2) What, in your opinion, would constitute evidence of demand dropping? Evidence I would look for is pulling of demand levers,

 

[traxila]

And BTW if their self driving gets to something like 99.99% perfect, which is not good enough for Level 5...

Umm, how is 99.99% not good enough? What does that even mean? Have you seen the driving taking place on a daily basis? How many of the human drivers on the road today qualify for 99.99% perfect? Lol maybe .01%...

 

[EVNow]

https://floridapolitics.com/archives/294598-house-av-pass

Interesting bill. Could be the first state to allow FSD cars without drivers ?

 

[azaz]

I'm mean, realize that Sorkin writes in "Billions" 2 anti-Tesla bits.

1) Assistant AG is pouring over financials of some shady company, says something like "these finances don't make any sense, like Tesla"

2) Has the charismatic founder of a rocket company (!!!) blown up on his own rocket ship.

How the eff is this guy allowed to act impartial to Tesla?

If only this were actually true?

Sorkin has had zero involvement in Billions since the Pilot.

 

[azaz]

1) Gene Munster, a long time Tesla bull, stated the order flow is dropping.
2) What, in your opinion, would constitute evidence of demand dropping? Evidence I would look for is pulling of demand levers,

A lot of these big funds have proprietary methods such as that short who has a way of measuring test drives

 

[Doggydogworld]

also they had a 200M$ regulatory credit windfall, much much higher than usual per car.

The "windfall" was more like 150m, based on prior GHG credit prices. @ReflexFunds thinks it was partly due to higher GHG pricing in 2019, if true that part would recur each quarter. FCA/EU pooling revenue will also recur, perhaps in the 100m per quarter range by itself.

The total mix COGS for the model 3 looks to be about 45k$ - 46k$.

I agree. That's why base Model 3 is effectively 39,500, which should clear cash COGS.​