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

[Cattledog]

There is one thing that bothers me that I kind of agree with the bears. Hope the smart folks here in TMC can provide some light into it:

Musk sometime in January said, there is a possibility that Tesla might see a small profit in Q1. That statement is totally disconnected with what happened with a whopping loss much severe than even the bears were predicting.

Was Musk totally removed from happenings on the ground that he didn't see such a big loss coming? Or was he just making up his own alternate facts?

The market rightly punished the stock, not just for the huge loss, but the total lack of credibility in Musk's observations and numbers he throws out without giving it a lot of thought. I mean, if a CEO can't even see such a huge loss looming ahead in just two months, but instead suggests that they could be mildly profitable - if I were in institutional investor, I will be pissed off too.

It’s why, paradoxically, an SEC wrist slap was actually ok in the long run. Tighten up the loose financial comments, whatever their source.

 

[Alexander Grea]

There is one thing that bothers me that I kind of agree with the bears. Hope the smart folks here in TMC can provide some light into it:

Musk sometime in January said, there is a possibility that Tesla might see a small profit in Q1. That statement is totally disconnected with what happened with a whopping loss much severe than even the bears were predicting.

Was Musk totally removed from happenings on the ground that he didn't see such a big loss coming? Or was he just making up his own alternate facts?

The market rightly punished the stock, not just for the huge loss, but the total lack of credibility in Musk's observations and numbers he throws out without giving it a lot of thought. I mean, if a CEO can't even see such a huge loss looming ahead in just two months, but instead suggests that they could be mildly profitable - if I were in institutional investor, I will be pissed off too.

If I would try to think of a worst case then I would guess he wanted to have a chance of price going over 360 for the bond payment to pay it at least partially in equity to reduce the need for a cap raise.

It didn't work out so now both the credibility is damaged and TSLA couldn't pay the bonds partially with equity as well.

Well worth the gamble though. If T Rowe had not sold so massive quantities then the price had probably climbed over 360 by March bond payment.

 

[Todd Burch]

What is your basis for saying "probably"? There is risk with all new technology, but Tesla has a well established track record of bringing leading edge tech to the market in its products.

Maxwell has DEMONSTRATED 1.2x capacity in 15 Ahr sided pouch cells. JB Straubel is personally involved with overseeing battery tech. Tesla would not be buying Maxwell if this tech wasn't real.

As Elon said on the Q1 call, Tesla will hold a "battery day" some time later this year or next. Thats the current expectations of us super bulls. "Probably doesn't even work yet" is a TSLAQ fantasy and FUD talking point.

Yes. Us old timers know that Tesla is ONLY interested in working battery tech. They have said as much. To paraphrase in the (rough) words of Tom Cruise: “show me the cells!”

 

[EVNow]

Well worth the gamble though. If T Rowe had not sold so massive quantities then the price had probably climbed over 360 by March bond payment.

Well, SP fell from 250 to ~300 after they announced job cuts and reduced Q4 profits, Q1 loss. Don't know how much T Rowe had to do with it, the 11 Million they sold over months is probably not that big a deal compared to daily volume.

 

[nursebee]

Applicable...

 

[omgwtfbyobbq]

They only missed by 700M$. If they had sold another 60k model 3's they would have hit guidance. Anyone can make that mistake.

It's S/X production that really hurt. Another 14k S/X deliveries is likely around $420 million in profit. The $120 million they missed from the 12k 3s they built but didn't deliver was less than ideal too. The $188 million in one time costs from writing off S/X trade-ins didn't help either.

On the plus side, if they can keep 3 production up they might be able to build 70k+ cars and deliver 60k+, and the S/X lines coming back up with more automation should improve margins and production.

 

[Doggydogworld]

I don't think just moving the driver from the car seat to an office is really "driverless".

It's the kind of baby step regulators like to see.

Waymo's strategy is actually very robotics and heuristics heavy rather than machine learning.

I don't see that. Angeulov's talks are very similar to Karpathy's. NNs for object ID, gradually moving into policy.

...a true machine learning solution needs billions of miles rather than thousand/millions of miles of data and Waymo currently has no obvious way of gathering this level of data without $100bn+ R&D investment in cars, sensors and test drivers.

$100 billion??? Seriously? I can think of ways to source data for a nickel a mile, and they're smarter than me.

But do you need a billion miles? (As an aside, why doesn't the whole "humans don't have lasers shooting out of their eyes" meme not also apply to training mileage? Humans train on <1000 miles, why can't Musk's SuperChip do it in a million?)

As a Karpathy slide shows, NNs improve a lot when you start adding data but incremental returns quickly diminish. A failsafe (e.g. LIDAR) to handle edge cases you like airborne cars the NN doesn't understand sounds like a reasonable way to achieve six 9s vs. collecting thousands (millions?) of instances of airborne cars for NN training.​

Agree it isn't a corner case, disagree that you don't need millions of instances to feed into the NN.

If you need millions of instances of left turns, don't you also need millions of instances of airborne cars? And millions more of cars skidding on their roofs? Why not?

You've REALLY lost me here. The NN will obviously be the main calculation behind all driving decisions.

​

Not in 2020.

The StarCraft training was actually somewhat similar to autonomous driving. They had thousands of prior matches to feed the NN with, and ****** the player inputs ******* of those matches.

AlphaStar only used imitation learning to bootstrap their agent. This got it to a basic level, quite a bit worse than the expert gamers it was imitating. They then used reinforcement learning - setting up a league of 10 agents with slightly different approaches which played each other in a continuous tournament. It was similar to AlphaGo, which used zero real-world data and beat the best in the world after 4 hours of training.

if you have about 500 Tesla engineering fleet cars running in a quiet Phoenix burb all day, ...

Waymo reports Phoenix disengagements to the state of CA?

 

[EVNow]

It's S/X production that really hurt. Another 14k S/X deliveries is likely around $420 million in profit. The $120 million they missed from the 12k 3s they built but didn't deliver was less than ideal too. The $188 million in one time costs from writing off S/X trade-ins didn't help either.

Yes, S&X falling by 50% was the big reason for the huge loss. Q1 Model 3 Deliveries was actually close to Q3 M 3 deliveries (51k vs 53k). Ofcourse, there was a 5% drop in margin because of lower production.

 

[K3jSe7Llp7]

You do know I hope that only if you buy during a share emission, you actually help the company? If you buy from the market, you just made a more prudent investor's day ;-)
I used to be all in on something I believed in and still do. Lost everything because of it anyway. Retail investors rarely beat inflation.

1. I understand your first point but there are subtleties, including supporting the stock price, and part ownership of what I care about. Putting my wallet where my mouth is, adding my name to the list.
2. I handily beat not only inflation but the indexes. Right now, for example, much of my non-tesla holdings are pure play alt energy yieldcos which pay among the highest dividends available, are defensive utilities, and have appreciated by 22-40% ytd believe it or not, which obviously won't last. I've even made money on tesla, setting aside the current unrealized loss, by trading a small fraction.
3. Your 'more prudent investors' must be the ones who sell me their stock when the price is low and buy it back when the price is high?
4. I'm sorry you had some misfortune on the market. I sure haven't always won.

 

[EVNow]

As an aside, why doesn't the whole "humans don't have lasers shooting out of their eyes" meme not also apply to training mileage? Humans train on <1000 miles, why can't Musk's SuperChip do it in a million?

Humans have already trained for 15+ years on object recognition, movement characteristics of various objects, collision avoidance etc. before they start learning to drive.

ps : Some of that is probably evolution too.

 

[techmaven]

After months without issues, it was actually very sunny today and my 3 slammed on it's brakes twice while approaching overpasses. I'm supposed to believe in robotaxis next year when my 3 currently cannot drive under overpasses without issues and also has fatal attractions to concrete dividers?

This is a sensor fusion problem and this blog post from Tesla explains some of it:
Upgrading Autopilot: Seeing the World in Radar

The initial implementation of AEB basically recognizes the back of a vehicle and tries to auto-brake in certain situations. If it cannot "see" the back of a vehicle it has been trained to recognize, it won't trigger. But they changed that to allow radar to be an only trigger, but that also means more likelihood of a false positive. They are supposed to have geocoded white list, so maybe the map tile for your location got wiped of geocoded locations.

Of course, this isn't the same software as the FSD software. Not only is the recognition software different, so is the way it triggers. AEB has a nasty problem set of interations with a human driver that isn't present for FSD.

Same thing with concrete dividers... we have not yet experienced that software. Critical parts of the software may be missing that is all the difference between working quite well and seemingly a hot mess.

 

[Cattledog]

Cathie really strutting her stuff there, superb!

She only needs to be half right for us all to grin ear to ear.

“...the longer the base, the bigger the breakout...”

Buckle up.

 

[Skryll]

Yes, S&X falling by 50% was the big reason for the huge loss. Q1 Model 3 Deliveries was actually close to Q3 M 3 deliveries (51k vs 53k). Ofcourse, there was a 5% drop in margin because of lower production.

I think there is a successful FUD strategy playing out, they spread the word that S/X will have a refresh soon and so more people may have deferred until after that. Now I see facebook posts from people saying that the refresh is nice but the real refresh would be to the other battery type with faster charging and you should still wait.

Partially teslas fault for leaking raven in public firmware releases.

 

[EVNow]

Partially teslas fault for leaking raven in public firmware releases.

Ideally Tesla would have updated S&X when 3 was released. But they don't have the manpower to do a lot of things in parallel. This definitely affected S&X sales. Ofcourse there was the tax credit cliff too.

Now I'm wondering whether people really shifted from S&X to 3. In the 4 quarters of 2018, 3 had no effect on S&X deliveries. As 3 deliveries increased, so did S&X. 3 deliveries were all in US, ofcourse. Will there be more of a shift in EU/China ? We have to wait and see - but 3 didn't seem to Osbourne S&X in the US.

 

[MarcusMaximus]

It's the kind of baby step regulators like to see.

I don't see that. Angeulov's talks are very similar to Karpathy's. NNs for object ID, gradually moving into policy.

$100 billion??? Seriously? I can think of ways to source data for a nickel a mile, and they're smarter than me.

But do you need a billion miles? (As an aside, why doesn't the whole "humans don't have lasers shooting out of their eyes" meme not also apply to training mileage? Humans train on <1000 miles, why can't Musk's SuperChip do it in a million?)

As a Karpathy slide shows, NNs improve a lot when you start adding data but incremental returns quickly diminish. A failsafe (e.g. LIDAR) to handle edge cases you like airborne cars the NN doesn't understand sounds like a reasonable way to achieve six 9s vs. collecting thousands (millions?) of instances of airborne cars for NN training.​

If you need millions of instances of left turns, don't you also need millions of instances of airborne cars? And millions more of cars skidding on their roofs? Why not?

​

Not in 2020.

AlphaStar only used imitation learning to bootstrap their agent. This got it to a basic level, quite a bit worse than the expert gamers it was imitating. They then used reinforcement learning - setting up a league of 10 agents with slightly different approaches which played each other in a continuous tournament. It was similar to AlphaGo, which used zero real-world data and beat the best in the world after 4 hours of training.

Waymo reports Phoenix disengagements to the state of CA?

There’s a lot wrong here:

1. The type of data matters. Said data needs to be video from the exact type/arrangement of sensors you plan to use in the final product(or from such a vast number of types/arrangement that the model can generalize, with some capacity cost). I don’t see any cheap way to accomplish that without doing what Tesla is doing and selling the cars with the FSD equipment already there to end customers who eat the cost.

2. Yes, you need a LOT of miles. The human brain has two major legs up here(leaving aside aspects of our learning we just don’t understand yet):
a. We have baked-in priors, with our genetics already determining a lot of the network contents and structure to make it easy to learn
b. We have >16 years worth of general knowledge and experience of the world to build on. We have good instinctual knowledge of basic physics, human behavior, etc before we ever sit behind the steering wheel.

3. Your improvements in success rate drop off asymptotically approaching 100%(obviously), but your error reduction rate doesn’t necessarily.

4. LiDAR doesn’t gain you any additional understanding to handle things like airborne cars. I’m not sure why you think it would?

 

[scaesare]

Ditto except 122K miles.

I raise to 139K!

 

[AcesDealt]

It's S/X production that really hurt. Another 14k S/X deliveries is likely around $420 million in profit. The $120 million they missed from the 12k 3s they built but didn't deliver was less than ideal too. The $188 million in one time costs from writing off S/X trade-ins didn't help either.

On the plus side, if they can keep 3 production up they might be able to build 70k+ cars and deliver 60k+, and the S/X lines coming back up with more automation should improve margins and production.

I agree it was mostly S/X that did it. It's an important to notice because most people will be focusing on model 3 which may have a moderately good sales ramp this year, but because it will also come with a decrease in per unit profitability, will not have as much of an effect on the income statement as simply getting S/X back to where it was before (if that's possible).

 

[3Victoria]

There’s a lot wrong here:

1. The type of data matters. Said data needs to be video from the exact type/arrangement of sensors you plan to use in the final product(or from such a vast number of types/arrangement that the model can generalize, with some capacity cost). I don’t see any cheap way to accomplish that without doing what Tesla is doing and selling the cars with the FSD equipment already there to end customers who eat the cost.

2. Yes, you need a LOT of miles. The human brain has two major legs up here(leaving aside aspects of our learning we just don’t understand yet):
a. We have baked-in priors, with our genetics already determining a lot of the network contents and structure to make it easy to learn
b. We have >16 years worth of general knowledge and experience of the world to build on. We have good instinctual knowledge of basic physics, human behavior, etc before we ever sit behind the steering wheel.

3. Your improvements in success rate drop off asymptotically approaching 100%(obviously), but your error reduction rate doesn’t necessarily.

4. LiDAR doesn’t gain you any additional understanding to handle things like airborne cars. I’m not sure why you think it would?

Too many sensors is like a committee with too many members, no good progress, and the final recommendation is a camel instead of a horse.

Lidar is a point cloud, one still has to interpret it. We think it is easy because we have an incredibly evolved visual cortex. Try interpreting the raw data that hasn't been transformed into a 'picture'.

 

[HG Wells]

It is a huge step in the wrong direction! Unless you think everywhere meaningful will be geofenced.

My real estate agent says she can get 10K more for a house with schools in the same geofence compared to a house just outside the fence.

 

[Pras]

$100 billion??? Seriously? I can think of ways to source data for a nickel a mile, and they're smarter than me.

A failsafe (e.g. LIDAR) to handle edge cases you like airborne cars the NN doesn't understand sounds like a reasonable way to achieve six 9s vs. collecting thousands (millions?) of instances of airborne cars for NN training.​

?

Sourcing any data (say phone or plug in telemetry recording or anything) vs. sourcing data from calibrated 8 cameras, mounted exactly the same on all similar cars, with additional data from all ultrasonic seniors and radar is different. Quality difference could be in “order of magnitude”.

Lidar is not a failsafe system, it is just a more accurate 3D mapping sensor. The car still need to decide how to anticipate the surrounding and do the driving. Lidar can’t defy laws of physics or predict human behavior, for that you need strong AI. Lidar is a redundant system, if not now maybe soon in future.