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

[adiggs]

Urm... HW 3 being more powerful helps with computer vision, but it isn't clear it really alleviates many of the problems.

"basically can drive on highway now" -- I did this a couple of weeks ago. Exciting for beta, monumentally unready for prime time. Basic things like not being able to handle many/most entrance/exits (veers right thinking it is staying in the middle of the lane), or steadily maintaining center lane despite being on the inside of a curve next to a semi (rational drivers will drift away from the semi, even if it means hitting the lane boundary).

I love it. It needs more than new hardware.

If you think "traffic lights and stop signs" are all that needs to be handled... wow: no need to listen for other vehicle horns? No need to deal with flagmen in construction zones? No need to properly handle traffic that doesn't follow expected lane markings (common for construction zones) or have lane markings (common following construction) or has conflicting lane markings (common for construction zones). No need to handle emergency vehicles (like the officer pulling over for a tail light being out, brake lights not working, etc.)

If FSD is officially released this year that will be very bad for Tesla, it is so far from ready that it isn't funny.

(And Tesla appears to be far ahead of the pack... no one is close)

I'm with @humbaba on the FSD thing:
1. What's the probability that Tesla will release full self-driving within a year or so?
-- we may be operating from different definitions of what "full self-driving" actually means. I take it to mean "come when summoned, drop me off at my destination, park yourself, figure everything else out on your own" (where that's written as if I'm talking to the car).

Given that definition, not in the next year or 2. As a buyable feature available to a consumer (such as myself), I expect to be waiting at least 5 years (so 2024 ish).

I also expect to see better and better "lab" instances of more and more situations being handled over these 5 years. I expect to see the driver assist to get outstandingly good over these 5 years.

I wouldn't be surprised if my Model X learns how to stop and start at stop signs/lights, make turns, follow navigation, and basically drive me Supercharger to Supercharger with me handling the parking lot, in the next few years. And me, the driver, being available to intervene in the driving with at least a 5-10s warning. So a car that will be responsible for the driving while on an interstate - I can see that. But that isn't FSD.


2. If they do release this feature, how good will it be? can it handle 99% of situations?
-- handling 99% of situations is abysmally bad. Handling 99% of miles is actually still pretty bad. For full self driving, the system will need to handle about 99,999 miles out of 100,000. Situationally though, 99.9999% of situations might still be be way too low. If you're driving in downtown San Fransisco, how many "situations" per block do you need to handle? It'll be more than 1 per mile. A FSD vehicle needs to drive around in downtown SF day after day, week after week, roughly 24 hours a day and never encounter a situation it can't handle.

And do it at speeds close to what a human can handle. If the car is driving 10 mph in an area that humans are driving 25, it's a hazard to traffic and that's not ready.


3. If this general AI based self-driving actually works, how will that affect the company's valuation?
-- Also with @humbaba, I see Tesla at least as far along and probably further. I understand that the theme from this year's CES is this tech as driver assist . Get something useful into cars now (as an industry), and use that to learn and make the systems better.

My valuation of the company doesn't need Tesla to win AI / FSD for the company to increase a lot.


My core valuation is quickly shifting to good old fashion EPS. With ~170M shares in circulation, every $170M in earnings is $1 EPS. If third quarter EPS of $3 (I round from $2.90) carries forward for 3 more quarters, then we own today a company that is producing $12/year in EPS. The closest comparison of size and growth rate I know of is Amazon, who gets better than a 100PE on much lower growth.

So my valuation today is $1200 (ok - in ~6 months). I only need for Q4/1/2 to average out to being as good as Q3 to get to $12 EPS. I think that's misleadingly conservative, and $15-$25 EPS is more likely - now I'm at a $1500-$2500 share price in ~6 months.

The company has so little stock in circulation, it only needs to generate $170M in earnings for each $1 EPS. This is a company that when it finally swung to positive earnings, it went from negative the previous quarter all the way to $3 in it's "first" (recent) profitable quarter.

The operating leverage is huge and not correctly valued yet - the recent (say last 3 years) trading history is being overvalued and the operating leverage / profitability is undervalued. (MHO). We don't need FSD for the company to be revalued bigly.

 

[humbaba]

Sorry, my riff was not really aimed at your post, just joking around.

If this guy can point you to public sources that can be evaluated, we can look into this. I have no idea what sort of data he has or how he is analyzing it. I am a statistician myself and can tell you that there many ways to draw faulty conclusions from data, especially if you are looking with some kind of biased agenda. So even if he has access to good data, there is no what to know that valid analysis has been performed. This is why we prefer publish work. The data and analysis needs to be open to scrutiny from those competent to review such work. Comparing one brand against another is problematic because there can be many other variables, from driver behavior to geographical differences that can cause differences in rates that have no real bearing on the underlying safety of the vehicle itself. So very good care must be taking in evaluating even basic statistics.

No apologies needed, you had -- I thought -- quite a good post.

And you make an excellent point about drawing faulty conclusions. I knew better and had failed to account when thinking about this case.

IMO, Tesla's published data puts other makers on notice, even if it is rather limited. AFAIK no one publishes accident (much less including accident-like) data on their vehicles. I'd love to see another manufacturer do the same.

 

[Lessmog]

What is Netflix's moat?

Uhh ... letter box?

 

[kyzhk]

They use a PTC air heater for the cabin, and an immersed electrical coolant heater for the battery. (In the S&X, the Model 3 uses the motor/inverter ran in an inefficient state to create heat for the battery.)

The problem is that in a Tesla you may need to provide heat to the cabin and cooling to the motor at the same time. (or vice-versa.) Not to mention that the AC is normally run with defrost and heat in the winter to dehumidify the air.

It's a solved problem in commercial heat pump application with variable refrigerant flow + valves. They can actually reuse heat from server rooms (handler in cooling mode) to heat other rooms connected to the same system, while still sourcing additional heat from outside world for supplement.

I suppose such system may not be able to be scaled down effectively to be put on car, but I wonder what GM bolt do? One poster mentioned bolt only uses a heat pump.

Market action comment:
Today's drop is quite peculiar to me as I would have thought some random price target update and advice from the "reputable" firms no longer carry much impact. There's really no substantial news today, well, I guess just tesla being tesla.

 

[adiggs]

If what the leak reported is true, that a typical commuting route will be mastered by HW3 FSD with no disengagements whatsoever, that will be a game changer.

....

This point I agree with. This might even be a corner case of FSD that we'll see get deployed as a buyable feature - the car watches me drive back and forth to work repetitively, and at some point it offers to make that drive on my behalf (ending up approximately in the parking lot where I normally end up, handling traffic and non-standard situations along the way).

And I agree that'll be a game changer. If I can turn my daily commute into a "get in the car, tell it to navigate to work, and then go back to sleep / fire up the computer / read a book / ignore the outside world" situation, then I've already got the house in the country that I'm ready to move to .

More broadly in the market, I can see that being a "must buy that car now" feature for a LOT of people.

 

[TradingInvest]

What limits do you mean? It appears the FSD team is testing it currently with no geofencing hacks. It's Andrej Karpathy's new unified vision neural network doing its magic on 10x-20x faster hardware.

Here's what Tesla's AI director Andrej Karpathy said about the new FSD neural networks three months ago, on the Oct 23 earnings conference call:

"Hi, everyone. My name is Andrej Karpathy. I'm the director of AI here at Tesla. And my team trains all of the neural networks that analyze the images streaming in from all the cameras for the Autopilot. For example, these neural networks identify cars, lane lines, traffic signs and so on. The team is incredibly excited about the upcoming upgrade for the Autopilot computer which Pete briefly talked about.
So in other words, we are currently at a place where we trained large neural networks that work very well, but we are not able to deploy them to the fleet due to computational constraints. So, all of this will change with the next iteration of the hardware. And it's a massive step improvement in the compute capability. And the team is incredibly excited to get these networks out there."​

So when Karpathy said that "we trained large neural networks that work very well", he was very robustly informed: they can both run the new network in shadow mode on AP 2.5 at a lower frame-rate, and they can also re-run captured video and sensor feeds on supercomputers simulating HW3.

The leak confirms that FSD is indeed working well on deployed HW3 as well.

I.e. it's the highest grade FSD solution "soon" on the market, including Waymo's.

Tesla's FSD is coming and it's the real solution. Waymo is on the wrong track, unless they have a separate program that I don't know.

 

[Lessmog]

The "threat" of more electric competition should not be seen as a risk to Tesla. It is a validation. That proposed VW plant won't be stealing sales from Tesla, it will be stealing ICE sales. If anything the investment into EVs by legacy makers should boost TSLA shares. It's not like VW will magically have better/more profitable EVs than Tesla. Best case for them is reasonable parity.

edit, no idea how I misspelled "VW" twice in the same post in two different ways.

I think the issue is more that, if the legacy members are seen as competitors to Tesla, then if they are perceived to be, for example, half as good, doesn't it therefore follow that Tesla's P/E should be twice theirs.

Congratulations, oh bird! This may be your very best post I have seen here so far. If only the rest of the wing-ed and other wild creatures would heed your sound advice ...

 

[Curt Renz]

Phil LeBeau CNBC Video - 10 minutes ago: Tesla competition heats up as big autos roll out electric car plans

 

[kyzhk]

Phil LeBeau CNBC Video - 10 minutes ago: Tesla competition heats up as big autos roll out electric car plans

Good progress, they are just about 10 years late, with couple more years to wait for their then outdated EVs tesla started selling couple years ago.

 

[anthonyj]

Phil LeBeau CNBC Video - 10 minutes ago: Tesla competition heats up as big autos roll out electric car plans

They’ve been saying this for ten years. They will never get serious about EVs.

 

[humbaba]

If what the leak reported is true, that a typical commuting route will be mastered by HW3 FSD with no disengagements whatsoever, that will be a game changer.

Remember how quickly Tesla HW2.5 Autopilot caught up to the MobilEye neural network ASIC solution and has no by most reports surpassed it in the spring of 2018 - using commodity hardware. With their own ASIC they'll be able to run much larger neural networks, which will be better for the existing functionality and smarter (new functionality).

Neural networks are improving a bit like the singularity, along an exponential path. For decades chess computers had no chance beating reasonably good human players - and then "suddenly", within a couple of years, the world chess champion was beaten by a computer chess player.

Automotive FSD will be very similar IMO.

Well, we fundamentally disagree about neural networks here. I wrote a longer piece about chess being a non sequitor, but I'll leave that out.

I will note that while there have been many valuable things enabled via neural networks (expert systems, etc.) they have yet to live up to the hype. People have been expecting too much of them for decades and thus constantly disappointed -- despite some really cool and useful things coming out of it.

IOW I agree that Tesla is making great progress, but disagree about how close it is to the goal, or even the relative rate of improvement. The saying that "the last 20% takes 80% of the work" doesn't even cover it when comes to FSD.

But, hey, maybe Tesla will release FSD in 2019 and it really will be ready. I'll be glad if so, but to be honest I think 2025 is very optimistic for full FSD. An FSD that is limited to certain summon situations or Interstate only driving is much more achievable and correspondingly likelier to happen earlier.

Summary: I think we basically agree on everything except timeline. Here's hoping you're right

 

[MP3Mike]

I suppose such system may not be able to be scaled down effectively to be put on car, but I wonder what GM bolt do? One poster mentioned bolt only uses a heat pump.

The Bolt uses three completely separate heating/cooling loops:

1. Immersed electric heater element to heater core in cabin for heating the cabin.
2. Cooling loop for the electronics
3. Cooling loop for the battery (with an Immersed heater)

Then they use an A/C pump with a heat exchanger to cool the battery, it isn't a heat pump in the sense that it can be reversed.

So in other words very much like Tesla, other than they heat glycol to heat the cabin instead of heating the air directly. And they have three coolant surge tanks instead of one.

 

[TradingInvest]

If what the leak reported is true, that a typical commuting route will be mastered by HW3 FSD with no disengagements whatsoever, that will be a game changer.

Remember how quickly Tesla HW2.5 Autopilot caught up to the MobilEye neural network ASIC solution and has no by most reports surpassed it in the spring of 2018 - using commodity hardware. With their own ASIC they'll be able to run much larger neural networks, which will be better for the existing functionality and smarter (new functionality).

Neural networks are improving a bit like the singularity, along an exponential path. For decades chess computers had no chance beating reasonably good human players - and then "suddenly", within a couple of years, the world chess champion was beaten by a computer chess player. Today you can beat the world chess champion with a chess program running on a laptop.

Automotive FSD evolution will be very similarly exponential IMO.

But yeah, I agree that Tesla should approach this carefully and systematically - and I think they are doing exactly that.

For example I'm pretty certain Tesla could have created a coast-to-coast FSD demo with zero disengagements in Q3 2018 already, using their HW3 prototype boards plus GPS geo-tagged adjustments like Waymo uses.

They didn't, because they want to release a robust solution with generic FSD.

That is the plan, most owners should be able to go to work and back to home through FSD using HW3. I hope Tesla don't do coast-to-coast demo yet, give us time to accumulate more shares.

 

[EVNow]

uh, no.

I've no idea what's going to happen - tomorrow's Brexit vote can take NASDAQ down a percent or two. That can push Tesla another $15 down. At that point who is betting on not hitting $300 ?

 

[Right_Said_Fred]

I've no idea what's going to happen - tomorrow's Brexit vote can take NASDAQ down a percent or two. That can push Tesla another $15 down. At that point who is betting on not hitting $300 ?

The outcome of the vote is already known. Don't expect it to move (US) markets much.

 

[TradingInvest]

OT

What is Netflix's moat?

If Netflix and Disney switch CEOs, Disney would become the new Netflix, and Netflix would lose a lot of subscribers.

Disney naturally has kids movies, on top of that, they can create whatever content they need. The CEO/founder can make a big difference. I would say Netflix's moat is it's CEO.

 

[EVNow]

The outcome of the vote is already known. Don't expect it to move (US) markets much.

Kind of. Depends on how many Tory rebels are there, will there be a no confidence vote that May loses etc.

 

[EVNow]

OT



If Netflix and Disney switch CEOs, Disney would become the new Netflix, and Netflix would lose a lot of subscribers.

Disney naturally has kids movies, on top of that, they can create whatever content they need. The CEO/founder can make a big difference. I would say Netflix's moat is it's CEO.

Entertainment is tough. Esp for Netflix where a hit doesn't bring in that much marginal extra revenue. Not sure Netflix has anywhere near the growth potential of Tesla. Tesla is at a stage where Blockbuster was wondering whether to open an online store.

 

[M3Rider]

I gave a variety of issues in another post, but a common theme is construction. Any long Interstate trip will almost certainly go through construction or encounter accident situations.

I expect the 10 fold improvement of HW3 to significantly improve computer vision, but I also seriously doubt it will be sufficient. A rock in the wrong place is currently invisible and I don't see HW3 being enough improvement to change that. Its supposed to handle four times the resolution at higher rate -- it will no doubt be better, but there is just so much gap. For driver assistance systems it is looking sharp and getting better. For even interstate limited FSD... it has a long ways to go.

[edit to add: no doubt the solution farthest along. Waymo is (currently) a joke, as is GM Cruise. When you are limited to a small number of fixed routes in a small area with a small number of destinations and still can't achieve full autonomy something is wrong.]

Keep in mind that cars with AP only will not get the upgraded HW3.

I think with govt regulations it may take a few years still. Though, I'm excited by the agile incremental improvements and they are learning through the software development iteration about hardware changes and new requirements to support FSD. The fact that they are training the neural network to learn the stop signs, etc. it's all in the right step. Lately, Elon has been delivering on his promises on the EAP + FSD and over the past few years I've noticed the improvements in my AP2.0 Tesla...phenomenal!

I barely drive on the highway (and use AP), but the few times I tried I had some issues:
- approaching a very sharp turn at the highway interchange(nav on AP) at 70mph and not reducing speed. This scared me, so had to take over and drop the speed.
- once I dropped the speed, turned AP on again and once inside the exit ramp the car veered to the outside of the turn and the left front wheel crossed the white line. There was a concrete barrier within a few feet of the line and this really scared my wife, so I took over. This was the most uncomfortable experience, especially considering what I had to hear afterwards.
- driving in the right lane an the highway, it so happened that in one spot the white line on the right disappeared(next to a rock wall) and the lane became twice as wide for a while. I believe AP didn't know what to do, I heard the "AP off" sound and it started slowing, so I had to take over. There was some message on the screen I didn't have time to read.
- another time I was in a right lane(multi-lane highway interchange), the lane was ending and there were arrows painted on the lane indicating "merge left". The car was going 70mph and not giving me any "confirm lane change to follow route". My lane was about to run out and with some cars in the lane next to me, the time to avoid an accident was quickly disappearing. I took over.

I don't know if AP would be able to resolve any of those scenarios when I took over, but I felt like it was not 100% ready.

Had another mishap with the summon when demo-ing the car to coworker. I summoned car backwards out of the regular parking spot in the garage halfway, then switched to "forward". With the car being perfectly in the middle of the parking spot between the white lines, it started turning the wheels before going forward, so once it went forward, it crossed the white line on the left. This was a bit embarrassing.

This all being a precursor to FSD, it feels like we are not ready for a prime time even with the existing stuff. Curious if anyone had spotless experiences with AP? Especially w/ nav on AP and less than straightforward interchanges?

 

[ggr]

Mod note: The very fine thread about Q3/Q4 financial projections has been renamed Near-future quarterly financial projections and made sticky, in case you were looking for it. --ggr.