Pras
Member
Don’t forget that they said that with design cost included the new chip is as expensive as the old nvidia. So cost saving is a big win. I mean give Elon the credit to dream and see this in 2015/16, to hire a team and get it done.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable
- Thread starter AudubonB
- Start date
I study computer science... and the thing people need to realize is that you need software to be optimized for hardware and vice versa. I hate Apple as a company, cuz they aren’t innovative, but it is a massive advantage to have hardware and software optimized. Reason is because if you have the pair, you can know what are the limiting factors and optimize around that. Could be ALUs, Cache, Instruction fetch decode execute, everything. That’s why it’s impossible for NVIDIAs shiz to work. Hardware wise no doubt they will match Tesla in two years. But it’s literally worthless without the software and neural net.
It falls right in line with his thinking all along. If a supplier, in this case Mobileye, either can't or won't produce what Tesla needs then they will do it in-house.
I actually couldn’t see them doing a chip specially for CNN inference.
Intel or Google or even Microsoft could be closer to doing that, but mainly for data centers, not for FSD cars.
NV’s strength is in training for now, but the Dojo project Elon mentioned today would probably take care of that too.
Ok I'll try to summarize what I could get out of this presentation with an attempt to make it somewhat palatable for non-technical folks.
1. What does "feature complete" mean: they've identified all/most of the "classifiers/predictors" (neural networks trained to solve specific problems) that are needed for FSD, and have an algo to make use of those classifiers to be able to autonomously drive a car given that classifiers are mostly classifying correctly. Of course that algo is going to be improved as well but they're confident it's good enough to drive, again given correct input from the machine learning systems that are supposed to both understand the environment and make predictions about its behavior.
2. The degree with which classifiers/predictors are correct determine how safe and efficient the overall system is. Once they have the right kinds of classifiers/predictors identified that would be necessary to solve FSD, they now are hard at work improving their accuracy. For that they have figured out a way to feed both successful human driving and unsuccessful Autopilot driving attempts as a feedback loop to re-train and make classifiers better. They also have a way to ask the fleet to send back real-world info about some very specific "interesting" objects/events/situations they need more data on to improve accuracy. This is key, most machine learning is about having labeled training datasets that paint a representative picture of what's out there, and Tesla is collecting such data at a rapidly accelerating pace (see #3)
3. This feedback loop will get quicker and quicker with more cars driving out there with AP shadowing and phoning data back that will be used to re-train their models.
4. Hardware problem is solved, they got a computer that is cheap and energy efficient enough to run all the inference and traditional algos from #1.
So basically it's now a matter of time to dot the i's and cross the t's, fundamentally a path to FSD have been found and all the critical pieces are already in place and proven to work to make it happen.
1. What does "feature complete" mean: they've identified all/most of the "classifiers/predictors" (neural networks trained to solve specific problems) that are needed for FSD, and have an algo to make use of those classifiers to be able to autonomously drive a car given that classifiers are mostly classifying correctly. Of course that algo is going to be improved as well but they're confident it's good enough to drive, again given correct input from the machine learning systems that are supposed to both understand the environment and make predictions about its behavior.
2. The degree with which classifiers/predictors are correct determine how safe and efficient the overall system is. Once they have the right kinds of classifiers/predictors identified that would be necessary to solve FSD, they now are hard at work improving their accuracy. For that they have figured out a way to feed both successful human driving and unsuccessful Autopilot driving attempts as a feedback loop to re-train and make classifiers better. They also have a way to ask the fleet to send back real-world info about some very specific "interesting" objects/events/situations they need more data on to improve accuracy. This is key, most machine learning is about having labeled training datasets that paint a representative picture of what's out there, and Tesla is collecting such data at a rapidly accelerating pace (see #3)
3. This feedback loop will get quicker and quicker with more cars driving out there with AP shadowing and phoning data back that will be used to re-train their models.
4. Hardware problem is solved, they got a computer that is cheap and energy efficient enough to run all the inference and traditional algos from #1.
So basically it's now a matter of time to dot the i's and cross the t's, fundamentally a path to FSD have been found and all the critical pieces are already in place and proven to work to make it happen.
I don't get it... NVIDIA's entire advantage right now is massively parallel floating point multiplication logic unit machines. That's what a GPU is. Why is Intel better?
While it does continue to improve with every new version, it was able to handle sharp curves as of June, 2018(I remember because the update went wide like 1 day after I left on a 3 week vacation). And staying in lane lines was pretty well solved in 2018.10.4 in March, 2018.
All this said, it’s a little bizarre to cite the time when they kicked off this particular round of training of their network as the point where they began development of the system. While we’re doing that, might as well say they’ve only been working on it for 1 day, since they’re very likely running a training round *right now*.
It still doesn't handle sharp curves on some roads. At all. It seems to be highly variable. Freeway onramps, it handles, but on many minor roads, it ignores the turn and keeps going straight, right across the double yellow line. My guess is that there's some special map data tagging, without which it fails miserably.
Singer3000
Member
Enjoyed the presentation. The vision is compelling and there at least seems to be a credible pathway to getting there. However as others have said, the market now doesn't believe Musk at all on his timelines.
I suspect the stock price is going to be far more fixated in the coming months with something very mundane in comparison:
BATTERIES!
Shifting demand towards SR+ because they still can't get enough cells. Vague timeline on 1GW project for Tesla Energy, and using cells made outside of the GF at that. The silver lining is that if Tesla, even after all the investment in GF1 with Panasonic, cannot get batteries fast enough, then the prospects for other auto companies looks dire. But overall quite bad for the speed of the EV revolution, as well as Tesla's cashflow. Did I mishear Musk say that basically all the R&D budget was now for FSD? Even if the trajectory of that project is a dead cert winner, they still need $$$ to get to the finish line.
I suspect the stock price is going to be far more fixated in the coming months with something very mundane in comparison:
BATTERIES!
Shifting demand towards SR+ because they still can't get enough cells. Vague timeline on 1GW project for Tesla Energy, and using cells made outside of the GF at that. The silver lining is that if Tesla, even after all the investment in GF1 with Panasonic, cannot get batteries fast enough, then the prospects for other auto companies looks dire. But overall quite bad for the speed of the EV revolution, as well as Tesla's cashflow. Did I mishear Musk say that basically all the R&D budget was now for FSD? Even if the trajectory of that project is a dead cert winner, they still need $$$ to get to the finish line.
I’m talking about exclusively freeways, since that’s where it’s primarily intended to be used in the current customer version.
I have mixed feelings about the presentation last night.
- Elon looked a bit nervous and sounded a bit desperate, like he was trying too hard to convince everyone how amazing FSD is going to be. This makes me worry that either demand has fallen off a cliff or there is a big battery bottleneck
- The presentation was really technical and geeky. I am not sure why they bothered with all that. Its loads of ammunition for half-wits to pick holes in it based on misunderstandings. A simple presentation of the car driving itself in all kinds of environments and Elon and the autopilot guys explaining how certain moves are possible would have been more spectacular and helped people believe what was coming
- Tesla Network sounds cool. Why not create it now but with drivers? Concentrate on a few cities with high density of teslas and then phase out the drivers when the time is right. That way the Tesla Network gets brand visibility and starts to grow a habit. Most people would choose a Tesla taxi drive over a regular Uber any day
Personally, I want to buy 10 Tesla Model 3s and get them set up and working (with drivers for now) in a city like Madrid. For a relatively small outlay, there is a massive opportunity to take over the city with robo-taxis. I would love to do that
- Elon looked a bit nervous and sounded a bit desperate, like he was trying too hard to convince everyone how amazing FSD is going to be. This makes me worry that either demand has fallen off a cliff or there is a big battery bottleneck
- The presentation was really technical and geeky. I am not sure why they bothered with all that. Its loads of ammunition for half-wits to pick holes in it based on misunderstandings. A simple presentation of the car driving itself in all kinds of environments and Elon and the autopilot guys explaining how certain moves are possible would have been more spectacular and helped people believe what was coming
- Tesla Network sounds cool. Why not create it now but with drivers? Concentrate on a few cities with high density of teslas and then phase out the drivers when the time is right. That way the Tesla Network gets brand visibility and starts to grow a habit. Most people would choose a Tesla taxi drive over a regular Uber any day
Personally, I want to buy 10 Tesla Model 3s and get them set up and working (with drivers for now) in a city like Madrid. For a relatively small outlay, there is a massive opportunity to take over the city with robo-taxis. I would love to do that
Pezpunk
Active Member
is this the first Tesla event you've watched? he literally always comes off that way. he's a stutterer, with a lot of nervous ticks.
I've watched every presentation from the past 3 years, ever since the Model 3 presentation. I could be wrong of course, and I hope I am!
Pras
Member
Not sure if you noticed, the cost of car with a “million mile life” battery is expected to be <38k. Assuming current batteries are cheaper, reasonable take rate for fsd, and a modestly higher option choices (color and wheels), they will make 20% margin on the SR+. If indeed FSD is real in the next year or two, all talk about margin will be useless. At that point Tesla should only sell high end car and use owned SR+ for taxis.
In terms of EV revolution, they still aspire to put 10 million cars in 10 years. If the car utility increases by 5 times, you are creating more impact.
Pras
Member
On the contrary, he was less nervous/more comfortable than the Y event. I think the difference was he let his team do the talking for most part, which is more reassuring to me as an investor (knowing his over optimistic views).
Singer3000
Member
Sure. But what's the use of 20% margins if you can only make 75k units per quarter. Am hoping for some really confident and clear explanations on how and when they can resolve the apparent battery bottleneck on Wed.
Causalien
Prime 8 ball Oracle
Ya their body language is mostly combative. The short attacks are probably getting to them. I think part of the fault is the early announcement that Q1 won't be cash flow positive. Essentially giving them about 6 months of time to attack based on that narrative and another 3 months until Q2 ER gets released.
Few cents, but mostly everything has been already said:
Now, if Elon could just focus on Fremont, GF1 and GF2, we would be very happy investors ;-)
- Jobs vs Elon: Elon can't do drama. He could have showed the demo video first, then let Karpathy speak, then hardware at the end. But he's Elon, so boring but crucial details first. Tesla should hire a PR chief that understands how Elon's mind work, and able to convince him on more impactful way of presenting/disclosing details. But it's a minor problem, for now.
- Hardware is a competitive advantage, software is a competitive advantage, fleet is single-point-of-success. They worked years to arrive here, with a fleet of sensors, one of the best talented team in the world (both hardware and software), with no legacy ICE problems. Cheers to Tesla: I'm no expert but it's evident to me that this is the way to achieve level 5. Incrementally, adding layers of complexity.
- The way they did it is "pure Elon": start from first principles, tackle the huge-but-single problem first. That problem is vision, because that's the way humans drive. That's very hard and very bold. Nobody had the guts or the means to do that. At Tesla, they knew they could eventually do it: they needed data (hence the HW2, few years ago), then enough power (Nvidia GPU, retrofittable), then good software (hire Karpathy and others). Then iterate, iterate, iterate. To me this seems the smartest way to do software, probably years ahead of others OEMs. Again, this is Elon MO: if it's theoretically doable, then we'll eventually get there.
- For me, this is probably Elon's best shuperhuman quality. He's cognitively fearless. Fearless. For him, if it's learnable, he will learn it. I don't have to tell you that most people don't work this way.
- Engineers at Waymo and OEMs do know this was the right approach. They did now at the time, but bowed before executives decisions of LIDAR, etc. They are probably wondering what it's like to work in a place where the boss knows all the details about ASIC and deep neural networks, and it's bold enough to build the all strategy with software in mind.
- EDIT: the idea of selling the hardware first and letting customers buy the software later is pure genius, and a learned lesson from SV. How many automakers do that? This will actually better margins of *cars already on the street*.
Now, if Elon could just focus on Fremont, GF1 and GF2, we would be very happy investors ;-)
Last edited:
Dan Detweiler
Active Member
My completely ignorant take of all this.
Presentation was overwhelming...but needed I think to show potential investors of the extent to which they have gone to reach their conclusions and instill confidence in the direction they are going.
FSD video was amazing but I'm sure will be slowed down and picked apart by the panick stricken TSLAQ crowd.
Stock price impact? Short term, the "explosion" video will have more impact because it requires zero intellect to understand and is completely reactionary. If you are a short term bull I don't think you are going to see much, if any relief. If you are a long term bull you feel a LOT better about your position and perhaps double down with more stick. If you are a bear.. especially a radical unreasonable bear (which seems to be more and more of the short population) you start flooding the internet with every conceivable negative you can come up with, including false stories, doctored videos, and FUD of all kinds in an effort to prolong the inevitable.
I see another 6 months or so of depressed stock price before the inevitable begins to take root. A weak earnings report this week won't help on that front either. The bears will be jumping all over that.
In the end...Tesla, just keep doing what you're doing. Truly amazing stuff.
Dan
Presentation was overwhelming...but needed I think to show potential investors of the extent to which they have gone to reach their conclusions and instill confidence in the direction they are going.
FSD video was amazing but I'm sure will be slowed down and picked apart by the panick stricken TSLAQ crowd.
Stock price impact? Short term, the "explosion" video will have more impact because it requires zero intellect to understand and is completely reactionary. If you are a short term bull I don't think you are going to see much, if any relief. If you are a long term bull you feel a LOT better about your position and perhaps double down with more stick. If you are a bear.. especially a radical unreasonable bear (which seems to be more and more of the short population) you start flooding the internet with every conceivable negative you can come up with, including false stories, doctored videos, and FUD of all kinds in an effort to prolong the inevitable.
I see another 6 months or so of depressed stock price before the inevitable begins to take root. A weak earnings report this week won't help on that front either. The bears will be jumping all over that.
In the end...Tesla, just keep doing what you're doing. Truly amazing stuff.
Dan
Similar threads
- Locked
- Locked
- Locked
