You are forgetting the first mover advantage and the IP portfolio they might carry. Elon might volunteer the battery/EV tech for faster adaptation but FSD may be not.
-
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
You are forgetting the first mover advantage and the IP portfolio they might carry. Elon might volunteer the battery/EV tech for faster adaptation but FSD may be not.
That’s what I thought, hence Elon’s snide remark that the Tesla FSD computer fits behind the glove box, and doesn’t take up half your trunk.
Hopefully, there will be some better tech pieces out in the next few days showing just how far ahead Tesla is in NN car compute and FSD than anyone else.
The NVIDIA Drive PX Pegasus reportedly does 320 TOPS, with TDP at 500 watts.
My impression is that they don’t hardcode the network architecture into the hardware... the hardware just provides blocks that are used in an order and with weights determined by the compiled network. So long a v2 still has the blocks to process v1’s network(with maybe some new ones and efficiency improvements to those old ones), it’d be backwards compatible.
mblakele
FSD Beta (99)
I think not. There's a training system that does the learning. That runs offline, likely "in the cloud". That builds the NN, which the in-car HW3 runs. You could think of the training system as a build system, except it's on a much larger scale than compiling "hello world". The HW3 engineer explained that they cross-compiled the HW2.x NN to run on HW3, without re-training. They can probably do that again for HW4.
Apple did something similar for m68k -> PPC, and again for PPC -> x86.
BOBTHEJOCKEY
Member
No you aren't missing anything. I don't own a Tesla but the optimism I am getting from my brother soulpedl on the HUGE surge in his confidence of autopilot recently leads me to believe that it won't be long before even a Trumpanzee will see that Uber and Lyft will be worthless very soon. Tesla will own autopilot like Google owns search. It's a foregone conclusion.
Question:
I’m not entirely sure what was meant when they mentioned unsupervised learning and "Project Dojo."
Were they referring to:
a. Some kind of bootstrapping with the training, they mentioned Self-supervision
b. Some kind of competitive network evolution, such as Generative Adversarial Networks or Genetic Algorithms
c. Unsupervised or reinforcement learning as might be found in biological brains, such as an Adaptive Resonance Theory based network or a behavioral conditioning circuit such as a Gated Dipole
Any thoughts or guesses?
I’m not entirely sure what was meant when they mentioned unsupervised learning and "Project Dojo."
Were they referring to:
a. Some kind of bootstrapping with the training, they mentioned Self-supervision
b. Some kind of competitive network evolution, such as Generative Adversarial Networks or Genetic Algorithms
c. Unsupervised or reinforcement learning as might be found in biological brains, such as an Adaptive Resonance Theory based network or a behavioral conditioning circuit such as a Gated Dipole
Any thoughts or guesses?
Last edited:
I feel that part of the reason of this analyst meeting was to make clear how nonsensical the valuation of Uber is, with a competitor that will be able to undercut Uber and Lyft on price, and be very profitable at it. And that Tesla deserves that Uber valuation.
And competitors are close behind? Not that I have seen at all, anywhere.
Doggydogworld
Active Member
And they have ML expertise that far exceeds Tesla. And they have an actual Robotaxi service. And they're legally allowed to operate with an empty driver's seat in CA. And they've driven thousands of miles on public roads without anyone in the driver's seat. And they have enough cash to fund any possible business plan.
Where they fall short is business plan. Tesla customers have for years paid thousands of dollars for a product that didn't even exist. Google/Waymo has nothing like that. They better get their a** in gear, or all the "Ands" listed above won't matter.
On the eve of a likely very disappointing Q1 call, to come with some vague specs and 21x as many frames, could well be seen as window dressing. Tesla perhaps had to do it, but so far nothing to really get its shareholders to double down perhaps? And why would they, if Tesla seems to be betting everything on this. It's a tricky company that can turn on a dime, Elon may as well say they've stopped putting steering wheels in cars. Those who invest to make a living have gotten plenty of reason to get very nervous over the past year alone, even with all the sales records being broken.
Pras
Member
The point was to compare current Nvidia processor not what Nvidia’s top offering. Plus the constraint was power usage (100 watt).
Just reading headlines on yahoo finance :
-
Elon Musk Boasts Tesla Built the ‘Best Chip in the World,’ Drops Nvidia
Elon Musk says Tesla driverless taxis coming next year, touts self-driving chip
Tesla promises ‘one million robo-taxis’ in 2020
The thing I think would help the share price is to see the same titles but with an external person/entity name instead of Elon/Tesla. At this point, it is just too easy to say that they are over-promising.
-
Elon Musk Boasts Tesla Built the ‘Best Chip in the World,’ Drops Nvidia
Elon Musk says Tesla driverless taxis coming next year, touts self-driving chip
Tesla promises ‘one million robo-taxis’ in 2020
The thing I think would help the share price is to see the same titles but with an external person/entity name instead of Elon/Tesla. At this point, it is just too easy to say that they are over-promising.
It’s a programmable CPU, with a small amount of instructions specific for neural network calculation (6 or 7 instructions, they were on a slide). The next slide was about their compiler that generates the specific set of instructions to calculate the entire network solution. (There was even a typo in the title stating ‘complier’ instead of ‘compiler’). They even mentioned some specific optimizations in their compiler. The next chip will have more instructions.
no, no , and because even I, a perma bull don't believe they can do that next year?
Don't get me wrong, the technical presentations are very impressive. Three things jumps out to me:
1. Karpathy also thinks LIDARs are dumb. He is not yes man to anybody.
2. Simulation vs real world data.
3. Semi-automated labeling.
I remember sometimes ago, I saw multiple presentation from Google self driving team talk about massive simulation infrastructure they are building to augment the limit amount of data from real world. They touch upon removing bias from simulated data but did not specify how. Today Karpathy says not only you have to deal with bias, you simply can not simulate what you don't know! This strikes me as a huge problem to simulation, and probably one of the reasons they use geofencing, as they would try to collect as much as they can about one location, and augment the simulation model based on what they learnt.
Another difficulty in AI is data labeling. Reinforcement learning does not require labeling, but its is extremely limited. It appears Tesla has developed a pretty good approach to automate large amount of labelling. Although fully automatic labelling is not there, this is already a big step ahead.
So the Tesla fleet is an invaluable asset to them, a very wide moat. They are very likely the front runner.
All that said, we just don't have enough evidence that they are almost there. I don't believe it, and I don't think wall street believes it.
on the other hand if one day Elon decided to becomes a super villain, he has a pretty powerful weapon.
Causalien
Prime 8 ball Oracle
I just checked the new Xavier architecture. For the Drive AGX. They are still reusing their graphics Tensor cores. As long as they are using these Tensor Cores, 1 TOPS in NVIDIA is inferior to 1 TOPS in tsla chips. I think the most important difference is that Nvidia's chips are stream processors. I am assuming latency galore in decision making.
A proper simulation of human neurons need a proper hardware specific design for it. Not saying that Tsla have it down (since nobody knows what they have), but at least theirs is probably optimized to use the best type of variable for the calculation. According to neuro scientists, a neuron has a voltage component, a frequency and a phase component to their signals. Technically possible to evolve by allowing the NN to evolve its own transistor resistor level logic. But Mongo told me that's not how it was designed.
Tesla is competing with ICEVs for now, BEVs hitting the market are sparse and that will change.
Tesla is working on FSD and there are others doing the same, well funded, no distractions.
The moment one other company manages to get Level 5 going, it's a price war right there. If Tesla sticks to its own fleet, it can be outnumbered 100:1 for robotaxis on the streets within a few years.
25-30% will not fly anymore. Taking money for FSD software may not fly anymore.
Imagine Tesla sticks to $35K+ cars and then $15K FSD capable econobox BEVs start hitting the market. How much will riders pay extra to be brought to the airport or work by Tesla whom they know make much more money, as the novelty wears off?
Todd Burch
14-Year Member
They are almost certainly saving all the training data. They will be able to reuse it.
Similar threads
- Locked
- Locked
