Incorrect. AlphaZero uses Monte Carlo tree traversal and evaluates 80 000 positions per second which is easily 5-6 orders of magnitude more than top human players.
-
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
Incorrect. AlphaZero uses Monte Carlo tree traversal and evaluates 80 000 positions per second which is easily 5-6 orders of magnitude more than top human players.
j6Lpi429@3j
Closed
TBH You have very valid concerns here, but don't forget that these situations are pretty are (well under 1% of cases) and that the car is both capable of remote observation, and has someone in it most of the time. If the car encounters something weird (traffic cop stood right in front of it, expected stop light showing no color), it could easily ping HQ for someone monitoring the fleet to look through the front camera and remotely tell the car what to do.
I cant recall being in the situations you describe in the last year, so maybe this happens 1 in every 365 days? and takes approximately 10 seconds to resolve. That really would not require much in the way of remote monitoring.
Plus if the situation looks weird, but the car right in front of the robotaxi just turned right, the car can take that as a very strong hint as to what to do, so its not even every car, only every occasional car.
BUT...
FWIW, the benefit of super-safe autopilot that is basically hands free for 99% of the time but still requires a driver is still of HUGE financial value, even before the 100% robotaxi future.
I think this cannot be repeated enough. Some people here think that FSD is binary and will provide 0 value until it works 100% perfectly under any circumstance. That's simply not the case. There is many use cases in which FSD will have significant value before that, e.g., for rental cars in semi-controlled environments (you arrive at the airport and your M3 rental car arrives autonomously, when you return it it goes automatically to the collection point, etc.)....
LN1_Casey
Draco dormiens nunquam titillandus
Hell, if that's all the FSD would be capable, it would STILL be worth it. You know how annoying it is to have to try and track down your rental agency from the airport in an area you've never visited? Especially if you have heavy luggage? If I could get the car to come to me and pick me up, I'd love it.
I drive a lot, perhaps more than most of you, at least 15K a year, but I come upon these incidents ALL THE TIME! Just two days ago I had to veer around some asphalt paving. There was a flagman telling me to take a wide circle around the paving machine. (FYI, I live in one of the most affluent areas in my state). Also I travel a lot. Do you know how many times I encounter a flagman for closed lanes? Every year! Every single year. And none of this...well, if you don't see a problem with FSD Tesla robottaxis refusing to obey cops or flagmen or whoever directing traffic, then Houston, we have a problem.
j6Lpi429@3j
Closed
Oh and parking... the ability to get to a car park, then jump out and go to a bar/restaurant and let the flipping car find a space...oh yes please.
We have two separate independent sets of analysis showing that COGS at a high production rate will be sub $30k. Tesla is not yet in the "squeeze a little out" stage, they are in the "substantial cost reduction due to economies of scale" stage.
If Panasonic can increase their production rate from 24GWh per year closer to the target 35GWh per year then their depreciation costs go down by around a third on the cells they are selling, this effect will apply to every bespoke supplied part of the M3 from every supplier.
Then there is the greater purchasing power of Tesla when they are buying 7k units per week v just under 5k. Then the greater purchasing power of Tesla suppliers when they are ordering materials to make 7k units per week instead of 5k.
So both raw material costs and depreciation costs reduce on a per unit basis up and down the entire supply chain as volumes increase. I have not heard of many non-commodity products where an increase in volumes by 30%-50% does not result in substantial cost reductions on a per unit basis.
Your questions were more of a technical nature, less about the investment side. And yes, there has been a lot of autonomy discussion here, too much in fact, so preferably it is not being promoted by open questions. Let's leave it at that.
Pras
Member
I know. I had to sell 30% yesterday as I had promised my wife to cut back at 180. Now this ****er AJ is back to pumping. Scoundrels.
Ah typical sell low from me but then the macro and Trump meant that the bottoms could be anywhere. I mean at one point 250 was ridiculous and it breached.
Well for me, I will buy back the sold position, if I see strong reversal today.
I prepared my wife that Elon will most likely buy Tesla out at around 100 so the worst we can lose is half our money...but the upside is still unlimited. She rolled her eyes and was like "whatever"....
aagosh
Member
I told my wife that in 3 years when our lease expire - I will buy out the Tesla MX from the gains made from the stocks. She rolled her eyes and was like "whatever"....
The problem with point no.1 is that it is largely irrelevant when we know that Tesla is only part way through their ramp and vehicle production is a game of extremely high operating leverage. The COGS figure you state is only for a point in time several months in the past.
re point no. 2. - you should be less confident.
LN1_Casey
Draco dormiens nunquam titillandus
Technique #3 – ‘TOPIC DILUTION’
Topic dilution is not only effective in forum sliding it is also very useful in keeping the forum readers on unrelated and non-productive issues. This is a critical and useful technique to cause a ‘RESOURCE BURN.’ By implementing continual and non-related postings that distract and disrupt (trolling ) the forum readers they are more effectively stopped from anything of any real productivity. If the intensity of gradual dilution is intense enough, the readers will effectively stop researching and simply slip into a ‘gossip mode.’ In this state they can be more easily misdirected away from facts towards uninformed conjecture and opinion. The less informed they are the more effective and easy it becomes to control the entire group in the direction that you would desire the group to go in. It must be stressed that a proper assessment of the psychological capabilities and levels of education is first determined of the group to determine at what level to ‘drive in the wedge.’ By being too far off topic too quickly it may trigger censorship by a forum moderator.
The Gentleman’s Guide To Forum Disruption - The Big Picture
Fact Checking
Well-Known Member
I agree in part (and thanks for the correction), AlphaZero uses a general purpose Monte Carlo tree search:
https://deepmind.com/documents/260/alphazero_preprint.pdf
"Instead of an alpha-beta search with domain-specific enhancements, AlphaZero uses a general-purpose Monte Carlo tree search (MCTS) algorithm. Each search consists of a series of simulated games of self-play that traverse a tree from root state s root until a leaf state is reached."
"Instead of an alpha-beta search with domain-specific enhancements, AlphaZero uses a general-purpose Monte Carlo tree search (MCTS) algorithm. Each search consists of a series of simulated games of self-play that traverse a tree from root state s root until a leaf state is reached."
Firstly, its evaluation speed is, according to the paper, not 80k/sec but 63k/sec:
AlphaZero: Chess: 63k/sec, Shogi: 58k/sec, Go: 16k/sec
I don't know what the basis of your 5-6 orders of magnitude figure is, 6 orders of magnitude suggests a difference of a factor of 100,000, i.e. your (slightly incorrect) 80k/sec figure becomes 0.8 positions/sec - do you really believe that human grandmasters can only evaluate 0.8-10 positions per second?
And while there are algorithmic shortcuts grandmasters use in endgame to see ahead 20-30 steps, the general depth in complex situations appears to be around 15:
'Magnus Carlsen claims that he can “see” 15 moves ahead, sometimes even 20.'
And note that he referenced 'full moves', which are actually 30-40 steps deep if we look at it as a tree search...
But even considering dynamic situations with no obvious forced moves, grandmasters are thought to be able to think ~3 full moves ahead, i.e. 6 position changes deep in the tree. Considering the average ~30 legal chess moves per position that's an effective evaluation speed of around 30^6 = hundreds of millions of positions per move - with the overwhelming majority discarded and not consciously evaluated. Considering the average chess move is ~135 seconds that gives a human grandmaster effective position evaluation speed that is better than AlphaZero's.
Anyway, my point was that AlphaZero is not using brute-force alpha-beta tree traversal like the other engines, but is using something a lot closer to what human grandmasters are doing.
Also note the limited size of AlphaZero's neural network:
"Apart from the representation of positions and actions described above, AlphaZero uses the same network architecture as AlphaGo Zero (9), briefly recapitulated here. The neural network consists of a “body” followed by both policy and value “heads”. The body consists of a rectified batch-normalized convolutional layer followed by 19 residual blocks (48). Each such block consists of two rectified batch-normalized convolutional layers with a skip connection. Each convolution applies 256 filters of kernel size 3 × 3 with stride 1. The policy head applies an additional rectified, batch-normalized convolutional layer, followed by a final convolution of 73 filters for chess or 139 filters for shogi, or a linear layer of size 362 for Go, representing the logits of the respective policies described above. The value head applies an additional rectified, batch-normalized convolution of 1 filter of kernel size 1 × 1 with stride 1, followed by a rectified linear layer of size 256 and a tanh-linear layer of size 1."
This is, effectively, a neural network of ~40 hidden layers with 3*3*256*256 (0.6M) weights in each layer - 23M weights total and maybe ~5K neurons-equivalent in a fully connected network. (If I calculated this correctly.)
Which is a fraction of the size of the networks Tesla's chip can handle: the Tesla firmware leak a couple of months ago suggested they have a visual network with 150 million weights. So it's also a reasonable inference that network size and quality can probably be increased at the cost of reducing MCTS search speed to move it even closer to human chess playing behavior.
We have only scratched the surface of machine learning and AIs.
Last edited:
Fact Checking
Well-Known Member
Note that what @Doggydogworld is referencing is the Model 3 cost of goods as visible in the quarterly earnings reports, which includes all sorts of other expenses that two tear-downs of the Model 3 very likely did not include: depreciation and amortization, stock compensation, warranty reserves and reserves for services provided in the lifetime of the car, such as connectivity expenses. These should be included in break-even steady state analysis, and they do add up to a value that is more than $30k for the $40k SR+, even with hypothetical perfectly optimized production lines.
So the ~$28k COGs from the German tear-down and from Sandy Munro isn't necessarily in conflict with claims of per units COGs being in excess of $30k.
My disagreement with @Doggydogworld is over including all GAAP expenses in the Model 3 COGs, which is invalid I believe: a lot of the depreciation costs are not genuine steady-state depreciation costs, but the result of past heavy investment cycles.
Last edited:
My whatever never talks on the phone. Phone sex is so teen!
Similar threads
- Locked
- Locked
- Locked
