Yes. Youtuber "greentheonly" has a bunch of interesting ones, e.g.
-
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
Yes. Youtuber "greentheonly" has a bunch of interesting ones, e.g.
I wrote the tweet about WS traders using cocaine and needing a dominatrix. Sorry dana
Last edited:
Doggydogworld
Active Member
Waymo did that in 2010. 10 loops of 100 miles each on a variety of roads. That was the milestone they had to clear to get approval to begin a self-driving development program.
j6Lpi429@3j
Closed
quote from Waymo CTO:
Have you heard of this guy Geohot? I bet he did it too.
It doesn't matter who solved that 1% case first, what matters is who put enough number of 9s in that 99.99...% reliability first, enough to deploy it on large scale.
Again, real word data is the key to make it work.
A side note:
The guy who led the Google self-driving car project in 2010 had left Google and went back to Stanford around 2015 or so (weirdly wikipedia didn't clearly state that, and I don't want to speculate the reason he left), current leadership of Waymo struck me more like bean counters instead of visionaries.
Last edited:
Even this isn’t particularly helpful, unfortunately. Tesla’s system today meets(and, actually, exceeds) that description right now, but my impression is that it doesn’t actually meet the true definition.
One of these days, I need to get a copy of the actual SAE documentation and look through for exactly which requirements there are in more precise terms, since many of the ones bandied about are clearly wrong. For example, the requirement that under no circumstance can control ever be handed back to the human without several minutes prior warning. Clearly, if someone targeted and damaged all sensors on the car at once, it couldn’t just manage blindly for minutes before handing back control.
Anstandswauwau
Member
Musk is doubling down on "500k in next 12 months"
Anstandswauwau
Member
Jesus, it stops nerve-rackingly late! Looks legit for an early public version, though.
Hope their demo software on the 22nd does better than that.
It only stands to reason that trains pulling 120 cars of goods are a lot more efficient than trucks weighing 35,000 kg or so including the weight of the truck. When I'm traveling, what I see are miles of railroad tracks that are seldom occupied and roads clogged with trucks. This seems to me to be a very inefficient way of delivering goods. Yes, trucks are needed at the end points, but a lot of shipping could be done by trains. This would not only reduce road congestion, but it would reduce road maintenance by a large amount.
Fact Checking
Well-Known Member
I was talking about the primary visual cortex, which is a small part of the brain:
The functional role is to turn a raster image into a vector image in essence - which is probably close to what the FSD computer's NN's first and largest layer does.
Full human vision includes functionality that the FSD computer doesn't have to do: for example object motion detection and prediction. It also includes a very large number of neurons that are hardwired to recognize things like human faces - again not something the FSD computer needs.
If we apply your calculation to the primary visual cortex:
- ~0.14 billion neurons,
- x ~10,000 synaptic inputs,
- x ~10 FLOPS,
- = 14 TFLOPS
BTW., why do you estimate ~10 FLOPS per synapse? AFAIK human synapse has about 0.1 bits of information content, while the FSD computer likely uses minifloats and is able to do a single inference calculation per cycle: multiplication with the weight and the application of the sigmoid function with hardwired functional units.
I.e. unless I'm missing something the correct calculation is more like 1 Tesla-FLOPS per synapse, and that's probably generous due to the low information content of human synaptic strength. I.e. the true comparison even based on your 10 billion neurons full visual cortex estimate would be a computing capacity of about 100 TFLOPs - close to the computing capacity of a single Tesla chip. (and there's two of them)
Isn't that a requirement that can never be met? No system can predict catastrophic failure, it can only act on catastrophic failure. If that's actually in the SAE docs, which I doubt, then the SAE docs were written with the intent of stopping any development.
Fact Checking
Well-Known Member
500k production in the next 12 months is very likely with the Shanghai Gigafactory scaling up.
Of course, all neurons are not created equal. White matter, for example, functions as a data bus. Then there's the fact that neurons are generally grouped into functional units of which all neurons are fed the same input data - in the cerebrum, the basic subunit is the cortical minicolumn. There's something like 200 million minicolumns in the brain (only a small fraction of which are dedicated to visual processing). There may be 120 or so neurons in a minicolumn, but a lot of the functionality is redundant (which should be obvious, as you can't have the death of a single neuron have a dramatic impact on the brain).
Then there's the "clock rate" of the brain. While computer clocks run at GHz speeds (and I'd imagine that the average simulated neuron may fire at variable "hundreds of MHz" rates), the average neuron in the visual cortex fires only on the order of 10 times per second (and isn't even capable of firing at over several hundred hertz). So there's a very large multiplier partially compensating for the fact that no processor is even remotely as parallelized as the brain.
Takes attention away from bad Q1 deliveries and reminds people that Tesla is more than just an EV maker. And the timing is perfect, since Uber and Lyft are doing IPOs, so focus on the "ride sharing" market is high. Redirecting some fraction of money that otherwise would have flowed into those IPOs to Tesla would be desirable.
And regardless, don't overread such videos. They're easy to "cheat without cheating" - record every test drive you do, and turn the one (out of hundreds) that went best into a PR video. Everyone who has a self-driving system does this same "trick".
Last edited:
What would you think the stock price will be at, if this turns out to be true a year from today?
Similar threads
- Locked
- Locked
- Locked
