Speaking of the long tail here (march of 9s), one would be lucky if it stayed linear. If the odds of an event are a uniform distribution, then you have a 1/x chance of capturing the event. If the event probability becomes 1/(10*x) then you need 10 times the samples to capture the same number of interesting events in the same amount of time as you used to. If you data collection increase rate is less than the event probability, you'll learn slower and slower.
-
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
Speaking of the long tail here (march of 9s), one would be lucky if it stayed linear. If the odds of an event are a uniform distribution, then you have a 1/x chance of capturing the event. If the event probability becomes 1/(10*x) then you need 10 times the samples to capture the same number of interesting events in the same amount of time as you used to. If you data collection increase rate is less than the event probability, you'll learn slower and slower.
Fact Checking
Well-Known Member
This assumes that computing power cannot reach the levels of processing of the human brain.
According to this estimate that's not the case:
The firing rate in the neocortex (which hosts 80% of the brain's neurons) is between 0.3 and 1.8 per second. With 80 billion neurons in the neocortex that's a firing rate of about 24-144 billion per second.
The average number of synapses per neuron is 10,000 - while the average information content of a synapse is 0.1 bits, or ~100 bits per neuron.
So the NN processing speed of the entire human neocortex is ~2,400-14,000 billion bits per second. Now the operations it performs per firing is addition and multiplication and capping - which we can recognize with a ~10x complexity factor, so the net speed is about 24,000-140,000 billion bits of simple arithmetic operations per second. (This is probably generous to the brain.)
The Tesla AI chip computes ~144,000 billion mini-floats per second (144 TOPS), where a mini-float is 8 bits. So the total processing power is ~1,152,000 billion bits of simple arithmetic operations per second.
So if we believe these estimates then the Tesla chip is already comfortably beyond the NN processing power of the human brain, by a factor of ~8x.
Put differently, every Tesla camera has as much NN computing power allocated as a single dedicated human brain watching that camera 24/7 ...
What the human brain arguably does much better is information storage: 1,000,000 billion synapses can store about 1,250 TB of data, which is a lot more than what Tesla can store in their NNs.
But if we accept that "legally safe driving" requires only a very small subset of the vast amount of data a human brain stores, then the Tesla AI chip can already do an order of magnitude better job, with vastly superior control latencies.
That is going to save lives, and this will be apparent from the accident statistics.
Last edited:
A good start. It's almost teenage-driver level!
There's a semi-blind corner on the right at the start of that video -- car needs to learn to slow down in anticipation. Second one later on.
While I don't see the Robotaxi's on the roads in 2020(maybe they get approval for certain urban cities where they only operate in that city), I do think the rate of improvement is going to be exponential. Not only because the number of Tesla will go from 500k to a million by 2020 but cause they'll likely start releasing some FSD features....but that they require human supervision...so city surface streets, stop signs, stop lights, roundabouts, etc...) Also the expansion to other regions(especially Europe) will help with edges cases(roundabouts galore).The samples of edge cases and interventions I think will increase exponentially more than the number of total cars out there.
Btw, thank god Q1 report is here. So tired of hearing about Tesla returning to losing money from here on out, not making money on SR, lack of demand, blah blah blah.
Btw, thank god Q1 report is here. So tired of hearing about Tesla returning to losing money from here on out, not making money on SR, lack of demand, blah blah blah.
Please tell me why this negates Buddha's and some 12 step observations that the only person you can change is yourself. Sticking to the error is important. Who makes it is a side issue—projection—fear of yourself making one like others? "Well, at least I didn't make that mistake."
Who cares if a rose has thorns, we all have fears. Does the rose care?
If TOPS was all you needed then the FSD computer/Tesla fleet will achieve sentience soon...
Wouldn't the system literally "evolve" because poor driving choices eventually lead to accidents and that's easy to filter?
RobStark
Well-Known Member
Civics with over 200k miles can still fetch $1500 because they can still be driven for another 10 years.
https://www.cars.com/for-sale/searc...t&stkTypId=28881&zc=90210&localVehicles=false
Rotte
Member
Now I'll in turn surprise you by agreeing... I suspect this is where the lion's share of the work needs to be done.
This would seem to be exemplified by the portion of the demo where clearly the vision network could identify the bike. The system then needed to understand that it was a bike on a car carrier, not a separate object to track. I found it interesting that they apparently trained the neural net to classify these as a single object. I was wondering if they would allow a later layer to simply "understand" the nature of "captive items".
I'd like to know what happens if that bike falls off the carrier. Or what the system thinks of a trailer being pulled by a truck, and what it will do when it sees it detached from the truck pulling it and become a sperate entity.
I think it is indeed reasonable to assume that the policy stuff is less mature, if for no other reason than it has to sit atop the visual/environmental portions of the stack. And they had to start over on those parts when they dropped MobilEye.
But interestingly, Karpathy did discuss where it males sense to use heuristics, vs moving stuff in to a NN at that layer. I'm sure there will be a lot of work, and potentially churn here. I think we may have differences on the meaning of "long time" or "not any time soon", however...
Last edited:
I guess I shouldn't be surprised that the earnings report is tomorrow and literally no one is talking about the stock in this thread?
humbaba
sleeping until $7000
what struck me about it was at the very end where the oncoming car looks to be on the center line and the projected path appears to be going into the oncoming lane. ouch!
I'm the one who makes the correct predictions about what timelines they're gonna succeed at and what they won't. I get my arrogance honestly in this area.
I've made more inaccurate predictions when it comes to the finances, which are apparently harder for me! (Well, my background was mostly STEM)
I guess was using "hard" in a rather harsh sense. Let's put it this way: Landing rockets on barges is to full-self-driving-everywhere as winning a spelling bee is to winning the Nobel Prize.
Last edited:
No, it doesn't. It assumes that we never figure out what the nature of the problem is. Computing power is wildly irrelevant.
Once an outlier is found, it's not hard to imagine similar but different cases, so the system can be gamed this way (this is one way where simulation can help because simulation helps if it can be imagined). Also, the fewer the instances of a particular category, the less likely they are to be encounters. If we take the lava flow example, the system doesn't know that it's a lava flow, but it does know the size (visual) that it's substantial (radar), and that it is stationary (visual), so it can fall back to the stop if something big is ahead and not moving. If there was thermal imaging, it would also know about heat. There is some question as to what would happen around a blind curve, but a human driver wouldn't fare any better and maybe worse due to reaction time.
Jesus Christ analysts are stupid.
If you look here, even those of us who are total "there are no robotaxis" skeptics think LIDAR is useless!!!
Wow.
I don't think they'd recognize a dollar bill if everyone else told them it wasn't a dollar bill.
Well, they're right about that.
Wow, wow, wow. OK, so this didn't manage to get anything through the brains of analysts. Wow.
Thanks for accumulating the comments from analysts who took the testdrives.
bdy0627
Active Member
I agree with that. I would think it will take easily 6 months just to get regulators to allow robot cars even after a massive amount of data is collected showing the relative safety compared with human drivers. If the tech actually rolls out by the start of 2020, figure at least 12-18 months of dealing with the initial rough edges/corner cases, and collecting the data. That would allow Tesla to then approach regulators perhaps as early as the end of 2020 or early 2021 with compelling data in a quest for allowing robot taxis. Nothing with Tesla goes perfectly, so figure this tech actually rolls out to the fleet (disguised as a seriously advanced driver assistance) sometime next year. If they manage to work through the majority of issues over 18-24 months, that would put us well into 2021 or even 2022 before starting to work on the regulation side. So, if everything goes fairly well, I would guess at least 2022 before robotaxis can become reality. Having said that, if this does go pretty well, there will be a LOT of enthusiasm generated about it well before 2022. In fact, even next year it may seem like it will be going live soon when in reality it will still probably take a few years from there.
sundaymorning
Active Member
Having watched the event yesterday in its entirety makes me a lot more confident about Tesla & the team. I love every speaker they brought to the stage (although Elon could show more patience when answering questions, and perhaps less optimism); overall, I think we’re on the right path and vastly ahead of everyone else in the FSD arena. From what Elon has said about “doing the right thing that you guys (analysts) want us to do..” my belief is that Tesla will raise capital.
Regarding the “game, set, match, Tesla...” comment I think Tesla is so far ahead of competitors that they are confident enough to show their cards on stage by delving deep into how their hardware is designed and software is implemented. Most companies would want these info top secret, but Tesla already knows they’re too far ahead so there is no use in keeping it a secret; it’s time to show everyone else their cards to demoralize the competition. If I was Waymo, Lyft or Uber I would be quite nervous about what Tesla’s technology will be able to do. Fundamentally, Tesla’s robotic fleet doesn’t need to go into every corner of the US, they just need to overtake the dense urban areas, starting with San Francisco or Silicon Valley and things should pick up from there. It won’t take long before Waymo and others realize their money losing operations won’t be able to keep up with Tesla and eventually sell their IP and write it off as a loss.
Concerning lidar, I’ve spoken to a computer programmer and he seems to agree with Elon, that lidar is crutch as it cannot “see” objects as well in terms of pixels, which will hard time to “predict” human and non human patterns of mobility. Having said that, he also thinks that cameras are a much better solution and thinks Tesla is also way past everyone else.
Regarding the “game, set, match, Tesla...” comment I think Tesla is so far ahead of competitors that they are confident enough to show their cards on stage by delving deep into how their hardware is designed and software is implemented. Most companies would want these info top secret, but Tesla already knows they’re too far ahead so there is no use in keeping it a secret; it’s time to show everyone else their cards to demoralize the competition. If I was Waymo, Lyft or Uber I would be quite nervous about what Tesla’s technology will be able to do. Fundamentally, Tesla’s robotic fleet doesn’t need to go into every corner of the US, they just need to overtake the dense urban areas, starting with San Francisco or Silicon Valley and things should pick up from there. It won’t take long before Waymo and others realize their money losing operations won’t be able to keep up with Tesla and eventually sell their IP and write it off as a loss.
Concerning lidar, I’ve spoken to a computer programmer and he seems to agree with Elon, that lidar is crutch as it cannot “see” objects as well in terms of pixels, which will hard time to “predict” human and non human patterns of mobility. Having said that, he also thinks that cameras are a much better solution and thinks Tesla is also way past everyone else.
GBleck
Member
I think that is telling of the general sentiment with TSLA earnings for tomorrow.
Similar threads
- Locked
- Locked
- Locked
