Wasn’t that remarkable! shortly after, they made AlphaZero which taught itself how to play better than any other prior computer engine—a remarkable algorithm…that being said, I think FSD is harder because we don’t know all the rules of the game, like you do with Go or Chess—these are environments constrained by specific rules and limited predictable results, so you can mini-max the best decisions every time….I don’t think you can mini-max FSD…or can you?Reminds me of the documentary AlphaGo, where the software improved over time and got increasingly better at learning and decision-making - to the point where it was able to master a game in which there are 10 to the power of 170 possible board configurations.
AlphaGo
Novel AI system mastered the ancient game of Go, defeated a Go world champion, and inspired a new era of AI.deepmind.com
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD Edge cases
- Thread starter gaspi101
- Start date
Humans are already very good at driving, but if we can create a computer that can perform like a human brain, we should be able to make it a responsible human brain that doesn't take the tired, lazy, dangerous and selfish actions that are the source of the large majority of collisions.
Stopped reading at the first paragraph. I am familiar with the theoretical differences between brains and computers. See Turing’s works. I said It’s going to be possible. I said nothing about a silicon version of our brains. Your response will be useful to someone else, just not me.
Terminator857
Active Member
Fernand
Active Member
I don't understand the fascination with "levels". I use NOA on freeways all the time, I don't care what level it is, it's very useful to me. Like riding a horse. It would be counterproductive to blindly let the horse make all decisions. That doesn't make it any less useful.
I've changed by interacting with the system for 3 years. It's adaptation of something like the dividing line between the autonomic and conscious nervous systems. I don't think about it much, I punch in and out of auto-steer as needed in contexts where human conscious decisions work better. Nicki and I together make a better driver than either of us alone. So why should I send Nicki out by him/her self as a "level 5" but less optimal device? We don't know how this whole vehicle autonomy story will develop. Maybe the environment will change to help the cars, as ALL cars evolve towards greater autonomy. For instance additional markers.
There's a trait that good programmers/designers have, and it's the optimism that allows them to get up in the morning and tackle "impossible" problems and schedules that many people see as madly unrealistic. They might be late, but they usually get there. You can't solve anything without that approach. The pedestrians who shout "fraud" can only imagine straight lines or criminal intent. They couldn't ever start, let alone solve, a tenth of what the Tesla team has accomplished. Or predict the outcomes. Wake up, it's never been done. It's a wonderful and exciting adventure.
I've changed by interacting with the system for 3 years. It's adaptation of something like the dividing line between the autonomic and conscious nervous systems. I don't think about it much, I punch in and out of auto-steer as needed in contexts where human conscious decisions work better. Nicki and I together make a better driver than either of us alone. So why should I send Nicki out by him/her self as a "level 5" but less optimal device? We don't know how this whole vehicle autonomy story will develop. Maybe the environment will change to help the cars, as ALL cars evolve towards greater autonomy. For instance additional markers.
There's a trait that good programmers/designers have, and it's the optimism that allows them to get up in the morning and tackle "impossible" problems and schedules that many people see as madly unrealistic. They might be late, but they usually get there. You can't solve anything without that approach. The pedestrians who shout "fraud" can only imagine straight lines or criminal intent. They couldn't ever start, let alone solve, a tenth of what the Tesla team has accomplished. Or predict the outcomes. Wake up, it's never been done. It's a wonderful and exciting adventure.
Daniel in SD
(supervised)
Basically the only meaningful difference between levels is liability, it doesn't tell me how sophisticated the software is. You could have a perfect driving agent that will drive down any road anywhere in the world, the AlphaGo of driving, and it's level 2 simply because the company doesn't want to become liable for accidents.
EVNow
Well-Known Member
On a podcast with the legendary linguist Noam Chomsky - he had a similar opinion about natural language capabilities of computers. He said they don't "understand" language.
More broadly, the basic idea of neural ntworks is that they work like biological neurons. So, anything human brain can solve, with enough training, they should be able to solve. This is the "first principle" that Musk keeps talking about.
But, biological neurons and NN have differences. So, we don't really know what the limitations of NN are - this is what I gather from the best minds in academia about it.
Here is a good podcast on differences between neural networks and the biological brain.
Transcript of interview to John Hopfield by Lex Fridman
John Hopfield: Physics View of the Mind and Neurobiology | AI Podcast #76 with Lex Fridman | Feb 29, 2020 Introduction by Lex Fridman The following is a conversation with John Hopfield, professor of Princeton, whose life's work weave beautifully through Biology, Chemistry, Neuroscience and Physics.
www.linkedin.com
Fernand
Active Member
Saying computers can't "understand language" because they don't follow context is no longer correct, because of the immense growth in processing power and memory.
It's largely a matter of memory, a) memory within a segment, reaching back to build context, e.g. a vector 3D representation, and b) predictive NN Deep Learning - memory of countless previous cases being applied forward.
Algorithmic autopilot does OK in simple cases. I believe that's what we saw with base AP driving on city streets. Where the rules applied it did better than the current "FSD Beta", which seems to be starting over - precisely in developing context, applying memory. It seems short on useful canned rules.
After 3 years of using NOA on Freeways I can anticipate where the system will have trouble, where its rules break down. It's become second nature, like I've internalized Nicki's rule set. Driving in NOA works great for me, with say 95% utility.
At this Beta stage I expected NOA on City Streets to be similar, though with at best 25% utility. But it's quite different. Aside from the obvious wide boulevards with no turns, I'm finding it harder to predict where the trustworthy 25% will be, and to anticipate problems. In a way it's all edge cases
It's largely a matter of memory, a) memory within a segment, reaching back to build context, e.g. a vector 3D representation, and b) predictive NN Deep Learning - memory of countless previous cases being applied forward.
Algorithmic autopilot does OK in simple cases. I believe that's what we saw with base AP driving on city streets. Where the rules applied it did better than the current "FSD Beta", which seems to be starting over - precisely in developing context, applying memory. It seems short on useful canned rules.
After 3 years of using NOA on Freeways I can anticipate where the system will have trouble, where its rules break down. It's become second nature, like I've internalized Nicki's rule set. Driving in NOA works great for me, with say 95% utility.
At this Beta stage I expected NOA on City Streets to be similar, though with at best 25% utility. But it's quite different. Aside from the obvious wide boulevards with no turns, I'm finding it harder to predict where the trustworthy 25% will be, and to anticipate problems. In a way it's all edge cases
Last edited:
I’m not sure I understand what you mean by no longer being correct...contextual information can only be analyzed with the use of the sum of human experience…if I see a stopped car and the driver is elderly and looks like their wheels are turned to the side, I’m going to infer that they’re having a hard time judging when it’s safe to make a left. If I’m behind them, perhaps I would decide based on those clues that maybe I shouldn’t zoom around them to get by. That might cause an accident, although FSD would analyze that it’s a stopped vehicle and we need to go around…there are millions of examples…what do you mean contextual clues are no longer problems?Saying computers can't "understand language" because they don't follow context is no longer correct, because of the immense growth in processing power and memory.
It's largely a matter of memory, a) memory within a segment, reaching back to build context, e.g. a vector 3D representation, and b) predictive NN Deep Learning - memory of countless previous cases being applied forward.
Algorithmic autopilot does OK in simple cases. I believe that's what we saw with base AP driving on city streets. Where the rules applied it did better than the current "FSD Beta", which seems to be starting over - precisely in developing context, applying memory. It seems short on useful canned rules.
After 3 years of using NOA on Freeways I can anticipate where the system will have trouble, where its rules break down. It's become second nature, like I've internalized Nicki's rule set. Driving in NOA works great for me, with say 95% utility.
At this Beta stage I expected NOA on City Streets to be similar, though with at best 25% utility. But it's quite different. Aside from the obvious wide boulevards with no turns, I'm finding it harder to predict where the trustworthy 25% will be, and to anticipate problems. In a way it's all edge cases
EVNow
Well-Known Member
Is this your take or the industry view ? Any papers you can refer me to ?
Otherwise I'll have to take Chomsky's words ...
L4-5 is only possible if
1. You remove all human drivers. Only self driving cars on the road.
2. You upgrade all the roads to have proper markings for FSD.
Then it is possible
EVNow
Well-Known Member
EVNow
Well-Known Member
So ... ?
On AI podcast by Lex Fridman - he has interviewed pioneers & legends in the field who will not say whether L4/L5 can be achieved.
BTW, Waymo is already L4.
ps : If what you are saying is the actual industry / SME consensus, billions would not be poured into AV industry.
This is just my personal informed opinion. I don't think L4-5 "everywhere" is possible without road upgrades, have you seen roads in Istanbul or Mumbai?
Waymo is only L4 in very few selected regions. Nowhere near working everywhere.
TESLA NOA can be considered L4 on highways with proper lane marks it is pretty good.
"Everywhere" is a completely different story.
This statement is false and dangerous. No Tesla vehicles ship with any L4 features today, and no filings with the government have been made from Tesla to indicate that they are even actively developing any > L2 features.
L4 has a strict definition and NOA does not meet it. Your post is dangerous because this sort of misinformation is liable to get somebody killed. In fact the belief that NOA is more capable than it is has caused unnecessary deaths already.
Please be more mindful of how you communicate about Tesla vehicle capabilities.
I appreciate your concern, but nobody will go risk their lives based on an online opinion post by someone they don't know.
People's understanding about things are shaped bit by bit from all of the information they may come across. Lies spread and gain traction the more frequently that they are repeated. In the case of Teslas, lies from an _owner_ carry more weight.
Why do you want to be part of the problem?
Similar threads
