EVNow
Well-Known Member
I guess the internal release is Wed-Thu.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: FSD Beta tweets
- Thread starter Terminator857
- Start date
Daniel in SD
(supervised)
AlanSubie4Life
Efficiency Obsessed Member
C'mon man, that was three weeks ago! 10.11.2 has been continuously learning and now, just 8 days ago, it is only hitting curbs at night in Las Vegas, a clear "corner case." Which is to say, it occurred at a corner. But this is off-topic here; will go back to the Elon tweets now. But basically the new version is going to fix all this, like Elon said (I'm paraphrasing).
How long? Elon: 2 weeks long.
AlanSubie4Life
Efficiency Obsessed Member
Daniel in SD
(supervised)
EVNow
Well-Known Member
That is the first principle !Whatever happened to first principles?
AlanSubie4Life
Efficiency Obsessed Member
Pretty sure that is not how things work, but I can understand Tesla not wanting to have a bunch of cars out there with different driving styles, continuously learning on their own, with unpredictable degradations in performance when crossing state and international borders. Not to mention the logistical and hardware difficulties of that approach.
EVNow
Well-Known Member
The logic goes like this.
To get FSD working you need to train NN …
- NN needs a lot of data
- To get to higher quality, NN needs a lot more data
- so, without Billion miles of data, FSD is not possible
ps : I remember a big data conference a few years back (‘15 ?) in which someone close to Elon laid out this exact logic and talked about how Tesla has the best shot at FSD.
Daniel in SD
(supervised)
Except we know from first principles that FSD is possible with only 100,000 miles of training data the same way we know purely vision based FSD is possible.
We have no idea whether or not the latest and greatest artificial NN architecture will be able to achieve FSD.
Elon's statement reminds me of the saying "If the only tool you have is a hammer, it is tempting to treat everything as if it were a nail"
Daniel in SD
(supervised)
There are many above average drivers with less than 100,000 miles of experience so we know that self-driving is possible without billions of miles of data. The same reason we know it’s possible to do without LIDAR.
AlanSubie4Life
Efficiency Obsessed Member
It’s going to be awesome when (if?) we figure out how we do this (and reproduce the process).
sleepydoc
Well-Known Member
You completely neglect the fact that even a below average human is capable of learning far more, far faster than the best AI algorithm.
Daniel in SD
(supervised)
Of course, but it seems just as silly to say that solving self-driving is impossible without billions of miles of training data as it does to say it's impossible without LIDAR. Both are crutches that are used because AI is still very primitive.
EVNow
Well-Known Member
First principles just means make the fewest assumptions.
Basic assumption is …
- NN can work like a human brain given enough data
But we do know one of the major differences between NN and human brain is the amount of training needed to learn. NN needs several orders of magnitude more data to learn. Perhaps because human brain has the benefit of a billion years of evolution.
And we keep discussing the differences between a camera and human eye.
Daniel in SD
(supervised)
That's a HUGE assumption. haha.
And does the human brain work by generating a vector space and then feeding it into a planner?
If the assumption were truly that a NN can work like a human brain then wouldn't you just use the billions of miles data to train an end-to-end NN? Then you don't even have to label anything!
Meanwhile FSD Beta is still hitting curbs. Maybe it's impossible to generate a vector space using current NN technology without billions of miles training data but it seems like a silly thing to say when it hasn't even been proven to be better than the approaches taken by other companies yet.
Similar threads
