I keep debating rather or not to push computer science on my oldest son because of this shit. He's still just a boy but getting closer by the day to being a young man. On one hand I think, he needs to be educated on the cutting edge of technology. On the other hand I think, by the time he's old enough to apply what he's learned it will be obsolete knowledge. What do you guys think?
In the same way a nigger invented streetlights. It found a performance improvement on existing algorithms. I can grab the relevant comments if you want me to, but what I got is in a three item list (a,b,c), b<=c and if this is false than the list is inverted but still sorted.
It didn't start with existing algorithms. It treated the process like a game. Every round, it started with an empty program, wrote the best sorting algorithm it could, scored itself by testing the performance, and learned from the result. Rinse and repeat until it spat out algorithms that perform better than what we had.
The resulting algorithm is substantially more complex than what you describe. In the case of a 3 item list, it uses 4 temporary registers, performs 3 comparisons, and has 6 conditional assignments, of which up to 4 may be executed for a given list. A list (a, b, c) which starts with c < b will still end up sorted ascending, small to large, and not inverted.
It didn't start with existing algorithms. It treated the process like a game. Every round, it started with an empty program, wrote the best sorting algorithm it could, scored itself by testing the performance, and learned from the result. Rinse and repeat until it spat out algorithms that perform better than what we had.
Yes but based on the information provided by the article, and comments, the result it arrived at was human equivalent with the exception of the efficiency it found. Which would imply it reinvented an already existing algorithm, with a single performance improvement for that algorithm. I'm not sure what sorting algorithm it implemented exactly as I can't read assembly.
I could be wrong, but I would like a proper sudo code for the five sort as that would show what method it is using.
[ + ] ItsOk2bArian
[ - ] ItsOk2bArian 1 point 10 monthsJun 10, 2023 04:18:49 ago (+1/-0)
What do you guys think?
[ + ] shitface9000
[ - ] shitface9000 [op] 1 point 10 monthsJun 10, 2023 12:40:15 ago (+1/-0)
[ + ] dontbeaphaggot
[ - ] dontbeaphaggot 1 point 10 monthsJun 9, 2023 23:53:02 ago (+1/-0)
[ + ] ItsOk2bArian
[ - ] ItsOk2bArian 2 points 10 monthsJun 10, 2023 04:14:53 ago (+2/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons 0 points 10 monthsJun 10, 2023 08:42:21 ago (+0/-0)
It's still bad, mind you. ChatGPT gets a lot of things wrong, for now. But often that's still an improvement.
[ + ] s23erdctfvyg
[ - ] s23erdctfvyg 1 point 10 monthsJun 9, 2023 21:52:06 ago (+1/-0)
[ + ] GloryBeckons
[ - ] GloryBeckons 2 points 10 monthsJun 10, 2023 03:46:37 ago (+2/-0)
It didn't start with existing algorithms. It treated the process like a game. Every round, it started with an empty program, wrote the best sorting algorithm it could, scored itself by testing the performance, and learned from the result. Rinse and repeat until it spat out algorithms that perform better than what we had.
See here: https://github.com/deepmind/alphadev/blob/main/alphadev.py#L943-L954
The resulting algorithm is substantially more complex than what you describe. In the case of a 3 item list, it uses 4 temporary registers, performs 3 comparisons, and has 6 conditional assignments, of which up to 4 may be executed for a given list. A list (a, b, c) which starts with c < b will still end up sorted ascending, small to large, and not inverted.
See here: https://github.com/deepmind/alphadev/blob/main/sort_functions_test.cc#L25-L41
[ + ] s23erdctfvyg
[ - ] s23erdctfvyg 1 point 10 monthsJun 11, 2023 04:00:43 ago (+1/-0)
I could be wrong, but I would like a proper sudo code for the five sort as that would show what method it is using.
[ + ] Mr_
[ - ] Mr_ 1 point 10 monthsJun 9, 2023 21:22:33 ago (+1/-0)