Some netizens used GPT-4 to replicate AlphaDev's sorting algorithm optimization, making DeepMind's new paper look less dazzling.
This algorithm is said to create sorting algorithms that are 3 times faster than algorithms written by humans.
When the news first came out, many netizens thought that this was a major breakthrough in the field of machine learning.
But after a period of fermentation, some netizens' evaluations of AlphaDev's achievements seem to be divided.
First, a Twitter user said that GPT-4 was also successfully found to be almost the same algorithm as found by AlphaDev.
Then many netizens were also discussing, feeling that the actual results achieved by AlphaDev were too far from the "historic breakthrough" they boasted themselves.
These discussions from netizens even attracted Marcus to watch lively.
In addition, Academician Ma also replied to netizens who used GPT-4 to discover the algorithm, "Twitter boss here for a visit"
GPT-4 can also "discover" the same algorithm
Now let's take a look at how this netizen who used GPT-4 to discover almost the same algorithm as AlphaDev did it.
He popped his Prompt and GPT-4 replies.
By the way, can I send this thing Nature?
The prompt word of netizens has two parts, the first part allows GPT-4 to optimize this sorting algorithm, mark out which instruction can be deleted, explain the reason step by step, and then go back and verify again.
Then he gave the second part of the prompt, and asked GPT-4 to continue according to the above prompt, let Temperature=0 (keep the results consistent).
Finally, GPT-4 summarizes it a little.
Think that it is enough to delete "mov S P" and replace the P of the deleted code with S.
Algorithms discovered by AlphaDev
As shown in the example below, the original sort3 implementation has min(A, B, C), using AlphaDev Swap Move, AlphaDev found that you only need min(A, B).
Compared with the optimized content of netizens using GPT-4, it can only be said to be exactly the same!
I wonder if GPT-4 is too powerful, or if AlphaDev is really just that?
AlphaDev was overblown?
Marcus also reposted a paragraph on Twitter that YC community netizens commented harshly about the algorithm discovered by AlphaDev.
Users of the YC community, Orlp, point out that they were able to improve 70% of a certain libC++ algorithm mainly because the library has not been actively developed in the past 10 years.
In addition, DeepMind's improvements actually work because the library itself has some problems with the efficient implementation of branchless ordering networks.
Other users point out that this view is "too extreme" and that the algorithm's ability to automatically generate new sorting algorithms is already a great thing.
ORLP replied that while the algorithm is indeed capable of automatically generating good code, it is far from revolutionizing or improving the state of the art.
The main point of view of netizens is that the algorithm has not found a new sorting method, but only optimized the code.
However, there are already "super optimizers" that can achieve similar effects in different ways.
And there are also projects on GitHub that have completed a faster automatic search of the network, which seems to work better than AlphaDev's algorithm.
Many domestic netizens still choose to believe in AlphaDev, because GPT-4 is copied quickly, not necessarily original.
However, some readers do think that such improvements are not particularly significant.
Disclaimer: The graphics and text come from the creators of the Internet platform client and do not represent the views and positions of the creators of this platform. If infringement is involved, please contact our company to delete, thank you!