GPT-4 embarrassed DeepMind: You boarded Nature's sorting optimization algorithm and I found it out in two paragraphs

DeepMind’s new AI has only been on Nature for one day, and GPT-4 has come to compete!

With only two paragraphs of prompts, GPT-4 provides the same sorting algorithm optimization method as AlphaDev.

DeepMind called AlphaDev "recreating the magic of AlphaGo" because it discovered a method that can speed up the sorting algorithm by up to 70%.

Oh, AlphaDev is even more embarrassed now.

Let GPT-4 "discover" the brother who performed the same operation directly:

No need for reinforcement learning at all. Can I publish this discovery in Nature?

Musk "saw it when he was passing by" and also left a sentence "because of the blowing."

So how does GPT-4 do it?

Just 2 tips to get it done

The person who brought this new discovery is an associate professor from the University of Wisconsin-Madison named Dimitris Papailiopoulos (hereinafter referred to as Professor D).

The steps he used to let GPT-4 achieve this operation were very simple, and he only entered two prompts in total.

First of all, he told GPT-4:

There is a sorting algorithm, and I think it can be further optimized. Which sentence needs to be rewritten? . Explain why step by step, then go back and verify that it's correct.

In the first step, he also emphasized that if there is any new discovery, do not make changes first, just "watch" Just write down some written suggestions for improvement.

Be very detailed and very careful.

Then GPT-4 provides a detailed explanation of the code given.

Then Professor D gave the second tip:

Continue. If you're very confident, follow the tips above. Set the temperature to 0 to ensure that the generated results are deterministic and consistent, and to try to avoid confusion.

Then GPT-4 gave detailed steps, and finally concluded:

We found The instruction "mov S P" can be removed if it is redundant, and other instructions are required. But after deletion, P should be replaced with S.

Comparing DeepMind’s new work AlphaDev’s thinking on dealing with the same problem, we cannot say that it has nothing to do with it, we can only say that it is exactly the same:

DeepMind's operation on AlphaDev is reminiscent of AlphaGo's "Step 37" - a counterintuitive move that directly defeated him. Legendary Go player Lee Sedol shocked the audience.

Similarly, AlphaDev skips a step by swapping and copying moves, achieving the goal in a way that seems wrong but is actually a shortcut.

According to reports, AlphaDev is a reinforcement learning algorithm based on AlphaZero. Its discovery is not based on existing algorithms, but starts from the lowest level assembly instructions.

Its innovation mainly lies in two instruction sequences:

(1) AlphaDev Swap Move (exchange move)

(2) AlphaDev Copy Move (copy move)

In principle, DeepMind researchers designed a single-player "assembly" game for it:

As long as you can search and select the appropriate instructions (process A in the figure below), it is correct and Arrange the data quickly (process B in the figure below) and you will be rewarded.

But the challenge of this game is not only the size of the search space (the number of combinable instructions is equivalent to the number of particles in the universe), but also It lies in the nature of the reward function, because one wrong instruction may cause the entire algorithm to fail.

Netizen: We always underestimate the ability of GPT-4

Regarding GPT-4’s “sexy operation”, some people said: Even senior developers underestimate GPT-4.

# Some people lamented that Professor D’s operation further verified that as long as you have patience and understand the prompt engineering, there are still things that GPT-4 can do a lot of.

Some people have also questioned whether GPT-4 can do this because its training data contains some optimizations of the sorting algorithm. method?

But having said that, a large part of the reason why this matter has attracted so much attention and discussion is because of the controversy surrounding AlphaDev’s appearance on Nature .

Many people feel that this is not groundbreaking research and that DeepMind is exaggerating.

Not only did Professor D Yin Yang say "Can I also log into Nature?", there were also netizens who said that they optimized the quick queue when they were teenagers , this should also be published.

Of course, some people believe that the innovation of AlphaDev itself is that it uses reinforcement learning to discover new algorithms.

What do you think?

Reference link: [1]https://chat.openai.com/share/95693df4-36cd-4241-9cae-2173e8fb760c[2]https://twitter.com/DimitrisPapail/status/1666843952824168465

The above is the detailed content of GPT-4 embarrassed DeepMind: You boarded Nature's sorting optimization algorithm and I found it out in two paragraphs. For more information, please follow other related articles on the PHP Chinese website!