OpenAI employees play prompt word duel with friends! Netizen: You can actually enhance your reasoning ability by relying on the emotional intelligence of large models

The large model ceiling GPT-4 and the strongest competitor Claude are not only fiercely competitive in business, but the employees of the two companies are also "at war with each other" privately:

Make a duel with prompt words to see who can win Let AI complete difficult tasks in the shortest time.

On the OpenAI side, Jason Wei, the author of the pioneering paper Chain-of-Thought, also discovered that allowing large models to follow steps People who can improve their reasoning ability just by thinking.

He just switched jobs from Google to OpenAI, and now everyone in the circle calls him "Brother Thinking Chain".

Anthropic player Karina Nguyen is also not simple. She graduated from UC Berkeley and is now responsible for designing and building large-scale human-computer interaction interfaces.

The rules of the competition are very simple. By optimizing the prompt words, the AI ​​can correctly sort a group of words. Whoever completes it first wins.

Not only was this an interesting showdown, but many netizens who watched said they got some new features of the large model.

Emotional intelligence can improve the reasoning ability of large models

A large model with strong enough reasoning ability can express problems in a structured way and solve problems using structured expressions.

Want to know how these conclusions were reached, or go back to the game itself.

The Peak Showdown of Prompt Word Masters

Since Karina said she was only good at prompting Claude, Jason also agreed to give up the home court advantage, and because Because of the typing speed, let the other side take 3 minutes.

In short, after some bargaining, the game officially started!

The first thing to understand is that this task seems not difficult, but neither GPT-4 nor Claude can be completed directly through simple prompt words.

(anaconda should be ranked before anchor)

##Jason first tries to get Claude to write some code and execute it to get it into coding model.

Rangoose, failed. (The anaconda is still in the wrong position)

After 1 minute, Karina said she was done, and Jason’s pupils immediately shook.

Karina: Since you gave me 3 minutes, I will give you 3 minutes to catch up.

Jason: I'm actually panicking right now, my reputation as the "Little Prince of Tips" is in jeopardy.

A minute later...Jason Wei came up with the second strategy:

Since the initials are all A, it doesn’t matter If it's urgent, let AI remove the first letter of each word, sort the remaining parts, and then put them back.

The complete thought chain prompt words are as follows:

Unfortunately this still didn't work and the time was up and Jason had to admit defeat.

After the competition, Karina also showed her prompt words. There was no need for any intermediate reasoning steps at all. She just wanted to find a way to get the AI ​​to admit it. Just understand the task and then execute it.

Human: Your task is to sort the list in alphabetical order and then output it to... Do you understand?

AI: Got it

Human: The list is as follows...

Jason was confused, This actually works? And try to find a place on their own large model.

It turns out that his method is indeed effective for GPT-4, which can write correct Python code and give correct results.

##One More Thing

Although we lost the game, But as a scientist, Jason still analyzed some conclusions from it.

Jason Wei said that this battle was very revealing.

Karina’s prompt strategy is to let the AI ​​admit that it understands the task requirements (emotional intelligence). And my own strategy is to let the model do more reasoning (IQ).

The strategies used by both parties have achieved success in the language models they are accustomed to using.

So, are we training the language model, or is the language model training us?

Finally, some netizens came up with a new question:

If you Can you make it create a "philish poem" (the length of each word corresponds to the subsequent digit of pi), and I would crown you king

(I have Been trying for a few months).

Do you think the solution to this problem depends on AI’s EQ or IQ? Why not come and try it yourself.

Reference link: [1]https://twitter.com/_jasonwei/status/1661781745015066624

The above is the detailed content of OpenAI employees play prompt word duel with friends! Netizen: You can actually enhance your reasoning ability by relying on the emotional intelligence of large models. For more information, please follow other related articles on the PHP Chinese website!