Let AI generate AI painting prompts. OpenAI's latest achievement, ChatGPT, was broken by netizens! I can also write code, fix bugs and compose poetry.

A new chat AI is going crazy among netizens.

It can directly generate code, automatically fix bugs, conduct online consultations, and imitate Shakespeare-style writing... It can hold various topics. It is just launched by OpenAI - ChatGPT.

Some imaginative netizens even use it to design games: first use ChatGPT to generate game settings, and then use Midjourney to produce pictures. It is simply a one-stop AI productivity.

Some people also compared ChatGPT’s responses with Google search results, and were surprised to find that ChatGPT is much more practical: Google is going to be cool.

Some people even predict that if Google does not make changes, it will soon be eliminated. By then, they are willing to spend 15-20 US dollars per month to use ChatGPT.

But the good news is that ChatGPT is currently in the free trial stage, and friends who have the conditions can try it quickly.

How amazing is ChatGPT?

Some netizens asked Google and ChatGPT respectively "How to use Latex to express differential results". Compared with Google's satisfactory answer:

ChatGPT gave it at once Out of the directly usable code, you only need to copy and paste:

Some people also asked ChatGPT to give the worst time complexity of bubble sorting, and asked It was said in the tone of "the fast-talking smart guy in the gangster movies of the 1940s", and was easily picked up by ChatGPT:

Directly ask what the bug is in the following code , ChatGPT can also explain it clearly to you:

In addition to computers, ChatGPT also performs well in other fields, such as AI online consultation.

When you tell it that you have a rash, have difficulty breathing, and feel weak, ChatGPT will analyze the cause of the disease for you like a doctor and suggest that you need to quit smoking and exercise more than taking medicine.

Compared with the medical advertisements in search engines that need to be carefully distinguished, it can be said to be very considerate.

There are also many smart netizens who use ChatGPT for design as mentioned above.

For example, this netizen asked ChatGPT to design his bedroom, then used the text it gave as a prompt word, and went to MidJourney to generate the image, saving even the money of hiring a design company.

So, after introducing so much, what is the principle behind ChatGPT?

Principle behind

As the name suggests, ChatGPT is fine-tuned on a model in the GPT-3.5 series, which has completed training in early 2022.

It interacts in a conversational manner, with the ability to answer questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests.

Like InstructGPT released by OpenAI earlier this year, ChatGPT also uses reinforcement learning for training, but it is different in data collection settings.

In model training, OpenAI uses supervised fine-tuning: the trainer provides dialogue samples and plays the two parties of the dialogue, that is, the user and the AI ​​assistant.

In addition to this, trainers can also access model writing suggestions to help them write answers.

To create a reward model for reinforcement learning, the research team needed to collect comparative data. To do this, they randomly selected the information written in the model, sampled the alternative results, and then had the trainer rank the samples.

The model was fine-tuned through the reward model and proximal policy optimization mentioned above, and multiple iterations of this process were performed to finally arrive at the model.

However, despite its excellent performance in many aspects, researchers also admit that there are still some bugs in ChatGPT at this stage.

For example, some netizens pointed out after many attempts that ChatGPT often overuses certain phrases, such as repeatedly stating that "it is a language model trained by OpenAI."

In addition, sometimes ChatGPT's answers are too lengthy, and even when netizens tell it that they need a shorter answer, it actually says more.

Some rational netizens pointed out that it is not that easy for ChatGPT to replace Google.

A netizen who works at Alphabet said that Google has not considered citing large models in search engines, but the biggest problem is that the cost is too high.

If users are charged, the price is likely to be as high as 150 to 200 US dollars per month. I am afraid that few people are willing to pay this money.

However, OpenAI CEO Sam Altman expressed confidence in the future development of ChatGPT:

He said that language Interfaces will be a future development direction, and ChatGPT is a sample for everyone. I believe that soon everyone will be able to use an AI assistant that can chat with themselves, answer questions and provide suggestions.

Having said that, what other interesting uses have you thought of for ChatGPT?

If you have the conditions, please click on the link below and try it~

Guide link: https://chat.openai.com/

Reference link: [1 ]https://twitter.com/sama/status/1598038818472759297[2]https://news.ycombinator.com/item?id=33817682[3]https://twitter.com/jdjkelly/status/1598021488795586561

The above is the detailed content of Let AI generate AI painting prompts. OpenAI's latest achievement, ChatGPT, was broken by netizens! I can also write code, fix bugs and compose poetry.. For more information, please follow other related articles on the PHP Chinese website!