What content has been updated in OPENAI's GPT-4?

overview

ChatGPT has become the hottest AI in recent months, none of them, and its previous GPT-3.5 ability is only used until 2021, and it is not the latest product. On March 14, OpenAI, the development agency of ChatGPT, officially released GPT-4. According to the official website, this is a "milestone work". Let's take a look at the new features of GPT-4:

1. Stronger ability to connect with context

In the past, GPT-3.5 limited 4096 tokens, while GPT-4 expanded to 32768. Here is a knowledge point for everyone. Many people will understand token as the number of characters, which is not completely accurate. A token is a parameter that divides a piece of text into smaller units, which is convenient for AI to calculate. It does not simply refer to the number of characters. In the previous GPT-3.5, we will find that if we chat with GPT for a long time, it will go off topic, because it has a limited scope of connection context, only 4096 tokens as parameters, while GPT-4 has expanded to 32768, which is estimated to be equivalent to 64,000 words or 50 pages of text, so that we can repeatedly discuss and communicate with GPT-4 on a question, and finally get the answer you want.

2. Stronger logic and reasoning skills

Let's look at an example given on the official website:

This is an example of calculating the overlapping time of three people. We can see that GPT-3.5 gave the wrong answer to this question, while GPT-4 can fully understand and reason and give the correct answer.

3. Add image processing capability

According to OpenAI's official introduction, GPT-4 is a super-large multi-modal model, that is to say, its input can be text or images, and it can output text uniformly after processing. In the official example, a picture is given and GPT-4 is asked what is funny about this picture. GPT-4 can accurately analyze that the joke lies in plugging the VGA video cable into the mobile phone for charging.

4. Higher professional standards

According to official data, in various tests, GPT-4 will achieve better results than GPT3.5. The green column in the figure represents the improvement of GPT-4 compared to GPT3.5. We can see that GPT-4 has greatly improved its ability in calculus, programming, AMC, lawyers, English, chemistry, physics, economics, statistics, writing, biology and other fields, and can rank among the top 10% in test results and human rankings. The improvement of psychology, history, and environmental science is not obvious, because the level of GPT3.5 in these fields is already very high.


And, it also acts as a mentor. When GPT4 solves a problem, it can guide the user to follow the human train of thought and gradually guide the train of thought instead of directly giving the correct answer.

5. Stronger multi-language support

GPT4 has greatly improved the multi-language processing ability of GPT3.5. The official test method is to select six tests for AI as benchmarks, translate them into multiple languages ​​and compare the accuracy. According to the data chart given on the official website, the scores of each language have been greatly improved, with an average increase of about 80%, and the improvement rate for Chinese is 80.1%. Compared with GPT3.5, GPT-4 is more friendly to Chinese, and it can even be used as a new translation tool.

6. Human Feedback and Continuous Improvement

Compared with GPT3.5, which only uses data until 2021, GPT4 has manual feedback channels and is constantly adjusted and updated. This allows it to be kept updated and maintained over time.