OpenAI releases new features: ChatGPT can now customize personality roles

① From the point of view of actual use, this function is equivalent to setting a "prompt word" at the system level;

② In addition to using the GPT model, customized instructions can also be used in patch mode.

Financial Associated Press, July 21st, on Thursday local time, OpenAI announced that it has added a new function called Custom instructions to ChatGPT: at the system level, it can customize some instructions for chatbots, making the robots more personalized and at the same time better close to the needs of users.

(Source: OpenAI)

After communicating with global users, OpenAI realized that in order to meet the unique needs of people from different regions and backgrounds, it is necessary to add some "manipulability" to chatbots. So this new feature is to allow users to set a series of requirements for chatbots themselves, and become the default background when opening a new chat.

To put it simply, this is to set the "prompt words" at the system level, which prevents users from having to adjust ChatGPT every time they open a new chat, including the two fields of "who is the user" and "what are the requirements for the robot".

In the official case, a teacher who teaches a third-grade science course can ask the robot to provide elementary school lesson preparation materials in the form of a table to clearly explain the advantages and disadvantages of discussing these topics in elementary school classrooms, such as whether elementary school students can understand these topics. Before this feature, ChatGPT would just throw out a bunch of proper nouns.

(Before and after using Custom instructions, source: OpenAI)

Similarly, users can also preset their identity and the instructions they expect the robot to follow in replying. For example, software programmers can ask ChatGPT to answer questions with specified codes (instead of the default Python) without explanation; users can also preset their own location, so that the robot will be more "down-to-earth" when recommending recipes.

OpenAI said that this feature has been launched on Thursday, and Plus users outside the European Union and the United Kingdom have already been able to experience it online, and it will be extended to all users in the next few weeks. In addition to using the GPT model, customized instructions can also be used in patch mode.

According to our actual measurement, the biggest effect of this function is to reduce the steps for ChatGPT to understand the results required by users.

For example, we let ChatGPT act as a psychological counselor with experience in the securities industry, and try to include charts in the reply. It was then posed a question about how to deal with investment losses.

In the case of not using customized instructions, ChatGPT will only say a bunch of words such as "you need to study more" and "focus on long-term investment value".

After the pre-set instructions, ChatGPT naturally assumed the identity of a "psychological counselor" and deepened its understanding of the securities industry. Of course, ChatGPT did not forget the instructions, and finally made a table to summarize the answers.

Of course, this is just a very shallow trial. In the comment area of OpenAI, a user said that this feature allows him to customize the writing style of ChatGPT and avoid copying and pasting a bunch of prompt words over and over again. There are also many users who expect that this is only the first step towards the personalized and customized route of ChatGPT in the future.