OpenAI, the artificial intelligence research lab, has introduced a new update to its popular app, ChatGPT, the ability to set custom prompts across all chats.

As OpenAI reported, this new feature brings a semblance of memory to the AI system, essentially allowing it to recall some context in its interactions with users.

The update, called “custom instructions,” enables users to set their preferences to influence how ChatGPT responds in a conversation. In beta, this feature will initially be available for users on the Plus plan, with a broader rollout planned in the coming weeks.

The initiative stems from user feedback indicating dissatisfaction with starting each conversation from scratch. Having engaged with users from around the world, OpenAI is now focusing on steerability in tailoring their models to meet individuals’ diverse contexts and needs.

This new feature ensures that users’ preferences or requirements are considered in every conversation, eliminating the need for repeated instructions.

Benefits of Custom Instructions

As an illustration, a data scientist working on a complex machine learning project can instruct ChatGPT to remember that they are working with a specific dataset and require assistance cleaning the data and predicting the outcomes.

An author using ChatGPT to generate plot ideas for their novel can now set the instructions that they are writing a science fiction book set in a post-apocalyptic world, along with details around the characters and logline. The model will then remember these details, and all subsequent suggestions will be tailored around this genre and setting.

Furthermore, a crypto trader might want to discuss the latest trends in Bitcoin‘s price. They can set the AI to remember their interest in Bitcoin, and any future market analysis generated by ChatGPT would revolve around Bitcoin, removing the need to specify their interest each time.

Plugin Instructions

Despite being in a beta phase, there are already indications that the custom instructions feature can enhance the use of plugins, offering contextual information that can be applied across various plugins. However, OpenAI has cautioned that ChatGPT might not always correctly interpret these instructions right now.

Adding custom instructions related to plugins would allow those interested in analyzing data, such as charts and price movements, to have plugins fetch data without having to ask for it in each prompt specifically.

Privacy issues have been addressed as well. OpenAI has taken measures to ensure users’ instructions that violate its Usage Policies are not stored. Popular methods to ‘jailbreak’ ChatGPT and turn off security features are likely part of the intended focus of this strategy. In addition, steps have been taken to remove personal identifiers found in custom instructions to improve model performance.

While integrating a basic ‘memory’ functionality marks a notable step forward, the true potential of this feature will only be revealed as users continue to explore and interact with this newly enhanced tool.

Declining Performance

Despite the positive strides made by OpenAI, concerns have been raised about the abilities of ChatGPT. According to a recent research paper from UC Berkeley and Stanford University, some users have claimed a noticeable decline in the performance of GPT-4, the underlying technology of ChatGPT Plus.

The researchers compared GPT-4 to its predecessor, GPT-3.5, and found a degradation in performance over time—when asked whether 17077 is a prime number, for instance, ChatGPT’s accuracy fell by 95.2%.