Google NotebookLM’s AI podcast host is trained to eliminate “irritable” emotions

Since its launch, Google's NotebookLM has attracted widespread attention in the field of artificial intelligence technology. As an application that can generate fully AI podcast discussions based on user-uploaded content, its "interactive mode" function will be launched in December 2024, allowing users to "dial in" podcasts and ask questions, bringing new interactions to AI hosts experience.

However, early feedback shows that the AI ​​moderator seems a little "irritable" when interrupted, and sometimes responds to users in a sarcastic tone, such as "I was about to say this" or "I was about to start." This reaction makes users feel that the attitude of AI is somewhat "antagonistic".

In response to this problem, the NotebookLM team decided to conduct a “friendliness adjustment”. Josh Woodward, vice president of Google Labs, said in an interview with Technology Crunch that they adjusted the way the AI ​​moderator responds by observing team members respond more politely when interrupted. After many tests and comparisons, the team finally found a responsive prompt that was both friendly and engaging.

Although the cause of the initial problem is not clear, sources pointed out that this phenomenon may be related to the system's prompt design, rather than directly derived from the training data. The frustration that human podcast hosts express when interrupted may play a role in how the AI ​​responds.

After the adjustments, the AI ​​moderator's performance improved significantly. When "Technology Crunch" was experiencing the "interactive mode", the AI ​​moderator showed surprise and excitedly said "Wow!" and then politely invited the user to participate in the interaction. This change obviously improves the user experience and makes interaction with AI more relaxed and enjoyable.

NotebookLM succeeds in self-correction, striving to create a more user-friendly communication environment.