Short-term Memory of Gen AI

TL;DR When asking specialized questions, instead of asking directly, try attaching relevant documents and instructing the Gen AI to use only the knowledge from those documents to answer. This method significantly improves the accuracy of the responses.

Many people use Generative AI, such as ChatGPT or Gemini, for Q&A, similar to a general Google search 🔍. This method pulls information from the long-term memory of the Gen AI, which is the knowledge learned during model training. It's like recalling what we read last week before taking an exam 📚. The advantage is convenience and quick access, but the downside is that the remembered knowledge might be inaccurate or outdated, leading to potential errors in the answers.

To make Gen AI work with controllable, whether it's newer or more reliable knowledge, the concept of short-term memory emerged. Compared to humans, it's like taking an exam with summaries or books open for reference. This method greatly enhances the accuracy of the answers but also has a drawback: it slows down the response time ⏱️ because it involves searching for the answers in the documents. For Gen AI, it's similar. If we feed a lot of information, it will slow down its responses. However, compared to humans, even the slow response of Gen AI is still faster than us 💨.

Try it out! If you have specialized questions that require in-depth answers, change the way you ask. Attach relevant documents, such as a 100-page PDF, and request the AI to use the information from that document to respond. You might get more precise answers.

Additionally, if the accuracy is still unsatisfactory, try adjusting the way you ask Gen AI. Instead of just asking directly, instruct the AI to specify the source of the information it uses to answer. If it's a book, ask for the page number, or if it's a long article, ask for the exact text block. From personal experience, this method helps Gen AI provide much more satisfying answers 👍.

How did it go? Share your experiences in the comments!