The use of open-ended Large Language Models (LLMs) in research poses significant risks, primarily due to their tendency to hallucinate or invent sources. To mitigate this, a closed-loop system has been proposed, where the model is trained solely on trusted raw source documents, effectively creating a smart search engine for a personalized library. This approach grounds all responses in verifiable documents, ensuring the accuracy and reliability of the research.

One method of implementing this closed-loop system is by utilizing an LLM Wiki, as suggested by Andrej Karpathy. Alternatively, AI knowledge bases like Recall can be employed to easily set up a closed retrieval system. This system guarantees that when an LLM responds to a question, its answer is strictly based on the uploaded PDFs and papers, thereby eliminating the risk of hallucinated sources.

The closed-system approach is essential for high-stakes research, as it provides a secure and trustworthy method of obtaining information. By limiting the model’s training data to trusted sources, researchers can ensure that their findings are accurate and unbiased.

Photo by Tima Miroshnichenko on Pexels
Photos provided by Pexels