Recently, LeCun, the father of deep learning, strongly recommended an experience sharing about using the large language model (LLM) tool GPT-4 for scientific research. This sharing explains in detail how to use LLM tools efficiently and is worth reading.
In addition, some pitfalls that need to be paid attention to when using LLM tools are also mentioned in the sharing, including:
Do not ask the LLM for information or tasks that you cannot verify yourself, the only exception being some critical tasks. For example, asking for ideas on apartment decor is fine, but for information that requires verification, you need to use literature review best practices, such as requesting a list of top reviewed articles, which can verify source and reliability.
Setting the context and using terminology and symbols allows the LLM to be biased towards the correct contextual information. In addition, explicitly tell the LLM what information should be used, such as using the Cauchy-Schwarz theorem to solve inequalities, etc.
Define the output format, such as code, mathematical formula, article, etc., and you can also ask for code, tables, plots, diagrams, etc. that generate specific content.
Finally, the output needs to be verified, including finding inconsistencies, obtaining supportable sources through Google search tool output terms, and writing code to test when possible. This is because LLMs often make mistakes that are inconsistent with their seemingly professional level.
In summary, efficient use of LLM tools can bring great help to scientific research, but it needs to be used with caution and the reliability of the output content must be verified.
