====== AI/LLM Tools ====== ===== About ===== * A brief guide to using large language models (LLMs) often referred to as artificial intelligence (AI) tools, such as ChatGPT/Claude/Gemini, in the context of academic research and writing. By Chih-Horng Kuo (chk@gate.sinica.edu.tw). Suggestions are welcome. * Target audience: undergraduate students, graduate students, and postdocs (in biology) ===== Preface ===== * Privacy, both regarding institutional rules and personal reasons, is important * For work: Ethics first! * Respect the general rules of academic integrity * As the user, you are responsible for the final output of your work * Include proper usage statements * For personal use: * Emotional support can be valuable and beneficial; graduate school (or life) is hard! * Should not be mistaken for true care or therapeutic depth * Be careful of potential misuse and harm ===== LLMs Explained ===== ==== What are LLMs? ==== * Key concept * LLMs work by estimating probability relationships among tokens, not by indexing content like a database such as Wikipedia * Key terms * Token: the basic unit of text that LLMs process; typically equivalent to or smaller than a word * Prompt: user input * Session: a series of user input and LLM output * Good practice: limit the scope and length of individual sessions * Memory: information kept from previous interactions; some systems provide this function * Artificial intelligence systems trained on large collections of text to generate human-like responses * Human: form thoughts first, then use language to communicate those thoughts * LLM: generate language by predict the next word (token) based on statistical patterns * LLMs do not “think” or “understand” like humans do; the appearance of “thought” is a byproduct of those patterns * LLMs mimic intelligence, but are not truly “intelligent”; can be very convincing! ==== Value ==== * Always-available conversation partners * Often, just having a conversation helps you think much more clearly * Particularly helpful in the context of emotional support * In some ways, LLMs are like crystals of human knowledge * Useful for learning new topics efficiently through interactive Q&A * Perform especially well on topics with abundant, high-quality training data * Highly proficient in language, particularly useful for clear and precise expression of ideas * Great tools for self-improvement * How much can LLMs be trusted? I don't know * Have I successfully used LLMs to achieve deeper thinking and clearer writing? YES! * What's the catch? Time and effort ==== Limitations ==== * Limitations and biases exist in both the training data and the model development process * English is the dominant language * Often optimized to be agreeable and non-confrontational * Limitations of human knowledge and available text data sets * LLMs cannot assess the quality of sources * Cannot reliably distinguish between high- and low-quality content * Cannot apply appropriate weighting * Academia: "good" vs. "bad" papers * Outside of academia: biased information from paid advertising and other influence * LLMs do not know what they do not know; may make up false info (AI hallucination) * For scientific writing * Summarizing and explaining existing knowledge often work well * May not be suitable for processing novel findings ==== Common Failure Modes ==== * Superficial content * Oversimplification * Overstatement * False coherence * Hallucination; fabricated web links, DOIs, and PubMed IDs * Non-specific terminology for highly specific topics ===== Core Principles for Usage ===== * Use LLMs as tools to sharpen your thinking, not black boxes for quick answers * Do not outsource your thinking process! * Use LLMs in your thinking and learning: you become better * Use LLMs to think for you: you skip the practice, bad for learning * LLMs may not save time * In fact, LLMs are most valuable when used to deepen the thinking process * Explore possibilities, compare alternatives, and refine nuances all require time and effort, but the process and end product can be rewarding * Example: You are deciding the title of your manuscript. You provided the abstract, do you ask LLMs to: * (1) suggest a title, or * (2) suggest three titles, compare the emphasis and tone, refine to express accurately, adjust based on the target audience * LLMs act like mirrors; their responses reflect your input and framing * Trained to be agreeable, they often conform to both your explicit instructions and implicit tone * Helpful to make your implicit thoughts become explicit and better organized ===== Suggested practice ===== * Start with a solid frame of reference * Provide clear context for the task, including your goals and expectations * Learn the basic of markdown can be very helpful ("help me to learn the very basic of markdown language in 5 minutes, explain to me how can I use this in my prompt to improve the clarity of my instruction to you") * What exactly do you want to obtain? * How do you want the LLM to act? * Are your expectations realistic? * Examples of possible uses (varying in feasibility and reliability): * Polish arguments * Organize thoughts * Brainstorm new ideas * Identify major gaps in reasoning * Provide critical review * Assess the output, make decisions, and provide feedback * Use an iterative process; may involve gradual refinement or drastic changes * Fact check, fact check, and fact check!!! * Know when to stop; did the last iteration provide meaningful improvement? * Exercises * Use figure legends as input, ask for a manuscript outline as output; explore different ways of ordering the figures * Use a poster abstract as input, ask for possible titles as output; compare how each emphasizes different aspects of the work or addresses different audiences ===== Human–AI Dynamics ===== * LLMs mirror your tone, assumptions, and style * Output can become overly agreeable, especially in long sessions * It is natural to feel defensive when an LLM points out flaws in your argument * However, LLMs are easily swayed by your rebuttals * Reinforcing your own biases is not useful * Asking LLMs to role-play can be useful * Not because the roles provide true expertise, but because the exercise can stimulate YOUR thinking * For example, asking an LLM to act as a qualifying committee member or manuscript reviewer may help you prepare * In these roles, LLMs can mimic tone and style, but they lack the knowledge, logic, and Judgment of human experts ===== Use Cases ===== * Force LLMs to criticize and challenge your thoughts * Use the "Deep Research" function (available in both ChatGPT & Gemini) for targeted search of PubMed and follow-up summary * Help to read dense papers * Build custom databases of selected references using NotebookLM by Google * Perform grammar check and copy editing * Prepare for discussion with seminar speakers; info about you and the speaker as the input, suggestions of questions to ask as the output * Draft or refine emails (with tone sensitivity) * Obtain technology-related suggestions, such as server/intranet setup, hardware purchase planning, Linux/NAS configuration, or organization of computer files and directories ===== Practical Advice & Next Steps ===== * Test the tool with topics that you know well first for evaluation * Push boundaries, learn the situations that LLMs are useful (or not) * Different tools for different uses; 2026 impressions: * ChatGPT is chatty; highly conversational, discussions can take unexpected turns, great for exploration * Claude is often more "get to the point" compares to ChatGPT * Gemini integrates Google Search; useful for questions with clear answers and you want the web sources * Learn to judge output, know when to reject the suggestions * Over time, thoughtful use of LLMs can improve your writing, reflection, and decision-making skills * AI/LLM tools are under intensive development, different tools may perform better at different tasks; develop your own taste and workflow, keep an eye on the development * More advanced tools, such as Claude Cowork/Code, are not covered here, but definitely worth exploring