In the surging wave of AI, whether sticking to doing things “by hand” is a form of romance or merely the stubbornness of using a horse carriage in the automobile era remains to be seen. However, if AI can foster the development and application of innovative thinking, its contribution to efficiency is undeniable, regardless of skepticism. This article offers some humble opinions on the use of various AI models based entirely on my personal experience.
Since the release of Gemini 2.5, Google’s AI has become the mainstay of my daily usage. Its knowledge base is stunningly rich. Beyond solving various research problems, I find it increasingly difficult to detach from it in daily life—whether for travel planning, product selection, or diet strategies.
However, in my experience, when I try to ask it for novel ideas, the results tend to be somewhat mediocre, and its ability to solve mathematical problems feels slightly inferior to GPT.
Verdict: Best used for learning new knowledge in research and as a partner for study discussions.
In the AI era, how to better utilize these tools is a topic of tireless discussion. Tools like Manus, through unique ideas and process control, can deliver impressive performance even with a weaker underlying model. The reason lies in how process control helps the AI better understand the problem and maintain consistent output.
How do we maximize AI utility? The key is “Understanding.”
To let AI understand, we must articulate requirements as clearly as possible and provide comprehensive prerequisite information (priors). AI needs to generate output while “knowing” the context. Fundamentally, LLMs operate within endless sequences of matrix information, transforming the information space of the prompt into the information space of the response. Countless multiplications activate pathways similar to neurons.
When we provide information highly relevant to the task, our goal is to activate those specific “neurons” that might otherwise remain dormant. This guides the transformation of the output space—forged by feeding it human knowledge—towards a refined final version.
With this understanding, one can master GPT. As one of the earliest famous LLMs, GPT is unique in the depth of its answers. We must use it well; much like the debate between GUI and Terminal, GPT is like the Terminal (though Claude feels more like it in experience, here I refer to the quality of output). We need to think carefully about how to structure our prompts to elicit profound responses. It may not be as “empathetic” as Gemini, but trust me, the payoff from using it well is immense.
Verdict: The key is asking the right questions. Best used for discussing new research directions and broadening horizons.
A formidable Coding Expert and the best assistant for code. But beyond that, what are its overlooked strengths?
Returning to coding: we know that great code requires a structured, clear, and definitive requirements document; otherwise, the result may be miles apart from the ideal. In the early internet industry, specific roles existed just to write requirements, bridging the gap between the client/demand side and the programmers. “Understanding” is truly humanity’s hardest goal, and it is also the key to conversing with AI—making the AI understand you.
In my view, Claude’s strongest capability is writing this “Technical Documentation”—understanding the task goal, decomposing the steps, and executing them without deviation. This ability allows it to dominate the coding world even if its base model isn’t the absolute strongest, effectively replacing many frontend developers because its ability to realize an ideal interface far exceeds most human workers. It relies on understanding the client’s needs and completing every step impartially. Keeping AI consistent is not an easy task.
So, answering the initial question: what is its greatest merit in research? It is the ability to execute commands with structural stability.
I use it to quickly summarize the writing techniques of the best literature in my field (the flow of the abstract, the bridge between title and content) and form a “technical guide.” I then use this guide to polish my own writing. We can even feed this guide back to Claude, asking it to analyze our own papers against these points and prescribe a remedy. This is truly doing more with less. Its ability to follow steps methodically also saves us from worrying about unstable AI outputs.
Verdict: The best assistant for paper revision and polishing. A tool that helps us understand why good papers are good.
“Talent emerges in every generation, each leading the way for hundreds of years.” Using AI well is absolutely the best way to help us conceive ideas and write excellent papers. Therefore, analyzing and learning these tools is necessary.
Science is a way of thinking, not just a natural discipline. Ge Wu Zhi Zhi (acquiring knowledge by investigating things)—I wish to encourage myself and all of you with these words.
P.S. This article was written and polished without the use of AI.