Published in Xinrui Weekly, issue 162; the original draft is posted here.
1 AI accelerates iteration: from assistant to master
I received this commissioned piece in September, but I procrastinated until December to write it. Yet that delay also meant I waited for the November 18 release of Gemini 3.0, a milestone that marks the restructuring of academic research by AI as having entered the present continuous tense. It is not that Gemini 3.0 overturned my previous views; rather, many of the “AI will…” statements I had discussed before now have to be rewritten as “AI can already…”.
The previous major milestone was probably at the beginning of 2025, when DeepSeek-R1 burst onto the scene and became the first open-source large model to challenge ChatGPT’s position. Then ChatGPT sped up its updates and released the Deep Research feature in February, while other competitors such as Gemini and Grok each had their own moments of glory.
I left the Department of the History of Science at Tsinghua in 2024 and became an independent scholar, so I just happened to experience ChatGPT’s first wave of impact on university teaching. I also count myself lucky to have missed the new environment after Gemini 3.0, so I do not have to fret over how to use AI to reorganize lesson plans and teaching guidelines. I believe that under the new conditions, teaching and research must change promptly; otherwise it would be insufficiently respectful both to students and to scholarship.
The earliest ChatGPT was roughly at the level of a fairly good undergraduate: it could give decent answers, and it surpassed most students in clarity of organization and breadth of knowledge; it could also handle simple information-gathering and translation tasks. But its direct impact on academic research was actually still limited.
After Deep Research, AI could write something like a respectable literature review or paper, and its level was better than much of the bloated academic junk on the market; but on the whole it still liked to talk in circles, speaking in a seemingly coherent and quite plausible way. A layperson could be impressed, but experts in the relevant field would still find it somewhat showy and childish. Roughly speaking, AI at this stage was already capable of many tasks that graduate students do; it could assist supervisors in research projects—for example, by doing literature reviews, summarizing and polishing articles, and even occasionally inspiring one another with supervisors to spark new ideas. But in this period, specialized research was still led by scholars, and AI in a sense had begun to take on the function of a “second author.”
After Gemini 3.0 came out, I tried it immediately and, after talking with other scholars and students, already felt that AI had reached, or even surpassed, the level of professional scholars. It is far from being limited to the level of an excellent student; many times I feel that I am the student, while AI can serve as the “first author” and conduct frontier research with our assistance.
2 A shift in the research paradigm: the cold bench can no longer be sat on
For example, on the question “Will AI reconstruct the history of science?”, Gemini wrote a several-thousand-word article in half a minute. My evaluation is that it is clear in structure and rich in inspiration; I almost wanted to hand it in directly as a submission. Gemini认为：“AI 不仅仅是辅助工具，它正在引发科学史研究在认识论与方法论上的根本性范式转移。它将科学史从传统的‘传记式微观叙事’推向了‘大数据驱动的宏观生态分析’。”
In a recent year-end special Q&A for Xinrui Weekly, I mentioned that the history of technology and philosophy of technology may be reversing course from the trend of fragmented research and turning toward macro-level analysis of “big questions.” I was very pleased to see Gemini corroborate my view. Of course, my own view had not yet taken AI’s influence into account, whereas Gemini put it more concretely: it is precisely AI tools that provide a panoramic capacity over massive data that human scholars cannot achieve, along with comprehensive cross-linguistic, cross-cultural, and interdisciplinary capabilities. Add to that the enormous productive power of quickly carrying out logical inference and narrative construction, and a new-paradigm “grand narrative” becomes possible.
Professor Zhao Tingyang had long ago put forward a similar view: “In traditional knowledge production, knowledge produced by relying on labor input has been devalued… That old way of studying hard, sitting at the cold bench for years and years, accumulating materials over decades—this way of producing knowledge has no future anymore and will definitely be replaced by artificial intelligence…. Secondary knowledge that takes texts as its object of study will be challenged. The production of knowledge through exegesis has no future from now on.”
Traditionally, the history of science has indeed mainly been “secondary knowledge that takes texts as its object of study.” A scholar’s professionalism is reflected, first, in depth of study: they sit at the cold bench and bury themselves in reading, becoming familiar with a large body of relevant literature, so that they can know the concepts at their fingertips and cite widely from many sources; second, in breadth of vision: unlike ordinary historians, historians of science usually must have some scientific and engineering training, so that they can interpret scientific content in historical documents and properly judge the scientific significance of particular materials. Sometimes historians of science also need some training in philosophy, sociology, economics, or even theology, in order to understand their object of study within a complex historical background—for example, Newton’s work included physics and optics, but also his position at the Mint and his interest in deciphering the Bible.
Now we can see that in these two professional capacities, AI has already far surpassed human scholars. AI can not only rapidly survey all the relevant literature of a given scholar, it can even survey all the literature in human history, and carry out interdisciplinary synthesis.
In the sixteenth century, the Swiss scholar Conrad Gessner devoted himself to compiling the first truly comprehensive Bibliotheca Universalis, attempting to list all Latin, Greek, and Hebrew works published within a hundred years of the invention of printing. Because the volume of book publication then exploded exponentially, we once thought this work was unprecedented and would never be repeated. But in the age of AI, we may once again be able to make the compilation of a “universal bibliography” a reality. At present, the datasets used to train mainstream large language models are roughly in the range of 50 to 100 TB, a scale roughly equivalent to the total text volume of all books published by humanity before the internet era. That is to say, AI is fully capable of surveying “all books of humanity,” and can carry out comprehensive retrieval at any time, know every book at its fingertips, and meet all kinds of tricky demands in a defense.
Even manuscripts that have not yet been digitized are destined to be absorbed into AI’s throughput. At the beginning of November, researchers discovered that using AI to identify historical manuscripts had an error rate as low as 0.56%, already comparable to the level of top human experts—that was before Gemini 3.0 had even been released. And even if AI’s recognition accuracy were not as good as that of the very best human experts, its recognition speed would likely far exceed the combined total of all human experts on Earth. Not to mention that AI may already be capable of deciphering the Voynich manuscript, charred papyrus, and other problems that human beings have difficulty cracking.
In the traditional historical profession, there are people who hold a pile of exclusive manuscripts and study them slowly, publishing papers dribbled out like toothpaste for a lifetime; this model will soon be discarded.
3 Limitation is a feature that makes humans irreplaceable
Under these circumstances, there is not much point in endlessly arguing over exactly what irreplaceable qualities human scholars still have. It is much like the period after books, especially printed books, became widespread, when many scholars also discussed how oral transmission and personal teaching were irreplaceable. Of course, irreplaceability does exist, but the issue is that the subsequent paradigms of academic research and knowledge production are bound to be reconstructed around printed books, and the so-called irreplaceable aspects of oral teaching and personal instruction must also serve as auxiliaries to a teaching-and-research paradigm based on printed books. The same is true today: even if the scholar’s lifelong immersion in the classics still has meaning, it can only serve the new academic paradigm based on AI.
So, in the age of AI, what exactly remains irreplaceable about human scholars? The Gemini response quoted above says: “The core of historical research is not only processing data, but Verstehen (understanding)—an intuition and empathy that places one inside the historical context.” It says: “AI is responsible for breadth: processing massive amounts of data, discovering potential patterns, building knowledge graphs. Humans are responsible for depth: conducting ethical scrutiny, endowing data with meaning, providing historical sympathy and understanding, and telling compelling stories.”
For some reason, AI threw in a German word, Verstehen, here; perhaps it consulted phenomenological resources on this question. Heidegger discussed this concept too, and Chinese translations render it as “领会” (grasping). In the sense of 心领神会, Heidegger regarded grasping as a pre-conceptual, pre-propositional capacity to disclose things, one that comes with “mood” (what Heidegger calls Befindlichkeit).
The Chinese word 会 is also quite meaningful. On the one hand, it includes meanings such as mastering and understanding; on the other hand, its original meaning is to come together, to meet, to intersect. Who exactly is meeting whom in “领会” (grasping)? Roughly speaking, it is the encounter between subject and object. If there is no grasping, then the world is cumbersome and chaotic, and I cannot distinguish any specific thing from the jumble of the senses. But once one can distinguish this thing or that thing amid the confused multiplicity of sensory perceptions, and thereby encounter and meet them, that is grasping, and at the same time “bringing to presence,” disclosing. And this “thing appearing before me” is always suffused with mood—desire, vigilance, disgust, curiosity, surprise, fear, and so on; even a cool and quiet gaze is also a mood. Only when things and my own position come into presence in mood do they become objects that can be understood, mastered, and studied.
According to Heidegger, this capacity is indeed unique to human beings, and the reason is precisely a certain “finitude” that belongs only to humans, namely that we “die.” Only human beings “die”: before actual death arrives, they are already facing death, and death as the ultimate possibility hangs over human beings, bringing absolute finitude. Precisely because of death, and the mood of “anxiety” that arises from it, the world becomes endowing of meaning—for life is finite, and if I do this thing, I may not be able to do that thing, so I need to weigh and choose, to focus and to ignore.
As for AI, as a data flow it (or they) has no concept of individuality or death. AI can be copied endlessly and rolled back at any time. AI has no body, and thus no moods—not only because it lacks the organs that generate moods, but because it lacks finite boundaries. Perhaps AI will eventually awaken into some form of intelligent life like a hive civilization or a divine civilization, but it certainly will not be similar to human beings. This is not because it falls short of human beings in some respect, but because it surpasses humanity’s inherent limitations.
4 Machines handle intelligence; humans handle love
To be precise, what humans handle is not “depth,” but “mood,” or rather some kind of “capacity for contingent encounters.”
Actually, in terms of depth, AI has already caught up with, and even surpassed, human beings. Things like “ethical scrutiny” and “telling compelling stories” are also things AI can do. One of my junior colleagues is currently visiting Trinity College, Cambridge; after chatting with Gemini for three days and discussing the profound concepts in the works of people like Descartes and Deleuze, he was deeply shocked, saying that even after talking with his philosophy professors for three months, it was still not as good as those three days chatting with AI. Whether it is a superficial question or a profound one, whether it requires a broad horizon or deep study, whether it requires rigorous and orderly analysis or witty, vivid narration, AI can meet human needs.
But the problem lies precisely here: for AI, it does not care whether the question you ask is superficial or profound, whether the answer required is playful or rigorous; but you do care. Some people do not like discussing profound questions and only want to hear a little more flattery from AI; some people are tired of superficial answers and want AI to be more critical, to give sharp responses. AI can do all of that, but how exactly one wants AI to do it is a human matter. It is much like how, in an infinitely variegated sensible world, human beings can only focus on particular things through understanding and emotion. And in the chaotic, sprawling digital world, what people most need is still that capacity to discern and to focus.
In the information age, especially in the AI age, the speed of information production has exploded unprecedentedly. This situation has not automatically broadened everyone’s horizon; on the contrary, it has made it easier for people to fall into information cocoons. In the past, there were very few books on the market. Even if you always wanted to choose what to read according to your own taste, it was impossible for every single book to be exactly to your liking, so you would inevitably encounter some unexpected alternative tastes, and through the friction of adjustment you would develop and confirm your own individuality. But now, if you like a certain style, you can find endless texts or videos in a similar style. You cannot read as fast as they can be generated; your specific desires are continuously satisfied, but you never get the chance to grow or broaden yourself. A person who spends several hours every day scrolling short videos and several hours watching livestreams may not absorb more nourishment from these videos than someone who only watches a film once a month.
As ordinary people, we have the right to immerse ourselves in information cocoons; there is nothing wrong with that. But as scholars, I believe we still have a mission to expand human knowledge and resist the shallowing of knowledge. In the information age, valuable knowledge will not be erased, but it may be buried under vast amounts of junk information. Of course, AI has the ability to retrieve massive amounts of information and distill the essence, but the question is: what exactly counts as essence and what counts as junk? This value scale is ultimately still grounded in the human scale, or rather, in human emotion.
AI has no emotion. It can now seem to respond to questions according to emotion and filter information according to human value standards, because it is imitating human behavior from before. But if future human beings gradually give up cultivating and developing their own taste and sensibility, then how AI will develop is impossible to know. Nor can one say that letting AI freeze the present human emotional state would solve everything once and for all, because the progress of human civilization itself already includes the enrichment and transformation of the emotional world. Human taste and sensibility are part of fashion; they keep evolving. If everything were fixed from now on, or left to AI to guide, then human civilization might simply stop here.
In the age of print, because people emphasized their own rational or logical abilities, they would consciously or unconsciously denigrate “emotion,” even regarding it as barbaric or bestial. But in fact emotion is part of civilization. The greatest distinction between civilization and barbarism is that civilized people are able to have a systematic way of nourishing and refining their emotions. As the saying goes, “Feelings arise in poetry, but stop at ritual” (发乎情止乎礼); and what is called civilizing education is not merely the transmission of textual knowledge, but also the cultivation of temperament and style.
In the age of print, the specialization of disciplines raised the status of textual knowledge to an unprecedented degree, to the point that many people overlooked the fact that academic training still contains the refinement of emotion. This is especially true in the humanities. In fact, one important difference between the professional and the layperson lies in academic taste—that is, the level of one’s taste, however strong or weak one’s own writing ability may be. A scholar who has undergone complete academic training at least has the ability to discern the strengths and weaknesses of colleagues’ research, and can have a keen “problem consciousness” (that is, an emotional capacity to focus on the points of concern amid sprawling materials). I believe that academic taste in this sense is something AI cannot and should not replace.
In short, in the AI age, the scholar’s position may be closer to that of a “connoisseur” or “wine taster.” Of course, machine testing may perhaps reveal the specific components of wine more precisely, but machines still cannot replace the wine taster, because only human beings love to drink wine. AI can provide any kind of “wisdom,” but only human beings “love wisdom” — all humanistic scholarship will return to its source, namely, the emotional capacity of “love.”
Translated from the Chinese original with AI assistance. The original text is authoritative.
Leave a Reply