在1024 模力工场 · AI社区日主论坛上，西湖大学工学院助理教授、博士生导师、高效智能计算实验室（ENCODE Lab）负责人王欢，以《AI 智能体对科研的影响》为题作主题分享。
王欢有两个重点研究方向：Efficient AI（高效人工智能），关注如何提高 AI 的运行效率、降低资源消耗；AI Evaluation（AI 评测），关注如何定义和衡量智能。
这次分享中，他结合计算机和 AI 研究的经历，讨论了 Agent 带来的两方面影响。一方面正好也关于“效率”：读文献、写代码、做实验等工作正在加速；另一方面则关于“评价”：论文更容易产出，也给成果鉴别、同行评审和科研评价带来了更大压力。
他认为，Agent 确实加快了科研进程：Agent 能自主决定下一步、调用工具执行任务，Harness 则提供上下文管理、工具调用等支持，让大模型能够持续开展工作；他自己每天早上 6 点，还会收到 Agent 筛选推荐的 10 篇相关文献。
但论文产出更快，不等于科学进步更快：科研旨在增进认识，结论需要接受检验，大量成果都声称有新发现时，如何判断它们是否可靠、究竟新在哪里，就成了更紧迫的问题。
与此同时，科研评价仍看重论文等量化成果，因此当一些老师倡议“慢科研”、呼吁重质量而非数量时，也有学生反问：老师自己发了那么多论文，为什么要求我们慢下来？
在王欢看来，Agent 的影响已超出效率提升，触及科研评价制度：论文越来越容易产出，评价体系该如何识别真正的贡献？
不过，他仍持乐观态度：AI 接手部分工作，也让人有更多空间思考，哪些探索和判断仍需要自己承担，以及人的独特价值在哪里。
以下为王欢讲演的完整速记，模力工场在不改变原意的前提下进行了订正和编辑。
Agent 加速科研，论文写得快了，成果更难评了
非常感谢主办方的邀请。这好像是我第一次在非学术会议上做分享。虽然现在是老师，但我以前也是程序员，所以也祝大家、祝我自己节日快乐！
今天，我想和大家分享的是智能体对科研的影响。
先做一个简单的自我介绍。我在浙江大学完成本科和硕士学业，之后去美国读博士，大约两年前回国，加入西湖大学。
这张来自我个人主页的图，概括了我的研究方向。其中一个关键词是“高效”，也就是 Efficient AI；另一个方向是 AI Evaluation，也就是 AI 评测。
“高效”就是大家理解的降本增效。我们会从算法、模型、系统和硬件等不同层面入手，这些都是我们关注的方向。
提到 AI 评测，大家可能会想到：开发一个模型，用各种 Benchmark 测试、打分。这是它的一种表现形式。但从我的角度来看，我们还需要追问更底层的问题。
Efficient AI 研究的是：已经有了一个 AI 模型、算法或系统，怎样让它运行得更快、更高效。而 AI Evaluation 关注的是：为什么我们认为一个系统比另一个系统更智能？我们到底观察到了什么，才可以得出这个结论？“更智能”究竟意味着什么？
这就是我为实验室规划的两个主要方向：一个让 AI 系统运行得更高效，另一个试图回答什么是智能，以及如何度量智能。
今天的分享不会涉及太多技术细节。科研是一个非常大的领域，我很难凭个人经验概括智能体对整个科研体系的影响。因此，这次分享主要立足于计算机领域，尤其是我们作为 AI 研究者所感受到的变化。
我会围绕三个方面展开：什么是智能体，它给科研带来了哪些变化，以及面对这些变化，我们应该如何思考。
首先，什么是 Agent？
这是一个非常热门的词，但我们需要想一想，它最根本的特征是什么。我们经常说某一年是某项技术的“元年”，那么，我们究竟看到了什么变化，才认为自己进入了一个新时代？
Agent 这个词并不只存在于 AI 或计算机领域，也不是一个新出现的技术词汇。在哲学文献中，它很早就已经被使用，但其含义与今天技术领域中的用法存在一些细微差别。
大家可以试着打开浏览器的无痕模式，搜索“Agent”，看看会出现什么图片。很多时候，搜索结果里最多的是“特工”。
从词源来看，Agent 的一个核心含义是“行动”，与 Action、Act 等词有关。在哲学意义上，行动的发出者或主体，就可以被称为 Agent。比如，我做出了某个行动，我就是这个行动的 Agent。
而在今天的讨论中，大家更强调的是：它仍然要采取行动，但不是由我亲自完成，而是由另一个主体替我完成。
这里就有一个问题：Agent 本质上仍然是代码或程序。很早以前，程序员就会写程序，帮自己完成某件事，为什么那些程序不被称为我们今天所说的 Agent？
其中一个重要区别在于，传统程序的运行规则和逻辑，主要是由人预先指定的。而今天的 Agent 在完成任务时，我们很难提前规定好每一步：先做 A，再做 B，然后做 C。它需要自己作出复杂决策。
所以，自主决策是一个非常核心的特征。另一个重要特征，则是在执行过程中调用各种工具，也就是 Tool Calling。
因此，今天我们讨论 Agent，主要是在讨论一种具备自主决策和工具调用能力的智能程序。
当然，智能程度并不存在一条截然分明的界线，并不是越过这条线就叫自主决策，没有越过就不算。它实际上是一个连续的谱系。
前面是在尝试比较严谨地讨论 Agent 的含义。对业界和工程实践来说，也有一个更简单的理解：Agent 是大模型加上外面的一层 Harness。
多数 AI 工程师或研究者，并不会像哲学讨论那样，精细地考察 Agent 的概念。他们更关心的是：原来的大模型主要用于对话，不能独立帮人完成任务；后来，我们在外面加了一层框架，让它能够管理上下文、记录已经做过的事情、调用工具，并接收工具返回的结果。
这层类似脚手架的东西，就是 Harness。这个词的字面意思是“马具”。
右边的截图来自 Hugging Face 的智能体课程，其中也使用了类似的表述。可以说，这是业内比较常见的一种理解方式。
讲完什么是 Agent，我们再来看，它会对科研产生什么影响。
刚才几位嘉宾分享了智能体在不同领域的应用。对于科研，它同样带来了很大冲击。但在回答这个问题之前，我们首先需要问：什么是科研？
提到科学研究，大家可能会想到博士生、研究员在实验室里拿着试管做实验；对于我们这个领域，则可能是坐在办公室里敲电脑、读论文。
我们会写论文、投稿、接受评审，论文录用后发表在期刊或会议上，再去参加学术会议。这些都是大家对科研的印象，也是科研的表现形式，但并不等于科研的本质。
假设一个人在知识上提出了新的东西，他没有发表论文，只是写了一篇博客，这算不算科研？大家是否认可其中的发现？
这样的例子说明，做科研与发论文，以及其他科研的表现形式之间，不能完全画等号。
严格来说，科研是一种遵循一定规范、围绕问题展开的认识过程，其目的在于增进人类对某个领域的理解。
那么，研究什么问题？科学问题。这句话听起来像一句废话，但“什么是科学问题”，其实是一个需要严肃讨论的问题。
日常生活中，我们说“这个不科学”，似乎是在说“这个不对”。但科学性并不等于正确性。
这里涉及一个我们可能在中学就接触过的概念：可证伪性，也就是 Falsifiability。卡尔·波普尔在《科学发现的逻辑》中对这一问题作过著名论述。可证伪性是一种被广泛讨论和接受的科学划界标准，但并不是唯一的标准。
按照这一思路，说一个命题具有科学性，并不意味着它一定正确，而是说，它能够接受经验检验，存在被证据推翻的可能。
我们收集证据，尝试检验甚至推翻它；如果目前还没有被推翻，就暂时接受它。不能通过这种方式检验的命题，不属于这一标准下的科学命题，但“非科学”也不一定意味着“错误”。
这是我们讨论科研时需要理解的一个基本问题。至于发论文等，则是科研的表现形式。
回到 Agent 时代，它为科研带来了什么？
首先，当然有很多积极变化。大家平时使用 Agent 写文档、写代码、做 PPT，这些工作我们同样需要完成。在构思研究想法时，我们也会与它对话。它在其他领域带来的效率提升，在科研中同样存在。
在座有很多程序员，大家可以想一想：自己是不是已经很久没有亲手写代码了？不少朋友这样跟我说。至少在计算机科研领域，情况也类似。很多学生反馈，自己已经很久没有直接写代码了，一个想法从提出到实现的周期大大缩短。
这里举一个例子。西湖大学一位从事自然语言处理研究的老师，其团队正在做 AI Scientist，也就是 AI 科学家。他们通过一个网站，将我刚才描述的这些过程集成起来。
这是一个代表性的例子，展示了 AI 如何加速科研工作中的多个环节。
那么，如果不只看这些工作环节，AI 对研究想法和科研本身又能带来什么变化？
另一个例子来自西湖大学 AI 方向的原发杰老师。就在几天前，他的团队在《自然》正刊发表了一篇论文：用户通过自然语言描述需求，模型就能自动设计相应的蛋白质结构。这是 AI 对科研产生积极影响的另一个例子。
如果从论文产出的流程来看，首先要阅读已有文献，包括 arXiv 等预印本网站上的论文。现在，AI 已经可以自动完成不少文献阅读和整理工作。
我自己每天 6 点都会让一个 Agent 帮我阅读相关文献，筛选值得关注的内容，并推荐 10 篇文章。
接下来是构思研究想法。很多基于文本的讨论，现在也可以借助 AI 完成。再往后，是写代码、做实验、写论文和审稿。
在这整个流程中，我个人觉得，可能只有实验环节，有时还需要自己动手。尤其在非计算机领域，需要使用试管、接触物理世界的实验，自动化的推进会慢一些。而在计算机领域，实验本身也可以被大幅加速。
所以，在数字世界中，也就是我所说的形式系统、符号空间或数字化空间里，原有的各种壁垒正在非常快地消失。
这是积极的一面。但成也萧何，败也萧何。当数字世界中的工作都被加速之后，另一个问题出现了：我们基于这些数字化产出的评价体系，正在承受巨大压力。
不知道大家有没有留意开场时循环播放的视频。今天的问题，不只是能不能做出很大、很新的成果，而是我们可以一下子生成大量“声称很大、很新”的成果。
谁来鉴别这些成果是否真的新？贡献究竟有多大？
在谈创新之前，我们首先需要有判别能力。今天一个很难的地方，恰恰不在于生成，而在于判别。
我举几个学术领域的例子。这些变化究竟是好是坏，有时很难简单判断。
以一个 AI 顶级会议为例。这个会议大约从 2013 年开始举办，前 14 年的投稿总量约为 5.5 万篇，而今年一年的论文取号量，就已经超过了这个总和。
需要说明的是，取号量并不等于最终投稿量，实际投稿可能会少一些，但这个数字仍然值得参考。
我们能不能据此认为，今年的科技进步相当于此前 14 年的总和？这很难说。但在这个领域里，我们知道一个重要原因：今天生产论文的成本实在太低了。
论文投出去以后，仍然需要人来审，需要博士生、导师和研究员逐篇评估。这给整个评审系统带来了非常大的压力。
另一个最近发生的事情，也很值得思考。
在科研中，我们评价一个学生或老师做得好不好，经常会参考发表了多少篇论文等量化指标。而 AI 智能体让这些数字化成果的生产速度大幅提高，也让竞争压力变得更大。
过去，很多人是读到博士阶段才正式进入科研。现在，本科生甚至高中生都已经开始做 AI 研究，有些高中生进入大厂，甚至去硅谷工作，大家可能也听过类似新闻。
一些学生刚进入大学，就去找博士生导师，希望进入实验室、发表顶级会议论文。有些老师觉得，不必这么着急，于是发表了关于“慢科研”的声明，呼吁大家不要过度追求数量，而要关注质量。
这份声明引起了很多讨论。许多老师签署了声明，其中包括我认识的一些来自国内外顶尖高校的老师。
但从学生的角度，也出现了另一种声音：有人统计签署声明的老师每年发表的论文数量，反问，老师们自己发表这么多论文，为什么又呼吁学生慢下来？甚至有人用“上车之后把车门焊死”来表达不满。
这些声音的出现，本身就是一个值得思考的现象。
我们可以说，AI 大大加速了科研的发展。但如果只有这一种评价，对事情的理解就过于简单了，甚至可能产生很大偏差。
类似的事情还在继续。比如，前几天 OpenAI 一次性发布了 700 多份与数学成果有关的论文。其涉及的内容和方向非常广泛，任何一位顶尖数学家，都很难独自完成全部验证。
这又回到了刚才的问题：当大量成果可以被快速生成时，我们如何判断它们是否成立，又有多少真正的新意？
所以，AI 和 Agent 在加速技术发展、让各种应用变得更加丰富的同时，也在猛烈冲击技术之外的东西，包括科研人员、科研体系、科研制度，以及科研精神。这种冲击的程度，甚至可能更大。
说到这里，积极影响和挑战都已经比较清楚了。
积极的一面是，科研所依赖的数字化材料和工作环节被大幅加速，甚至研究想法的形成也能从中受益。与此同时，科研体系如何评价成果、科研共同体如何坚持原有的规范和精神，都面临新的压力。
下一步会怎样？坦率地说，我们现在还不知道。变化正在发生，整个领域也在重新思考如何评价这些变化。
一些科研工作者可能会因此感到悲观：AI 已经这么发达了，科研接下来怎么办？
但我个人其实相当乐观。当一件事情已经可以由 AI 完成时，它可能恰恰说明，这件事并不构成人之为人的本质。我们也因此有更多时间去思考：有哪些事情是人能够做、而 AI 还不能做的？那些事情，可能更接近对“人”的真正定义。
前不久，《奥德赛》这部电影很受关注。借用这个比喻，今天 AI 高歌猛进，我们仿佛正在书写希腊英雄远征特洛伊的史诗《伊利亚特》。
但与此同时，我们也需要一个像《奥德赛》那样的“返乡”过程。而且，很可能不是先出征、再返乡，而是出征与返乡同时进行。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
At the main forum of the Ai Community Day Forum, Wang Jing, Assistant Professor, Doctoral Tutor, Head of ENCODE Lab, University of West Lake, shared the theme " The Impact of AI AI agent on Scientific Research ".
Wang Jing has two main areas of study: Efficent AI (Efficient Artificial Intelligence), which focuses on how to improve the efficiency of AI operations and reduce resource consumption; and AI Evaluation, which focuses on how to define and measure intelligence.
In this sharing, he discussed the two implications of Agent in the context of his computer and AI research experience. On the one hand, the work on “efficiency”: reading literature, writing codes, doing experiments, etc. is accelerating; on the other hand, “evaluation”: papers are easier to produce and put more pressure on results identification, peer review and scientific evaluation.
In his view, Agent did speed up the scientific process: Agent was able to decide next step on its own, call tools to carry out his mission, and Harnesse provided context management, tool call, etc., to enable LLM to continue its work; he himself received 10 relevant literature recommended by Agent for screening every morning at 6 a.m.
However, the output of the papers is faster than that of scientific progress: scientific research is aimed at increasing awareness and conclusions need to be tested, and it becomes even more urgent to determine whether or not they are reliable and where they are new when a great deal of results claim to be new.
At the same time, scientific evaluation still values the results of quantization, such as dissertations, so when some teachers advocate “slow science” and call for quality rather than quantity, others ask: why do teachers ask us to slow down when they have published so many papers?
In Wang Jing ' s view, the impact of Agent has exceeded efficiency by touching on the system of scientific evaluation: papers are becoming easier to produce, and how should evaluation systems recognize real contributions?
However, he remained optimistic that AI would take over part of the work and give more room for reflection, what exploration and judgement still had to take on itself, and where the unique value of human beings was.
The following is a full summary of Wang Jian's speech, which was revised and edited without changing the original intent.
Agent, speed up the research. The papers are written fast. The results are even harder to evaluate.
Thank you very much for the invitation. It's like the first time I've shared it at a non-academic conference. I'm a teacher now, but I used to be a programmer, so I wish you all a happy holiday!
Today, I want to share with you the impact of AI agent on scientific research.
Make a simple introduction. I completed undergraduate and master ' s studies at Zhejiang University and then went to the United States of America for a doctorate, returned home about two years ago and joined West Lake University.
This is a map from my home page, outlining my research direction. One of the key words is "efficient," that is, Efficient AI; the other is AI Evaluation, that is, AI Assessment.
“Efficiencies” are understood to be reduced efficiency gains. We will start at different levels, such as algorithms, models, systems and hardware, which are all directions of our concern.
Referring to the AI assessment, one might think of developing a model, testing and scoring with various Benchmarks. It's an expression of it. But from my point of view, we also need to ask questions at the bottom.
Efficient AI studies how an AI model, algorithm or system already exists to make it work faster and more efficiently. And AI Evaluation is concerned: Why do we think one system is smarter than another? What exactly do we see to arrive at this conclusion? What does "more intelligent" mean?
This is the two main directions I plan for the lab: one to make the AI system work more efficiently, the other to try to answer what intelligence is and how to measure it.
Today's sharing will not involve much technical detail. Scientific research is a very large area, and it is difficult for me to sum up the impact of AI agent on the entire scientific system with personal experience. Therefore, this sharing is based mainly on changes in the computer sector, especially as we are AI researchers.
I will focus on three areas: what is AI agent, what changes it brings to scientific research, and how we should think about them.
First, what's Agent?
This is a very popular word, but we need to think about what its most fundamental features are. We often say that a given year is a “middle-year” of a particular technology, so what changes do we see before we think we are entering a new era?
The word Agent does not exist only in AI or computer fields, nor is it an emerging technical vocabulary. In the philosophical literature, it has long been used, but its meaning differs somewhat from that of today ' s use in the technical field.
You can try to open the browser's unearthed mode, search Agent, and see what's coming. In many cases, the search results were most frequently “agents”.
From the source of the words, one of the core meanings of Agent is “action”, which is related to the words Action, Act, etc. In philosophical terms, the originator or subject of the action can be called Agent. For example, I've made an operation, and this is me, Agent.
In today's discussion, it is even more emphasized that it remains to act, not by myself, but by another subject.
Here's the question: Agent is still essentially a code or program. A long time ago, programmers wrote programs to do something for themselves. Why aren't those programs known as Agent today?
One important difference is that the rules and logic of the operation of traditional procedures are primarily pre-designated. And today's Agent, when it's done, it's hard to set every step ahead of schedule: do A, then B, and then C. It requires complex decision-making on its own.
So autonomous decision-making is a very central feature. Another important feature is the use of tools during implementation, namely, Tool Calling.
So today we are talking about Agent, primarily about an intelligent process with the ability to make autonomous decisions and tool call.
Of course, there is no clear line of intelligence that goes beyond what is called autonomous decision-making and does not count. It's actually a continuous spectrum.
The first is an attempt to discuss the meaning of Agent more precisely. There is also a simpler understanding for industry and engineering practice: Agent is LLM plus the outside Harness.
Most AI engineers or researchers do not examine the concept of Agent as carefully as philosophical discussions do. They were more concerned that the original LLM was used primarily for dialogue and could not be used independently to carry out the task; later, we added a framework outside that would allow it to manage context, record what had been done, call tools and receive the results of its return.
Something like a scaffold, that's Harness. The word literally means "pony."
The right screenshot is taken from a AI agent course in Hugging Face, where similar expressions are used. This can be said to be a relatively common way of understanding within the industry.
And when we're done talking about Agent, let's see what it's gonna do to science.
Several guests just shared the applications of AI agent in different fields. It also has a great impact on scientific research. But before answering that question, we need to ask first: What is scientific research?
Referring to scientific research, one might think that doctoral students and researchers are experimenting in labs with test tubes; for our field, it might be sitting in the office knocking on computers and reading papers.
We write papers, contribute to them, are reviewed, are published in periodicals or conferences and attend academic conferences. These are the impressions and expressions of science, but they are not the essence of science.
Assuming that a person has something new in his knowledge, he doesn't publish a paper, he just writes a blog. Do you recognize the discovery?
Such examples show that there is no complete equivalent between doing research and publishing papers and other forms of scientific research.
Strictly speaking, scientific research is a normative and issue-oriented process of awareness aimed at improving human understanding of a given field.
So, what's the problem? Scientific. That phrase sounds like a piece of crap, but “what is a scientific issue” is actually a serious discussion.
In everyday life, we say “this is not scientific”, and it seems to say “this is not right”. But scientificity is not the same as correctness.
This is about a concept that we may have come across in secondary school: perjury, that is, Falsifiability. Carl Popper has made a famous statement on this issue in The Logic of Scientific Discovery. Perjury is a widely discussed and accepted scientific delimitation standard, but not the only one.
In this line of thinking, to say that a proposition is scientific does not mean that it must be correct, but that it is capable of being tested empirically and that there is a risk of being overturned by evidence.
We gather evidence, try to test it or even overthrow it; if it is not overturned at this time, we accept it for the time being. The proposition that cannot be tested in this way is not a scientific proposition under this criterion, but “non-scientific” does not necessarily mean “mistake”.
This is a fundamental issue that we need to understand when discussing scientific research. The publication of papers, etc., is an expression of scientific research.
Back to Age, what did it bring for science?
First, of course, there have been many positive changes. You usually use Agent to write documents, write codes, do PPTs, and we need to do the same. We will also engage in dialogue with research ideas when they are conceived. The efficiency gains it brings in other areas also exist in scientific research.
There are a lot of programmers in this room, so let's see if it's been a long time since you wrote the code. A lot of friends told me that. At least in the field of computer science, the situation is similar. Many students have responded that they have not written the code directly for a long time, and the cycle between the introduction and the realization of an idea has been significantly shortened.
Here is an example. A teacher at the University of West Lake who is working on natural language processing is working as an AI Scientist, or AI scientist. Through a website, they have integrated the processes I have just described.
This is a representative example of how AI accelerates multiple aspects of scientific work.
So, what can AI do about research ideas and research itself, if not just about these work links?
Another example comes from the former teacher of the U.S.A.A.'s direction. Just a few days ago, his team published a paper in Nature's Official: Users describe demand in natural languages, so that models can automatically design the corresponding protein structure. This is another example of the positive impact of AI on scientific research.
When looking at the process of the output of the papers, the first step is to read the available literature, including papers on the pre-printed website, such as arXiv. Now, AI has been able to automatically complete a lot of literature reading and sorting.
I myself get an Agent to help me read the literature at 6 a.m. every day, filtering out interesting content and recommending 10 articles.
And here's the idea of research. Many text-based discussions can now be completed with AI. Later, it was code writing, experimenting, writing papers and readings.
In this whole process, I personally feel that there may be only experimental components, sometimes on my own. In particular, in the non-computer field, experiments involving the use of in vitro tubes and exposure to the physical world are required, and the advance of automation is slower. And in the computer field, the experiment itself can be greatly accelerated.
So, in the digital world, i.e. the form system, symbol space or digital space that I call, the old barriers are disappearing very quickly.
That's the positive side. But it's not as bad as it is. When work in the digital world has accelerated, another problem has arisen: we are under great pressure to evaluate systems based on these digitalized outputs.
I don't know if you're paying attention to the video replaying at the opening. The question today is not just whether or not we can do something big and new, but we can produce a lot of “proclaimed big and new” results at once.
Who will be able to ascertain whether these results are really new and how much?
Before talking about innovation, we need first to be able to judge. A difficult place today lies not in generation, but in judgement.
I would like to give a few examples in the academic field. Whether these changes are good or bad is difficult to judge.
Take an example of an AI top-level meeting. The conference started in 2013, with a total of approximately 55,000 contributions in the previous 14 years, and the number of papers published this year has exceeded that sum.
It should be noted that the numbering does not amount to a final submission and that the actual contribution may be less, but this figure is still worthy of reference.
Can we assume that this year's progress in science and technology is equivalent to the sum of 14 years? It's hard to say. But in this area we know one important reason: the cost of producing papers today is so low.
Once the paper is published, there is still a need for people to examine it and for a full assessment by doctoral students, mentors and researchers. This puts a very heavy strain on the entire evaluation system.
Another recent development is also worth thinking.
In scientific research, we evaluate how well a student or teacher is doing, and often refer to quantization indicators, such as the number of papers published. AI AI agent, on the other hand, has led to a significant increase in the rate of production of these results and has increased competitive pressures.
In the past, many people did not officially enter science until they reached the doctoral level. Now, undergraduates and even high school students have started doing AI research, and some high school students have gone to large factories, even to work in Silicon Valley, and you may have heard similar stories.
As soon as they entered the university, some students went to the doctoral tutors in order to enter the laboratory and publish their top conference papers. Some teachers felt that there was no need to rush and then issued a statement on “slow science” calling for quality rather than excessive quantitative pursuit.
That statement generated much discussion. Many teachers signed the statement, including some of the teachers I knew from the top universities in the country and abroad.
But from a student's point of view, another voice emerged: The number of papers published each year by the teachers who signed the statement was counted, and why did the teachers call for the students to slow down? There are even those who express their dissatisfaction by “welding the doors after getting in the car”.
The emergence of these voices is in itself a reflectional phenomenon.
We can say that AI has greatly accelerated the development of scientific research. But if there is only such an assessment, the understanding of the matter is too simple and may even lead to significant deviations.
Something like that continues. For example, the previous day, Open AI had a single publish of over 700 papers related to mathematical results. It involves a wide range of content and direction, and it is very difficult for any top mathematician to complete all of this alone.
This goes back to the question just now: how can we judge whether a great deal of results can be produced quickly and how many real innovations can be made?
So, AI and Agent, while accelerating technology development and enriching applications, are also outside the technology of shock, including researchers, scientific systems, scientific systems, and the spirit of research. The scale of such shocks may even be greater.
The positive effects and challenges are now clearer.
On the positive side, the digital materials and work links on which research institutes rely have been greatly accelerated and even the development of research ideas can benefit. At the same time, there are new pressures on the scientific system to evaluate the results and on the scientific community to adhere to the original norms and spirit.
What happens to next step? Frankly, we do not know yet. Changes are taking place and the whole area is rethinking how to evaluate them.
Some scientists may feel pessimistic about this: AI is already so developed, what happens next?
But I'm actually quite optimistic personally. When something can already be done by AI, it may be precisely what it means to be human. So we have more time to think about what people can do and what AI can't do. Those things may be closer to a true definition of “person”.
In the recent past, the film Odyssey has received considerable attention. With this metaphor, today, AI sings a great advance, as if we were writing The Elyot of Troy's Expedition, a Greek hero.
But at the same time, we need a “return” process like Odyssey. Moreover, it is likely that the march will not take place first, but in parallel with the return.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.