11-19 更新 实时
首页/ AI 大模型/ 正文

AI 智能体加速博士研究什么,不能做什么

AI 大模型 来源:Towards Data Science 2026-11-13 21:30 0 阅读 机器翻译
AI 智能体加速博士研究什么,不能做什么

LLM 如何改变 AI 实验室的研究方式

我是一名 AI 实验室的博士生,研究领域处于机器学习与生态学的交叉地带。读博三年,我的日常工作与当初入学时的预想已经大不相同。

刚起步时,上手一门新的编程语言或框架,意味着要花上几小时、有时是几天读文档,才能做出点有用的东西。一个简单的 bug 就能耗掉我几个小时。想搞清楚如何高效部署一个模型,本身就是一轮漫长的试错。如今,LLM 会解释陌生的工具、找出 bug、并为实验提出新思路,往往只需过去一小部分时间。

这不仅仅是省时间的问题。它说明 AI 实验室的日常工作正在被重新组织。本文是我对哪些在变、哪些没变的看法。

真正变快的是什么

流行的说法是,AI 现在干起了研究生的活。这话有一定道理。物理学家 Scott Dodelson 最近在《Physics》上撰文指出,AI 智能体可以承担许多传统上交给研究生的任务,比如读论文、处理数据、写程序,耗时以分钟计而非以月计。他描述了首次使用这些 AI 工具时的“惊叹”感。但他也发现,这些工具会画出一张偏差达 20 个数量级的图,而且在人类指出之前毫无察觉。他的结论是,他仍然需要研究生,无法想象没有他们怎么做物理。

在我自己的工作中,我注意到研究的执行层面出现了巨大的提速:

代码。搭建带有数据加载器、快速训练、验证和可视化脚本的样板代码,所需时间只是过去的一小部分。

实验。一周之内你可以跑更多实验,于是更多想法得以及早被检验、及早被淘汰。

写作辅助。起草和编辑能获得很大的起步优势,尤其是当你害怕盯着空白页面、觉得把想法讲出来比写下来更容易的时候。

如果研究全部是执行,那故事到这里就结束了。但事实并非如此。

我的很多研究都始于一个非常简单的问题:这个物种真的存在于这里吗?

想象在森林里放置一台录音机,用来监听某一种特定的鸟。连着五个早晨它什么都没听到。这只鸟不在吗?也许。但也可能下了雨,鸟保持安静,而它其实一直都在。生态学家把这称为“不完全检测”。未观测到一次,并不等于不存在。

现在把这个数据集交给你能想象到的最快的 AI 智能体。它会乐呵呵地拟合一个分类器,通宵调参,然后报告在“存在”与“不存在”上的出色准确率。每一行代码都正确。但结论仍然可能是错的,因为标签本身就在说谎。

再快的速度也解决不了这一点。能解决它的是了解这些数据如何诞生的人:录音机放在哪里,当时天气如何,这个物种有怎样的行为习性。这类知识会导向一个不同的模型,比如占域模型,它把“它在那里吗?”和“我们会注意到它吗?”区分开来。

我认为这种模式在 AI 实验室里远比我们愿意承认的更常见。Alex Sen Gupta 最近在《Nature》上的一篇文章提到,以野外工作为基础的博士研究更难自动化。在我看来,只要一个项目的数据带有某种“身世背景”,它就很难被自动化,而大多数有意思的项目都是如此。

我对 AI 实验室将如何变化的看法

以下是我的三个预测。

1. 每位研究者都会变成团队负责人。研究生现在已经可以把编码、文献检索和初审等工作同时交给 AI 工具处理。问题也随之从“我这周能完成多少事?”转变为“我该提出什么需求、它是否值得推进、结果是否正确?”研究生不会消失,但最有价值的技能变成了引导工作方向和判断工作成果。

2. 领域知识成为优势。当每个人都拥有同样强大的工具时,实验室之间的差异就在于它掌握而工具并不掌握的东西。在我的领域,这就是生态学:数据是如何收集的,哪些问题对预测和在未来 5 到 10 年保护物种最重要,以及什么样的说法会让生物学家觉得不可信。

3. 实验室变得更加跨学科。如今,一名机器学习方向的学生可以比以往快得多地熟悉气候科学、医疗健康、教育、半导体或金融领域。最好的 AI 实验室不会只充斥机器学习专家,而会汇聚能从各个角度理解问题的人。

如果执行变得廉价,还剩下什么?

对问题的好奇心和品味。提出任何工具都想不到要问的问题,并判断哪一个值得你投入一年的生命。

人与人之间的信任。老实说,领域专家不会仅凭模型的说法就采信它,至少目前如此。当一个结果挑战了数十年的领域经验时,专家希望有人为此负责。机器无法为自己的工作背书,研究者可以。

AI 正在改变我们做研究的方式,但没有改变我们做研究的初衷:好奇心、韧性、协作与影响力。

高斯过程回归结合现代算力,可实现近乎实时、可扩展且样本高效的预测。

深入探讨如何保障 Model Context Protocol(MCP)部署的安全:威胁模型、真实世界的攻击利用,以及面向智能体系统的架构最佳实践。

学习离散节点嵌入

在本文中,我想让你充分了解 DINO 论文的工作原理以及它为何出色。我尽量让这篇文章……

数据科学家、数据分析师还是数据工程师?你如何知道自己适合哪一个?

来自一位管理 8 人团队的人的经验教训

AI 自食其尾:生成式系统中的模型崩溃风险

离开学术界后,你正在寻找自己的第一份数据科学工作吗?这可能没有你想的那么容易。

一个模型统治一切

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(What AI Agents Speed Up in PhD Research, and What They Can't)

How LLMs Are Changing Research in AI Labs

I'm a PhD student in an AI lab, working at the intersection of machine learning and ecology. Three years in, my everyday work looks very different from what I anticipated when I started.

When I started, picking up a new programming language or framework meant hours or sometimes days of reading documentation before I could do anything useful. A simple bug would cost me hours. Figuring out how to deploy a model efficiently was its own long round of trial and error. Today, LLMs explain unfamiliar tools, spot the bug, and suggest new ideas for experimentation, often in a fraction of the time.

This isn’t just about saving time. It’s a sign that the everyday work of an AI lab is being reorganized. This article is my view of what is moving and what is staying put.

What actually got faster

The popular story says AI is now doing the work of graduate students. There is some truth in it. Physicist Scott Dodelson recently wrote in Physics that AI agents can take on many tasks traditionally given to graduate students, like reading papers, crunching numbers, and writing programs, in minutes rather than months. He describes the "wow" factor of using these AI tools for the first time. But he also found that they will produce a plot off by 20 orders of magnitude and not realize it until a human points it out. His conclusion is that he still wants graduate students, and can't imagine doing physics without them.

In my own work, I've noticed a tremendous speed-up in the execution layer of research:

Code. Setting up the boilerplate code with data loaders, quick training, validation, and visualization scripts takes a fraction of the time.

Experiments. You can run more experiments in a week, so more ideas get tested and discarded early.

Writing support. Drafting and editing get a big head start, especially if you dread staring at a blank page and find it easier to talk through ideas than to write them.

If all of research were execution, this would be the end of the story. It isn't.

A lot of my research starts with a very simple question: is this species actually present here?

Imagine a recorder placed in a forest to listen for a particular bird. For five mornings it hears nothing. Is the bird absent? Maybe. Or maybe it rained, the bird stayed quiet, and it was there the whole time. Ecologists call this "imperfect detection." A missing observation is not the same as an absence.

Now hand that dataset to the fastest AI agent imaginable. It will happily fit a classifier, tune it overnight, and report excellent accuracy on "present" versus "absent." Every line of code will be correct. And the conclusion can still be wrong, because the labels themselves lie.

No amount of speed fixes that. What fixes it is someone who knows how the data was born: where the recorders sat, what the weather was like, how the species behaves. That knowledge leads to a different model, such as an occupancy model that separates is it there? from would we have noticed?

I think this pattern is far more common in AI labs than we admit. A recent Nature article by Alex Sen Gupta mentioned that field-based PhDs are harder to automate. In my opinion, any project where the data has some sort of backstory to it is harder to automate, and that describes most of the interesting ones.

How I think AI labs will change

Here are my three predictions.

1. Every researcher becomes a team lead. Grad students can already hand off coding, literature searches, and first-pass reviews to AI tools, all at the same time. The question shifts from "How much can I get done this week?" to "What should I ask for, is it worth pursuing, and is the result right?" Grad students aren't going anywhere, but the most valuable skills become steering the work and judging it.

2. Domain knowledge becomes the advantage. When everyone has the same powerful tools, what sets a lab apart is what it knows that the tools don't. In my field, that's ecology: how the data was collected, which questions matter most for forecasting and protecting species over the next 5 to 10 years, and what a biologist would find implausible.

3. Labs get more interdisciplinary. An ML student can now get conversant in climate science, healthcare, education, semiconductors, or finance much faster than before. The best AI labs won’t just be full of ML experts. They’ll bring together people who understand the problem from every angle.

If execution is getting cheap, what's left?

Curiosity andtaste in problems. Asking questions no tool would think to ask, and knowing which one is worth a year of your life.

Trust between people. Honestly, a field expert won't take a model's word for it, at least for the time being. When a result challenges decades of field experience, experts want someone accountable for it. A machine can't stand behind its work. A researcher can.

AI is changing how we do research. It hasn't changed why we do it: curiosity, resilience, collaboration, and impact.

Gaussian Process Regression coupled with modern computing enables for near-real-time, scalable, and sample-efficient prediction.

March 24, 202111 min read

A deep dive into securing Model Context Protocol (MCP) deployments: threat models, real-world exploits, and architectural best practices for agentic systems.

August 6, 202525 min read

Learning discrete node embeddings

November 29, 202212 min read

In this story, I would love to give you a a good idea of how the DINO paper works and what makes it great. I've tried to keep the article...

December 1, 20216 min read

Data Scientist, Data Analyst, or Data Engineer? How do you know which one is right for you?

January 23, 20217 min read

Lessons from someone who manages a team of 8

April 16, 202416 min read

AI eating its own tail: the risk of model collapse in generative systems

July 30, 20248 min read

Are you looking for your first data science job after leaving academia? It may not be as easy as you think.

December 29, 20209 min read

One model to rule them all

November 3, 202311 min read

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 Towards Data Science(What AI Agents Speed Up in PhD Research, and What They Can't),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗