EP31: AI如何加速科学发现? 对谈Isomorphic Labs工程师
中场沉思录
本期嘉宾Jan，前Spotify资深工程师，现就职于Isomorphic Labs（Alphabet旗下AI药物设计公司）。从CERN的物理学实习，到Spotify九年的推荐系统与基础模型研发，再到如今投身AI for Science，Jan分享了他跨越研究机构与工业界的职业路径。我们聊了传统推荐模型与大型基础模型的差异、AlphaFold如何加速药物发现、AI代理如何自动化实验，以及在一个AI能写代码的时代，工程师最该修炼什么。最后还聊了聊：如果不用为钱工作，你会做什么？AI会先让我们永生，还是先毁灭我们？如果你在科技行业工作或对AI For Science感兴趣，相信这一期会给你很多启发。
Guest: Jan, former Staff ML Engineer at Spotify, now at Isomorphic Labs (Alphabet’s AI drug design company). From a physics internship at CERN, to nine years at Spotify working on recommender systems and foundation models, and now to AI for Science, Jan shares his career path across research institutions and industry. We discuss the differences between traditional recommendation models and large foundation models, how AlphaFold accelerates drug discovery, how AI agents automate experiments, and what engineers should focus on in an era where AI can write code. Finally, we talk about: if you didn’t have to work for money, what would you do? Will AI make us immortal first, or destroy us first? If you work in tech or are interested in AI for Science, this episode will give you a lot to think about.
高光内容
1. 从CERN到Spotify再到AI制药：每一次跳跃都是主动选择
Jan在CERN实习时感受到科研机构的慢节奏，转而投身科技公司。在Spotify九年，从数据工程师转到ML，做过消息推送、首页推荐、基础模型。离开是因为“九年了，我想做点不一样的事”，而AI制药是他眼中下一个有巨大影响力的领域。
2. 推荐系统的范式转移：从小模型到大模型
过去Spotify训练成千上万个小模型，每个只做一件事，训练快、部署易。现在转向少数超大模型，能同时处理多个任务，但推理成本高、延迟大。Jan认为大模型在推荐领域有潜力，但需要大量A/B测试验证。
3. AI如何加速药物发现：AlphaFold与模拟实验
Isomorphic Labs的目标是把尽可能多的湿实验搬到计算机模拟中。AlphaFold预测蛋白质结构，帮助判断分子能否与靶点结合。药物发现全流程仍可能长达十年，但AI有望将临床前阶段从数年缩短到数月甚至数小时。
4. 用AI代理跑实验：把无聊且易错的工作交给Agent
Jan分享了他现在的工作方式：不再手写代码，而是用编码Agent；甚至用AI代理来运行实验——自动配置参数、调用工具、生成图表。虽然偶尔有bug，但比手动更高效、更不易出错。
5. 未来工程师最重要的品质：品味、领域知识和主动性
写代码和调库的重要性在下降，而架构判断力、对业务领域的理解、以及主动探索的驱动力（agency）越来越关键。Jan以自己为例：曾经讨厌生物，现在却对药物发现充满好奇，因为“深入了解后，好奇心自然就来了”。
Highlights
1. From CERN to Spotify to AI Drug Discovery: Every Jump Was a Deliberate Choice
During his internship at CERN, Jan felt the slow pace of a research institution and decided to move into tech. At Spotify for nine years, he transitioned from data engineer to ML, working on messaging, homepage recommendations, and foundation models. He left because “after nine years, I wanted to do something different,” and AI drug discovery is, in his eyes, the next field with enormous impact.
2. The Paradigm Shift in Recommender Systems: From Small Models to Large Models
In the past, Spotify trained tens of thousands of small models, each doing one thing, fast to train and easy to deploy. Now it’s shifting to a few very large models that can handle multiple tasks at once, but with high inference costs and latency. Jan believes large models have potential in recommendation, but need extensive A/B testing to validate.
3. How AI Accelerates Drug Discovery: AlphaFold and Simulated Experiments
Isomorphic Labs aims to move as much wet-lab experimentation as possible into computer simulations. AlphaFold predicts protein structures, helping determine whether a molecule can bind to a target. The full drug discovery pipeline can still take up to ten years, but AI has the potential to shorten the preclinical stage from years to months or even hours.
4. Running Experiments with AI Agents: Handing Boring and Error-Prone Work to Agents
Jan shares how he works now: he no longer writes code by hand, but uses coding agents; he even uses AI agents to run experiments, automatically configuring parameters, calling tools, and generating charts. Although there are occasional bugs, it’s more efficient and less error-prone than doing it manually.
5. The Most Important Qualities for Future Engineers: Taste, Domain Knowledge, and Agency
Writing code and tuning libraries are becoming less important, while architectural judgment, understanding of the business domain, and the drive to explore proactively (agency) are increasingly critical. Jan gives his own example: he used to hate biology, but now he’s curious about drug discovery, because “once you understand something deeply, curiosity comes naturally.”
时间轴
* 00:00｜开场介绍
* 01:16｜从CERN到Spotify：研究机构与工业界的差异
* 03:24｜早期对AI/ML的兴趣，在Spotify从数据工程师转向ML
* 05:12｜在Spotify的ML工作：消息推送、推荐系统、基础模型
* 08:14｜传统推荐模型 vs 大型基础模型：优势与挑战
* 13:56｜为什么离开Spotify，加入Isomorphic Labs
* 18:40｜Isomorphic Labs是什么，以及Jan的角色
* 22:27｜药物发现全流程：AI能加速哪部分，时间线有多长
* 28:14｜AlphaFold介绍：如何预测蛋白质结构，训练数据来源
* 33:48｜AI for Science：AI如何成为科学发现的加速器
* 37:51｜AI能否做出相对论级别的原创发现？
* 39:07｜最激动人心的AI for Science领域
* 43:34｜AI for Science的主要挑战
* 45:24｜普通人如何参与AI for Science
* 49:06｜在Isomorphic Labs与Spotify的工作差异
* 51:19｜如何使用AI工作：编码Agent与实验自动化
* 56:31｜工程师哪些技能在贬值，哪些在升值
* 59:18｜主动性与好奇心：是天赋还是可以培养？
* 1:03:40｜如果不用为钱工作，你会做什么？
* 1:05:27｜最关心的问题：什么能真正改变生活
* 1:06:30｜抗衰老与长寿：十年内能实现吗？
* 1:09:44｜AI风险：灭绝概率与人类未来
Timeline
00:00 | Introduction
01:16 | From CERN to Spotify: Differences Between Research Institutions and Industry
03:24 | Early Interest in AI/ML, Transitioning from Data Engineer to ML at Spotify
05:12 | ML Work at Spotify: Messaging, Recommender Systems, Foundation Models
08:14 | Traditional Recommender Models vs. Large Foundation Models: Advantages and Challenges
13:56 | Why He Left Spotify and Joined Isomorphic Labs
18:40 | What Isomorphic Labs Is, and Jan’s Role
22:27 | The Full Drug Discovery Pipeline: What AI Can Accelerate, and How Long It Takes
28:14 | Introduction to AlphaFold: How It Predicts Protein Structures, and Training Data Sources
33:48 | AI for Science: How AI Becomes an Accelerator for Scientific Discovery
37:51 | Can AI Make Original Discoveries on the Level of General Relativity?
39:07 | The Most Exciting Areas in AI for Science
43:34 | Main Challenges in AI for Science
45:24 | How Ordinary People Can Participate in AI for Science
49:06 | Differences Between Working at Isomorphic Labs and Spotify
51:19 | How He Uses AI at Work: Coding Agents and Experiment Automation
56:31 | Which Engineering Skills Are Depreciating and Which Are Appreciating
59:18 | Agency and Curiosity: Talent or Can They Be Cultivated?
1:03:40 | If You Didn’t Have to Work for Money, What Would You Do?
1:05:27 | The Issue He Cares About Most: What Can Truly Change Lives
1:06:30 | Anti-Aging and Longevity: Achievable Within Ten Years?
1:09:44 | AI Risks: Extinction Probability and the Future of Humanity
