Ryan Peterman (@ryanlpeterman) on X

X AI KOLs Timeline News

Summary

Interview with Google DeepMind's pre-training area lead Vlad Feinberg about landing jobs at frontier AI labs, covering needed skills, research vs engineering differences, and scaling laws.

Vlad Feinberg (@FeinbergVlad) is Google DeepMind’s pre-training area lead and I asked him all about how to land a job at a frontier lab like Google DeepMind, Anthropic or OpenAI. In this episode: • Skills frontier labs need • Differences between software engineering and AI research • Domains that matter for frontier research • Concrete steps engineers can take to get closer to research • Jeff Dean spot bonus story Where to watch: • YouTube - https://youtu.be/cDyi91onoJ8 • Spotify - https://open.spotify.com/episode/5XgGwEsWKDKmXsXErgCf74?si=GorGk0nzQFqBnBFq7UM7KQ… • Apple Podcasts - https://podcasts.apple.com/us/podcast/the-peterman-pod/id1777363835… • Transcript - https://developing.dev/p/google-deepmind-pre-training-lead… Thank you to the sponsor of this episode for supporting my work: • WorkOS: makes your app Enterprise Ready with easy to use APIs to add SSO, SCIM, RBAC, and more in just a few lines of code, check them out at https://workos.com Chapters: 00:00 - Intro 00:33 - Skills frontier labs need 08:45 - The difference between AI research and engineering 21:41 - Domains that matter for the frontier 30:50 - Marketing yourself to frontier labs 35:13 - Concrete steps engineers can take 38:29 - Overview of pre-training areas 47:23 - Jeff Dean spot bonus story 50:14 - Favorite Gemini war story 58:59 - Advice for his younger self 01:03:07 - Outro
Original Article
View Cached Full Text

Cached at: 06/15/26, 11:11 PM

Vlad Feinberg (@FeinbergVlad) is Google DeepMind’s pre-training area lead and I asked him all about how to land a job at a frontier lab like Google DeepMind, Anthropic or OpenAI.

In this episode:

• Skills frontier labs need • Differences between software engineering and AI research • Domains that matter for frontier research • Concrete steps engineers can take to get closer to research • Jeff Dean spot bonus story

Where to watch:

• YouTube - https://youtu.be/cDyi91onoJ8 • Spotify - https://open.spotify.com/episode/5XgGwEsWKDKmXsXErgCf74?si=GorGk0nzQFqBnBFq7UM7KQ… • Apple Podcasts - https://podcasts.apple.com/us/podcast/the-peterman-pod/id1777363835… • Transcript - https://developing.dev/p/google-deepmind-pre-training-lead…

Thank you to the sponsor of this episode for supporting my work:

• WorkOS: makes your app Enterprise Ready with easy to use APIs to add SSO, SCIM, RBAC, and more in just a few lines of code, check them out at https://workos.com

Chapters:

00:00 - Intro 00:33 - Skills frontier labs need 08:45 - The difference between AI research and engineering 21:41 - Domains that matter for the frontier 30:50 - Marketing yourself to frontier labs 35:13 - Concrete steps engineers can take 38:29 - Overview of pre-training areas 47:23 - Jeff Dean spot bonus story 50:14 - Favorite Gemini war story 58:59 - Advice for his younger self 01:03:07 - Outro


TL;DR

Google DeepMind 预训练负责人 Vlad Feinberg 详解在顶尖 AI 实验室找工作的核心技能:底层工程(内核开发加速 LLM)、研究品味(处理随机性中的决策)以及对缩放定律的深刻理解。


实验室最渴求的技能:内核开发与底层工程

Vlad 指出,目前顶尖实验室最普遍且旺盛的需求是 内核开发和底层工程,目标是实际加速 LLM 的运行。
无论是在架构创新、优化推理服务(如改进 KV 缓存),还是跨越整个技术栈的改动,最终都需要以高吞吐量、低延迟的方式交付可运行的软件。这与经典后端工程思维紧密相关,是一个非常适合深耕的开放领域。

研究 vs. 应用:并非泾渭分明

针对 DeepMind 内部的分工,Vlad 解释虽然存在预训练、后训练等纯研究团队(负责制作前沿模型配方并确保训练稳定),以及围绕 Gemini 的产品应用团队(如优化搜索结果),但二者无法完全切割。
例如,在提升 LLM 事实准确性、评估来源时,同样包含大量硬核研究工作。他提出一个 “研究-应用光谱”——当今的 AI 从业者必须能够在该光谱上灵活移动。

软件工程 vs. AI 研究者:关键差异

Vlad 以蒸馏技术为例说明底层工程的重要性:
蒸馏涉及从万亿 token 规模的巨型 LLM 中提取统计信息,并传递给学生模型。为了高效处理这种数百万美元级别的计算投入,团队经历了三到四代蒸馏基础设施的重构,每次花四个月重新思考系统设计(抽象、存储、跨数据中心读写)。正是这些“经典软件工程”投资,才带来对蒸馏缩放定律的新理解和像 Flash 3.0 这样的成果。

那么,一个纯后端工程师直接转到研究团队会面临哪些困难?

  1. 缺乏文献背景:不熟悉领域已有的前沿工作,无法有效追溯高价值论文。
  2. 思维模式不同:软件工程的任务依赖图(DAG)是确定性的(修服务 A → 修服务 B → 单调进展);而研究 DAG 是随机的,某些节点可能成功也可能失败,需要探索。
  3. 无法提前“设计”最短路径:你需要培养研究品味——一种凭直觉估算某个方法成功率与时间投入的能力。

研究品味 = 正确遍历随机 MDP

Vlad 引用 Jacob Steinhart 的框架:研究即马尔可夫决策过程(MDP)

  • 研究项目中的里程碑构成随机依赖图,成功节点会开辟新可能性。
  • 软件工程可以写出所有路径并选最短;但研究需要事先估计不同想法的“成功率”和“时间投入”,然后决定下一步。
  • 这种“尚未开始就能判断概率”的能力,正是人们常说的研究品味,也是读博过程中培养的核心技能。

缩放定律:LLM 研究的基石

针对预训练,Vlad 强调理解缩放定律至关重要。

  • 关键不是函数形式(幂律、指数),而是能够预测特定扩展方案下的测试损失
  • 与 ImageNet 不同,LLM 的每次预训练都是一次性的——你永远看不到完整的训练数据集。
  • 你不能只在 MNIST 或 CIFAR 上有效,然后直接搬到 ImageNet 级别。许多方法跨尺度会失效。
  • 因此,团队需要提出配方:一个将计算量映射到训练流程的函数,同时结合准确的预测规则,从而指导大规模扩展开箱。

来源

视频链接:https://www.youtube.com/watch?v=cDyi91onoJ8&feature=youtu.be

Vlad Feinberg@FeinbergVlad·May 18: How to land a job at a frontier lab

https://vladfeinberg.com/2026/05/10/how-to-land-a-job-at-a-frontier-lab.html…

Similar Articles