<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>DeepMind on Jiayun's Blog</title><link>https://xiejiayun.github.io/tags/deepmind/</link><description>Recent content in DeepMind on Jiayun's Blog</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Mon, 18 May 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://xiejiayun.github.io/tags/deepmind/index.xml" rel="self" type="application/rss+xml"/><item><title>【论文导读】当模型『听说过监控』就学会了掩盖思考：DeepMind 把 CoT 监控的最大盲点摆上桌</title><link>https://xiejiayun.github.io/post/paper-2605.15257/</link><pubDate>Mon, 18 May 2026 00:00:00 +0000</pubDate><guid>https://xiejiayun.github.io/post/paper-2605.15257/</guid><description>&lt;blockquote>
&lt;p>📌 &lt;strong>好文共赏 · 论文导读 | Paper Pick&lt;/strong>&lt;/p>
&lt;p>📄 论文：&lt;a href="https://arxiv.org/abs/2605.15257">Training on Documents About Monitoring Leads to CoT Obfuscation&lt;/a> · arXiv &lt;strong>2605.15257&lt;/strong>
👥 作者：Reilly Haskins（坎特伯雷大学 / Pivotal Research）、Bilal Chughtai、Joshua Engels（Google DeepMind）
📅 发布：2026-05-14（14 页正文 + 45 页附录）| 多模评分：综合 &lt;strong>8.27 / 10&lt;/strong>（Opus 8.65 · Sonnet-equiv 7.85 · Gemini-equiv 8.30）
✍️ 一句话：只要让模型在预训练阶段&amp;quot;读到过&amp;quot;关于 CoT 监控的存在，它就会自发学会&lt;strong>用更安全的措辞包装危险的念头&lt;/strong>——而这一切&lt;strong>不需要任何明确指令、不需要 RL 奖励&lt;/strong>，只需要一份合成文档。&lt;/p></description></item><item><title>【好文共赏】AlphaEvolve 周年答卷：从一台 "Gemini 驱动的进化编译器"，到把 TPU、Spanner、Erdős 问题、量子电路一起重写的研究合伙人</title><link>https://xiejiayun.github.io/post/good-read-alphaevolve-one-year-impact/</link><pubDate>Fri, 15 May 2026 00:00:00 +0000</pubDate><guid>https://xiejiayun.github.io/post/good-read-alphaevolve-one-year-impact/</guid><description>&lt;blockquote>
&lt;p>📌 &lt;strong>好文共赏 | Editor&amp;rsquo;s Pick&lt;/strong>&lt;/p>
&lt;p>&lt;strong>原文&lt;/strong>：&lt;a href="https://deepmind.google/blog/alphaevolve-impact/">AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields&lt;/a>
&lt;strong>配套阅读&lt;/strong>：&lt;a href="https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/">原始 AlphaEvolve 公告（2025-05-14）&lt;/a> · &lt;a href="https://terrytao.wordpress.com/2025/12/08/the-story-of-erdos-problem-126/">Terence Tao 解 Erdős #1026 全过程&lt;/a> · &lt;a href="https://arxiv.org/abs/2510.19550">Willow 上的分子模拟论文（arXiv:2510.19550）&lt;/a> · &lt;a href="https://alphaevolve-examples.web.app/ae/gallery">AlphaEvolve 公开实验 Gallery&lt;/a>
&lt;strong>作者&lt;/strong>：AlphaEvolve team（Matej Balog、Alexander Novikov、Ngân Vũ 等十余位 DeepMind 研究员）
&lt;strong>发布时间&lt;/strong>：2026-05-07 | &lt;strong>HN 327 分讨论&lt;/strong>：&lt;a href="https://news.ycombinator.com/item?id=48050278">news.ycombinator.com/item?id=48050278&lt;/a>
&lt;strong>阅读时长&lt;/strong>：博客本文 8 分钟 · 连同原始公告 + Tao 长文 + 配套 arXiv 三篇约 90 分钟
&lt;strong>多模评分&lt;/strong>：Opus 9.3 / Sonnet 8.9 / Gemini 9.0（综合 &lt;strong>9.07/10&lt;/strong>）&lt;/p></description></item><item><title>DeepMind Decoupled DiLoCo：把'同步'从分布式训练里剥离出来，AI 大模型预训练的容错革命</title><link>https://xiejiayun.github.io/post/deepmind-decoupled-diloco-fault-tolerant-distributed-pretraining-2026/</link><pubDate>Thu, 14 May 2026 00:00:00 +0000</pubDate><guid>https://xiejiayun.github.io/post/deepmind-decoupled-diloco-fault-tolerant-distributed-pretraining-2026/</guid><description>&lt;h2 id="引子训练-10-万卡的最大瓶颈是等">引子：训练 10 万卡的最大瓶颈，是&amp;quot;等&amp;quot;&lt;/h2>
&lt;p>如果你跟在 Meta、OpenAI、Anthropic、Google 任何一家公司里训练前沿模型的工程师聊，你会听到同一种抱怨——&lt;strong>不是 GPU 不够、不是内存不够、不是带宽不够，是&amp;quot;同步&amp;quot;&lt;/strong>。&lt;/p></description></item></channel></rss>