Dwarkesh Patel

Ilya与我探讨了SSI的战略、预训练存在的问题、如何提升AI模型的泛化能力，以及如何确保AGI顺利发展。 𝐄𝐏𝐈𝐒𝐎𝐃𝐄 𝐋𝐈𝐍𝐊𝐒（节目链接） * 文字稿：https://www.dwarkesh.com/p/ilya-sutskever-2 * Apple播客：https://podcasts.apple.com/us/podcast/dwarkesh-podcast/id1516093381?i=1000738363711 * Spotify：https://open.spotify.com/episode/7naOOba8SwiUNobGz8mQEL?si=39dd68f346ea4d49 𝐒𝐏𝐎𝐍𝐒𝐎𝐑𝐒（赞助商） - Gemini 3是我用过的第一个能发现我未曾预料的关联的模型。我最近写了一篇关于强化学习信息效率的博客文章，Gemini 3帮助我理清了所有思路。它还为我生成了相关图表，并零错误地运行了机器学习小实验。立即在 https://gemini.google 试用Gemini 3。 - Labelbox帮助我创建了一个转录我们节目的工具！过去我在转录方面一直很挣扎，因为我不仅需要逐字稿，还需要经过改写、读起来像文章一样的转录稿。Labelbox帮助我生成了为此所需的*精确*数据。如果您想了解Labelbox如何帮助您（或者您想亲自尝试转录工具），请访问 https://labelbox.com/dwarkesh - Sardine是一个AI风险管理平台，它汇集了数千个设备、行为和身份信号，帮助您评估用户欺诈和滥用的风险。Sardine还提供一套智能体来自动化调查，以便在欺诈者利用AI扩大攻击规模时，您也能利用AI扩大防御规模。了解更多信息，请访问 https://sardine.ai/dwarkesh 要赞助未来的节目，请访问 https://dwarkesh.com/advertise 𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒（时间戳） 00:00:00 – 解释模型的锯齿状特性 00:09:39 – 情绪与价值函数 00:18:49 – 我们在扩展什么？ 00:25:13 – 为什么人类比模型泛化得更好 00:35:45 – 直击超级智能 00:46:47 – SSI的模型将从部署中学习 00:55:07 – 对齐 01:18:13 – “我们是一家不折不扣的研究时代公司” 01:29:23 – 自我对弈与多智能体 01:32:42 – 研究品味

安德烈·卡帕西特辑。在这次访谈中，安德烈解释了为什么强化学习很糟糕（但其他方法更糟），为什么通用人工智能只会融入过去约两个半世纪2%的GDP增长中，为什么自动驾驶花了这么长时间才取得突破，以及他眼中教育的未来。与他交谈非常愉快。 𝐄𝐏𝐈𝐒𝐎𝐃𝐄 𝐋𝐈𝐍𝐊𝐒 * 文字记录：https://dwarkesh.substack.com/p/andrej-karpathy * Apple播客：https://podcasts.apple.com/us/podcast/andrej-karpathy-agi-is-still-a-decade-away/id1516093381?i=1000732326311 * Spotify：https://open.spotify.com/episode/3iIYVmmhXwh3fOumypWVpC?si=33d37708b2b44e2f 𝐒𝐏𝐎𝐍𝐒𝐎𝐑𝐒 * Labelbox帮助您获取比默认情况下更详细、更准确、信号更强的数据，无论您的领域或训练范式如何。立即联系：https://labelbox.com/dwarkesh * Mercury帮助您更好地运营业务。这是我们播客使用的银行平台——我们喜欢它能在一个地方查看账户、现金流、应收账款和应付账款。几分钟内在线申请：https://mercury.com * Google的Veo 3.1更新是对本已出色模型的显著改进。Veo 3.1的生成内容更连贯，音频质量更高。如果您有Google AI Pro或Ultra计划，今天即可在Gemini中试用：https://gemini.google 赞助未来剧集，请访问：https://dwarkesh.com/advertise 𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒 00:00:00 – 通用人工智能仍需十年 00:30:33 – 大语言模型的认知缺陷 00:40:53 – 强化学习很糟糕 00:50:26 – 人类如何学习？ 01:07:13 – 通用人工智能将融入2%的GDP增长 01:18:24 – 超级人工智能 01:33:38 – 智能与文化的演变 01:43:43 – 为何自动驾驶耗时良久 01:57:08 – 教育的未来

理查德·萨顿是强化学习之父、2024年图灵奖得主，也是《苦涩的教训》一书的作者。他认为大语言模型是一条死胡同。在采访他之后，我对理查德立场的最佳辩护如下：大语言模型无法在工作中学习，因此无论我们如何扩展规模，都需要*某种*新架构来实现持续学习。而一旦我们拥有了这种架构，就不再需要专门的训练阶段——智能体将像所有人类、甚至所有动物一样，在过程中即时学习。这种新范式将使当前的大语言模型方法过时。 在采访中，我尽力代表这样一种观点：大语言模型或许可以作为经验学习发生的基础……火花四溅。衷心感谢阿尔伯塔机器智能研究所邀请我前往埃德蒙顿，并允许我使用他们的演播室和设备。敬请欣赏！ 𝐄𝐏𝐈𝐒𝐎𝐃𝐄 𝐋𝐈𝐍𝐊𝐒 * 文字记录：https://www.dwarkesh.com/p/richard-sutton * Apple播客：https://podcasts.apple.com/us/podcast/richard-sutton-father-of-rl-thinks-llms-are-a-dead-end/id1516093381?i=1000728584744 * Spotify：https://open.spotify.com/episode/3zAXRCFrHPShU4MuuIx4V5?si=c9f4bf24fb4c43e3 𝐒𝐏𝐎𝐍𝐒𝐎𝐑𝐒 * Labelbox 使得在超逼真的强化学习环境中训练AI智能体成为可能。凭借经验丰富的应用研究团队和庞大的主题专家网络，Labelbox 确保你的训练能反映重要的现实世界细微差别。在 https://labelbox.com/dwarkesh 将你的演示项目转化为工作系统。 * Gemini Deep Research 专为深入探索难题而设计。在本期节目中，它帮助我从早期的策略梯度方法追溯到当前的方法，结合清晰的解释和精选的示例。在 https://gemini.google.com/ 亲自尝试。 * Hudson River Trading 不会将团队孤立起来。相反，HRT 的研究人员在一个单一代码库中公开交流想法并分享策略代码。这意味着你能够以惊人的速度学习，并且你的贡献将对整个公司产生影响。在 https://hudsonrivertrading.com/dwarkesh 查找空缺职位。 要赞助未来的一期节目，请访问 https://dwarkesh.com/advertise 𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒 00:00:00 – 大语言模型是死胡同吗？ 00:13:51 – 人类会进行模仿学习吗？ 00:23:57 – 经验时代 00:34:25 – 当前架构在分布外泛化能力差 00:42:17 – AI领域的意外 00:47:28 – 通用人工智能之后，《苦涩的教训》还会适用吗？ 00:54:35 – 向AI的传承

I recently got to interview Robert Caro, who is my favorite author of all time. This gave me the excuse to reread his biographies of Lyndon Johnson and Robert Moses. I decided to compile their lessons about power into this video essay. Thanks to Max Farrens, who paired with me in writing the script for this video essay, and to Aaron and Ishan, who produced this video essay. Music by Tom Fox @tomfoxcatalog If you just want to read the essay: https://www.dwarkesh.com/p/lessons-in-power

I spent two days touring Intel’s most advanced fab and talking to its technical leadership to understand what the chips of the future will look like. A lot of what's coming next – gate-all-around transistors, backside power, CFET, hybrid bonding – is hard to follow without actually seeing it, so we made all the awesome animated explainers that I wish I had while I was prepping. I also asked Intel’s leadership how the semiconductor industry will cope with the enormous shock in demand from AI. AI compute has grown 3x a year for the last decade, but according to Intel's CTO, Intel could only double leading-edge wafer production with today’s global stock of chipmaking tools. What kinds of tricks can foundries pull to extend this crossover point where demand surges way ahead of supply? This was a totally new format for us - hope you enjoy! I also made flashcards for the concepts I had to learn to prep for this episode – link below. 𝐄𝐏𝐈𝐒𝐎𝐃𝐄 𝐋𝐈𝐍𝐊𝐒 * Flashcards: http://flashcards.dwarkesh.com/intel * Transcript: https://www.dwarkesh.com/p/intel-tour * Apple Podcasts: https://podcasts.apple.com/us/podcast/what-will-computer-chips-look-like-in-2045/id1516093381?i=1000797633702 * Spotify: https://open.spotify.com/episode/1kycOQ4f8ZHsXMOvUujdq6?si=NVKBI5NCRqirns5WPXzepQ Music for this episode composed by Tom Fox @tomfoxcatalog 𝐒𝐏𝐎𝐍𝐒𝐎𝐑𝐒 * WorkOS helps you do business with enterprises. It gives you access to all the features most commonly required by enterprises: SSO, SCIM, RBAC, audit logs, and more. It’s easy to plug these features in — it only takes a few lines of code — and WorkOS is releasing new functionality all the time, including things like auth for agents. If you want to join the 3,000+ companies that rely on WorkOS, go to https://workos.com * Jane Street built a powerful tool called magic-trace to help analyze and debug their trading systems. It only exists because one engineer noticed a feature on Intel chips called Processor Trace and thought it’d be valuable to build on top of. Now, traders and developers across Jane Street use magic-trace every single day! If you want to be trusted to pursue the ideas you think matter most, Jane Street is hiring. Go to https://janestreet.com/dwarkesh * Antithesis’ testing platform lets you perfectly reproduce every single bug you find. That’s because it runs on a fully deterministic computer, which strips out all the unpredictability that can make certain bugs flaky and hard to pin down. To build it, Antithesis found every source of non-determinism in Intel’s x86 architecture and either sealed it off or replaced it with a controlled simulation. Learn more at https://antithesis.com/dwarkesh To sponsor a future episode, visit https://dwarkesh.com/advertise. 𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒 00:00:00 – Intro 00:00:47 – How Intel brings up a new process node 00:03:22 – What Intel learned from 10nm 00:06:28 – From planar transistors to gate-all-around 00:13:13 – Why High-NA EUV pays off for Intel but not TSMC 00:19:20 – Could chips have evolved completely differently? 00:23:33 – Inside Intel's "Hot Topics" meeting 00:29:30 – Replacing copper wires with ruthenium 00:32:41 – Why stacked transistors take a decade to ship 00:39:45 – Hybrid bonding and the chip-to-chip bottleneck 00:48:00 – There aren't enough tools to 10x chip production 00:53:14 – Can Intel catch up to TSMC? 01:01:02 – Terafab, and rethinking the fab from scratch