Dwarkesh Patel Says Human Data Shortage Caps AI Efficiency Progress

Dwarkesh Patel称人类数据短缺限制AI效率进展 当前AI的进展依赖于扩大定制化人类数据分布，而非样本效率的根本性改善，这使得模型扩展无法匹敌人类的学习速率。 该文章指出，近期AI的进展并非由样本效率的突破所驱动，而是得益于数据分布的拓展、算力扩展以及对人工标注数据的大规模投入。相比仅凭极少数据便能学会复杂任务的人类，AI模型需要消耗数万亿Token，且运行在一条完全不同的扩展曲线上，宛如缝合而成的“科学怪人”。尽管效率极其低下，该技术在商业上仍具有高度可行性，因为训练成本可以分摊给数十亿并发用户。最终，克服分布外局限性可能需要借助AI来实现AI研究本身的自动化。 近年来的 AI 进展主要由更庞大的数据分布和算力驱动，在样本效率上并未出现实质性突破；当下的强化学习很大程度上是在依赖海量人类专家数据来建立先验概率。 (0:00) 人类从出生到成年大约仅接触 2 亿个 token，而前沿 AI 模型需训练数十万亿至上百亿万 token，两者在数据效率上存在近百万倍的巨大差距。 (3:52) 针对生物演化等同于预训练的辩驳，人类基因组仅 3GB 且编码蛋白质的部分仅占 1% 至 2%，根本无法存储预训练权重，大脑连接组完全是在生命周期内从头构建的。 (4:39) 尽管 AI 学习极其低效，但其高昂的训练成本可通过服务数以亿计的并发会话得以摊销，使白领工作自动化在商业上依然极具价值。 (8:46)

Dwarkesh Patel argues that recent advances in artificial intelligence reflect an aggressive expansion of curated data and compute rather than genuine progress in sample efficiency, masking a profound architectural divergence from human intelligence. Data Expansion Rather Than Sample Efficiency (0:00) While intelligence can be defined as sample efficiency—the volume of data needed to reach competency in a domain—recent model gains stem from widening training distributions and scaling compute rather than algorithmic efficiency. Reinforcement learning essentially functions as synthetic data generation where massive compute tests outputs against rubrics or verifiers. For this mechanism to work, models require initial prior probabilities, necessitating massive amounts of bespoke human expert demonstrations across fields like legal diligence, financial filing, and technical writing. The labeling and reinforcement learning environment industry that aggregates this expertise earns billions of dollars annually and is on track to reach decabillions. Patel likens modern models not to unified learners, but to a Frankenstein's monster constructed from a billion sewn-together graphs of curated examples. Because data constitutes the primary driver of progress rather than closely guarded hyperparameters or architectures, open-source competitors have been able to close the performance gap quickly. Citing Epoch, Patel notes that open models lag frontier systems by roughly four months, largely by distilling public frontier model APIs. The Scale Disparity with Human Learning (3:52) The data consumption required to train frontier models exposes a massive gap when compared to biological learning. A human hearing and seeing an average of 2,000 words per hour absorbs roughly 200 million tokens between birth and adulthood. In contrast, frontier models consume tens to hundreds of trillions of tokens—approaching a million-fold difference. Furthermore, while human students might solve a problem once or twice, training methods such as Group Relative Policy Optimization (GRPO) force models through hundreds to thousands of rollouts per task to resolve credit assignment. Physical and embodied domains highlight the same disparity. A human can learn to teleoperate a humanoid robot or robotic arm within hours, whereas AI models struggle to execute reliable, open-ended tasks despite millions of hours of demonstration data. Similarly, a teenager learns to drive safely with roughly 20 hours of practice, utilizing three to four orders of magnitude less data than autonomous driving stacks at Tesla or Waymo. Rebutting Evolutionary and Architectural Counterarguments (4:39) Patel rejects the view, previously articulated by Andrej Karpathy, that biological evolution serves as a massive pre-training phase that invalidates direct comparisons to randomly initialized models. The human genome contains roughly three gigabytes, with only 1% to 2% being protein-coding, which Patel notes is mathematically insufficient to encode the structural parameters of a neural network; evolution provided loss functions and hyperparameters, while the connectome is built from scratch. Furthermore, even if evolution were treated as pre-training, an already-educated AI still requires vast bespoke datasets to acquire a single marginal skill, whereas an educated human does not. Other objections prove equally fragile under examination. Appealing to raw multimodal sensory streams fails because deaf and blind individuals reach full general intelligence while consuming far less data than the 200-million-token baseline. Appealing to parameter scaling laws is also insufficient: applying Chinchilla scaling laws shows that scaling parameters to infinity while holding compute optimal only reduces data requirements by roughly a factor of ten, confirming that humans operate on an entirely different scaling curve. Commercial Amortization and the Automation Bottleneck (8:46) Despite extreme sample inefficiency, frontier labs remain economically viable because training expenses can be amortized across billions of simultaneous user queries. If a human had to read every repository on GitHub to learn programming, the investment would be ruined by sequential human labor limits; an AI model can ingest the data once and serve massive concurrent traffic. Consequently, tasks common enough to be incorporated into training distributions can yield immense commercial returns even without resolving the underlying sample efficiency bottleneck. The real operational frontier lies in tasks demanding out-of-distribution reasoning. While structured roles have historically been automated, complementary effects can increase human labor demand elsewhere, leading Patel to forecast higher overall demand for human software engineers in 2027 than today. To bridge the remaining gaps, AI labs intend to automate AI research itself, betting that automated researchers can eventually resolve the sample efficiency problem. Patel concludes that current AI models can generate immense economic value on in-distribution tasks through sheer brute-force data ingestion, but achieving full general capability hinges on whether automated AI researchers operating on these very architectures can ultimately overcome their own sample inefficiency.

Dwarkesh Patel 认为，近年来人工智能的显著飞跃并非源于样本学习效率的质变，而是一场依赖海量数据分布扩张和强化学习合成数据的工程推进；当前模型并非掌握了通用学习能力的人类式智能，而是依靠数以万亿计的专家数据样本缝合而成的“弗兰肯斯坦怪物”，其技术内核仍是一个庞大的数据黑洞。 样本效率停滞与专家数据黑洞 (0:00) Patel 指出，若将智能定义为掌握某领域技能所需的样本效率，过去几年的 AI 研发几乎未见显著突破，模型能力的提升几乎完全由扩展数据分布和堆叠算力所驱动。强化学习（RL）在本质上是一种合成数据生成机制，即利用庞大算力在验证器或评分规则下摸索何为“良好数据”；然而，这一过程的前提是模型本身必须已具备产出正确解法的先验概率，因而极度依赖海量的人类专家真实轨迹。在 Scale AI 和 Surge AI 等数据标注平台上，从处理遗留文档、起草并购尽调报告到撰写管理咨询方案，各垂直技能都需要成百上千名行业专家编写思维链和参考规范。这一专有数据行业目前的年收入已达数十亿美元，并正快速迈向百亿美元量级。 百万倍人机数据差与开源快速追赶 (2:40) 人类与现有 AI 在数据利用效率上存在悬殊差距。一个成年人在成长过程中接触的自然语言词元总计约 2 亿个（以每小时 2000 词推算），而前沿 AI 模型的训练量高达数十万亿至上百兆词元，两者消耗相差近百万倍。在具身智能领域，人类仅需数小时即可学会遥操作机械臂，或用约 20 小时学会驾驶汽车；相比之下，Unitree G1 等机器人与 Waymo、Tesla 的自动驾驶系统在消耗了数百万小时演示和数个数量级更多的数据后，依然难以稳定应对开放式复杂任务。模型在运用如 GRPO 等算法时，针对单一任务需反复尝试数百至数千次推演才能完成信用分配。Patel 强调，这种对数据的绝对依赖正是开源模型能紧随其后的根本原因：Epoch 报告指出开源模型仅落后前沿模型四个月，因为技术护城河主要在于数据，追赶者只需通过公开 API 蒸馏数据即可迅速抹平差距。 驳演化预训练假说与缩放定律局限 (4:39) 针对 Andrej Karpathy 等人提出的“演化是人类漫长预训练阶段”这一观点，Patel 反驳指出人类基因组大小仅为 3GB 且编码蛋白质区域仅占 1% 至 2%，从数学上看根本不可能预存千亿级网络参数，演化筛选出的仅是超参数与损失函数，大脑神经连接组依然完全是在人类个体寿命周期内从零构建的。此外，演化论无法解释为何已有完整认知的人类学习一门新编程语言只需极少资料，而受训充分的 AI 学习一项边际技能仍需海量专家演示。至于感官多模态的辩解，视听障碍群体在接收远少于 2 亿词元的情况下仍能拥有完整的通用智能。若根据 Chinchilla 缩放定律推算，即使将参数规模推至无穷大以减少训练数据，所需数据量也仅能缩减为原来的十分之一，这证明人类大脑在学习效率上运行在截然不同的缩放曲线上。 极低效率下的商业摊销与研发自循环 (8:46) 尽管 AI 的样本学习极其低效，但各大实验室仍押注其能实现白领工作的自动化，因为分析师、会计和软件工程师的大量工作属于常见的高频工作分布，极易被纳入训练集。人类若需通读全部 GitHub 开源仓库才能学会编程将毫无经济价值，但大模型消耗这些数据后，其获得的编程能力可以在数十亿次并发用户会话中被摊销，从而创造出惊人的商业回报。核心的瓶颈在于工作对“分布外思维”（OOD）的要求有多高；Patel 预测，由于 AI 会作为互补工具提升产出，到 2027 年市场对人类软件工程师的总需求可能反而高于当下。面对顽固的分布外挑战，前沿实验室的最终战略是让大模型率先实现 AI 科研本身的自动化，再由这些自动化研究员去攻坚样本效率难题，从而在现有大模型的特有智能形态之上开启极速迭代周期。 Patel 总结指出，当前 AI 的突飞猛进建立在对海量人工数据和算力的过度依赖之上，这一低效模式凭借商业摊销在特定标准任务中极具威力，但大模型与生物智能在缩放曲线上存在根本差异；若无法在机理上解决样本效率与分布外泛化问题，AI 自动化闭环能否顺利攻克真正未知的科学创新仍充满不确定性。

## Background 本期内容由 Dwarkesh Patel 于 2026 年 6 月 19 日在 RSS 及 Web 平台的个人播客发布。节目简介指出，本篇音频随笔探讨了驱动当前 AI 进步的真正内核、人类与 AI 之间悬殊的样本效率差距，以及这一效率瓶颈是否会阻碍 AI 实验室的终极愿景，并简要介绍了赞助商 Mercury 推出的财务自动化工具 Command。 ## TL;DR 文章指出，近年来 AI 的爆发并非源于样本效率的实质跃升，而是依靠海量专家标注轨迹与算力穷举生成的合成数据进行硬性“缝合”，导致其在语言、具身与驾驶等任务上与人类存在高达数万至百万倍的样本效率鸿沟。针对“生物演化即预训练”或“依据 Chinchilla 定律堆砌参数即可追平效率”等流行辩解，作者从人类基因组信息量上限、盲聋人群体的通用智能事实以及参数扩展的数据缩减极限进行了逐一驳斥。从经济学视角来看，即便训练过程如同消防水龙头般极度低效，AI 也能通过吞食全网数据并在数十亿并发会话中无限分摊成本，从而在常规白领工作上释放惊人的商业价值。然而，面对软件工程等深度依赖“分布外思考”的领域，行业押注用现有 AI 自动化自身研发的设想，仍必须直面一个缺乏类人样本效率的系统能否解决未知科研难题的核心死结。 ## Atmosphere & style 全篇呈现出高度凝练、冷静且极具批判精神的学术独白风格，彻底剥离了科技圈常见的盲目造神狂热。讲述者如手术刀般自如调动信息论、Chinchilla 标度律与生物物理指标，层层拆解 AI 行业神话背后的“数据黑洞”，论证逻辑严丝合缝且充满思辨锋芒。整场倾听体验兼具冷峻的现实主义质感与强烈的智性冲击力，让听众在颠覆直觉的数据反差中重新审视前沿技术演进的真实代价与路径。 ## Bulletin - `00:00` 前沿 AI 的飞跃并非源自底层算法的样本效率提升，而是依托各领域定制专家轨迹与大规模强化学习穷举构建的“数据黑洞”；Epoch 统计显示开源社区仅落后前沿闭源 4 个月，主因是数据极易通过公共 API 快速蒸馏。 - `03:11` 人类成年前仅接触约 2 亿个语言 Token，而大模型需灌入数十万亿甚至上百万亿 Token，二者样本效率相差近百万倍；具身操作与自动驾驶等现实任务同样印证了人类与模型间数个数量级的数据效率鸿沟。 - `05:45` 人类基因组有效编码区仅几十兆字节，在信息论上根本无法储存大脑突触权重，“演化预训练论”不成立；同时 Chinchilla 标度律证明无限堆叠参数最多只能缩减约 10 倍数据需求，参数扩展无法弥合人类高出数百万倍的样本效率。 - `08:46` AI 训练虽极其低效，但其掌握的常规任务能力可在数十亿并发会话中无限分摊成本；而对于高度依赖分布外推理的软件工程等领域，AI 作为互补要素将刺激需求，预计到 2027 年人类程序员岗位需求将不减反增。 ## Key takeaways - `00:00` AI 近年的性能突破并非源于模型样本效率的跃迁，而是通过投入爆炸性算力与覆盖各垂类专家轨迹所构建的“数据黑洞”强行拉升。 - `00:00` 强化学习在本质上扮演了合成数据过滤器的角色，单任务需要生成成百上千条展开轨迹以解决信用分配问题，与人类极低练习频次形成鲜明反差。 - `00:00` 开源模型追平前沿闭源模型仅需约 4 个月，印证了数据蒸馏才是抹平性能差距的核心抓手，而非架构超参数等隐秘训练技巧。 - `03:11` 成年前人类仅接触约 2 亿个语言 Token 即可形成通用智能，而大模型预训练语料高达数十万亿 Token，两者样本效率存在约百万倍的差距。 - `03:11` 具身与自动驾驶展示出严重的样本效率停滞：人类数十小时即可掌握驾驶或遥操作，而顶尖模型吸收数百万小时演示和数十亿英里行驶数据仍无法在开放场景泛化。 - `05:45` 人类基因组中有效蛋白质编码区仅数十兆字节，从物理与信息论底层上彻底否决了“演化已预先编码并储存了人脑全部神经网络权重”的假说。 - `05:00` 自然进化仅提供了超参数与损失函数等骨架机制，人类大脑连接组完全是在出生后的生命周期中以极高效率从零自发构建的。 - `05:00` 模型即便完成海量预训练，每学会一项新边际技能仍需海量数据喂养，暴露出大模型无法像人类一样举一反三的根本缺陷。 - `05:00` 盲聋人群在感官输入被大幅剥夺的前提下依然具备完整通用智能，证伪了“人类依靠海量物理模态 Token 弥补语言数据不足”的辩解。 - `06:30` Chinchilla 标度律的数学推演表明，将模型参数量扩充至无穷大也仅能将所需数据量压缩约 10 倍，参数扩展绝无可能追平人类数百万倍的样本效率。 - `08:46` 尽管 AI 训练过程极其低效，但其习得的常规工作技能可通过数十亿个并发会话无限分摊成本，从商业与宏观经济学角度来看具有极高净收益。 - `10:00` 白领工作自动化取决于任务对“分布外（OOD）思考”的依赖度，纯规则性与可预测的工作早已被早期技术自动化，剩下的核心阵地普遍要求应对未知边界。 - `10:00` 由于 AI 作为互补要素将大幅降低开发门槛并引爆软件项目规模，预计到 2027 年全社会对人类软件工程师的总需求将超过现在。 - `10:00` 让缺乏样本效率的 AI 去攻克自身底层的样本效率难题面临逻辑死结，看待 AI 递归进化应摒弃“毫无作为”与“降生神迹”的两极化认知。 ## Detailed walkthrough ### `00:00` AI 进步背后的数据黑洞：被忽视的样本效率停滞 衡量智能的核心标准之一是“样本效率”（sample efficiency），即系统在某一特定领域达到流利且胜任的操作水平究竟需要摄入多少数据。回顾过去几年的 AI 突破，业界在训练的样本效率上其实并没有取得人们想象中那样显著的实质进展；模型能力的提升，更多是依靠大幅拓展并改善了数据分布，以及为了生产出这些高质量数据而投入的爆炸性算力扩展。 强化学习（RL）是推动这一轮进化的主导力量。从本质上讲，RL 可以被视作一种高强度的合成数据生成机制：开发者将海量算力砸向验证器（verifier）或充当裁判的判别模型（rubric / judge），以便在茫茫可能的轨迹中筛选出真正优质的行为路径。随后，模型被训练去预测这些正确的展开过程（rollouts），这与最初训练模型预测互联网文本中的下一个词别无二致。然而，这套机制要能运转，前提是模型本身必须具备一定的先验概率去“猜出”正确解。正因如此，如果想要模型在各个具体领域表现得游刃有余，就必须预先喂入数量极其惊人的、跨越各行各业的人类专家操作轨迹。 这种专家级数据的任务针对性与定制化程度令人咋舌。如果在 Mercor 或 Surge 等数据标注平台的招聘页面浏览，便会看到大量令人意外的职位需求：有专门负责将杂乱遗留文档排版成精致 Word 文件的“Word 专家”，有撰写高难度并购尽职调查（M&A diligence）与证券备案文件的资深法律顾问，还有负责产出标准化市场调研的管理咨询师。每个垂类技能背后，都需要数百名专业人士撰写示范答案、制定评估细则并详尽解释解题思维链（Chain of Thought）。这也正是为什么生产此类高质量专家标注与提供沉淀技能的 RL 环境的数据标注行业，年营收能迅速从数十亿美元向数百亿美元狂飙。 这种学习方式与人类有着天壤之别。试想一下，如果一个人需要经历数十年的全日制课程、由数百名教授同时授课、做上数百万道练习题，才能学会把一个 Word 文件排版得体，那是怎样的场景？而且仅对比练习任务的数量还会低估差距：人类学生做一道课后习题可能只尝试一两次，而大模型在采用 GRPO（组相对策略优化）等算法时，为了解决信用分配问题（credit assignment problem），针对每一个任务都需要生成几百甚至上千条轨迹进行筛选。今天的 AI 绝不是一个像人类那样掌握了多项通识技能的单一实体，它更像是一个“科学怪人”（Frankenstein's monster）——由数十亿个精心构建的样例组织切片生拼硬凑缝合而成。 近期 Epoch 报告指出，开源模型落后顶尖前沿模型的时间差仅有 4 个月左右。开源社区和追随者之所以能在短短几个月内迅速抹平与前沿模型的差距，根本原因正是数据才是推进性能的核心驱动力：数据可以通过公共 API 被极为便捷地蒸馏出来，但超参数细节、训练技巧和底层架构优化却很难被直接窃取。如果过去的进步真的主要源于架构算法或优化技巧，追赶的门槛绝不可能像现在这样低。人们仰望 AI 时，看到的是群星璀璨般的全能能力；但在那片星系的引力中心，潜藏着一个肉眼不可见、吞噬一切的“数据黑洞”。 ```mermaid flowchart TD accTitle: 现代前沿 AI 模型的“数据黑洞”生产闭环 accDescr: 图表展示了由海量专家轨迹冷启动，通过 RL 算力穷举展开，再经蒸馏驱动追赶的完整数据黑洞机制。 subgraph ColdStart ["1. 专家先验冷启动"] A["垂类人类专家<br/>(法律/咨询/Word排版等)"] --> B["精细化思维链与操作轨迹<br/>(Mercor / Surge 标注生态)"] end subgraph RLSearch ["2. 算力穷举与合成筛选"] B --> C["基础模型先验预热"] C --> D["大规模 GRPO 展开<br/>(单任务成百上千次 Rollouts)"] E["自动化验证器 / Judge Rubric"] --> F["筛选正确轨迹<br/>(解决信用分配难题)"] D --> F end subgraph Assembly ["3. 模型缝合与扩散"] F --> G["多领域数据硬性缝合<br/>('科学怪人'式模型架构)"] G --> H["前沿 API 开放"] H --> I["开源社区快速蒸馏<br/>(4个月追赶周期)"] end ``` - **开源与前沿差距**：Epoch 统计显示开源模型仅落后前沿闭源模型约 4 个月，证明底层数据蒸馏在能力复现中起到了决定性作用。 - **标注市场估值**：支撑 RL 训练的专家标注与环境搭建生态目前年营收已达数十亿美元，并正在迈向数百亿美元级别。 - **学习机制差距**：GRPO 等强化学习算法在单个任务上需进行数百到数千次 rollout 试错来克服信用分配难题，而人类学习只需极低频次的练习。 --- ### `03:11` 百万倍的差距：人类与模型在现实世界的学习效率对比 为了量化 AI 与人类在样本效率上的悬殊鸿沟，可以从人类日常摄入的数据尺度切入对比。假设一个人在成长过程中平均每小时听到或看到 2,000 个单词——这是一个相当慷慨的估计——那么从出生到步入成年，他所接收的语言数据总量大约为 2 亿个 token（200 million tokens）。然而，当前处于技术前沿的大型语言模型，其训练数据吞吐量普遍在数十万亿到上百万亿 token 之间。人类终其一生所见的文字信息，与模型灌入的语料相比，有着接近**一百万倍**的数量级差异。 这种样本效率的极端低下在具身智能与机器人领域体现得尤为刺眼。任何一个普通成年人，通常只需要短短几个小时的简单教学，就能掌握对任意人形机器人或机械臂的遥操作技能。如果 AI 能够具备人类这种级别的即时学习效率，机器人行业现在就会是一个规模达数十万亿美元的庞大市场，全世界早已布满了数不胜数的宇树 Unitree G1 等机器人不知疲倦地处理各类生产力工作。现实却截然相反：即便研究人员收集了数百万小时的真实遥操作演示数据，模型依然无法胜任稍微复杂或开放性的日常物理任务。 在自动驾驶领域，对比同样令人警醒。一个普通青少年通常只需要大约 20 个小时的实际道路驾驶练习，就能考取驾照合法上路。即使把一个人前 16 年成长过程中用来理解物理世界运转规律、建立空间感与直觉的全部生活经验全部计入折算，其所消耗的环境交互信息量，依然比 Waymo 和特斯拉用来训练自动驾驶全栈大模型的数据量低了足足三到四个数量级。 | 对比维度 | 人类学习者 | 前沿 AI / 自动驾驶模型 | 数据需求量差距 | |:--- |:--- |:--- |:--- | | **语言与通识学习** | 成年前累计接触约 2 亿 tokens（按 2,000 词/小时估算） | 数十万亿至上百万亿 tokens（预训练与后训练语料） | 约 **1,000,000 倍**（模型需吞噬百万倍数据） | | **机械臂与具身操作** | 几个小时上手遥操作任意复杂机械构型 | 汇集数百万小时人类演示，仍无法稳健应对开放环境 | 质的鸿沟（海量数据仍难泛化） | | **道路驾驶技能** | 约 20 小时路考练习（加上约 16 年的常识物理先验） | 特斯拉 / Waymo 累积数十亿英里全天候车队行驶数据 | 约 **3 到 4 个数量级**（几千至上万倍） | --- ### `05:45` 反驳“进化预训练论”：人类基因组装不下神经网络参数 针对上述样本效率鸿沟，AI 圈内普遍存在一种常见的反驳意见，安德烈·卡帕西（Andrej Karpathy）在做客 Dwarkesh 播客时也表达过类似观点：人类不能算作真正的“冷启动”。他们认为，人类在出生时绝非一张白纸，背后经历了地球数十亿年生物进化史的“预训练”；因此，拿人类在短短十几年寿命里见过的少量数据，去对比那些从彻底随机初始化的白纸状态（cold start）开始学习的底层 LLM，本质上是不公平的。 但深入检视生物学硬指标便会发现，这种“进化等同于大规模预训练”的论调根本站不住脚。人类基因组的总数据容量大约只有 3GB，这其中真正负责编码蛋白质的外显子区域仅仅占了整个基因组的 1% 到 2%。 区区几十兆字节的有效编码信息量，从信息论的物理底层上，就根本不可能储存一个被数十亿年进化预先调优好的大脑神经网络所有参数权重。人类大脑之所以能用微不足道的数据飞速掌握语言、物理法则与驾驶技能，必定依赖于某种在根本机制上比当前深度学习要高效得多的样本泛化结构，而非靠基因组暗中携带了庞大的预训练模型权重。 ### `05:00` 反驳演化与多模态反论：人类通用智能无需海量感官 Token 对于“人类之所以学得快是因为演化早已完成预训练”的反驳，更贴切的比喻应当是：演化只是找到了合适的超参数（hyperparameters）和损失函数（loss functions），而在个体的生命周期中，人类依然是从零开始构建大脑的连接组（connectome）——这才是真正等同于神经网络权重和参数的东西。退一步讲，即便承认模型在预训练中吞食的数百万亿 Token 只是在“追赶自然演化”，也依然无法解释为什么模型每获得一项新的边际能力（marginal capability），都必须砸入天文数字的数据。一个受过基础教育的人类，在学习一门新的编程语言时，绝不需要请一百位教授轮番灌输几百万行代码；然而大语言模型即便在预训练彻底完成后，每学会下一项边际技能，依然需要海量微调数据的粗暴填喂。 另一个常见的辩解是“多模态数据差异”——有人主张人类在成长过程中通过眼耳感官接收了巨量的物理世界信号，从出生到成年折合下来可能包含数百亿乃至上千亿 Token 的感官信息。然而，盲人和聋哑人的存在彻底击碎了这一假说。先天失明或失聪的群体哪怕被彻底切断了视觉或听觉的高通量感官输入，依然拥有无可争议的通用智能（general intelligence）。尤其是听障人士，他们无法听到语音 Token，只能依靠手语和阅读来认知世界，其一生摄入的语言 Token 数量甚至远低于此前估算的 2 亿上限。这强力证明了那些所谓的数十亿物理感官 Token 根本不是让人类变聪明的核心所在，也暗示人类与 AI 之间数百万倍的样本效率差距不仅没有被夸大，反而可能被严重低估了。 - **连接组从零构建**：人类大脑的突触连接是在后天发育中形成的，演化并未预设所有权重，仅提供了学习架构。 - **边际学习的巨大鸿沟**：人类掌握先验后学习新语言只需极少样本；AI 掌握先验后掌握新技能仍需庞大数据。 - **感官剥夺反例**：盲聋人群体在感官数据大幅缺失的情况下依然具备完整通用智能，证明海量模态输入并非智能的必要条件。 --- ### `06:30` 参数扩展的数学死局：Chinchilla 定律与样本效率的鸿沟 第三种反驳认为这纯粹是模型规模尚未见顶的问题：依据标度律（scaling laws），更大的模型具备更高的样本效率；人类大脑拥有约 100 万亿（100 trillion）个突触，而当下的前沿模型参数量仅在 5 万亿（5 trillion）左右，因此只要把模型再扩大一到两个数量级，或许就能自然追平人类的样本效率。但只要仔细审视标度律的底层数学结构，就会发现这个愿望完全脱离了现实。 在经典的 Chinchilla 标度律方程中，参数项（parameter term）和数据项（data term）对损失（loss）的贡献是独立相加的。假设要训练一个模型并追求极致的样本效率，试图通过堆砌无限多的参数来把所需数据压缩到最低，代入 Chinchilla 论文中的实证常数计算就会发现：即便将参数量提升至无穷大（infinity），为了达到相同的损失值，所需的数据量最多也只能缩减大约 10 倍（a factor of 10）。而人类的样本效率足足比现有模型高出数千到数百万倍，仅仅依靠扩大当前架构的模型体量，根本无法填平这一天壤之别的沟壑。这无可辩驳地表明，人类大脑正运行在一条完全不同的标度曲线上。 ```mermaid flowchart TD accTitle: 参数缩放与人类样本效率对比 accDescr: 展示无限增加参数量所带来的数据缩减上限与人类样本效率之间的巨大鸿沟。 subgraph ScalingLimit["Chinchilla 标度律边界"] A["前沿模型 (~5T 参数)"] -->|"参数量扩展至极限 (N → ∞)"| B["数据需求仅下降约 10 倍"] end subgraph RealityGap["现实鸿沟"] B -.->|"无法填补 10³ ~ 10⁶ 倍差距"| C["人类标度曲线"] C --> D["数十万至数百万倍更高样本效率"] end ``` --- ### `07:34` 商业赞助插播：Mercury Command 自动化财务流水 在推进高深理论探讨的同时，日常商业运转对自动化也有着极其实际的需求。主持人提到，自己在赚取收入后希望尽快将其投入运营，但又必须预留足够的储备金以应对即将到来的运营支出与预估税费。为了精确测算出需要预留的金额，他使用了 Mercury 银行平台内嵌的原生 AI 功能——Command。 Command 已经整合在 Mercury 体系内，因此天然拥有执行任务所需的一切上下文。用户只需给出一个目标日期，Command 就会自动读取当前账户余额，计入该截止日前所有待收的应收账款（invoices），并调取过去六个月的交易流水，扣除月均支出与已排期的款项。如果平台外部有即将发生的突发支出（例如“注意，7 月份有一笔 12,000 美元的承包商款项需要支付”），只需在对话框中随手备注，Command 便会将其动态纳入测算。由于所有输出都在聊天界面完成并附带底层数据的穿透链接，用户可以极快地交叉验证结果，确认无误后直接下达指令让其起草盈余资金向个人账户的转账。 - **上下文无缝打通**：基于用户完整的历史银行流水与发票系统实时推理，免除手动导入账目的繁琐。 - **自然语言交互校对**：支持对话式临时变量注入（如 12,000 美元承包商预付款），并提供底层数据溯源以供审核。 --- ### `08:46` 样本效率真的致命吗？白领工作自动化的“高吞吐分摊”逻辑 抛开所有技术层面的对比，核心问题在于：样本效率低下真的会阻碍 AI 实验室实现其宏伟目标吗？当前前沿实验室背负着两大核心诉求：第一是实现白领工作（white-collar work）的自动化，第二是实现 AI 研究本身的自动化。就白领工作而言，各大实验室正押注一个核心事实——软件工程师、财务分析师或会计师平日里处理的大多数任务高度常见。因为常见，所以这些任务很容易被大规模抓取并强行塞入训练分布（in-distribution）之中。从各家实验室近期暴涨的营收曲线便能看出，哪怕无法破解人类学习的深层秘密，单靠把这些常规工作分布化，就已经释放出了惊人的商业价值。 从经济学视角来看，用极其低效的数据训练 AI 去做这些工作完全无伤大雅。人类寿命的物理极限根本无法承受 AI 所经历的训练广度与深度——如果一个人患有某种奇特的学习障碍，必须“读完 GitHub 上的每一个公开代码仓库”才能成为一名合格的软件工程师，那么社会根本无法承担培养这个人的成本；当他好不容易完成早期学习时，恐怕早就到了领社会保障金（Social Security）的退休年龄，且毕生一次只能处理一个项目。然而 AI 截然不同：即便训练过程如同消防水龙头喷水般“荒谬地低效”（ludicrously inefficient），AI 习得的能力却可以在数十亿个并发会话中无限分摊（amortized），从而在商业账本上实现极其惊人的净收益。 ```mermaid flowchart LR accTitle: 人类与 AI 训练学习的经济学模型对比 accDescr: 对比人类单体串行学习与 AI 暴力训练后并发分摊的商业逻辑。 subgraph HumanModel["人类学习模式"] H1["单体学习容量有限"] --> H2["无法遍历全部代码库"] H2 --> H3["单线程工作产出"] end subgraph AIModel["AI 暴力分摊模式"] A1["低效海量数据灌注"] --> A2["吞下全网代码库"] A2 --> A3["数十亿并发会话分摊成本"] A3 --> A4["商业回报大幅跑赢投入"] end ``` --- ### `10:00` 分布外难题、软件工程的未来与 AI 自我迭代的迷思 然而，这引发了一个关于工作本质的根本问题：白领员工在日常工作中究竟需要进行多少无法预先训练的“分布外思考”（out-of-distribution thinking）？这取决于具体工种。那些纯粹机械、高度可预测的岗位，早在现代 AI 浪潮来临前就已经被自动化技术吞噬，比如银行柜员与旅行中介。但另一类工种每天都需要处理远离既有数据分布的棘手难题，软件工程正是其中的典型。尽管人们普遍认为软件开发是第一个会被 AI 全盘接管的岗位，但事实恰恰相反：“我敢打赌，到 2027 年，全社会对人类软件工程师的总需求将超过现在”，这在很大程度上得益于 AI 作为互补要素（complementary input）所带来的需求爆发。 面对这类依赖分布外推理的高难度领域，各大实验室的终极应对方案是：先用现有手段自动化 AI 研究本身，再让被自动化出来的“AI 研究员”回过头来攻克底层的样本效率难题。但这随即引出了最核心的死结——一个自身缺乏人类级样本效率的 AI，到底能不能解决通往类人智能路上的未知科研难题？当前舆论对此的讨论显得异常笨拙（clumsy）：人们要么彻底否定 AI 加速 AI 研发的可能性，要么简单粗暴地假设代码跑通之后瞬间会凭空降生一个全知全能的“造物主”（some kind of God pops out）。真正的现实很可能介于两者之间：在一个特定时期内，AI 研发速度确实经历了前所未有的大幅跃升，但这完全是建立在大语言模型自身特有的智能形式之上。关于这一复杂图景的具体推演，将在后续的专题文章中展开。 - **工种分化与分布外挑战**：机械性工作早已被规则代码解决，而软件开发深度依赖应对未知边界情况的能力。 - **2027 年软件人才需求预言**：得益于 AI 赋能带来的开发门槛降低与软件项目爆炸，人类程序员的需求不仅不会缩减，反而将迎来净增长。 - **AI 递归进化的两极谬误**：打破“毫无作为”与“奇点降神”的非黑即白思维，聚焦于基于 LLM 本身特性的新型加速周期。 ## Notable quotes > `00:00` 衡量智能的核心标准之一是“样本效率”（sample efficiency），即系统在某一特定领域达到流利且胜任的操作水平究竟需要摄入多少数据。 — Dwarkesh Patel > `00:00` 今天的 AI 绝不是一个像人类那样掌握了多项通识技能的单一实体，它更像是一个“科学怪人”（Frankenstein's monster）——由数十亿个精心构建的样例组织切片生拼硬凑缝合而成。 — Dwarkesh Patel > `00:00` 人们仰望 AI 时，看到的是群星璀璨般的全能能力；但在那片星系的引力中心，潜藏着一个肉眼不可见、吞噬一切的“数据黑洞”。 — Dwarkesh Patel > `03:11` 人类终其一生所见的文字信息，与模型灌入的语料相比，有着接近一百万倍的数量级差异。 — Dwarkesh Patel > `05:45` 区区几十兆字节的有效编码信息量，从信息论的物理底层上，就根本不可能储存一个被数十亿年进化预先调优好的大脑神经网络所有参数权重。 — Dwarkesh Patel > `05:00` 演化只是找到了合适的超参数（hyperparameters）和损失函数（loss functions），而在个体的生命周期中，人类依然是从零开始构建大脑的连接组（connectome）——这才是真正等同于神经网络权重和参数的东西。 — Dwarkesh Patel > `06:30` 即便将参数量提升至无穷大（infinity），为了达到相同的损失值，所需的数据量最多也只能缩减大约 10 倍（a factor of 10）。而人类的样本效率足足比现有模型高出数千到数百万倍，仅仅依靠扩大当前架构的模型体量，根本无法填平这一天壤之别的沟壑。 — Dwarkesh Patel > `08:46` 即便训练过程如同消防水龙头喷水般“荒谬地低效”（ludicrously inefficient），AI 习得的能力却可以在数十亿个并发会话中无限分摊（amortized），从而在商业账本上实现极其惊人的净收益。 — Dwarkesh Patel > `10:00` 我敢打赌，到 2027 年，全社会对人类软件工程师的总需求将超过现在 — Dwarkesh Patel > `10:00` 人们要么彻底否定 AI 加速 AI 研发的可能性，要么简单粗暴地假设代码跑通之后瞬间会凭空降生一个全知全能的“造物主”（some kind of God pops out）。真正的现实很可能介于两者之间 — Dwarkesh Patel ## Mentioned Dwarkesh Patel, Mercury, Mercury Command, Mercor, Surge, Epoch, GRPO (组相对策略优化), 宇树科技 Unitree G1, Waymo, 特斯拉 (Tesla), 安德烈·卡帕西 (Andrej Karpathy), Chinchilla 标度律, GitHub ## Video description <p>Read the <a target="_blank" href="https://www.dwarkesh.com/p/the-sample-efficiency-black-hole-2">transcript</a> here.</p><p>Thanks to <a target="_blank" href="https://mercury.com">Mercury</a> for sponsoring this essay!</p><p>Mercury just released a new feature called Command, which gives me AI right in my banking platform. And since I use Mercury to run basically my entire business, Command has access to all the info it needs to get real work done. I can ask it to send invoices, or categorize expenses, or even transfer money… and Command just handles it. Learn more at <a target="_blank" href="https://mercury.com/command">mercury.com/command</a></p><p>Timestamps:</p><p>(00:00:00) – What is really driving AI progress?</p><p>(00:03:11) – Comparing human vs AI sample efficiency</p><p>(00:08:46) – Does sample efficiency matter?</p> <br/><br/>Get full access to Dwarkesh Podcast at <a href="https://www.dwarkesh.com/subscribe?utm_medium=podcast&#38;utm_campaign=CTA_4">www.dwarkesh.com/subscribe</a> [Source](https://api.substack.com/feed/podcast/202734201/6bfb394237f089ea0355b8632dfa3537.mp3)

[0:00] Dwarkesh Patel: 衡量智能的核心标准之一是“样本效率”（sample efficiency），即系统在某一特定领域达到流利且胜任的操作水平究竟需要摄入多少数据。 [0:00] Dwarkesh Patel: 今天的 AI 绝不是一个像人类那样掌握了多项通识技能的单一实体，它更像是一个“科学怪人”（Frankenstein's monster）——由数十亿个精心构建的样例组织切片生拼硬凑缝合而成。 [0:00] Dwarkesh Patel: 人们仰望 AI 时，看到的是群星璀璨般的全能能力；但在那片星系的引力中心，潜藏着一个肉眼不可见、吞噬一切的“数据黑洞”。 [3:11] Dwarkesh Patel: 人类终其一生所见的文字信息，与模型灌入的语料相比，有着接近一百万倍的数量级差异。 [5:45] Dwarkesh Patel: 区区几十兆字节的有效编码信息量，从信息论的物理底层上，就根本不可能储存一个被数十亿年进化预先调优好的大脑神经网络所有参数权重。 [5:00] Dwarkesh Patel: 演化只是找到了合适的超参数（hyperparameters）和损失函数（loss functions），而在个体的生命周期中，人类依然是从零开始构建大脑的连接组（connectome）——这才是真正等同于神经网络权重和参数的东西。 [6:30] Dwarkesh Patel: 即便将参数量提升至无穷大（infinity），为了达到相同的损失值，所需的数据量最多也只能缩减大约 10 倍（a factor of 10）。而人类的样本效率足足比现有模型高出数千到数百万倍，仅仅依靠扩大当前架构的模型体量，根本无法填平这一天壤之别的沟壑。 [8:46] Dwarkesh Patel: 即便训练过程如同消防水龙头喷水般“荒谬地低效”（ludicrously inefficient），AI 习得的能力却可以在数十亿个并发会话中无限分摊（amortized），从而在商业账本上实现极其惊人的净收益。 [10:00] Dwarkesh Patel: 我敢打赌，到 2027 年，全社会对人类软件工程师的总需求将超过现在 [10:00] Dwarkesh Patel: 人们要么彻底否定 AI 加速 AI 研发的可能性，要么简单粗暴地假设代码跑通之后瞬间会凭空降生一个全知全能的“造物主”（some kind of God pops out）。真正的现实很可能介于两者之间

[2:40] The correct way to think about these models is not like a human who has learned all these different skills that you see these models displaying. It's more like a Frankenstein's monster, which has been built out of a billion graphs of carefully constructed examples all sewn together.

[0:00] So one definition of intelligence is sample efficiency. That is to say, how much data do you need in a given domain to operate fluently and competently? And it's actually not clear that we've made that much progress in training sample efficiency over the last few years. It seems like more so we've just dramatically widened and improved the data distribution. The main way that AIs have been getting better is from adding more and better data and scaling the compute required to develop that data in the first place. Obviously, RL is the main way that this has happened. You can think of RL as basically a kind of synthetic data generation, where you dump a ton of compute against a verifier or a rubric if you have another one as a judge, and you do this in order to find out what the good data is in the first place. And then you train your model to predict these correct rollouts, much in the same way that you might train that model to predict the next word in internet text. For this process to work, the model must have at least some prior probability to anticipate the correct solution in the first place, which is why you need mind-stretching amounts of human expert trajectories in every single field and skill that you want the model to eventually be competent in. It's hard to overstate how task-specific and bespoke this human expert data is. If you want some intuition I recommend checking out the job descriptions on Mercore or Serge websites There are listings for word specialists who will convert legacy documents into polished word files and legal experts who will write realistic M diligences or securities filings and management consultants who will write up template market research. And it is not only that the data have to be so domain-specific, but there has to be so much of it. Each skill corresponds to at least hundreds of human experts who are generating example completions, writing rubrics, and explaining their chain of thought. There's a reason that the data industry that is producing these expert labels and the RL environments in which these meticulously cataloged skills can congeal is earning billions a year in revenue, soon to be decabillions. Now imagine if it took a couple decades worth of courses with hundreds of concurrent professors and millions of practice tasks for you to learn how to polish a word file. Even the task count difference here understates the gap because the models have to grind their far more numerous tasks, each far harder. Whereas a human student might practice a textbook problem once or twice. With GRPO, these models are generating hundreds to thousands of rollouts per task, and they need to to solve the credit assignment problem. The correct way to think about these models is not like a human who has learned all these different skills that you see these models displaying. It's more like a Frankenstein's monster, which has been built out of a billion graphs of carefully constructed examples all sewn together. Epoch recently reported that open models lag state frontier models by four months I think the reason it is relatively easy for open source and previous laggards to catch up to within months of the frontier is that data is the real driver of progress And data can be easily distilled from public APIs, whereas hyperparameters and training tricks and architectural optimizations cannot. And if the latter were driving most of the progress, then catching up would be far harder than we are observing it to be. It is easy to forget how much data these models are trained on and how much more it is than what we humans see in our lifetimes. We see these AIs as a galaxy glittering with capabilities. But at their center, invisible to the naked eye, holding all the constellations together is an unimaginably massive black hole of data. Just a couple of points of comparison to help drive home how big this difference is. Here's one. If a person sees and hears on average, let's say generously, 2,000 words an hour, then between the time they're bored and the time they're an adult, they'll see about 200 million tokens. Now, by contrast, these frontier models are trained on somewhere between tens to hundreds of trillions of tokens. That is close to a million-fold difference. Here's another point of comparison. If you wanted to you could learn to teleoperate any random humanoid or robot arm within hours and if we could get AIs to learn just as fast robotics would be a decatrillion dollar industry and you have an endless army of Unitree G1s doing all kinds of useful work in the world. But the reason we can't do this is that our AIs learn much less efficiently than we do. And even with the millions of hours of demonstrations that we collected, this is not enough to allow them to perform complex open-ended tasks. And a final point of comparison, a teenager can learn to drive a car with about 20 hours of practice. And even if we include their 16 years of growing up and understanding how the world works and building physical intuition. There's still three to four orders of magnitude less data than Waymo and Tesla are using to train their self-driving car models. Now, I want to deal with a couple of common responses and objections that people have to these kinds of comparisons. One thing people will say, and I think Karpathy said this when it came on my podcast, is that for humans, many billions of years of evolution had to go into basically pre-training us. And so we're being unfair when we're comparing how little data we see within our lifetimes to what these cold-started LLMs, who are just starting off with a totally random initialization, have to learn from. I think this is not the right way to think about it. Our genome is only three gigabytes big, and only one to 2% of it is protein coding. And that is simply not enough space to store the parameters of this network that supposedly evolution has pre-trained. [5:00] The closer analogy is more that evolution found the right hyperparameters and the right loss functions, and that within our lifetime, we are still from scratch building up the connectome in our brain, that is to say, the analogous thing to the weights and parameters of the neural net itself. And even if you granted this comparison and you said, yes, the hundreds of trillions of tokens that these models see to get pre-trained is similar to just catching up to evolution, that still doesn't explain why any new marginal capability that you want to give these models takes so much data. So once you haven't educated, again, you don't need a hundred different professors to teach you how to learn a new programming language. But these AIs, even once they're pre-trained, still require enormous amounts of data to learn the next marginal skill and the next marginal skill after that. Another objection to this kind of comparison is that we're not including multimodal data that we're seeing in our lifetimes. So we include all this sensor information that we see from birth to adulthood that's probably tens to hundreds of billions of tokens of data and my response to this objection is simply that blind and deaf people who have been cut off from all the sensory information still have general intelligence and that suggests to me that all these billions of sensory tokens are not really the thing that is making humans smart and in fact deaf people who don have the ability to hear any tokens who just have to consume them via sign language and reading are probably ingesting far less than the 200 million language tokens that we ballparked earlier, which suggests that even the million-fold difference that we calculated earlier might be an understatement. Okay, the third common objection people make is that we just haven't scaled enough. We have these scaling laws. They tell us the bigger models are more sample efficient. The human brain we know is about 100 trillion synapses. And we have frontier models that are currently around 5 trillion parameters. And so maybe we could just achieve human level sample efficiency if we made these models one to two orders of magnitude bigger. The reason this objection is off mark is actually quite interesting. So if you look at the way the scaling loss equations work, they tell you that the parameter and data terms are added to the loss independently. So suppose you have a model and you've trained it to compute optimally, and you say, I want to be sample efficient. I want to use as little data as possible, and I'll throw in as many parameters as is necessary to make that happen. So take the constants from the chinchilla scaling law paper. Even if you increase the number of parameters by infinity, that would only decrease by a factor of 10 the amount of data that you need in order to keep the same loss. Humans are somewhere between thousands to millions of times more sample efficient than these models So scaling the size of current models simply can make up for that discrepancy And this really does suggest that humans are in a different scaling curve altogether As soon as I earn money, I want to put it to work. But I also need to save for things like upcoming expenses and estimated taxes. So to figure out exactly how much I need to set aside, I ask Command. Command is AI that is built into Mercury, which is my banking platform. And since I already use Mercury to run my entire business, command has access to all the information it needs to get work done. I just tell command the date I'm interested in and it does the rest. It takes my current balance and adds whatever invoices we do by the cutoff. Then it reviews my last six months of transaction history. So I can subtract out my monthly average expenses along with any scheduled payments. And if there's anything relevant coming up that's not in Mercury yet, I can just flag it. Things like heads up, there's a $12,000 contractor payment that's slated for July. And that gets included in the final output. Because this is all happening in chat and every answer has links to the underlying data, I can easily double check command's work. And once I'm convinced, I can just tell command, all right, that looks good. Just transfer the surplus to my personal account. And you will immediately draft the transfer for me to approve. Command is live now. Visit mercury.com slash command to learn more. Mercury is a fintech company, not an FDIC insured bank. Banking services provided through Choice Financial Group and column NA members FDIC AI generated responses and suggested actions may vary and are not guaranteed Okay all these nerdy comparisons aside you might ask why do we even care about sample efficiency? Is this actually necessary for the labs to achieve the two overarching objectives they have, which are, one, automate white-collar work, and two, automate AI research itself. The bet that the labs are making with white-collar work is that the common tasks that a software engineer or analyst or accountant needs to do are common. And as a result, you can bring them into the training distribution quite easily. If you look at the revenue curves of these labs over the last few months, it does suggest that there's an enormous amount of value from bringing into distribution these kinds of common tasks, even if we can't replicate whatever is making human learning so special. And it might be more inefficient to train AIs to do these kinds of tasks than it is to train humans. But so what? Human lifespan simply does not allow for the the quantity and the breadth of training that these models experience. If you, as a human, had some weird learning disability where you needed to read through every public repository on GitHub before you could be a competent software engineer, then it would simply not make sense to train you up. You'd be on Social Security by the early stages of your education, and even once you were trained, you would only be able to work on one project at a time. But AIs can learn these skills by firehosing [10:00] training at a time, and what they learn can be amortized across billions of sessions at once. So we can be ludicrously inefficient in training them up and still be wildly in the green. And then there's a question of, well, how much out-of-distribution thinking do white-collar employees need to do that you simply can't train for in advance? This is more a question about the nature of different jobs than it is a question about AI research. And it also depends on which job you're talking about. Some jobs are so mechanical and predictable that we were able to automate them long before the modern era of AI. For example, bank tellers or travel agents. But there are other jobs which require dealing on a daily basis with problems that are quite distant from the data distribution. I think software engineering is probably one such. This is the job that AIs are supposed to take first, but I would be willing to bet that there's overall more demand for human software engineers in 2027 than there is right now, largely due to the complementary input of AI. The last plans for this latter category of jobs is first to automate AI research and then have the automated AI researchers solve the sample efficiency problem. So then the question is, can AIs, which do not have human-level sample efficiency, nonetheless solve the remaining research problems that stand on the way of human-like intelligence and learning? This is a very complicated question, and I'll have to address it in a much longer future blog post. But just to tease it a bit, I think that the way that people currently think about an intelligence solution is very clumsy, because either people dismiss the possibility of AI speeding up AI progress altogether, or they assume that some kind of God pops out the other end. They don't reason carefully about what it looks like to have a period where AI progress is much faster than usual, but have that happen atop LLMs and the particular kinds of intelligences that LLMs are. But I'll save that for next time. In the meanwhile, if you want to read this blog post or all the other blog posts I write or be alerted when I write a future blog post, go sign up for my newsletter at my website, dwarkesh.com. All right. I'll see you later.