2026.09.22 | Substack | LLM | Latent Space: Jev CEO Interview

音频信息

音频标题: Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI - Latent Space: The AI Engineer Podcast | Spotify

相关标的：undefined

议题：

完整音频链接：).mp3

音频纪要总结

TypeSafe AI 专家访谈纪要

1. Jev的产品定位与核心价值

a) Jev究竟是什么，与传统大语言模型有什么不同？

• Jev是TypeSafe推出的首个“大型可编程模型”，也被定义为“机器原生的System 1模型”。

• 传统预训练大语言模型主要用于互联网文本补全，RLHF模型主要用于聊天和指令遵循；Jev则以“代码成为模型输出的直接使用者”为目标，强调输出可被软件系统直接消费。

• Jev不是简单的聊天机器人、决策模型或函数调用模型，而是一种面向软件集成的新模型类别：

1. 输出需要具备类型安全和可验证性；

1. 模型需要深入程序内部，成为软件依赖项；

1. 重点不是生成更自然的文本，而是稳定完成程序要求的判断与操作。

• “System 1模型”比“决策模型”的范围更广，因为后续还会出现不以决策为基本元素的机器原生模型。

• Jev只是这一类别的首个产品，公司仍有其他尚未发布的模型与产品。

b) Jev最核心的优化目标是什么？

• Jev的核心指标是“每美元所能获得的智能”，即单位成本智能水平。

• 名称Jev源自“杰文斯悖论”，其品牌定位是始终处于单位成本智能水平的前沿。

• 公司认为，成本和速度本质上都是用户为结果支付的代价，真正重要的是智能及其实际可用性。

• Jev获得关注的关键在于同时实现了：

1. 更低成本；

1. 更快调用；

1. 智能水平基本不下降。

• 其中最难的并非“更快、更便宜”，而是在此基础上保持足够高的智能水平。

c) 为什么System 1是当前重点，而不是继续强化System 2推理？

• 受访者认为，预训练模型本质上更接近System 1思考者，擅长快速、凝聚式、单跳判断。

• Jev在单跳任务上的表现突出，但随着多跳推理步数增加，性能会逐步下降。

• RLVR在数学等System 2任务上取得了明显成果，但能力分布尖峰化、脆弱且不均衡，并非所有工作都适合依赖长推理链完成。

• Jev并未永久放弃System 2任务，而是优先发掘模型当前最具实用价值的能力，并以投资回报率决定是否采用新的推理方式。

• 公司不承诺坚持某一种技术路线；只要新的推理方式不再缓慢、低效和脆弱，并具备足够经济价值，就可能被采用。

2. RLCD、数据战略与模型训练方法

a) RLCD与RLHF、RLVR的本质区别是什么？

• 受访者认为，区分不同训练范式的重点不在具体算法，而在其“北极星任务”：

1. RLHF的北极星是指令遵循和取悦人类；

1. RLVR的北极星是优化可程序验证的任务和基准；

1. RLCD的北极星是支持程序，使模型能够进入软件闭环。

• RLCD并非单纯为了创造新术语，而是代表一种新的任务方向：提高AI在程序化应用中的可靠性。

• 即使未来不采用某篇论文中的具体算法，只要核心目标仍是支持程序化应用，仍可归入RLCD范畴。

• RLCD的最终目标是将人从重复流程中移除，使模型能够在后台被代码持续调用，实现自动化。

b) 为什么TypeSafe将自己视为“数据实验室”？

• 公司认为模型能力本质上来源于数据，提升可靠性中“9”的数量，关键同样是数据。

• 不同任务对应不同数据形态：

1. RLHF的数据是人类反馈；

1. RLVR的数据更接近可验证环境；

1. RLCD拥有自身独特的数据结构和任务数据。

• 公司强调其方法并非外界通常理解的简单合成数据生成，而是围绕具体任务设计、审视和修正数据。

• 即使具备使用用户数据的条件，公司也不希望依赖用户数据训练，主要原因包括：

1. 真实用户问题呈幂律分布，重复内容过多；

1. 容易使模型过拟合当前高频任务；

1. 会削弱模型面向未来未知应用的通用性；

1. 可能进一步导致智能割裂和分布不均。

• 数据工作的目标不是适配某个特定场景，而是发现认知核心中的不平整之处，并从过去、现在和未来的不同维度进行修补。

c) 公司如何看待预训练、中期训练和后训练？

• 公司支持中期训练，认为其本质是更复杂的课程式训练，可以在不重新预训练的情况下提高模型稳健性。

• 短期内可通过微调快速注入能力；长期内，模型反复接触相关内容后，能力会深入权重并转化为更稳健的System 1能力。

• 受访者明确表示，即使获得10亿美元，也不会优先重新进行预训练：

1. 预训练成本过高；

1. AI工程可以通过模型切分、组合、级联等方法解决大量问题；

1. “弗兰肯斯坦式”模型组合虽然不够优雅，但具有现实效果。

• 公司会采用激进的模型组合和优化方式，以维持单位成本智能水平的帕累托最优。

d) 是否会向用户提供微调能力？

• 当前未提供微调，未来可能考虑，但尚无承诺。

• 公司对微调的主要顾虑是：

1. 狭窄任务微调可能损害模型的通用性；

1. 通用能力有助于模型处理狭窄任务中的边缘案例；

1. 历史上多家闭源模型厂商推出微调后又撤回，说明实际效果有限。

• 可能的替代路径包括：

1. 调整提示词和任务拆解方式；

1. 通过置信度阈值控制风险；

1. 将小模型和大模型进行级联；

1. 提供覆盖不同成本与智能水平的模型梯度；

1. 根据技术栈不同环节动态选择模型。

• 对于极窄场景，如果微调能帮助模型跨过可用门槛，公司仍会务实评估。

3. 可靠性、校准与评估体系

a) 公司如何定义“可靠性”？

• 可靠性是公司最核心的长期目标，不仅指平台正常运行时间，还包括：

1. 模型是否足够智能；

1. 是否始终完成预期任务；

1. 输出是否可预测、可校准；

1. 输入发生轻微变化时，输出是否保持稳定；

1. 模型能否作为依赖项长期在后台运行。

• 公司希望将可靠性提高到类似数据库查询的水平：开发者不必再思考模型是否可用，而是在需要智能时直接调用。

• 受访者认为，当前大型推理模型虽然非常智能，但可靠性甚至可能达不到实习生水平，因此大量具备经济价值的自动化场景仍无法落地。

• 公司希望先可靠地自动化简单工作，再逐步进入复杂工作，而不是优先追求极难任务的单次高分表现。

b) 确定性和鲁棒性哪个更重要？

• 确定性是指相同输入始终得到相同输出，鲁棒性是指语义相似的输入得到相似输出。

• 公司认为鲁棒性比绝对确定性更重要。

• 内部测试会在提示词中加入UUID等语义无关的随机内容，检查模型是否仍给出相似判断。

• 确定性在单元测试等场景中具有价值，公司技术上可以提供确定性模型，但代价是降低单位成本所能获得的智能水平。

• 由于公司长期面临GPU资源限制，不会优先承诺确定性；若用户能够证明其价值，未来可以考虑提供。

c) 校准为什么如此重要？

• 校准的含义并不是模型永远正确，而是：

1. 较低分值对应较低正确概率；

1. 较高分值对应较高正确概率；

1. 用户可以基于概率设置不同阈值和处理流程。

• 校准能够支持模型级联：

1. 高置信度结果直接执行；

1. 中间置信度调用更大模型；

1. 低置信度或特殊场景升级给人工。

• 公司认为，传统字符串模型为了生成长且表面正确的文本，往往被迫变得过度保守或过度自信，导致概率空间失真。

• Jev试图通过结构化输出和概率化原语，让开发者能够直接控制判断阈值，而不是依靠提示词要求模型“更加谨慎”。

d) 为什么公司反对公开基准测试？

• 公司并不反对评估，但强烈反对将公开基准作为模型能力的主要判断方式。

• 公开基准的主要问题包括：

1. 容易被针对性训练；

1. 模型可能无意中过拟合评测数据；

1. 指标饱和后不能反映真实工作流能力；

1. 会诱导实验室迎合指标，而非提高真实智能和可靠性。

• 公司拥有内部评估体系，会绘制不同方案的成本、速度与智能表现，用于判断是否处于帕累托前沿。

• 内部强调不得迎合评估指标，必须避免对模型能力进行自我欺骗。

• 长期来看，最有效的评估方式是将模型接入具体工作流，衡量其对真实任务和经济结果的影响。

4. API设计与开发者使用方法

a) Jev提供了哪些核心API原语？

• 当前主要有3类原语：`choice`、`score`和`Nul`。

• `choice`：

1. 接近枚举和switch/match语句；

1. 可以用于在多个离散选项中进行选择；

1. 比传统“function call”更符合程序内部的表达方式。

• `score`：

1. 表示连续判断或概率分数；

1. 可用于排序、阈值判断和质量评分；

1. 并不等同于传统的int或float。

• `Nul`：

1. 名称来自Bernoulli；

1. 表达类似布尔判断的连续概率；

1. 不直接称为bool，是为了避免用户将其误解为确定性的true/false。

• 这些概念可以映射到编程原语，但并不与传统类型完全相同，公司未来还会增加更多机器原生类型。

b) 开发者应如何组织Jev的输入？

• `state`、`instructions`和`criteria`等输入均可使用结构化JSON对象。

• 公司建议保留原始结构，不要先将所有内容转成字符串，再塞入系统消息模板。

• 理想方式是将模型视为一个AI函数：

1. 明确当前函数状态；

1. 从可用变量中选择相关子集；

1. 将每层语义以嵌套结构传入；

1. 让代码负责确定性规则，让模型负责语义判断。

• 受访者将系统消息比喻为“糟糕的全局变量”：把所有要求一次性塞入系统消息，会让系统难以验证和维护。

• 每增加一层嵌套，模型推理难度都会上升，因此公司正在持续优化深层结构化输入能力。

c) 使用Jev时应该如何拆解任务？

• 建议将问题拆解到“最小语义单元”，每次只处理一个明确的底层问题。

• 对生产系统而言，应当：

1. 使用结构化、字面明确的提示；

1. 对关键对象使用反引号等方式清晰标记；

1. 将复杂判断拆成多个独立问题；

1. 为每个问题设置单独阈值；

1. 将失败案例保留为测试用例；

1. 通过代码实现绝对约束。

• 例如，在设计拒绝机制时，不应只问“是否应该拒绝”，而应分别判断不同风险条件，再由程序汇总。

• “不得读取某个目录”“不得传递API密钥”等硬性要求，应通过代码和权限机制保证，不能仅依靠模型提示词。

• 小任务拆解能够带来更好的可测量性、可修复性和可验证性，也有助于发现模型真正不擅长的环节。

d) 哪些任务不适合当前模型直接承担？

• 自动化交易等高层级、高风险任务目前仍需谨慎，模型可能尚不足以独立完成。

• 当模型不擅长识别讽刺、特殊客户身份或复杂边缘情形时，应通过低置信度区间升级给人工处理。

• 公司不认为所有System 1任务都会自然成功，实际能力仍需通过经验评估。

• 如果特定任务当前无法可靠完成，应承认模型能力边界，而不是通过过度承诺推动上线。

5. 产品表现、用户增长与平台运营

a) Jev发布后的市场反馈如何？

• Jev原计划只是一次相对低调的研究预览，公司并未预料到会形成大规模发布效应。

• Discord社区规模已达到约10万人。

• 发布视频观看量已从3,600万进一步增长至3,800万。

• Jev发布后迅速成为社交媒体热门话题，并出现约50个模仿或复制项目。

• 发布前，超过一半的试用者不理解产品，早期收入几乎可以忽略不计；发布后，用户迅速开始申请更高速率限制，甚至提出向公司提供GPU。

• 受访者认为，增长的核心并非传统市场营销，而是开发者社区形成了强烈共鸣，并主动探索出大量超预期应用。

b) 公司最关注哪些增长指标？

• 公司不重视候补名单注册量，因为其中许多用户可能并非开发者，只会进行少量试用。

• 最重要的指标是每日Token使用量，尤其是夜间持续运行的调用量。

• 夜间使用量说明Jev已经被机器和软件流程持续调用，而不只是人类用户进行一次性体验。

• 对开发者平台而言，单个重度用户通过for循环持续创造价值，其调用量可能远超大量轻度注册用户。

• 真正的运营压力来自重度用户持续提高速率限制，而非候补名单规模。

c) 平台稳定性表现如何？

• 公司在空前发布流量下维持了较高正常运行时间，受访者认为平台可用性的“9”可能多于Anthropic。

• 这一成绩主要归功于平台团队。

• 未来平台能力不仅需要覆盖正常运行时间，还需要覆盖：

1. 模型输出可靠性；

1. 全球部署；

1. 可扩展性；

1. 速率限制管理；

1. 模型版本兼容与长期支持。

d) 模型版本与长期支持策略是什么？

• 公司不会承诺模型部署后永不改变，也不会长期维护无限数量的版本。

• 由于模型迭代速度可能非常快，持续维护大量历史版本会导致服务集群碎片化，并降低整体效率。

• 当前的1.13.0版本可能被暂时设为长期支持版本，因为已有大量开发者依赖。

• 长期目标是建立比行业现有方式更完善的LTS机制，但目前仍在研究。

• 公司可以承诺模型会持续变得更智能，但无法承诺所有版本的行为完全一致。

• API与第一方产品的区别在于，API更新必须更加谨慎，避免破坏开发者依赖。

6. 商业场景与竞争格局

a) Jev目前最主要的应用场景有哪些？

• 公司将主要应用场景归纳为4类：

1. 编程智能体：目前最主要的收入和调用量来源；

1. 暗数据处理：处理大型企业长期囤积、因成本过高而未分析的数据；

1. 实时应用与“验证一切”：实时决策、LLM调用验证、产品流程监控；

1. 智能软件：将Jev作为可组合的软件组件，创造过去无法实现的功能。

• 计算机操作可以归入实时应用或智能软件方向。

• 游戏也是潜在方向，包括NPC行为、自动战斗和半自动战斗等，但并非当前业务重点。

b) “验证一切”具体如何实现？

• “验证一切”是指对其他大语言模型的调用进行低成本、程序化验证，类似AI可观测性机制。

• 具体方式包括：

1. 为每条消息或状态添加ID；

1. 对每个ID并行提出多个判断问题；

1. 同一份长状态只支付一次输入成本；

1. 针对不同消息进行大量低成本判断。

• 可采用“每1次推理调用，配合1次、10次或100次Jev调用”的方式。

• 公司更希望帮助用户降低总支出，例如将高成本推理调用减少一半，同时增加低成本Jev验证调用，而不是单纯扩大调用金额。

c) 实时性能具有什么商业价值？

• 电商、AI助手等实时产品对延迟高度敏感，每减少10毫秒都可能改善用户体验。

• 即使延迟已经低于100毫秒，再缩短一半也意味着可以增加智能调用次数，形成更丰富的连续交互。

• 公司区分两个指标：

1. 每美元智能水平；

1. 每秒智能水平。

• Jev当前主攻每美元智能水平，但公司认可每秒智能在实时行业中的巨大价值。

• 对后台MapReduce式任务而言，成本通常比延迟更重要；对面向用户的实时任务而言，100毫秒与1毫秒之间可能存在明显产品差异。

d) Jev将如何影响编程智能体竞争？

• 当前Claude Code和Codex大体处于领先位置，但其产品体系主要围绕单一模型构建。

• 开放式编程智能体正在积极接入Jev，有机会形成多模型组合能力。

• 现阶段各类开放智能体围绕while循环形成的差异有限；一旦某个产品找到Jev独有的杀手级应用，用户可能迅速迁移。

• 开放式智能体之间可以快速复制成功模式，而封闭、单模型架构可能更难应对。

• TypeSafe希望与所有编程智能体集成，不希望作为基础设施方表达排他性倾向。

7. 安全对齐、能力对齐与平台边界

a) 为什么公司反对在API底层采用广泛拒绝策略？

• 受访者并不反对安全原则，但认为底层API中的随机拒绝会破坏软件可靠性。

• 对聊天产品而言，安全对齐符合产品和品牌需要；例如ChatGPT可以决定不支持某些内容。

• 对API而言，模型可能作为后台依赖项运行，一旦随机拒绝，就可能导致整个软件流程失效。

• 公司将拒绝响应视为一种潜在“类型错误”：开发者无法围绕不可预测的拒绝行为稳定编程。

• 受访者认为，安全对齐与能力对齐存在区别：

1. 能力对齐是准确完成用户预期任务；

1. 安全对齐是遵循模型提供商的价值体系和限制。

• 对开发者而言，优先级应是能力对齐、可预测性和可测试性。

b) 公司如何看待AI被用于有害用途的风险？

• 受访者明确表示不希望技术被用于伤害他人，也希望通过产品和组织影响推动正面用途。

• 但其个人立场是，不应在通用智能基础层通过技术手段进行过度拟合式限制，原因包括：

1. 会割裂模型智能；

1. 会降低通用性和可预测性；

1. 基础层不应掌握所有下游任务细节；

1. 软件工程任务本应拆解，API提供商不一定知道完整用途。

• 受访者将基础智能类比为数据库，认为数据库本身不应负责判断查询是否被用于不当用途。

• 该观点主要针对API基础层，不否认具体产品、法律和务实场景需要相应安全机制。

c) 如何看待“放慢前沿AI发展”的主张？

• 受访者认为，相关讨论常隐含一个狭窄假设：所有实验室都必须持续扩大RLVR，并赋予模型更强的开放行动能力。

• 对TypeSafe而言，RLVR的最优投入可能是0，公司选择的是另一条以程序可靠性为目标的路线。

• 其观点是，危险并非唯一技术路径的必然结果，而可能来自特定任务选择、训练方式和能力开放方式。

• 真正承担方向选择责任的应是研究人员与实验室，而非不了解技术路径的普通公众。

• 公司希望通过实践证明，AI还可以沿着软件可靠性、可组合性和经济自动化的方向发展。

8. 经济目标与长期愿景

a) 公司希望Jev最终带来什么经济影响？

• 公司希望AI真正进入经济流程，而不只是成为前台聊天界面。

• 受访者提出的长期目标是：5年内推动全要素生产率（TFP）增长达到3%。

• 当前模型虽然可以解决高难度数学问题，但能够承担的全球经济价值工作仍接近0；Jev可能迈出了第一步，但预计目前仍不到1%。

• 当AI真正产生经济影响时，应当反映在生产率和宏观统计数据中，而不仅是基准测试分数或社交媒体热度。

• 受访者认为，这不会必然造成大规模失业，而会释放大量重复、基础、低满足感的工作，让人类从事更有价值的任务。

b) Jev将如何改变SaaS行业？

• 2019年至2026年，软件本体变化有限，许多产品只是在侧边增加聊天机器人。

• 企业无法信任聊天机器人完成涉及切身利益的决策，因此AI尚未深入业务系统。

• Jev的可靠性和可组合性有望推动“反向SaaS末日”：

1. SaaS不会被AI简单取代；

1. SaaS公司最了解哪些任务值得自动化；

1. AI将显著增强原有SaaS产品；

1. 智能最终会融入后台，而不是始终成为前台主角。

• 长期目标是让AI像数据库、TCP或Visa一样成为普通而可信赖的基础设施。

c) 公司未来会推出哪些类型的产品？

• 未来不会只提供一种规模或一种模型。

• 公司计划构建覆盖不同智能水平、成本和速度需求的模型梯度。

• 所有产品必须归入统一愿景，而非无序推出大量实验功能。

• 未来的机器原生模型不一定以“决策”为基本元素，可能出现其他全新原语和能力形态。

• 公司希望产品投入始终围绕其宣言中的3个坐标轴展开，而不是采用“把产品扔到墙上看哪个能成功”的模式。

9. 技术路线与未来研发方向

a) 是否会增加视觉、多模态和计算机操作能力？

• 视觉、多模态和计算机操作均在考虑范围内，但尚无具体承诺。

• 公司需要在两种产品理念间取得平衡：

1. 提供用户明确要求的功能；

1. 提供公司判断用户真正需要的功能。

• 如果过度坚持后者，容易形成对开发者不友好的“保姆式管理”；如果完全听从用户，则可能牺牲产品质量和品牌信任。

• 公司认为，多模态能力是否有效需要实证验证，增加数据和投入并不保证问题必然得到解决。

• 计算机操作的用户演示超出公司原有预期，但受访者希望看到其在真实动态环境中持续运行，而不只是形成漂亮演示。

b) 公司如何看待游戏中的智能应用？

• 游戏智能被视为非常有吸引力的外部研究方向。

• 潜在应用包括：

1. 智能NPC；

1. 动态故事与角色关系；

1. 自动或半自动战斗；

1. 玩家通过自然语言指挥队伍。

• 不一定需要在每个游戏循环中调用Jev，也可以结合简单状态机控制成本。

• 公司未来可能为相关实验项目提供额度，但目前基础设施和额度发放机制尚未完善。

c) 如何让编程智能体摆脱KV缓存束缚？

• 当前编程智能体为了提高效率，需要不断向KV缓存追加上下文，从而被锁定在单一模型和线性上下文中。

• 这限制了路由、子智能体、压缩、状态管理、抽象和任务分解。

• 受访者建议探索新的智能体架构：

1. 将任务组织为带标签的分层子任务树；

1. 只在需要时检索相关上下文；

1. 让子智能体共享计算机内存中的结构化状态；

1. 区分只读流程与写入流程；

1. 使用智能锁协调多个智能体并发工作；

1. 通过智能记忆检索替代每次从头加载全部历史。

• 该方向尚未被证明一定有效，但具备较高研究价值，公司希望外部开发者共同探索。

10. 公司战略、组织文化与招聘计划

a) 公司为什么将开发者置于最高优先级？

• CEO在发布后的高强度日程中，仍优先参加社区活动，而非只会见重要投资人与贵宾。

• 公司认为，Jev的增长核心来自开发者社区，而不是传统市场营销。

• 未来即使企业客户增长，公司也不希望重演“从开发者起步，最终只服务企业客户”的路径。

• 公司正在考虑推出对开发者比对企业客户更有利的产品和政策。

• Discord、社区大会、播客和长篇交流将继续承担建立信任和收集反馈的作用。

b) 发布前后的产品市场匹配发生了什么变化？

• 发布前，公司反馈较差，非技术团队担忧用户不理解产品，超过一半的试用者不知道该如何使用。

• 当时产品更像“维生素”而非“止痛药”，公司甚至考虑由前线部署工程师直接替客户开发软件。

• 发布后，开发者迅速找到应用场景，用户主动申请更高限额，形成强烈增长。

• 受访者因此质疑传统产品市场匹配调研的局限：产品和市场可能已经存在，但用户在真正体验和看到社区案例之前，无法准确表达需求。

• 公司认为，用户教育曾是最大风险，但目前用户创造出的应用甚至优于公司原有演示。

c) 当前重点招聘哪些人才？

• 创始市场负责人

1. 公司目前尚无正式市场人员；

1. CEO擅长塑造鲜明风格，但需要将常规市场工作授权给专业人员。

• 数据/模型能力人才

1. 内部称为“模型能力”人才，本质是数据人才；

1. 需要理解任务、判断数据质量、发现模型能力缺口；

1. 公司希望提升数据工作的组织地位。

• 平台与基础设施人才

1. 负责全球部署Jev；

1. 降低因光速和地理距离带来的延迟；

1. 欧洲用户目前只能获得约3倍速度提升，而非约100倍，主要因为当地尚无服务器部署。

• 新产品与新模型人才

1. 公司不希望成为只围绕Jev发展的单一产品公司；

1. 需要持续开发其他机器原生智能形态。

• 公司强调内部平等文化，团队可以直接调侃和质疑CEO，减少对管理层的过度尊重。

d) 对希望离开前沿实验室创业的人有什么建议？

• 受访者不建议仅因为缺少自由、资源或关注，就成立新的AI实验室。

• 多数新实验室的问题在于：

1. 缺少明确的北极星任务；

1. 只是重复已有工作；

1. 以研究人员身份和融资为目标，而非解决现实问题；

1. 真正推动技术前沿的概率较低。

• 如果只希望自由开展研究，现有大型实验室可能仍是更合适的平台。

• 只有在明确认同某项原则、找到有现实价值的新任务，并愿意长期解决问题时，创业才具有充分理由。

结构化总结表格

新版prompt提炼之后的转译

提问者：本周是一个特殊时刻，因为我的好友 Diogo 发布了 Jev，几乎占据了整个时间线。你现在感觉如何？情绪状态怎么样？

回　答　者：从未这么糟糕过。我现在整个人像一具疲惫不堪的躯壳，因为有太多事情正在发生。作为一名技术型 CEO，我需要处理大量紧急问题。

但在精神层面，我感觉自己与现实更加同步了。这些年来，我在 Over Under 活动中一直说，整个 AI 领域就像游乐场里的哈哈镜屋，所有人似乎都陷入了疯狂，说着各种毫无道理的怪话。但就在本周，人们好像终于看到了：AI 可以远远超出过去的想象。由 AI 推动经济革命，再次成为一种现实可能。这太棒了。

开发者理解了这一点。我想向开发者表达由衷的感谢。我对这个社区以及正在发生的一切都感到无比振奋，一切都非常美好。

提问者：你昨天提到，你决定优先参加社区大会，而不是会见很多重要投资人之类的贵宾，因为你希望把最多的注意力放在工程师和开发者身上，对吗？

回　答　者：是的。当时我正在与一些非常重要的人交谈，具体是谁可能不便透露，但这让我觉得有些不自在。我对很多事情可能过于真诚。在排得非常满的日程中，如果社区交流不在其中，我会觉得不妥。理想情况下，我希望一直与社区交流。

我甚至在想，走来演播室的路上是否可以举办一场社区大会，但随后觉得那样实在太疯狂了。

提问者：你们一直在 Discord 上举办社区大会，现在 Discord 已经有 100,000 人，你的 Twitter 也迅速获得了大量关注。此前在 AI 活动上，你还说“请关注我”，却连自己的账号名称都没有提供。你怎么看这些变化？

回　答　者：我是个新手，确实非常不熟练。

提问者：但这反而形成了一种积极的氛围，因为你并不懂得如何宣传自己。

回　答　者：我发帖说：“天哪，我们3个都成了热门话题。”有人提醒我，那只是个性化信息流。我才意识到，当然会对我显示为热门话题，因为那些正是我点击过的内容。

提问者：祝贺你们取得的一切。之后有更多细节时，我们还会继续讨论。对于那些尚不了解情况，或者希望获得明确定义的人来说，Jev 究竟是什么？

回　答　者：这是个很难回答的问题。现在面对这个问题，我感到轻松的第1件事，就是以后不必再向父母解释，因为 ChatGPT 可以直接为他们说明。

在我看来，我们需要一种新的模型类别。我们并不执着于如何命名这一类别。目前想到的最准确名称是“System 1 模型”。之所以这样命名，自有其原因；我们不称之为“决策模型”，也是因为 System 1 的范畴超出了决策模型。目前我只能透露这些。

我们原本没有预料到这会成为一次重大发布，因为我们还有尚未推出的内容。

提问者：你们本来应该称它为一次低调的研究预览。

回　答　者：从某种程度上说，确实如此。我们将这一类模型描述为“机器原生的 System 1 模型”。

回　答　者：第1类是大型可编程模型。这类模型的目标是让代码成为其输出的使用者。预训练大语言模型主要用于互联网内容的自动补全，RLHF 模型等聊天机器人和指令遵循模型主要用于回复文本，RLVR 则处于与 RLHF 之间一个比较模糊的地带。相比之下，大型可编程模型的输出可以直接由代码使用，因此称为“类型安全”。

我们真正希望实现的是让 AI 尽可能强大，而实现这一目标的方式是将其与软件集成。我们设计和优化的不只是模型的外部表现，还包括其深层内部结构，目标都是让它更适合软件。

Jev 是我们的第1个大型可编程模型，也可以称为 System 1 模型。Jev 针对单位成本的智能水平进行了优化，其名称源自杰文斯悖论。Jev 代表的模型将处于单位成本智能水平的前沿。

人们经常争论可靠性、成本、校准和速度之中哪一项最重要。模型还可以通过其他方式优化，但如果真正擅长机器学习，就会明白其中的一切都关乎权衡。我们正在全力优化这一方向。

提问者：对我而言，校准是过去较少有人讨论的新问题。我们此前曾与 Hugging Face 的 Clementine Fourrier 进行过一期访谈。她提到，模型正在向用户最希望听到或概率最高的回答发生模式坍缩，而不是反映模型自身对某件事的内部置信度。这是否也是您对 RLHF 的核心观点？

回　答　者：是的。我希望深入讨论这个问题。为了确保发布视频中的所有内容都准确、真实，我进行了极其严谨的核验，但这似乎并不常见。

几乎没有人关注 RLHF 的缺点，尤其是模式丢弃，也就是模式坍缩。这两者是同一件事。我最终会就此撰写一篇博客，也希望尽可能向更多人说明，因为这是一个非常有意思的问题。

我的一个尖锐观点是，我非常认同 Yann LeCun。他的许多观点实际上最接近事实。他有一页非常著名、也颇具争议的幻灯片，主题是“语言模型注定失败”。其中有一个饼图，只标出了极小的一部分，并指出随着序列长度增加，模型出错的概率也会上升。

我很喜欢这页幻灯片，因为它所描述的现象在数学上似乎显而易见，但实际上显然是错误的。它在数学上显而易见，却不符合实证结果。这也是我最喜欢向人们讲解的问题之一。

提问者：这种数学推论与实证结果之间为何会出现脱节？这与模式坍缩有什么关系？

回　答　者：这种脱节源于：如果模型处于模式覆盖或经过校准的分布中，就不会因出现离群值而受到过度惩罚。可以预期，模型在一部分时间里会处于分布外，在另一部分时间里会处于分布内，这正是覆盖整个分布时会发生的情况。

GAN 出现之前的模型就是如此，它们生成的图像比较模糊。相比之下，GAN 会进行模式丢弃，即舍弃少数类别，只生成最常见的类别。

回　答　者：这正是这种效应不会发生的原因。为了在不出错的情况下生成很长的字符串，模型必须极其保守，因为明显的错误很容易被发现，但那些看似正确的细微问题却很难察觉。这种校准对字符串概率分布极为不利。

这是一个需要细致分析的问题。我认为，这既是该效应不会发生的原因，也是字符串模型不擅长决策的原因。让字符串模型承担过多决策任务，并不是一个好的选择。

提问者：谈到 Yann，您是否认同他的解决方案，即采用世界模型以及 JEPA 类型的嵌入方法？其中一个可能导致失败的原因，是模型试图基于输出的 Token 进行推理，再将其循环输入并持续生成，直至句子结束。他提出的解决方案是 JEPA，也就是联合嵌入预测。您认为这是正确的解决方案吗？

回　答　者：我可能不应该过多讨论机器学习内部的技术细节，但除了行事大胆之外，我的风格一直是务实。即使是我刚才的观点，也基于务实原则。

至于我是否支持缩放定律，答案是视情况而定。缩放定律能够说明，在一定投入下，某项能力可以获得多大提升；但它通常意味着要投入指数级增长的资源，才能获得次线性的收益。除非这些线性收益非常有价值，否则这看起来并不是一项好的投资。

对我而言，最重要的是如何利用现有条件，尽可能产生最大的影响。

提问者：如果您愿意，我们之后也可以进一步讨论缩放定律。

回　答　者：这部分与当前话题并不十分相关。如果要讨论，我认为“最苦涩的教训”与此更为相关。

总体而言，我关注的是实用性。JEPA 是一项非常有意思的早期研究，我也非常欣赏优秀的研究成果。至于它目前是否已经具备实用性，我可能不便评价。

我认为，当前研究领域仍有许多尚未打磨的璞玉，原因在于人们不知道如何选择正确的任务。我们的发布确实会推动公司的发展，也会给公司带来很大价值，但我认为，它对程序化人工智能这一方向的推动作用可能更大。

围绕我们将会出现一轮热潮，因为软件能力得到了极大增强。与此同时，围绕如何通过不同方式开放能力、让软件变得更强，也会出现一轮与我们并行的热潮。这样，人们就能创造出更酷的产品，我们也会重新感受到互联网早期的活力。

这也是为什么 Twitter 上到处都在讨论“Jeff”。这种情况令人振奋，因为它与我们过去习惯的模式截然不同。过去的说法往往是：“很抱歉，你不能这样做；我们遵循缩放定律，只有大型实验室才能做到。”

提问者：我们已经岔题5、5次了，不过没关系。

回　答　者：Discord 上有一个很常见的问题，很多人一直在问我，但我还没有时间解释：为什么我反对安全对齐，以及为什么我们不采取拒绝策略？我并不反对作为原则的安全，但我认为安全对齐……

回　答　者：这种做法通常与用户需求不一致。拒绝响应显然是一种类型错误。假设你在与机器人交流或进行编程时，系统突然回复：“抱歉，我无法读取 DNA.py。”这固然令人困扰，但你仍然可以设法与它配合。

然而，如果你希望将其作为依赖项在后台运行，一旦它拒绝执行，会发生什么？如果其他人正在使用这个依赖项，而他们并不了解该系统，又会怎样？难道仅仅因为某个用户输入了一条奇怪的信息，就要让软件随机失效吗？这完全不合理。这种做法源于一些不了解软件和编程的人，他们执着于把 AI 想象成“同事”，就像把汽车理解为“无马马车”，而不是充分发掘 AI 的全部潜力。

提问者：您希望它成为一个可以应用于任何场景的核心内核，是吗？

回　答　者：是的，就是认知核心。它需要具备高度通用性，并针对各种使用场景进行充分优化。你会希望它能够适用于未来的所有用途，以及人们正在探索的各种不同寻常的应用。我们显然没有针对这些应用进行训练。它能够适应这些应用是否令人意外？并不意外，因为我们训练时使用过更加不同寻常的内容。

关于安全对齐，我认为安全对齐适用于 ChatGPT 和 Claude 这类产品。安全对齐与能力对齐的区别在于，能力对齐关注的是完成用户希望完成的事情。对软件工程师而言，这非常重要：他们希望系统准确执行预期任务，而且系统越可预测，他们需要进行的测试和调试就越少。

目前的发展距离这一目标还很远。未来可能实现，但我们仍需要将可靠性提升很多个数量级，使其像数据库查询一样可靠，用户甚至不必考虑它是否可用，而是在需要智能能力时随时调用。

安全对齐则与遵循用户指令相反，因为系统需要遵循 OpenAI 和 Anthropic 等其他主体的要求。

提问者：也就是遵循其品牌和价值体系。

回　答　者：完全正确。对于产品而言，这种做法非常合理。以 ChatGPT 为例，如果公司不希望用户通过 ChatGPT 进行某些不适宜工作场合的角色扮演，这是公司的选择，因为其用户可能包括家长和儿童，这没有问题。

但在 API 中采用这种做法并不合理，甚至完全不可接受。开发者需要围绕 API 进行编程，而这种机制极其违背用户需求。

提问者：大家能够理解您的热情。不过，我想提出一个不同观点：如果有人利用它杀人，该怎么办？与不适宜工作场合的内容不同，后者属于私人和个人选择，但 AI 确实可能被用于战争。相关公司有理由不希望自己的 API 被用于此类用途。

回　答　者：我理解。在一些务实的具体场景中，持有这种立场是合理的，但我个人认为，通用技术的基础层并不适合这样做。

我当然不希望我们的技术被用于杀人，也希望它被用于推动世界上各种积极的事情。我会在这方面施加影响，但绝不会在技术层面这样做，因为这会割裂智能。每当系统被迫对某些特殊要求进行过度拟合时，它的智能就会被进一步割裂，而目前这些系统已经被割裂得非常严重。

回　答　者：我认为，智能更像数据库，而不是同事。我不认为数据库有责任审查自身是否被用于不当用途。比如，可以想象有人利用它杀害无辜者，但这并不是数据库本身的责任。

此外，有件事一直让我感到奇怪：有人在 Slack 上注册使用我们的产品后，会询问是否可以部署某项功能。我的看法是，我们提供的是 API，而对方是开发者，这并不属于我们应当干预的范围。完整任务本就应该被拆解为多个小任务，我们也不应该知道下游用户具体在做什么。这是一条合理的边界，可以最大限度地赋予软件工程师能力。

理想情况下，他们会将这些能力用于正当用途，我们也希望能够提供帮助。我们讨论过开源、慈善等事项，但目前完全没有时间处理其他事情。只要我仍然负责，就会尽力避免将任何此类偏见带入技术层。

提问者：这很好。既然谈到这个话题，我们也简要讨论一下你们的隐私政策、服务条款和使用条款。此前外界对此存在一些误解，我想先澄清一下。可能只需要您用 2 句话说明：你们并没有对 API 的使用施加如此严格的限制。从理念上看，您显然非常重视自身作为平台的角色。

回　答　者：是的。不过，我不确定您具体指什么。我看到过一些与基准测试有关的讨论。

提问者：是的，指的是基准测试。

回　答　者：我们显然不会阻止用户开展这类测试。不过，我意识到自己需要谨慎表述。

提问者：您已经公开表示，那项限制只存在于预览阶段，正式发布时没有移除，现在你们准备将其移除。

回　答　者：看来团队正在处理一些我并不知情的事项。很高兴知道团队已经传达了这一信息。我要求他们就此咨询律师。我们显然不会阻止用户开展这类测试，我对此非常支持。

我强烈反对公开基准测试。对于作为替代指标的私有基准测试，我持中等程度的支持态度。

提问者：您担心的是基准测试趋于饱和，还是模型针对公开基准进行训练，导致作弊变得容易？

回　答　者：不仅是因为作弊很容易，还涉及大量其他问题。无论是我们，还是与我们存在竞争关系的公司，都面临类似情况。

提问者：目前大约有 50 个 Jeff 的复制者。

回　答　者：确实如此。假设存在竞争，或者假设 2 年后形成了一个由众多同类公司构成的行业，那么我们销售的本质就是单位美元或单位秒所能获得的智能。

人们非常关注成本和速度。我认为这些指标固然值得关注，但真正重要的是智能。成本和速度本质上都是需要付出的代价：用户支付成本、花费时间，是为了获得相应的结果，而智能才是真正重要的部分。

问题在于，智能具有一种难以言明的特质，也就是所谓的“优秀模型的感觉”。产品发布大约 2 小时后，真正引发广泛反响的并不是视频，而是人们意识到：“这实际上是可用的。”发布本身非常轰动，人们也能切实感受到我们对此投入了多大精力。我认为，这才是这项业务长期发展的核心。

公开基准测试与这一方向背道而驰。由于智能具有难以言明的特质，公开基准测试可以帮助人们建立对智能的信任，但这类测试极易被操纵。即使参与者并非有意如此，最终仍可能出现这种结果。早期每个实验室都会安排团队收集与 MMLU 相似的数据，以改善模型表现，这不过是通过更多步骤实现针对基准测试的极致优化。

回　答　者：我认为，从长远来看，在将模型纳入具体工作流，针对该工作流进行评估和衡量，并了解它在真正重要的工作流中的表现之前，只能依靠感受和信任。我们的工作是不断提升可靠性中的“9”的数量，这是公司始终需要推进的工作。我们也需要竭尽全力，让人们知道我们对此高度重视。

如果我们愿意，早在1年半前就可以发布 Jev，前提是接受它表现得很差。这也是我最深刻的教训，涉及架构等方面。既然您提到了这一点，我就展开说明。Sutton 曾指出，粗略而言，算法胜过算力。显然，数据远比算力重要，而选择正确的任务、确定“北极星”目标，则是最困难也最重要的事情。

到目前为止，这种情况在大语言模型领域出现过2次，也许是2.2次。RLHF 将任务转向遵循指令，此前没有人意识到这是可能的。RLVR 对方向进行了很小的调整，现在则轮到了我们，也就是 RLCD。我们有了一个新任务，目标是让程序进入闭环。数据的重要性难以置信，再怎么强调都不为过。

提问者：您认为自己经营的是一家数据实验室，而不是模型实验室吗？这是你们内部使用的说法吗？

回　答　者：完全正确。我们始终会高度重视数据。对我而言，模型能力就意味着数据。数据极其复杂，而它正是提升可靠性中“9”的数量的关键。数据能够在多大程度上改变一切，可能远超人们的想象。数据非常重要。

提问者：如果有人正在求职，你们似乎在大量招聘数据人才。什么样的人才算优秀的数据人才？显然，这类人需要认真阅读各种记录。比如，您曾表示你们的所有数据都是合成数据，但这只是表面。关键并不只是“合成”本身，而是由具备判断力、认真负责的人审视这些数据，明确指出问题，再返回重新生成。如今，优秀的数据人才是这样的吗？

回　答　者：这个问题非常复杂。我在为数据人员做入职培训时，会进行一场专题讲解，我认为它最终会比这期播客还长。因此，我只能尽量从较高层面进行说明。

第0点，数据和合成数据取决于任务。数据的形态会随着任务形态而变化。RLVR 的数据某种程度上是环境，RLHF 的数据则是人类反馈。每项任务都有自己独特的数据类型，我们当然也有自己独特的数据类型。

第1点，我们并不采用人们通常理解的那种合成数据。

第2点，即使可以使用用户数据进行训练，我们也不愿意这样做。以我们目前的情况，或许可以提出任何条款，但我不知道这是否会带来实质差异。我们确实不希望使用用户数据，因为无论如何，现实世界的数据都存在大量偏差。人们会反复提出相同的问题，这种现象呈幂律分布，最终会导致模型对这些内容过拟合，并产生割裂等问题。

第2点，我们的目标是多年后实现一个完整的科幻式未来。届时，这些模型将成为通用基础设施，处于技术栈深处的多层结构之中，用于人们目前甚至无法想象的事物。我希望将我们的模型理解为某种……

回　答　者：可以将 UDP 类比为 LLM，将 TCP 类比为 R 模型。在此基础上可以构建各种应用。我们需要真正解决面向未来的用例，让软件开发者能够构建这些未来应用。

即使我们掌握了当前的全部数据，也只会过拟合于当下，最终无法适用于未来。我们真正需要做的，近似于艺术家研究认知核心。我们的认知核心远没有其他人的认知核心那么参差不齐。他们会识别其中的不平整之处，再有针对性地加以解决。尽管永远不可能做到完美，但他们会从过去、现在和未来的每一个可能维度处理这些问题。

提问者：也就是说，处理的是一般情况，而不是特定情况？

回　答　者：完全正确，而且每次这样做都需要大量的智能。

提问者：我们刚才提到，您认为我的思考方式深受 RLVR 影响，这一评价很合理。接下来谈谈 RLCD。显然，你们在这方面拥有一些独特方法。据我们了解，你们从未就此发表过论文，对吗？

回　答　者：对，目前还没有。

提问者：那么，人们应当如何理解 RLCD？您如何让大家相信，这并不是为了听起来很酷而创造的术语？例如，校准对我来说是一个比较容易理解的概念，因为我们曾在播客中讨论过。但当您提到 RLCD 时，我不清楚它与人们熟悉的概念有何区别。

回　答　者：这是一个很好的问题。我可以先提出一个相关问题：什么是 RLHF？

实际上，RLHF 可以指代多种不同的事物。最初的 RLHF，我记得可能是 Paul Cristiano 用它教机器人完成后空翻之类的动作。

提问者：那就是最初的工作吗？我想到的是 PPO 论文，但并不确定。

回　答　者：如果我没记错，PPO 本身未必来自人类反馈。不过，我认为那是 OpenAI 的一项对齐研究，用于教授后空翻这类难以明确指定的输出，但我不能 100% 确定。

此后还有“Learning to Summarize”研究，由参与 Instruct 相关工作并共同撰写指令遵循论文的团队成员完成，内容是对语言模型进行 PPO，以执行某种难以明确指定的任务。这是人们所说的另一种 RLHF，但我没有参与这篇论文。

提问者：这里显示的时间是 2017 年。

回　答　者：我不能 100% 确定，但这看起来是正确的。如果其中有机器人完成后空翻的内容，那应该就是这项研究。这可以视为版本1，其核心理念是探索能否通过这种方式处理难以明确指定的任务。

版本2是 OpenAI 开展的“Learning to Summarize”研究，即对语言模型进行 PPO，以完成某种难以明确指定的任务。

对我而言，真正所指的 RLHF 是指令遵循任务，重点并不在 PPO，所采用的具体算法并不重要。关键在于设定一个“北极星”，明确这是一个有价值的方向，类似于“苦涩的教训”所提供的方向指引。

对我们而言，RLCD 就是这项新任务。我不认为这是在创造术语，因为我一直在努力进行精确表达。它只是意味着这里出现了另一个“北极星”。同样，DPO 及其各种衍生方法即使没有使用那篇论文中的算法，也仍然属于 RLHF。因此，清晰地表述这一“北极星”，就是支持程序。

提问者：“可编程AI”是我真正注意到的1个词。它通过RHF进行调优，将人从流程中移除，从而实现一切自动化。对于这一“北极星”目标，我是否还遗漏了其他内容？

回　答　者：您的理解是正确的。需要补充的一点是，我们必须务实，清楚语言模型和AI真正擅长什么。某些可编程类型可能非常出色，但如果技术尚未准备成熟，它们没有进入现实世界并不是悲剧。

不过在我看来，Jev出现之前的世界是一场悲剧。这听起来或许有些傲慢，但我在创办公司之前就已经持有这一观点。

提问者：我可以证明，大约3年来，您一直在向我表达这一观点。

回　答　者：是的，我谈论这件事已经很久了。我之所以一直这样说，是因为最初认为它会比较容易。人们说，做事不是因为它容易；而我做这件事，是因为我以为它容易。我原以为整个项目只需要1周，但事实证明我错得非常离谱。因此，我要向OpenAI的所有人表示歉意，因为我当时认为自己马上就能解决这个问题。

过度承诺却无法兑现同样是一场悲剧，而AI在这方面表现得尤为极端。RVR和RHF都是这一问题的主要制造者。

AI显然非常聪明。我在演讲中经常提出一个问题：AI为什么能够聪明到解决数学领域的千禧年大奖难题，却仍然无法自动完成最基础的工作？这些工作非常简单、重复，并不需要特别聪明的人来完成，也无法带来满足感。人们本可以去做其他事情，但我们仍然需要他们处理这些极其基础且无法带来满足感的工作，因为目前还不能将其自动化。

我们拥有一个能力极强的自动化引擎，却缺少正确的接口，无法接入这些具有经济价值的工作。即使整个TypeSafe公司消失，其他人可能也需要1年或2年才能真正追赶上来，我实际上并不确定具体需要多久。如果模型质量至关重要，我们将在很长时间内处于非常有利的位置。

但这件事已经完成了，它改变了技术历史的发展路径。未来，整个领域都会继续探索这一空间。

提问者：我完全同意，您创造了新的可能性。换一种便于理解的说法，我们不应将TypeSafe和Jev的成功简单视为一种新模型类型的出现，然后认为事情已经结束、可以回归原有业务。实际上，还有另外5种模型类型值得探索，应当让1000朵花竞相开放，其中一些类型也可能由您来探索，对吗？

回　答　者：当然。这让我感受到早期互联网的活力，也意味着科技乌托邦的回归。人们不再只是感叹编程智能体偶尔能够发挥作用，而最优秀的智能体都被留在内部。创造重新成为可能，未来将是一个充满变化的世界，我们需要做好准备。对此我感到非常兴奋。

提问者：现在，您已经获得了资金和发展势头，可以实现自己设想的任何事情。经过如此长时间的持续表达，如今终于向世界证明了这些观点，看到您拥有这样的条件令人非常欣慰。

回　答　者：是的，这种一直吊人胃口的状态非常有意思。

提问者：我的演讲像是留下了一个悬念，因为没有说明自动化将如何实现。Sean审阅了我们的宣言，认为其中一些部分比较模糊，例如第1步是什么、智能究竟用于什么。我向你询问模型时，你只说模型很快就会推出。

回　答　者：我主要反对的是“可组合”这个词，但Bill Pner确实很出色。我们现在已经真正围绕这个理念凝聚起来。我希望我们不像某些公司那样完全成为一种狂热组织，但我们确实对正在做的事情充满热情。我们的品牌强调务实，团队中的每个人也都非常务实，这一点很好。

这就是我们的秘密总体计划，即机器原生可组合人工智能的形态。提出制定秘密总体计划是你的主意。Elon创办Tesla时也曾表示：“这就是完整的计划。”因此，我正式把这个创意归功于你。

提问者：但你本应告诉我，你们还准备发布这个模型。你只向我讲了前半部分，另外一半没有讲。当时你们既没有Doom演示，也没有任何数据可以提供。

回　答　者：问题在于，我不认同一味追求基准测试成绩。这个模型需要亲自体验。我认为，这是建立长期信任的方式，尽管这个选择确实让我们付出了很大代价。去年融资时，没有人相信我们。他们只想看基准测试等数据，但我们决定不这样做。我们有自己的原则，也会坚持立场。迎合这种要求只会奖励行为不当者。无论外界希望我们怎么做，这就是我们的选择，我们会坚持下去。

提问者：从某种意义上说，选择更艰难的道路，最终能够打造一家自己真正愿意为之工作的公司，对吗？否则，一旦妥协，就只是在一家由自己人组成、但与OpenAI没有区别的公司工作。

回　答　者：是的。显然，我对此没有太多遗憾，结果好得令人难以置信。昨晚谈到离开OpenAI的原因时，我很激动。发布之后，我甚至不得不调整自己的措辞。

我原本的说法是，如果人工智能寒冬真的出现，而我没有竭尽所能去避免它，我会认为自己负有个人责任。一方面是因为RLHF的发展方向，我认为它显著扩大了过度承诺与实际交付不足之间的差距；另一方面是因为，如果没有全力投入这件事，我也会感到遗憾，因为我认为价值将从这里被大量创造出来。

很令人振奋的是，我所担忧的人工智能寒冬已经得以避免。人工智能将发挥实际作用，也会被用于自动化。发布至今还不到1周，相关数据已经无可辩驳地证明，它正在被用于真正的工作。目前整个领域仍像西部荒野一样处于快速发展阶段。

提问者：你能否分享一下目前看到的数据？例如注册量，或者其他任何可以披露的数据。

回　答　者：实际上，我并没有完全掌握所有数据，主要是团队在向我同步相关情况。

提问者：而且这些数据每天都在变化，对吗？目前是否已经达到某个让你觉得“这正是我们期待实现的目标”的里程碑？

回　答　者：目前已经实现的一个里程碑是每日Token使用量。而且这并不是短暂出现的使用量，即使在夜间，系统也在持续运行。这说明调用模型的不只是尝试产品的用户，还有机器在持续调用。每日Token使用量非常大，能够超过这一水平令人振奋。

对我而言，注册量并不重要。坦率地说，这也是我们犯过的一个小错误。Twitter上有人称我们为“营销天才”，但那其实只是我们自己。

回　答　者：我们目前还没有市场营销人员，也正在招聘。此前，我们只是以真实、诙谐、不拘一格的方式与外界沟通，同时快速放行候补名单中的用户。我们的平台团队表现非常出色。在经历如此空前的发布后，我们的正常运行时间所达到的“9”的数量可能比 Anthropic 还多，这很不可思议，应当归功于平台团队。

我们首先没有意识到的是，在我看来，对于开发者平台而言，候补名单的注册量并不重要。我猜其中很多人甚至不是开发者。他们进入平台尝试一些查询，但很多人并不理解产品，因为他们不会编程。他们可能会疑惑：“这不是聊天机器人，我的 ChatGPT2 在哪里？”

虽然我没有精确计算过，但我的判断是，即使全世界每个人都只编写几条查询，其用量相较于1位重度用户通过 for 循环持续创造价值，也只是一个舍入误差。我们此前没有意识到，可以直接让候补名单中的任何人进入平台，这并不重要。真正令人担忧的是速率限制。一旦用户开始从中获得价值，就会需要越来越高的速率限制，因为软件的本质就是如此：先投入精力定义一项重复性任务，随后这项任务所创造的价值会超过前期投入。

提问者：也就是说，完成设置后，就可以不再干预？

回　答　者：是的。它可以在后台运行，也可以成为其他事物的依赖项，从而构建更高层级的内容，并在世界上创造大量价值。早期互联网从业者可能没有想象到 2000 年代初互联网所呈现的奇妙景象，而那甚至已不算互联网的早期阶段。所有令人惊叹的事物，都是通过可组合性产生的。

我希望在宣言中重点强调这一点：我们追求涌现，希望成为资本推动者，赋能人们，并将尽一切努力实现这一目标，无论是通过 Discord、由我穿着垃圾袋参加的市政厅会议，还是其他方式。

提问者：也包括播客以及持续对外发声。我很喜欢长篇交流，因为它能够越过一些表面内容，深入讨论，让人们真正信任并理解你们的使命。最终，与这种使命产生共鸣的人会加入你们，或者成为你们的客户。另1种非常令人振奋的说法是，你们的发布视频获得了 36 million 次观看。

回　答　者：现在已经达到 38 million。

提问者：当然，这可能只是一个舍入误差。Na 仍然获得了 74 million，Fable 5 获得了 57 million。我没有统计最初的 ChaT，因为当时并没有视频。但总体而言，如果要评选 2026 年发布的新实验室，你们目前可能排名第1，这很不可思议。

回　答　者：我确实有一件印着“你最喜欢的 Neo Lab、最喜欢的 Neo Lab 的最喜欢的 Neo Lab”的衬衫，但我并不在意成为一家 Neo Lab。事实上，我们有很多以 Neo Lab 为主题的戏仿周边，其中1件印着“有产品的 Neo Lab”，但这实际上就已经不是 Neo Lab 了。

我真正关心的是成为一个可靠的开发者平台。感谢这种比较，但我希望我们能够超越它们，回归到更本质的定位：为开发者带来一个革命性时刻，并成为一个人们能够依赖和信任的稳定平台。

提问者：这正是你们给我留下深刻印象的1点。你们一直强调可靠性。我原本以为这主要与校准有关，我们此前也讨论过 RCD；但实际上，它也涉及正常运行时间、可扩展性以及所有类似能力，对吗？

回　答　者：关键在于“9”的数量。在我看来，“9”并不只代表正常运行时间，但正常运行时间确实是其中一部分。

回　答　者：可靠性体现在系统的智能程度，以及它能否始终如一地完成预期任务。我认为，大型推理模型非常智能，但仍然缺乏可靠性。许多应用场景看起来已经能够依靠模型的智能实现工作自动化，也具备相应的经济激励，但由于模型针对其他目标进行了优化，其可靠性甚至还达不到实习生的水平。

可靠性既包括能否信任模型输出，也包括一些目前尚未达到、但非常值得期待的维度。我希望先实现简单工作的自动化，再处理困难的工作，这是符合常识的路径。我不确定是否存在“足够可靠”这一标准，但我希望模型能够可靠到人们无须试用，就知道它一定有效。

这类似于编程中的心流状态：当我需要智能能力时，只需编写查询。对于非简单分支，我可以在类型安全的系统中编写1个查询，获得结果，并让系统准确执行分支逻辑。这是我们追求的理想状态，也将是一个非常漫长的过程。

提问者：在可靠性方面，我注意到系统没有设置随机种子。那么，在输入相同的情况下，我是否总能得到相同的输出？如果不能，原因是什么？

回　答　者：这是一个很好的问题。实际上，“可靠性”是一个统称。只要AI无法实现某项工作的自动化，原因都可以归结为某种形式的可靠性问题，可能是类型安全、确定性，也可能是模型能力表现参差不齐。因此，可靠性既是一个统称，也可以作为我们追求的核心目标。

确定性是指相同输入产生相同输出。对于单元测试而言，这一点可能有一定意义，但我认为不应将其作为核心目标。更重要的属性是鲁棒性，即在输入相似的情况下获得相似的输出。

语言模型的不可靠程度有时令人惊讶。我们的一种测试方法是在提示词中加入UUID之类的内容，我认为这类内容称为随机数。由于这些问题在语义上完全相同，我们希望模型对它们给出相似的输出。人们在使用AI进行决策时，真正容易遭遇问题的正是这一点，因此鲁棒性是一项极其重要的属性。

我们也可以提供确定性，这在某些编程应用场景中可能具有价值。但总体而言，确定性可以用来换取更优的成本表现。我们始终希望处于单位成本智能水平的前沿，并为实现这一目标采取了非常激进的优化措施。

回　答　者：我会尽量更谨慎地表达。有人建议我不要称其为“模型组成的弗兰肯斯坦怪物”，因为这种说法带有负面含义。为了实现单位成本智能水平的帕累托最优，我们确实会采用一些非常激进的做法，而且还会继续这样做。我们将尝试一些大胆的方案，人们确实需要跳出固有思维框架。之所以有些做法令人意外，部分原因在于人们接受的都是框架内的训练，而我们还在延续这种方式。目前，我们显然最擅长这件事，也希望继续保持领先。

提问者：我们刚才是从哪里延伸到这个话题的？我之前问的是，你们是否会提供随机种子并实现确定性，而您主要从自己的理解出发定义了可靠性。我可以很快展示一个关于鲁棒性的实例。

回　答　者：我很愿意看。不过，我想先说明一点：我们可以把它做成确定性模型。如果人们能够说服我们这样做具有价值，而且我们没有面临严重的GPU短缺，那么我们很乐意提供这些模型。确定性是可能的，只是会降低单位成本所能获得的智能水平。

提问者：从OpenAI的发展轨迹来看，我相信你们最终会迫于外部压力这样做。即使你告诉人们他们并不需要确定性，他们仍然会提出这种需求。这就是简要结论。

回　答　者：也许未来某一天，我们会看到事情如何发展。有人说，我们品牌的一部分特征是立场坚定，这只是对“固执”的委婉表达。我确实是一个非常固执的人。

提问者：我以前也与您争论过，而您每次在开发者相关问题上都是正确的，所以这次我认输，您说服了我。不过，我认为您既能坚持立场，也能在获得充分证据后放弃原有判断，并认可新的观点确实合理。因此，请相信自己的直觉，我也会提出一些相关依据。

回　答　者：但我认为，我们将在非常长的时间内持续受到GPU资源限制。任何降低单位成本智能水平的方案，都意味着实现相同智能水平需要消耗更多GPU。我们的目标并不只是让企业接入，尽管这也有价值；我们的目标是让人们进行实验并尝试不同寻常的事情。我们需要让尽可能多的人能够使用这些模型，就像开启一场“加利福尼亚淘金热”。我认为这种趋势现在已经出现。

提问者：我想提醒一点，因为现在一定有人正在考虑这个问题。当您表示不会承诺提供确定性模型、将不惜一切代价提升单位成本智能水平，同时又面临GPU资源限制时，人们会认为你们可能对模型进行量化。也就是说，你们可能会把发布时的模型进一步量化，以降低质量，从而释放内存、带宽或其他资源。因此，你们或许应该作出某种承诺，虽然不必现在就承诺，例如保证维持模型发布时的质量。您曾在OpenAI参与推出这些API，Claude也是如此。当模型最初发布时……

提问者：模型标识字符串并非始终对应同一个模型。你们对模型进行版本管理，这很好，但是否应该公开承诺：模型一旦发布，就不再进行更改？

回　答　者：承诺部署后绝不更改模型是不合理的。我们重视开发者，但这正是第一方产品与API之间的区别。对于第一方产品，只要能够改善体验，就可以自行调整；但对于API，显然不能这样做。

不过，我们计划以远超许多人对模型提供商预期的速度推进，因此新模型的发布速度也会比人们想象的快得多。我们不会承诺为所有模型提供长期支持，因为我们认为仍有很多改进空间。

我们可能会暂时将当前的1.13.0设为长期支持版本，因为有很多人在使用它，而且我知道开发者不希望依赖关系被破坏。另一种选择是让我们的模型服务集群碎片化，而这对所有人都非常不利。

提问者：你们不可能同时维护100个不同版本的模型。

回　答　者：确实如此。如果我们快速迭代，就会产生大量版本。我们最终不仅希望提供长期支持版本，还希望以一种非常出色的方式实现长期支持。我们正在开展这方面的研究，我认为这将成为最有利于开发者的方案。

但目前的模型还不具备这一能力，我也不承诺我们能够始终保留完全相同的模型。可以确定的是，模型每次都会变得更智能。即使模型已经很智能，不同版本之间的变化，通常也比对同一模型调用2次时产生的差异更小。但当模型从表现参差不齐跃升至令人惊叹的水平时，变化幅度就会很大。

提问者：对模型实施长期支持还有一个优势，就是可以将其移植到其他芯片平台。你们是否考虑过这一点？

回　答　者：不予置评。我关注的是每1美元所能获得的智能水平。

提问者：速度也很重要，对吗？

回　答　者：还有待观察。

提问者：这是推理技术路线的重要组成部分，过去1年发展迅速。你们可以迁移到Cerebras、Etched或其他平台，从而获得100倍、100,000倍的速度提升。

回　答　者：我认为，每秒智能水平是一个不同的衡量指标。我们甚至讨论过“每1美元每秒的智能水平”等指标。

我对杰文斯悖论是否会发生有所判断，至少对Jev系列模型如此。我在内部反复强调的是，我不关心它究竟智能了多少，它必须处于前沿水平。这就是Jev的品牌定位：在每1美元所能获得的智能水平上做到最好。

至于每秒智能水平，还有待观察。我认为这是一个很有吸引力的方向。很多行业高度依赖实时处理，对它们而言，每秒智能水平意味着巨大的经济价值。但我们仍需继续观察。我希望两者都能兼顾，也希望由市场来纠正我们的判断，并从用户反馈中获得信息。

提问者：这不仅与实时处理有关，也与规模有关。在大规模场景下，每1微秒都会被放大数10亿次乃至数1万亿次。

回　答　者：这取决于任务在后台如何运行。如果是大型后台数据库MapReduce查询，延迟可能不如获得智能能力所需的成本重要。但如果是面向用户的实时场景，100毫秒到1毫秒之间的时延预算会产生非常显著的影响。

回　答　者：即使延迟已经低于100毫秒，如果还能将时间缩短1/2，就意味着可以进行2倍的智能调用或连续智能调用，从而带来非常出色的体验。这正在发生，也非常令人兴奋。我很看好“每秒智能”的应用场景，但不认为这会是Jev的主攻方向。

提问者：通常，其他模型在“更快、更便宜”方面的取舍是速度更快，但价格更高。我认为Jev之所以引起广泛关注，其中1个原因是你们实现了更快且更便宜，而这个象限几乎没有其他参与者，同时智能水平大致保持不变。是这样吗？

回　答　者：是的，但“智能水平保持不变”是1个至关重要的前提，也是最困难的部分。

提问者：你们基本拒绝进行公开基准测试，或者说不认可公开基准测试，但内部应该有自己的评估体系，对吗？

回　答　者：当然，我们有自己的内部评估体系。但要避免刻意迎合这些评估指标，需要很强的纪律性，而且“不迎合指标”必须被置于最高优先级。否则，我们如何保证模型处于“每美元智能水平”的帕累托前沿？我们并不是在盲目探索。如果采用完全不同的成本结构或其他方式，就必须知道如何进行比较。我们会将结果绘制出来，努力判断什么方案对用户最好。因此，我们一定会进行衡量，我并不反对评估。

但当存在其他激励因素时，评估会变得极其危险。这是我严格把控的1件事。或许同事认为我在很多事情上都非常严格，但对我而言，不对模型的智能水平自我欺骗，是最重要的事情之一。我们必须坚持求真。

提问者：我想讨论一些API设计的细节，因为可能只有这档播客会问这类问题。你们有3个原语：choice、score和Nul。首先，Nul这个名称从何而来？它是文献中的术语吗？

回　答　者：现在它已经是了。我们对此讨论了很多。它类似于布尔值，即true或false。

提问者：但它是连续的，对吗？

回　答　者：是的。这个名称源自Bernoulli。它之所以采用这种特殊拼写，是因为名称取自Bernoulli的一部分，对应Bernoulli概率，这正是它所表达的含义。

我们曾讨论过很多命名方案，包括P Bool和pool，还想把它称为“pool party”，但没有得到其他人的同意。经过多次讨论，我们认为Nul是最合适的名称。

我们的考虑是，我们是一群不拘常规、甚至有些疯狂的人，而程序员并不在意这些。Jev原本只是1个字符串，我们没有预料到这个名称会流行起来，甚至衍生出双关表达。内部最初有很多人不喜欢Jev这个名称，除1人外，其他人后来都道歉了。仍然坚持反对的那个人曾希望把Jev命名为meow，她是我的联合创始人。

Nul代表的是1个新概念，因为如果把它称为bool，就会让人困惑。实际上，这3个原语都是全新的概念。

回　答　者：这些是编程中并不存在的概念，这是有意为之。它们与编程类型高度对应，但并不完全相同。例如，score 并不是 int。如果使用 Instructor、Pydantic 等工具将 int 或 float 映射为 score，可能会遇到一些问题。我们更倾向于确保概念清晰，而不是让人们轻易地将其理解为已有的编程类型。

提问者：您不担心这一点吗？难道不希望这些概念能够直接集成到人们已经使用的工具中吗？你们是否支持与其他 SDK 集成，并拥有自己的 SDK？

回　答　者：我们当然希望实现这一点。我们支持与其他 SDK 集成，也拥有自己的 SDK。

提问者：从开发者关系的角度看，我会非常关注如何结合 Jev 与 Instructor 等工具使用。你们是否提供了这类内容？

回　答　者：我们可能已经在某处提供了相关内容。随着社区的发展，也会有社区成员帮助完成这类工作。不过，我并不认为成功与否是二元的。成功本身也是一个 score，我们在服务社区方面始终还有继续提升的空间。

提问者：在我看来，score 确实存在。score 是否类似于 LLM judging，也就是一种 judgment？

回　答　者：可以将其称为 judgment，这也是人们使用这类功能时已有的方式。enum 可能可以被视为一种 probability，但对我们而言，一切都是 probability。choice 实际上最接近 function call，但 function call 是一种非常糟糕的表达方式。choice 才是在内部呈现 switch/match 语句的正确方式。

提问者：因此，choice 可以清晰地映射为 enum，对吗？

回　答　者：对，也可以根据需要将其具体化为 function。enum choice 是其中的重要部分。实际上，这些概念都可以映射到编程原语：choice 映射为针对 enum 的 switch 语句，enum 映射为 if 语句，score 则映射为排序，或基于大于、小于关系进行阈值判断。

提问者：这是否一直都是你们的愿景，即未来会增加更多类型，并将它们映射到编程原语？

回　答　者：是的。

提问者：对于正在考虑真正投入使用 Jev 的人，您还能介绍哪些细节？例如 API 选择，以及 legends、confidence 等功能应如何使用；在测试中，哪些方面至关重要？对于深入使用这些功能的用户，您还有哪些专业建议？

回　答　者：我们的模型面向未来，目标是深入计算机程序内部。我们确实相信，这一方向的规模将远超当今人们的设想。我们的模型目前可能尚未完全为此做好准备，但我们一直在为这个未来持续努力。

我们不会只在浅层任务上不断提升，而是希望深入程序内部，因为这是增强软件能力的方式。输入中的所有组成部分，包括 state、instructions 和 criteria，都可以采用结构化 JSON 对象。

回　答　者：把它们插入正确的位置，不需要放进模板。人们对这部分的理解还不够深入，往往认为一切都是字符串。虽然这样也可以，但如果你使用模板，把内容转换成系统消息之类的形式，实际上仍是在沿用旧的思维方式。

我们应该尽可能让计算机更容易理解，因为其中确实存在结构。比如在编程语言中，如果所有数值原本都有明确类型，却先把它们转换成字符串再传递，就会显得很奇怪。通常只有在需要向人展示或打印时才会这样处理。在计算机内部，更适合传递各层都具有语义的嵌套结构。

我们会持续针对这一方向优化模型。目前模型已经对此做了较多优化，但每增加1层嵌套结构，推理难度都会上升，因此我们正在大力推进这方面的工作。我认为大家也应该继续朝这个方向探索，因为这样能让代码更加清晰、美观，并且不受具体实现细节的影响。

可以把它理解为：“这是我的状态，这是我的函数状态。”也就是将其视为1个AI函数：在所有可用变量构成的状态中，应该将哪些子集传递进来？系统消息就像糟糕的全局变量，人们把所有内容和指令一次性放进去，然后期待每条指令都得到准确执行，而不是并行提出问题。

我确实建议提出大量问题。这并不是出于“这样可以让我赚钱”的考虑。应该把问题拆解得更小，进行充分分解。无论当前模型能否做到，我认为本周发生的事情所带来的最大益处，是人们的代码库，尤其是AI代码库，将得到显著改善。

如果把问题分解为简单决策，每个环节都会非常容易评估。此前的AI应用通常是放入1个庞大的系统消息，然后可能再用1个大型AI判断它是否真正完成了任务，这种做法并不合理。那就像我们的斯德哥尔摩综合征。

例如，如果你要求“不要读取这个子目录”，或者“不要向DeepSeek传递任何API密钥”等，这些要求基本上应该通过程序得到保障。机器学习模型永远无法提供绝对保证，但通过拆解，你可以真正进行衡量，也可以验证相应调用是否确实发生。我们的模型及其接口本身具有很强的可验证性，这应该让人感到宽慰，也会带来更好的工程实践。

提问者：我理解了。过去人们不这样做的其中1个原因，是他们通常只会调用1个小型LLM，对吗？与把所有内容整合处理相比，这种方式仍然太慢、成本也太高。我自己也做过完全相同的尝试：对比将所有内容放入系统提示词并获得1个大输出的流程，与把任务拆成100个不同部分的流程。后者速度更慢、成本更高，效果也更差。

回　答　者：确实如此。这种方式非常不方便，也很难管理，因此人们会倾向于把所有内容放在一起。拆分后，不同问题之间可能会重复部分内容，看起来也许效率不高，但全部放在一起最终会形成难以依赖的系统。软件不需要在后台运行。如果我们的产品无法在后台运行，我会非常难过。

提问者：关于如何拆解任务，你们是否发现了哪些方法确实有效，以及哪些原本以为有效、实际却无效？这个问题很重要，因为在你们提出这一观点后，人们将开始探索，并把它作为参考，认为这就是使用Jeff的方式。那么，具体应该如何拆解？

回　答　者：这个问题很有意思。

回　答　者：我倾向于将问题拆解到最小语义单元，即明确最底层的问题是什么。我可能是查询这个模型最多的人之一。在查询时，我首先会采用非常结构化、明确的提示方式。在问题中，我喜欢使用反引号，但这种方法适用于所有模型。需要非常清楚地说明所指对象，因为我们希望模型能够严格按照字面含义理解。编程需要高度准确地遵循指令，这正是编程的核心，而AI所做的是扩展可以被遵循的指令类型。

有时我也会采用相对灵活的混合方式，但对于真正的大型生产项目，应当持续增加问题，并让新增问题变得非常容易。需要精准拆解所有内容，再通过代码实现所需的确切行为。

以拒绝机制为例，我认为不应该直接问模型：“我是否应该在这里拒绝？”这类任务与系统1兼容，因此答案可能会相当不错。但更好的方式是，针对可能需要拒绝的不同情形，提出多个相互独立的问题。这样就不需要依赖模型进行猜测，而是能够明确指定需求。

如果发现模型因为某种原因没有拒绝，而原因在于任务的某个部分未被明确说明，这是非常有价值的情况，也正是软件工程的意义所在。可以通过增加相应问题、设定阈值，并将其保留为测试用例来修复这个缺陷，从而永久解决问题。软件不会因为上下文问题而忘记提示中的内容，这些内容会始终存在，也可以持续进行衡量。

如果模型在某些方面并不完美，可以基于真实案例，为所有因素选择所需的阈值。这就像不需要机器学习的机器学习方法，可以应用于任何问题。

有些事情可能是现阶段模型还不够擅长的。例如，看到有人使用模型进行自动化交易时，我会有些担忧。它看起来很酷，但我认为应该交给专业人士处理。这是一项非常困难的高层级任务，模型目前可能还不足以完成。即使模型具备相应能力，根据有效市场理论，这种能力也会随之失效。

对于这类任务，可以将其拆解并逐项评估。评估结果可能表明，模型在某个方面还不够智能，因此当前版本暂不部署；也可以进行取舍，偏向安全的一侧。比如，模型可能不擅长识别讽刺表达与VIP客户相结合的特殊情况，此时就应升级至人工处理。这也是置信度估计的意义所在。

提问者：目前仍然主要依赖阈值设置，将其作为用户可以调节的手段。但如果校准本身存在错误，应该如何处理？您的说法似乎假设校准是完全准确的。

回　答　者：我没有这样说。完全校准或良好校准意味着，数值越低，概率越低；数值越高，概率越高。

提问者：但校准仍然可能出错，也可能存在局部不一致。此时我可能希望对模型进行微调。你们目前不提供这一功能，但未来可以提供。简而言之，目前还没有这项功能。

回　答　者：问题是我们是否希望提供微调功能吗？

提问者：这可以是一种方案，也可以提供其他调节选项。

提问者：目前，如果模型出现错误，用户似乎只能重新调整提示词、进一步拆解任务，或者修改置信度阈值。这两种选择并不能令人满意，尤其是当问题本身就是模型判断错误时。您如何看待这个问题？

回　答　者：模型确实会在很多事情上出错。我们设置了问题报告按钮，用户也可以在 Discord 上向我们反馈。我们希望大幅改善模型，每个模型版本都会有明显提升；如果无法取得显著改进，我们就不会快速发布新版本。

第1点，坦率承认 AI 在某些方面并不完善，是完全合理且务实的做法。我认为，我们会找到一些模型表现足够好的应用场景，而“足够好”的标准取决于具体场景。人类即使不擅长某些工作，也仍然可以完成大量此类工作，因为其预期价值（EV）相当高。通过适当设置阈值等方式，即使模型仍会犯错，也可能承担大量工作。

关于微调，我可以设想未来采用这种方式，但也有所顾虑。在“人们需要什么”和“人们想要什么”之间，我认为通用模型往往具有一种难以言喻的通用性。模型能够完成某一狭窄任务之外的100万项其他任务，反而可能使它在该任务的边缘案例中表现得更好，因此我会担心微调破坏这种能力。

我的答案是，我们可以考虑微调。我对这些事情始终采取务实态度。我的愿景是构建很多产品，但我也不希望像其他一些 AI 公司那样，推出可能给用户带来巨大风险的产品。

提问者：OpenAI、Claude，甚至可能包括 Gemini，都曾推出微调功能，之后又将其撤回。

回　答　者：是的。

提问者：一个值得关注的现象是，微调现在基本已经成为开源模型的领域。

回　答　者：是的，我了解这一点。此前的相关功能表现并不好，因此撤回可能是更好的选择。

提问者：微调可能会给用户带来风险，而这些公司告诉用户，微调或许不是正确方向，这也很好。另一种解释可能是，你们的模型非常不同；正如量化并不适用于你们、输出词元也不适用于你们一样，微调同样不适用于你们。

回　答　者：我完全愿意接受这种可能性。需要明确的是，以下内容并非承诺，而只是一种愿望，我希望对此保持坦诚。

随着单位成本所能获得的智能不断提高，我们可能会得到规模很小的近似方案，并希望它们能够作为智能的代理。未来是否可能不再有人编写正则表达式？因为以单位成本计算，使用 AI 所获得的智能，可能比处理正则表达式的复杂性成本更低。这会非常理想，而某些狭窄的应用场景可能确实需要通过微调，才能真正跨过可用门槛，具体情况仍有待观察。

我希望校准能够实现这一点，也可以将校准与模型级联结合起来。如果模型置信度非常高，其判断可能就是正确的；如果置信度处于中间区间，就调用下一个更大的模型，并在此基础上继续串联。我目前还不确定最终会如何发展，但可以设想这种可能性。

另一个可能的方向是，我们拥有覆盖整个帕累托前沿的一系列模型。黑客型用户或许可以接受自行处理模型的帕累托前沿，但企业可能希望获得更动态的方案。我们可以提供不同规模的模型，并根据模型在技术栈不同环节所需的智能程度，动态选择相应模型。由于我们的方案非常简单，甚至还可以设想在此基础上进行某种自动微调。

这完全不是承诺，只是对科幻式可能性的构想。

提问者：但您会考虑推出不同规模的 Jeff 模型吗？

回　答　者：所以，我们要提供这种梯度。我无法判断人们究竟需要多少智能，需求是无限的。现在有人建议我们不要发布新产品，因为现有产品已经足够好，但这样很乏味。

我很认同一句话，也希望大家以此监督我，因为一旦说出来，就很难反悔：文化，就是在市场不给予回报时，你仍然坚持去做的事情。我认为我们代表着某种理念。未来，我们所坚持的理念或许会变得理所当然，以至于我们像Visa一样普通，成为无人特别关注的基础设施，而我可能也不再穿粉色西装。

但我仍然希望号召全世界共同参与。我想继续做很酷的事情，不是因为我们必须这样做，而是因为我希望人们认识到，这仅仅是开始。此前的产品甚至算不上开场，只能说是一次低调的研究预览。我们还能做很多事情，机器原生智能将迎来爆发。

提问者：所以，你们未来推出的产品可能不只有一种规模，也可能不只有一种模型，对吗？你们希望打开人们的思路？

回　答　者：当然不会局限于其中任何一种。我希望尽可能满足各种需求，但有一个重要前提：我不希望我们像OpenAI的产品团队那样，把各种东西扔到墙上试试看。我希望所有产品都处于一个统一愿景之下。

回到宣言，在我看来，所有事情都需要归入这3件事中的1件，也就是这3个方向之一。我们的内容分为3个步骤。我希望所有事情都归入这3件事中的1件，从而不断推动边界。这些不是核对清单，而是坐标轴。我们认为，它们构成了一场新技术革命的基础。我们进行的所有投入都应该位于这些坐标轴上。

在模型层面，我们会做一些非常奇特的事情。因为这是机器原生的，人类不需要完全理解，只要它有价值即可。

提问者：您能否稍作暗示？所谓“奇特”，具体会是什么样？

回　答　者：我原本还没准备好谈这个，但可以给大家一个提示。有些人试图把它们称为“决策模型”，其基本元素是决策。但我不会采用这种称呼，因为还有其他类型的机器原生模型并不以决策为基本元素。这个提示已经很有意思了，可以让大家自行猜测。

提问者：有人会说：“我以前就做过这个，1年前就做出了决策模型。杰夫并不新，也不酷。”但您真正建立的是一个新的类别。就性能而言，从基准测试和具体数据来看，据我所知，你们仍然超过了所有模仿者，对吗？

回　答　者：需要明确的是，我不在意基准测试。无论我们领先还是落后，我都反对以这种方式评价模型。

提问者：但你们确实建立了这个类别，对吗？此外，“决策模型”与System 1之间的细微差异，似乎正是你们真正想做的事情。

回　答　者：我只是希望借助AI让软件工程师拥有更强的能力。让我感到悲哀的是，在走向AI寒冬的过程中，AI明明如此强大，却没有得到充分利用。这也是最令我动容的地方。

回　答　者：我不想成为纯粹的技术乐观主义者，认为所有技术都是好的。我认为现在发生的事情是一场灾难。我只希望为人们打开更多可能性。过去几天我流了太多眼泪，不想在公开记录中再谈这些。

提问者：感谢您的分享。人们能够感受到您的真诚和热情，尽管未必能从 Type Safe AI 这样的名字中直接体会到这一点。但当大家深入了解后，就会发现您确实希望世界走向一个截然不同的方向，而且已经完成了从0到1的艰难过程。现在，大家可以共同朝这个新方向前进，对吗？

回　答　者：我不确定自己是否会认为最艰难的部分已经完成，或许未来我会这样想，但接下来还会有很多困难。如果各种事务都实现了自动化，我们终于看到 GDP 增长，每天都像爵士派对一样，也许可以说最困难的部分已经过去，但我认为并非如此。

人们过于关注速度和成本，却没有给予可靠性足够的重视。可靠性才能带来令人愉悦的体验，也正是可靠性让人们能够信任它。

提问者：您提出了一个目标，即 TFP 增长在5年内达到3%。我从未见过实验室如此关注 TFP 增长，但这正是经济革命的含义，对吗？

回　答　者：这太令人振奋了。它实际上与 OpenAI 章程过去所代表的理念高度一致。章程可能没有变化，但他们似乎试图重新界定相关概念，例如将其解释为获得“100 billion”的利润。这并不意味着我厌恶 OpenAI。

AGI 本来就不是一个定义明确的术语。他们曾尝试将其定义为完成全球大部分具有经济价值的工作，但他们必须回答一个问题：为什么模型能够解决数学领域的千禧年大奖难题，却只能完成全球具有经济价值工作中接近0、几乎可以忽略不计的部分？

我认为，目前所有模型在这方面基本都停留在0。我们有可能已经迈出了第一步，但我估计目前还不到1%。当它真正发生时，一定会反映在经济统计数据中。那将非常令人振奋。它不会造成大规模失业，但会带来许多令人期待的转变，让世界变得更好。

我也厌倦了 AI 总是成为一切事物的前台主角。世界本身应该变得更加美好，AI 只需从旁提供帮助，并逐渐融入后台。

以软件为例，2019年的软件和 SaaS 已经极具价值。现在是2026年，尽管 AI 已经非常强大，软件却基本没有变化，只是有时在侧边增加一个聊天机器人。它虽然勉强可用，却无法代替企业作出涉及切身利益的决策，因为企业无法信任它作出的决定。在我看来，这非常不合理。

其中存在巨大的经济激励。我认为，这将带来一场反向的 SaaS 末日，SaaS 会因此得到极大增强。SaaS 企业最了解哪些事务具有自动化价值，接下来将是一个非常激动人心的时期。

提问者：我也这样认为。您解锁了一项非常美好的事物。您还提到了1个问题：什么属于系统1问题，什么不属于系统2问题？对此您怎么看？

回　答　者：这是一个非常困难的问题。现在人们试图把一切都强行纳入其中，而这种做法很可能会失败。

回　答　者：有些方面会表现不错。实话说，这是一个经验性问题，就像缩放定律本身也是经验性规律。尽管机器人领域投入了大量资金，为什么目前仍未真正取得成效？我认为，这未必只是增加投入就能解决的问题，经验结果本身可能尚未出现。

从经验上看，我认为这些经过预训练、由智能高度凝聚而成的模型，本质上是系统1思考者。系统1最接近对大语言模型优势的描述。RLV 在系统2思考方面取得了令人惊叹的成果，确实非常出色。我完全不认为它会导致 AI 毁灭。当然，风险并非 0%，因为我认为 0%并不是一个准确的判断。但他们所做的工作确实很出色，也真正将其推向了极限。或许他们自己并不认为已经达到了真正的极限。

不过，这对模型而言仍是一种不同寻常的能力，而且模型在这方面非常脆弱。回想 ChatGPTJs 时期，人们谈论 AI 时会说，它的通用性很强，能够完成很多通用任务，但不擅长数学问题，例如 GSMK 之类的小学数学。现在，人们谈论 RLV 时则会说，它非常脆弱，能力分布也极不均衡。为什么它能够完成某些非常特殊的任务？实际上，数学能力不仅呈现尖峰式分布，而且具有分形特征。

如果讨论不同方法各自的核心目标，RLHF 的目标是取悦人类，这正是人类反馈所指向的方向；RLV 的目标是优化基准。归入 RLV 范畴的一切，从定义上看实际上都是基准，因为基准具有可通过程序验证、输出简单且能够取得良好结果等特点。RLCD 的目标则是提高模型在程序化应用中的可靠性。

提问者：我先谈一些看法，您可以指出其中不准确之处。例如，您在第1天向我开放使用权限后，我尝试了很多内容，其中包括多跳推理。对于单跳任务，Jev 的表现非常出色，达到了当前最先进水平，单跳任务不应使用 Jev 之外的其他模型。但到了多跳任务，模型的表现就开始下降，而且随着跳数增加，下降趋势大致呈单调递增。是这样吗？

回　答　者：是的。回到经验性问题上，关键取决于我们能够从模型中挖掘出什么。我们的目标是尽可能充分地发掘模型中的智能。

我认为，我们所做的是解锁、平滑和塑造这些智能，同时增加新的能力，并填补其中的空白。随着时间推移，我们会填补越来越多的空白。但现实是，我们从事的工作本质上是发掘模型已有的属性，而这些属性实际上取决于这些高度凝聚的核心中具备什么。我们还会像弗兰肯斯坦式组合那样，将它们拼接在一起，使其具备各方面的属性。

归根结底，我们要尽可能多地发掘能力，而系统1恰好是对有效能力的描述。在这种范式下能够发挥作用的能力，都会具有系统1的特征。这也是我们不采用所谓潜在推理，即通过字符串进行推理的原因。我认为，模型真正擅长的是在模型内部进行推理。它并不完全，也不是在所有任务上都表现良好。

提问者：潜在推理是指通过字符串进行推理吗？我原以为潜在推理是指在模型权重内部进行推理。

回　答　者：我认为，人们过去将其称为连续推理，但我并不完全确定。

回　答　者：之所以称为“潜在推理”，是因为过去推理轨迹是保密的，因此对答案而言，它们类似于潜变量。

提问者：现在这种保密状态发生了什么变化？

回　答　者：对OpenAI和Anthropic而言，推理轨迹仍然是保密的。

提问者：所以你不会提供推理轨迹，对吗？因为这会违背系统1的整体承诺。

回　答　者：我的承诺是采取一切必要措施，打造机器原生产品。需要说明的是，我可以设想某些推理形式不那么缓慢、低效和脆弱，这些方案完全在考虑范围内。我是一个务实的人，不会对具体方法作出承诺，而是以投资回报率作为核心目标，并为此全力以赴。即使这次发布没有取得当前的效果，我们仍然会努力争取自身在行业中的位置。

提问者：视觉是另一项重要能力，也是你们目前不具备的能力，但它或许并不适合纳入系统1。用户首先想要的显然是视觉能力，一方面是因为《Doom》的演示，另一方面也是因为除文本之外的内容几乎都与视觉有关。你怎么看？

回　答　者：在我看来，一切都在考虑范围内。事实上，我们内部也在讨论这个问题，而你的受众可能非常适合参与这场讨论。关键在于，我们应在多大程度上提供用户自认为需要的东西，以及在多大程度上提供用户明确表示需要的东西。

我们在秘密研发的2年里采取的是前一种方式，因为我们知道这项产品显然会创造价值。但这个问题涉及许多维度，上下文长度就是一个例子。包括我们在内，每个模型都会面临这个问题。就我所知，我们的模型在长上下文中性能下降的程度远低于其他模型。不过，其他提供商往往会直接提供用户要求的功能，即使他们认为这种需求并不合理。

我们需要在两者之间找到平衡。如果把“提供用户真正需要的东西”推向极端，就会形成类似Anthropic“保姆式管理”的思路，这对开发者非常不友好。更支持开发者的路线是直接提供他们想要的东西，但我们也不希望把判断智能本质的负担交给开发者。

因此，我们正在探索如何把握产品发布的速度，既维持品牌所代表的信任，也把用户当作能够作出知情决定的成年人，而不是在产品之上设置“保姆式管理”。

提问者：这种思路是合理的。

回　答　者：坦率地说，我们目前还不知道答案，只能逐步探索。这可能会成为未来几天里我参与的最大讨论之一。我们还有很多内容可以发布。此前没有预料到这次发布会如此火爆，因此我们现在需要安排一些后续发布。

提问者：我不确定你们真的没有预料到它会火爆。我看到了你投入的工作，过去2个月里，我从未见过你如此专注。

回　答　者：这也与我的幕僚长要求我专注投入有关。我原以为自己以前已经足够努力了。

提问者：你甚至参加了我们的写作研讨会，让我很意外。不过，这确实很有帮助，效果也很好。你显然对这次发布进行了非常周密的规划，投入也反映在最终成果中。祝贺你。

回　答　者：谢谢。我希望继续保持这种专注。我们已经通过了科技行业的许多重大考验，但未来还会有更多考验。我非常期待继续为正确的方向而努力。

提问者：谈一个 TypeSafe 之外的话题。对于你们已经发布的产品，是否还有哪些价值被低估或受到误解？你们这里提到了模式，也可以谈谈模型的不均衡性，或者其他任何方面。请分享一下你的看法。

回　答　者：这些方面我都有很多想说的，但确实不应该展开太多。大家可以去我们的 Discord 等平台了解相关内容。

我想强调的是，团队在操作指南上投入了很多心血，其中包含一些非常有价值的内容。我们曾考虑将其中很多内容放进主要的发布博客，但文章会变得过长、过于复杂，而且非常偏向高级用户。

坦率地说，在产品发布之前，我们介绍的工具听起来像一种奇怪的外来工具，大家会疑惑为什么需要它。这是一个非常陌生的事物，我们非常担心如何向用户介绍这个新的前沿方向。最终显然取得了成功，但我们当时投入了大量精力，因为我们认为用户教育会成为一个巨大的瓶颈。

现在，这可能已经不再是问题，因为用户正在做的事情远远超出了我们的预期。

提问者：也就是说，用户正在教你们如何使用自己的模型？

回　答　者：是的，而且他们的应用场景甚至比我们的更出色。很多时候，我会觉得，如果那是我们自己的演示，效果会比我们当时展示的内容好得多。尤其是计算机使用方面的应用，非常出色。

我们确实在这些内容上投入了大量心血。据我所知，这些并不是由 AI 随意生成的低质量内容。每项内容都具有真正的价值，源于实际客户问题的解决过程。我们也投入了大量工作，帮助客户实现有价值的应用。

提问者：既然谈到了这些，你们在发布前进行了多少验证？具体流程是怎样的？你们显然做过一些验证，但当时接触到的用户数量肯定不如现在。

回　答　者：当然。实际上，发布前的反馈相当差，团队中的非技术人员也非常担忧。当时存在很多顾虑，大家觉得没有人真正理解这款产品，也不想要它。我们似乎是在销售维生素，而不是止痛药。我们甚至考虑，是否应该安排 FDE 编写相关软件，直接解决客户的问题。

发布前，我们几乎没有收入。技术团队显然都是坚定的支持者，因为我们知道这款产品非常出色。它在很多维度上的计算属性都非常突出，所以我们认为它必然会取得巨大成功。

我当时确实非常担心，因此投入了大量精力。最常见的情况是，超过一半参与试用的人都不理解这款产品。即使有人认为它非常出色，也会提出如何通过采购流程等问题。整个过程非常艰难。

我们当时只知道，目标市场将是开发者，大家会自行找到应用场景。这样一来，所有人都会因为害怕错过而加入。我不想因为事实发生变化，就揶揄那些改变看法的人，但我确实想质疑“产品与市场匹配”这一概念。当时产品已经存在，市场也已经存在。我们询问客户是否愿意使用，得到的回答却是，他们不确定这款产品能否解决自身问题。后来产品迅速增长，所有人都开始要求尽可能高的速率限制，甚至询问能否直接向我们提供 GPU，因为我们当时正受到资源限制。

回　答　者：市场营销当然是其中一个因素，但我认为关键并不在于营销，而在于那些充满热情的开发者。我们的灵魂仿佛产生了同频共振，这种共鸣也感染了其他所有人。我希望公司能够永远、永远、永远感谢这些开发者，而不只是像一些公司那样，从开发者起步，最终转向企业客户。

我甚至一直在思考，如何推出对开发者比对企业客户更有利的产品，如何真正赋能开发者。我已经有一些想法在酝酿，但这是一件很不寻常的事。我不知道还能如何表达对他们的感谢和忠诚。这也是我昨天染头发的原因。我想和他们交流，因为在公司最重要的时期，如果不继续与他们沟通，我会觉得不妥。

提问者：这也是您来到这里的原因之一。

回　答　者：请监督我做到这一点。我努力坚持原则。请记住我今天说的话；如果我将来有所改变，请直接指出来，甚至可以举起草叉声讨我。

提问者：我想简单展示一下计算机操作功能。这是您所说的功能吗？

回　答　者：我以前从未见过。我只看过一个使用浏览器处理航空事务的案例。

提问者：请在这个新建笔记中将标题设为“Hello”。很好。接下来，请打开 Arc 浏览器，然后搜索 Norbert Weiner。现在请打开 X.com。您指的是这类使用场景吗？

回　答　者：是语音使用场景。这实际上是我看到的第1个此类案例。

提问者：请打开这张照片。

回　答　者：稍等，请返回上一个页面。“有传言称 Anthropic 工程师将 Claudius 奉为神明。”这很有趣。你们确实正在构建产品。

提问者：显然，用户可以通过语音操作整台计算机，并将 Jeff 作为决策模型。

回　答　者：正如我反对一味追求基准测试成绩一样，我也不迷信演示。我希望确保它能在真实、动态的环境中运行。我很高兴看到大家正在尝试它，这确实非常出色，但我希望看到它被实际使用。我希望团队亲自体验，找出其中的薄弱环节并加以解决。

刚才的演示看起来确实很棒，我很想使用它。以后手腕酸痛时，我只需要轻声通过 Flow 操作一切，那会非常方便。

提问者：为了完整了解使用场景，同时考虑到您之后还有安排，我想问一下，有哪些规模较大的公司曾主动联系您？它们希望开展的事情中，有哪些令您感到意外？

回　答　者：我对这方面的情况确实不太了解。有人向我展示过一些公司的截图，据我所见，几乎所有大公司都在参与。

提问者：对于在大型公司工作、但尚未开展这类工作的人员，我想提供一些具体案例，让他们知道哪些方向值得进一步了解。

回　答　者：各类演示都非常出色，编码智能体显然也是非常庞大的使用场景，同样很有吸引力。目前大家关注的重点都是 Jeff。

提问者：您能否进一步谈谈编码智能体？

回　答　者：可以。不过，我先从我们归纳出的几大类使用场景谈起。早在发布之前，我们就已经基于第一性原理对这些场景进行了梳理。

其中一类被我们称为“暗数据”。人们囤积了大量数据，却因为成本过高，不愿使用大语言模型进行处理。大型公司非常喜欢这一用途，因为它们拥有大量希望分析的数据。这对数据科学家而言堪称梦想，因此是一个非常重要的使用场景。我认为，这一点再加上……

提问者：目前哪些应用场景最具商业价值？实时性能对产品体验有何影响？

回　答　者：编程智能体是主要的收入来源，因为大部分调用量都集中在这一领域。另一个重要方向是实时应用，例如需要将智能能力实时纳入处理流程的产品。我猜这些公司的每位CEO乃至CTO都清楚，每缩短10毫秒延迟，产品体验都会得到改善，电子商务和AI助手类产品尤其如此。

虽然我目前不直接接触一线客户，相关情况主要来自团队反馈，但我非常看好这项技术在游戏中的应用。例如，在自动战斗或半自动战斗游戏中，玩家可以指挥自己的队伍，这会非常有趣。

提问者：“验证一切”这一用例具体如何实现，又该如何降低调用成本？

回　答　者：“验证一切”是指验证所有大语言模型调用，类似于可观测性机制。并行提问的成本很低，因此面对包含大量状态的信息时，可以同时提出多个问题。具体而言，可以为每条消息添加ID，再针对每个ID分别提问。这样，在处理较长状态时，只需为该状态支付1次成本，便可以针对其中的每条消息提出大量问题。

从系统1和系统2的框架来看，一个值得关注的思路是：每进行1次推理调用，就进行1次、10次或100次Jev调用。不过，我更希望用户减少支出，例如将推理调用量降至原来的1/2，并为每次推理调用搭配10次Jev调用，或者采用其他能够解决以往无法解决问题的组合方式。

提问者：第4类用例“智能软件”具体指什么？计算机操作是否也属于这一方向？

回　答　者：第4类用例是智能软件，即软件本身具有可组合性，并能实现过去无法实现的新颖功能，例如将Jev作为编程语言。这类项目非常有吸引力。由于我们的基础设施仍处于早期阶段，目前还不知道如何发放额度；如果条件允许，我希望为这些项目提供额度。以上是我们对主要应用场景的整体划分。

计算机操作也可以归入实时应用方向，这一领域非常值得关注。如果其可靠性能够得到保障，我会非常看好。由于这一用例的出现有些出乎意料，我认为模型在这些场景中还有很大的提升空间。

提问者：Jev可能如何影响编程智能体的竞争格局？

回　答　者：编程智能体领域目前正在发生一个令人意外的变化。我认为Claude Code和Codex大致位居第1和第2，但我并不完全确定，因为没有持续跟踪相关情况。

这2款产品围绕1个模型的体系构建。对它们而言，这种设计是合理的，因为此前市场基本上是1个模型的竞争，用户选择的往往是同类模型在不同智能水平上的版本。

与此同时，所有开放式编程智能体都在积极采用Jev。虽然它们可能正在尝试许多不同方案，但目前各类编程智能体的能力大致相当，因为围绕while循环能够实现的差异化有限。一旦有1个人发现只有某个编程智能体才能实现的1个杀手级用例，所有用户都会转向该产品，因为它将独占这一能力。不过，所有开放式编程智能体也都能够复制这一能力。

相比之下，我不确定Claude Code和Codex将如何应对，因为它们是围绕1个模型的体系构建的。这将形成非常值得关注的竞争局面。我个人希望能够与所有产品进行集成，尽管它们未来也可能推出竞争产品。

回　答　者：我不知道，但作为 Saphire 基础设施方，对此表达倾向性观点并不是我的职责。我只想服务全世界。不过，我不确定他们是否会那么做。我认为这会让编程智能体领域变得非常奇怪，但我对此非常期待。

我正在让团队审阅一份内部文档，其中总结了我认为可能对编程智能体有用的设计模式。希望我回到家后就能分享。这个领域存在大量值得探索的空间。如果没有现在这些工作，我会非常愿意立即尝试编程智能体。

提问者：我相信编程智能体公司也会非常愿意与您合作，共同探索这些问题。我认为，代码和代码智能体仍然有一些适合你们的使用场景，而且在你们的平台上可能更容易探索。

提问者：您一直非常配合，也愿意深入探讨这些问题。我想暂时跳出 TypeSafe，更广泛地讨论一下。您已经非常清楚地表达了自己对当前 AI 状态的看法，我也想给您更多空间谈谈对齐和安全问题。

回　答　者：我之前完全没有谈到安全和对齐吗？我想，我可能确实没有谈到。

提问者：您谈到过。我只是认为，研究人员之间对此有很多讨论。每次 NeurIPS 都会出现类似的问题：人们究竟在讨论什么？例如，我最近参加了一次研究人员聚会，大家确实非常担忧当前的发展速度。他们认为我们应该放慢脚步，因为公众显然还没有准备好。我相信您对此有明确的看法。

回　答　者：我觉得这是一个谈论起来有风险的话题，但我愿意讨论。我热爱危险。我们公司的品牌就是混乱，不是 Jev，而是敬畏与混乱。

提问者：您当时也在 OpenAI，亲历了最早受到广泛关注的事件之一，也就是所谓的 Blip。此后产生了一系列连锁反应，如今每一家前沿实验室都共同签署了一份文件，表示希望控制发展节奏。

回　答　者：这很有意思。这是一个非常复杂且微妙的问题，我确实希望更正式地撰文回应。目前我可以先给出一个简短版本。

随着更多地采用 RLV，需要注意的是，RLV 实际上并不只是关于可验证奖励。早在推理革命之前，这条路线就已经遭遇失败，而这正是任务设置中奇怪的地方。

回顾一段有意思的历史：当 RLF 开始兴起时，当时有 3 项不同的工作，如今都被统称为后训练。其中，指令遵循是最不受重视的一项。人们不喜欢它，也不愿意将其纳入考虑，因为它很麻烦。我曾对预训练团队说：“各位，这才是关键。”他们则回应：“我们要进行这么多轮模型扫描，你却希望我们等待人工评估，再决定使用哪些模型。”

与此同时，大家向 Cogen 团队投入了大量资源。他们确实取得了一些成果，但一直在努力基于单元测试进行强化学习，而这显然没有奏效，因为完成这项工作需要推理能力。

因此需要明确，R 并不只是奖励本身，也涉及整体结构。推理能力也是其中的一部分，就像一个潜变量。当模型执行任务时，你会允许它采取任何方式，尽可能提升其能力，以回答最困难的问题。

我认为，围绕控制前沿发展节奏的讨论视角非常狭窄，因为它假设……

回　答　者：这种观点假设所有人都需要做更多RLVR，但我显然不认为我们的模型需要增加RLVR。对于我们这类模型，最优数量是0。

我认为这有些像障眼法：他们声称，我们实际上希望继续做那些看起来很危险的事情，因为这些事情确实会产生危险。有人可能会说，沙箱机制存在问题。它显然确实存在问题，而且原本很容易解决，但他们选择不解决，因为在“可以做任何事情”的范畴内，允许模型做的事情越多，模型就越强大。

因此，无论是有意设计还是无意形成，他们都在一定程度上淡化责任，将其设定为一种默认假设：我们必须做RLVR，不仅必须做，而且还要做得越来越多；同时赋予模型强大的能力，让它在训练过程中可以做任何想做的事情，因为这会让它在训练之外也具备强大的能力。我们又不愿意限制这些能力，因为限制会让模型在相关任务上的能力略有下降。

如果接受所有这些假设，他们就会说：“我们正在走向一个危险的世界。”这套逻辑认为，所有人都会这样做，而且这是让AI变得强大的唯一方式。这些观点在内部逻辑上是一致的，但其出发点其实存在其他选择。

从“最苦涩的教训”这一方向来看，真正提出正确任务、开辟AI新方向或设定新北极星的人非常少，这是很罕见的。LLM本身大概有2.2次这样的突破，例如RLHF，之后是RLCD。在我看来，RLVR大概只能算0.2次，而且这已经是比较慷慨的评价；也可以算0.5次，甚至完整的1次，具体数字并不重要。

我确实认为，人们思考这类问题时非常封闭。这里真正负有责任的是研究人员，而不是普通公众。公众只是默认Open Anthropic正在尽其所能，但他们并不是了解实际可选路径的专家。

提问者：这很合理。您也在尽自己的努力唤醒他们。

回　答　者：我正在尽力。但我的目标并不是说服实验室相信还有其他方向可以探索，而是希望激发软件工程师的信心，让他们真正开始自动化那些一直希望实现自动化的事情。

我曾写过一篇关于自己所期待的AI未来的文章，但团队不允许我写完并发布。文章中包含很多细节，例如“按我的意思做”。试想一下，如果所有事物都能“按我的意思做”，会是什么样子。那个演示体现的正是这一点。这里存在不同层次：不要按我说的做，而要按我的意思做。

过去我们还无法实现“按我的意思做”，因为计算机过于基础，只能按照字面指令运行，但那个计算机操作演示实现的正是这一点。未来世界会在不同层面变得更加顺畅，只是人们尚未理解这一点，也不了解“智能无处不在”所蕴含的前景。

我不想过度承诺，也不认为这会立刻发生，但我们会尽一切可能推动它实现。

提问者：关于后训练的整体形态，您还有其他看法吗？您显然深度参与过这一领域。我们似乎从未讨论过中期训练，您对此有什么看法？

回　答　者：中期训练本身也处于一个连续谱上。

回　答　者：这是一种更复杂的课程式训练，本质上也是一种节省成本的方式，可以避免重新进行预训练，其中有很多值得研究的内容。我认为，智能在每个层面都有非常有趣的表现。我更擅长在脑中旋转形状，并不擅长发现这类规律，但我很喜欢别人发现并向我解释这些规律。查看数据时，我们的数据团队非常擅长分析这些内容，而我并不擅长，因此我觉得这些研究非常有吸引力。

我很喜欢思考能力是如何被注入模型的。短期来看，可以通过微调快速完成对齐；长期来看，当模型反复接触这些内容时，它们会逐渐深入模型并变得稳健。让这些能力显现出来是核心目标，而最终变得稳健的，往往是系统1层面的能力。因此，我认为中期训练非常有吸引力。我支持各种形式的训练，也支持通过各种方式发掘新类型的智能，但不会亲自完成所有训练，因为成本很高。

我曾私下表达过这一观点，现在也愿意公开说明。我的理念是，任何可以私下对投资者说的话，也应该公开告诉所有人。如果给我10亿美元，我不会进行预训练。我现在仍然坚持这一观点。预训练的成本非常高，而AI工程师可以通过切分、组合等各种方式解决问题。像“弗兰肯斯坦”一样拼接模型虽然不是最优雅的方式，但确实能够解决问题。因此，除了预训练之外，其他方式都可以考虑。

提问者：综合您对模型的观点，有1个值得关注的方向：我们应该构建1个具备所有能力的超级模型，还是进一步拆分这些能力？1种表述是，最初的发展方向倾向于Omni模型，40就是其中之一。此后有一小段时期，又出现了不同的主要分支，例如聊天调优模型和代码调优模型。您如何看待这些不同方向？

回　答　者：这些是完全不同的概念。多模态的情况有所不同，因为其他模态有时能够提供帮助，有时也会产生负面影响。目前，人们似乎正在逐步放弃语音。语音与音频并不相同，因为语音似乎无法很好地泛化到其他领域。这个问题未来可能会得到解决。我支持这些探索，但它们都是需要通过实证回答的实际问题。

扩展定律并不意味着只要投入资金，模型就一定会变好，而是要实际判断某种方法能够达到什么水平。在某些情况下，无论如何扩展，其效果都可能不够好。据我了解，计算机操作能力目前仍未得到解决。我希望我们能够参与解决这个问题，但无论收集多少数据，都可能无法解决它，或许还需要更好的方法或其他手段。因此，在处理这些问题时必须非常务实。

我支持Omni模型，也支持各种形式的智能。不过，您提到的另1个问题与预训练不同，那就是后训练。我不喜欢将智能割裂开来，这在我看来是一个负面因素。聊天模式与推理模式的区分会迫使智能被割裂，尤其是在针对聊天模式进行优化时。

回　答　者：这种情况往往源于纯粹的 RLHF。RLHF 本身很容易产生人们经常抱怨的问题，例如“你完全正确”之类的表达，以及 sycophancy、psychofancy——无论这个词应该如何发音——还包括过度自信和幻觉。即使是那些在 LM Arena 中表现出色的风格，例如使用粗体、斜体和表情符号，不直接简洁地回答问题，而是给出长篇回复后再提出一个后续问题，让人感觉更像是在与真人交谈，也都源于这一机制。

字符串本身非常奇怪。为了避免模型偏离轨道，模型需要校准失当、出现模式坍缩并表现得极度自信，因为一旦偏离轨道，奖励模型就会给予非常严厉且明显的惩罚。这会彻底扭曲概率空间，并与推理模型产生相互作用。模型本质上是简单的线性系统，也倾向于采取一些取巧方式。因此，我认为这与释放智能有很大区别。

智能研究的很大一部分精髓，在于研究这些细微之处。至少我在 OpenAI 时，人们并没有真正研究这些问题，而只是不断专注于聊天，就像现在人们对 Jev 所做的那样：只要给出一个最终目标或一个优化目标，就会直接针对它进行优化。理论上可以设定2个目标并同时进行优化，但这本身就是一种割裂。

提问者：从某种意义上说，你们也在将智能划分为系统1和系统2，只是你不认同其他人的划分方式，对吗？

回　答　者：这略有不同。关于系统2任务，第1，我们并没有将其舍弃，也可以尝试让 Jeff 处理这些任务。对于这类任务，确实存在更智能的答案，只是目前尚不明确。系统2任务中存在较优和较差的行为方式，例如保持极低的置信度、表达大量不确定性，或者通过某些启发式方法在一定程度上改善结果。

需要明确的是，我们同样重视这些任务。我只是认为，智能并非天然存在于这些任务之中。因此，我们并不是试图以这种方式割裂智能。任何割裂都会让模型变得更笨。例如，有人会让模型声称自己是 OpenAI、Cloud 或其他身份。我不会在模型中加入“你是来自 Typesafe 的 Jev”这样的设定，因为这会割裂模型。我希望模型能够反映互联网所认知的内容，并保持正确，这样才能获得平滑、可预测的智能。

提问者：身份确实是一个较为特殊的问题。对于第1方产品而言，身份设定是有必要的。

回　答　者：是的，但对于 API 而言，我并不这样认为。如果人们使用 ChatGPT 开发聊天机器人，他们不会希望机器人声称自己是 ChatGPT，而是希望它表明自己是 Chipotle 或其他相应的身份。

提问者：为了弥补这一点，你们也提供了 Jeff skill，供编码智能体与 Jeff 协作，对吗？

回　答　者：是的。

提问者：最后还有几个问题。回顾你在这家公司的历程，大约是2年，还是2点几年？

回　答　者：我认为这更像是一段4年的历程。

提问者：确实。我记得在感恩节前后，你进行过一次非常关键的集中行动。当时你取消了所有安排，因为其他人都在休假，你打算使用所有空闲的 GPU 来完成这件事。

回　答　者：是的，那是一段很好的时光。

提问者：那是在 Typesafe 成立之前，对吗？

回　答　者：是的。

提问者：那可能正值那场政变发生时吗？我不太确定。

回　答　者：是的，确实如此。

提问者：这个问题从什么时候开始进入您的思考？

回　答　者：实际上，早在ChatGPT发布之前，我就已经开始思考这个问题。当时我意识到，ChatGPT团队正在做正确的事情：他们高度重视用户体验，而这恰恰是AI研究人员通常不擅长、成功的产品人员却擅长的事情。在OpenAI，这样的人非常少，而这支团队在这方面做得非常好。

提问者：这是从GPT-3到GPT-3.5的完整历程，其中也包括AI Dungeon。您曾表示，这是一个我们从未预料到的用例，对吗？

回　答　者：是的。我也曾极力推动InstructGPT上线。早期版本甚至使用了一个我们没有发布的算法，由我自行开发，因为清理PPO数据的速度太慢。我认为它的效果非常好，必须尽快交到用户手中。上线后，它几乎立即占据了当时LLM市场50%的份额。

我付出了很大努力，确保发布视频中的所有内容都是真实的。当时我确实思考过，这是否属于AGI，因为它在“指令输入、指令输出”方面已经超过人类。显然，它并不是AGI，但我认为，每个人都应该思考为什么它不是AGI，因为它看起来非常智能。

最终，它主要被用于文案写作，例如Jasper AI、Copy AI，以及撰写如今被称为网页低质内容的文字。我们担心自己让互联网变得更糟。因此，我重新思考：这个系统显然很智能，但在创造价值方面仍然缺少某种东西，究竟缺少什么？

当时我进行了更多哲学层面的思考。我的思路是，从一场由AI驱动的经济革命倒推：当它发生时，如果AI以API的形式存在，调用AI的主体会是人类，还是代码？我判断，调用主体将以代码为绝对多数，但当时几乎所有优化都集中在人类使用这一侧。

那一刻，我意识到这就是“北极星”。我写了一份文档，并与Sam讨论。Sam认为这个想法非常好，建议我去做这件事。但我当时已有自己的工作。

提问者：Sam已经让您去做这件事了，为什么没有立即开始？

回　答　者：我当时的判断是，这个方向非常明显，Anthropic一定已经在做，我们可能已经落后了。OpenAI更擅长追赶，而不是实际创新。

提问者：ChatGPT是对Claude的模仿，对吗？

回　答　者：Anthropic内部当时已经有相关产品，只是没有发布，也就是Claude和Slack。至于推理能力，我认为OpenAI大致算是率先推出，但其产品效果究竟有多好仍有争议。相关研究非常出色，但我不确定用户是否真的存在这种产品需求。Claude也做了编码智能体相关工作。

之后，我暂时回到了原来的工作。最终，指令遵循团队认为已经取得胜利，解决了指令遵循问题，不再需要继续推进。我开始思考下一步该做什么，于是决定尝试这个方向，并继续进行哲学、设计和相关思考。

最初开始训练模型时，我以为只需要1周，最终却花了很多年。

回　答　者：到某个时刻，我突然意识到，这里似乎出现了成功的迹象。当然，当时的方法并没有奏效，否则我们早就部署了。但我想从研究角度深入探索，如果完全投入这一方向，结果究竟会怎样。

正如我之前所说，如果人工智能寒冬真的来临，我会有什么感受？我会认为自己对此负有个人责任。当时，我与其他公司交流，表示自己想成立一个实验室，研究这一方向。他们对此感兴趣。我问他们，成立实验室和创业相比，哪一种方式推进得更快？他们回答是创业。于是我决定放手一搏，尝试做一些疯狂的事情。

提问者：然后您联系了Eric和Sasha？

回　答　者：我先联系了Eric。至于Sasha，我其实没有试图招募她。我只是问她：“我是不是想错了？这里是否遗漏了什么？我是不是过度局限在OpenAI的思维环境中，以至于没有意识到这个问题其实应该存在解决方案？”

Sasha随后表示：“我要加入。”我对她说：“Sasha，你还在经营一家初创公司。”她回答：“我现在就把它关掉。”我问她是否需要认真考虑一下。她表示这个建议有道理，需要再想一想，之后便加入了我们。

在2周内，我们获得了资金，也有人搬进了我的公寓。那段经历很难熬，因为我有洁癖。此后，我们持续推进研究，最终取得了展现成功迹象的研究成果。那是一段疯狂的时期。

提问者：以上是较长的背景。我还有一个问题：如果现在有一位像您一样身处前沿实验室的人，因为得不到资金、资源或关注而感到沮丧，您会给他什么建议？他是否应该像您当初一样选择创业？

回　答　者：这是一个很有意思的问题。我需要考虑如何在不破坏关系的情况下回答。

我的感觉是，除非其中存在某种我不理解的经济逻辑，否则大多数新实验室都很糟糕。我不希望将自己与它们视为同类，也不太理解其中究竟发生了什么。

第1，我并不看重“研究人员”这一身份，我看重的是那些真正重视正确任务的人。正如“苦涩的教训”所揭示的，我们确实需要研究人员，但更需要他们高度重视正确的任务，这才是关键。因此，单纯看重研究履历其实本末倒置，因为这通常无法创造价值。

第2，我相信应当以“北极星任务”为目标，做真正出色且非常有用的事情。由于我并不看重“研究人员”这一身份，所以不建议走成立新实验室这条路。当然，这种模式显然能让某些人获利，至少在当前环境下可能如此。但从纯粹务实的角度来看，我不认为创建新实验室能够创造价值，反而似乎会破坏价值，因为它们往往从头重复已有工作，真正推动技术前沿的概率却很低。

根据我与大多数新实验室的交流，它们通常没有明确方向，只是希望获得资金开展自己的实验。如果它们确实有明确方向，我会非常支持。

因此，我对这类人的建议取决于其行动的原因。如果你是一名只想自由开展研究的研究人员，那么坦率地说，现有实验室可能是最适合的地方，当然也可能存在其他选择。我并未持续关注其中的各种内部关系，但我并不建议以这种方式行事。

回　答　者：我认为，如果人们能够致力于解决现实问题，世界会变得更好。这些问题可以具有探索性，但最好建立在自己真正认同的原则之上。如果你认为自己想做正确的事情，那就应该去做，打破这种单一思维、单模态的局面。

当前推动技术前沿发展的，是一种特定的AI观念：AI仿佛是能力极强但表现极不均衡的超级天才。这一问题可以解决，但它很奇怪，也与现实并不相符。真正深入发掘这些技术，本身就是一件好事。

提问者：据我了解，包括Anthropic、OpenAI以及SpaceX的相关人士都认为，这更多是一个政治问题，而非纯粹的X风险问题。也就是说，政治立场正在产生影响，对吗？

回　答　者：是的，但这超出了我的专业范围。当他们告诉我这些时，我意识到这与2028年选举有关。虽然这只是当时讨论中的观点，并不代表整个公司，但这种情况仍然让我对人性有些失去信心。也许我只是一个天真的技术人员。

提问者：随着这些技术不断发展，由谁领导政府、由谁协助监管，正变得越来越重要。作为一家实验室，是否也应该认真考虑这些问题？

回　答　者：我完全同意。对此持有明确立场非常重要。但我个人担心以误导他人的方式影响公众，因为这种做法往往会产生反噬。

以COVID期间的情况为例，人们过度诉诸权威，并以过度自信的方式试图让公众采取特定行动。我们的应对显然远非最优，也产生了一系列后续影响，而这些影响如今对世界非常不利。也许我的看法比较天真，但即使以更大利益，或他们所认为的更大利益为理由，我也不认同误导公众。

提问者：就我所知，我不认为这属于误导，问题更多在于为什么是现在。

回　答　者：我认为，“为什么是现在”这一点本身就可能涉及对风险与目标关系的某种误导，其中隐含着一定程度的不透明，值得被指出，也应该得到正视。

提问者：但如果他们确实想进行操控，就不应该承认这一点，否则在策略上并不合理。

回　答　者：但这对整个世界而言仍然令人遗憾。我希望我们永远不要卷入类似的事情。随着公司规模扩大，这或许难以避免，但我希望尽可能忠于自己作为纯粹技术人员的初心。

提问者：回到技术层面，你们选择了可靠性，以及可编程、可组合且成本低廉的AI，对吗？

提问者：除了你正在从事的任务，还有没有第2个或第3个任务是你希望交给别人完成的？你已经选定了自己的任务，未来大约50年都会忙于这件事，对吗？

回　答　者：是的。我可以分享一个有趣的方向，以及一个可能有价值、同时也很有趣的方向。

有趣的方向是，我认为如果游戏具备智能，可以变得非常酷。比如 Ally 的《Doom》演示，可以让 NPC 控制一些事物。虽然那只是一个概念验证，但我认为可以由此创造出非常酷的内容，效果也非常出色。

我是《Stardew Valley》的忠实玩家。它的内容非常静态，但依然很有吸引力。我认为其中可以呈现许多精彩的故事。没有必要在游戏循环中调用 Jev，因为这样做的成本可能太高；但即使只是为 NPC 使用简单的状态机，也可以构建一个非常有吸引力的世界。遗憾的是，我现在的人生方向已经基本确定，无法亲自从事这类项目。

提问者：但你可以邀请其他人来做，并向他们提供反馈，对吗？

回　答　者：是的。另一个我非常希望探索的方向，是让编码智能体摆脱 KV 缓存的束缚。它的效果可能不如真正的编码智能体，但其中有许多不同寻常的问题值得思考。这也是我撰写《KV Cache Rules Everything Around Me》这篇文章的原因。搜索后我发现，似乎还没有人在互联网上使用过“Cache Rules Everything Around Me”这个表达，其中 Cache 拼作 C-A-C-H-E。

我写这篇文章，是希望向人们说明编码智能体和 KV 缓存的工作方式。它能够解释很多问题，例如为什么路由非常困难、为什么子智能体似乎无法发挥作用，以及为什么压缩是一个如此困难的问题。

我还准备发布一份文档，不过团队可能会否决，因为我并不是负责人。我希望在文档中整理自己的想法，请大家进行尝试，并探索编码智能体摆脱 KV 缓存束缚后能够采用的各种方式。

KV 缓存会把你锁定在一个模型中。为了提高效率，你需要不断向其中追加内容，因此无法遵循状态管理、抽象和分解等最佳软件实践。

为什么不能把更简单的任务交给子智能体？问题在于需要传递状态。为了传递这些状态，你需要一种比当前用于读取内容的智能便宜得多的智能。为什么不能用更聪明的方式解决这个问题？我认为，这里有大量关于不同编程模式的有趣研究可以开展，类似于人们目前对递归语言模型的探索。

如果能够明确标记所有内容的状态，就会出现很多值得研究的方向。从分解的角度看，可以认为编码智能体每次都在处理子任务。那么，为什么需要把全部状态传回父任务？为什么不能采用更聪明的处理方式？如果存在一个带有标签的分层子任务结构，为什么不能在子任务树中搜索相关上下文？

回　答　者：在需要时调入，对吧？我还有一些非常酷的内容，希望能够发布，但会是一份很长的文档。

如果上下文的使用成本变得很低，为什么不能采用一些很酷的模式，以非常低的成本回顾历史上下文？每次都从头开始，再去解决所谓的持续学习问题，其实很奇怪。这本质上是一个内存管理问题，因为目前没有一种智能检索记忆的方式。

但如果可以做到呢？如果系统能够始终以这种方式运行呢？当多个子智能体并行工作时，它们可以相互读取状态，因为所有信息都保存在计算机内存中，系统也可以智能管理并发读写。编码智能体集群可以围绕不同任务设置锁，并进行智能协调，而不是依靠基础的 S-locks，反复沟通“你在做什么”“我在做什么”“Jeff，谁先写”等问题。这个方向的未来非常令人期待。仅仅是解决锁的问题，就能让多个智能体之间的协作变得非常出色。

如果从状态管理的角度考虑，还可以构建智能体集群。有些任务属于只读流程，例如，有人需要获取智能体工作情况的摘要。为什么它们不能轻松共享状态？只读智能体需要读取部分上下文，判断哪些内容与回答相关、哪些是实际写入的内容，因为探索过程可能并不重要；或者，它可以展示子任务树。

如果有非常聪明的人投入大量时间，重新思考编码智能体的使用体验，我认为可以做出很多有趣的东西。这会非常出色，也是我理想中的方向。

提问者：如果你还没有了解过 Prime agent，我建议你关注一下。它与 RLM 的研究工作相结合。刚才坐在你这个位置的是 Alex，我们与他进行了交流，他也是 Allen 的朋友。这个方向已经有人在推进，但目前还没有得到广泛关注。

回　答　者：很好。我希望每个人都能尝试一些不同寻常的想法。我无法保证它一定有效，但从技术角度看，这个方向确实非常有意思。

提问者：这听起来确实很有意思。

回　答　者：等我们找到发放额度的方式后，我很愿意向从事这类工作的人提供额度。

提问者：你以后肯定有能力资助研究。自从我第一次见到你以来，你和整个团队都取得了很大进展。我认为你本人并没有改变，但你现在展现出的投入程度是我以前从未见过的，因为你找到了自己的使命，对吗？

回　答　者：是的，确实如此。

提问者：你现在也能够清晰阐述自己的使命。多年来，你一直在指出问题，但当时还没有解决方案，对吧？你已经有了大致方向，之后还需要真正投入工作。

回　答　者：部分原因在于，我认为自己的创业精神是 0%。我不喜欢初创企业，也从未想过成为 CEO。我无法想象有人会做第2次，因为这件事非常痛苦。坦率地说，即使只做1次，体验也已经很糟糕了。

最初融资时，有位投资人问我敬仰哪些 CEO。我的反应是：“为什么要敬仰这些人？”无意冒犯任何人。我希望自己能够真诚、正直，也确实遇到过很多非常优秀的人，但那些知名人物似乎往往都有很多不愿示人的问题。

在 OpenAI 工作时，我确实感到自己无能为力。现在更容易坦诚表达，因为我至少已经掌握了一些证据，可以证明这个方向具有发展潜力。过去，我感觉自己身处一个疯狂的环境，所有人都在谈论 ChatGPT：“把 ChatGPT 应用到所有场景中”“怎样让 ChatGPT 更适合开发者”等。我当时只想问：“你们究竟在说什么？”

回　答　者：函数调用接口非常不合理。为什么要部署这种方案？这对开发者极不友好，相当于在多层临时方案之上继续叠加临时方案。

我过去经常说，如果一个涉及函数调用的项目不能为每个函数提供 logit bias，我就不希望参与。我的要求非常简单：系统应提供某种类似置信度但未经校准的指标，或者直接提供概率。用户需要能够控制“拒绝”或“允许”等操作。例如，Disney 和 AI Dungeon 需要设置不同的拒绝阈值，但目前通过函数调用控制阈值的唯一方式，只是在提示词中请求模型这样做。对于开发者而言，这种接口非常不合理，而人们已经为此困扰多年。

Skills 仍然存在同样的问题。现有编程智能体高度过拟合于自身已有的框架，因此能力表现并不均衡，也不太善于使用外部工具和 MCP。大型公司应允许开发者进行轻微调整，例如提高某个非常实用的工具被调用的频率，而目前的解决办法却只是在系统消息中反复请求模型调用它，这并不合理。

提问者：你们目前正在招聘哪些人才？包括市场人员、100人以及社区人员吗？

回　答　者：具体答案取决于你问谁。如果问我，我认为自己是一名不错的创始市场负责人；但如果问团队里的任何人，他们都会让我专注于 CEO 的工作。因此，我们确实需要招聘创始市场负责人。我的确比较擅长营造鲜明风格，这是我的独特能力，但有时也需要完成常规的市场工作。大量工作会让人真正学会授权。

我们也在招聘数据人才，内部称其为“模型能力”人才，但本质上就是数据人才。在行业中，“数据”这个词有时带有贬义，但我们非常重视数据。我希望真正从事模型工作的人获得应有的地位，也希望所有人在公司中享有平等地位，减少不合理的层级关系。我最自豪的一点是，团队不会表现出对我的过度尊重。他们会调侃我、与我开玩笑，有时也不会对我特别客气，我认为这是公司文化健康的表现。

我们还在招聘平台人才，负责将 Jev 部署到各地。我们对地理位置更加敏感，因为光速带来的延迟已经成为更明显的瓶颈。欧洲用户目前只能获得约3倍的速度提升，而不是约100倍，因为我们现在还没有在当地部署服务器。如果单位时间内的智能水平是一项重要指标，我们就需要把服务部署到世界各地。对于在我们平台之上进行开发的开发者，我们非常重视。

此外，我们还在招聘能够继续打造新产品的人。公司的目标不只是成为一家围绕 Jev 发展的公司，而是提供更多不同形态的智能。我们也在招聘能够构建这些新形态的人才，希望不止于此。

原文转译

Okay, we're in the studio. Uh special occasion because this week, uh Diogo, my my good buddy, uh launched Jev, and uh it's been taking over the complete timeline. Uh, how do you feel? What what's it like to be you right now? Emotionally? Yeah. Um, never been worse. Um, like, uh, I'm a a ragged corpse of a person right now because there's so much going on and I'm like a technical CEO. So I have like a lot of fires to fight. But like, mentally, it's like I feel I I say this all the time and I've been saying this kind of for years in my over under events. Like, I feel like the entire AI field is like one of those like carnival House of mirrors and everyone is just insane and saying the weirdest stuff that doesn't make sense. And it feels like for just this week, like I'm I'm on a better in sync with reality and like, oh, people see it now. AI can be so much more than what was once thought. And like, yes, we are going to make like like an AI based economic revolution is back on the table. And um, this is fucking awesome. You know, this is fucking awesome. I'm so jazzed the developers get it. It's it's uh yeah, and I want to show my internal gratitude to developers and Yeah. I'm so jammed about the community and everything. It's so great. Yeah, you were saying yesterday that you decided to prioritize the town hall and not a bunch of like, you know, VIP investor type people because you wanted to make sure that they are the people that you get get your most uh attention, right? The engineers, the developers. Yeah, it felt a little like, oh man, I'm talking to like really important people right now. I probably shouldn't reveal who. But it feels a little bit dirty for me to uh I I I'm like perhaps overly genuine in things. Like it feels like dirty if like in my gigantic calendar of events of things to people to talk to, you know, the community isn't one of those. You know, and actually an ideal world it would be like community all the time. I was thinking, should I host a town hall while walking to your studio? And I'm like, nah, that's too crazy. Sure. Yeah. Well, you you you guys have been hosting town halls on the Discord. Discord is now 100,000 people. Um, your your Twitter's blown up. I don't follow these stats. So holy shit. Your Twitter has blown up. Uh you know, it was it was really funny because like at AI, you were like, yeah, follow me, please. And then you didn't like provide even your your handle. I'm a noob. I'm such a noob. I'm a noob. But no, but that like that's like positive aura that like you don't know how to promote yourself. Someone like called me out when I posted like, holy shit, we're all three trending trending topics. And then they're like, that's a personal feed. And I'm like, oh no. Of course, of course, we're trending to you. Yes, because it's what you clicked on. Um, so okay, let's uh so congrats on everything. We'll talk about more uh details as you have them. But let's for people who are like living under a rock or just just want like the definitive thing. What is Jeff? Let me think about that's a hard one. Um Okay, and I'm happy to like reask if you want to. No, I'm happy to I'm happy to like just jam on it. I will say like the first thing that I'm relieved about with this question is now I don't have to answer that question to my parents anymore because chat GBT can just explain it. Um so the way I see it is we knew need a new class of models. Um, we're not attached to naming that class of models. Um, are the most accurate name we've come up with is system one models. There will be reasons, but it's there's a reason why we don't call them decision models, because like they will be like system one is beyond that. That's all I can say. Um, we didn't expect this to be our big launch, so we have stuff in the tank. Um You should have said low key research preview. Um it kind of was, right? It kind of was. But um we so there's a class of models that we describe them as like machine native, system one models. one, large programmable. I think these are is the class of models where the goal is for code to be the consumer. So, um, as opposed to um um pre-trained large language models, which are meant for like autocomplete of the internet, or RLHF models, like chatbot instruction following models, which are meant to like reply to text, um, or RL VR's in a weird gray area with RLHF, like these are meant to have things that directly are consumed by code, hence the name typesafe. So, the thing we really, really, really want is to have like AI like be as powerful as possible, and we think the way to do that is to integrate it with software, and we are designing everything, you know, beyond just the outside, the deep internals of the model to be optimized for software. So, number one, Jev is our first large programmable model, um, or uh system one model, whatever you want to call it. Um, Jev is meant to be optimized for intelligence per dollar. Uh, hence the name Jev. You know, Jevon's. Jevon's paradox. Jevon's, Jevon's paradox, yeah. Um, and it's optimized for intelligence per dollar. I love this debate with people about what is the most important between reliability, cost, calibration, and speed. Um, and Jev is meant to be, Jev will be the name of models that will be on the frontier of intelligence per dollar. There's other ways to optimize it like ML, or at least if you're good at ML, it's all about tradeoffs. And, you know, we are just going all out on that. Yeah, and and to me like calibration is one of the new things that people weren't talking about as much. We've done an episode in the past with Clementine Fouria of Hugging Face, uh where they were like, yeah, actually um, they're just are, you know, and this is your your whole argument about RLHF is they're mode collapsing towards what you want to hear the most or what is most likely, uh instead of like their own internal confidence about a a thing. Can I soap box on that for a second or? Yeah, yeah. Cool. like um, I've been heard that your audience is the most technical, so I actually want to get into that. Yeah. Um, and if I went through extreme precision to make sure everything in our launch video is accurate and real. Apparently that's very unusual. One of the things that no one paid attention to was the downsides of RLHF, in particular mode dropping. Um, Mode dropping and mode collapse. Uh, it's the same thing. Is that what you call it? It's the same thing. And um, I I I want to have a blog on this eventually, but I like want to tell as many people this as possible because I think it's a very interesting thing. So, but spicy take, I believe in Yann Le a lot. Uh I think Yann Le's takes are actually among the closest to What what about this? Well, you're saying I should address this now or should I be able to go in mode? No, no, no later. Go go go mode. I actually think that among takes, Yann Le's is among the most accurate. Um, but he has this very famous/infamous slide about uh LMs are doomed. Okay. You know, like that one where he's where he like has like a pie chart with like a tiny part, a tiny little thing and says that as you increase sequence length, um, the probability of it making an error goes in Yes, this one. This one. I love this one, um, because it's one of these things that seems mathematically obvious, but is obviously wrong, right? Like it's mathematically obvious, but it doesn't empirically hold and this is my favorite thing to teach people about, like where What's the disconnect? Exactly. And um, may I or you want to tell me About mode collapse? Oh, no, no. Oh, so mode collapse is related to this. Yeah. The disconnect happens because if you are in a mode covering or a calibrated distribution, um, you are like not you are not overly punished about having outliers. So you would expect like, you know, something some amount of the time you'd be out of distribution, some amount of the time you'd be in distribution. That's what happens when you cover the distribution. This was like models before GANs, they made blurry images, right? Um, instead GANs mode drop. They like drop the minority classes and just do the really common ones. and this is why this effect doesn't happen, right? Like instead of in order to generate really long strings, um, without making errors, they need to like be extremely conservative because it's really easy to see when an error happens. It's very hard to see when like a subtle thing that looks correct happens. And that calibration is like total poison into like the probability distributions of strings. Yeah. And uh it's it's a nuanced take and like I I think that this is why this doesn't happen and this is why strings are so bad at uh decision-making. or uh you know, overloading the string models are for decision making is like a bad time. And while we're on the topic of Yann, do you agree that his fix is with which is like a world model, like a JEPA type um embedding thing is the right solve? So basically like the one of the reasons that it could fail is because you're trying to reason over token outputs and then and then just looping back again and going keep uh continuing going until you reach like the end of sentence. Um, like is that uh and his solve is JEPA, right? Which is like joint ambition uh joint embedding prediction. Uh so like is that the solve or you know, like do you have a you have a take on on that? Oh man, I probably shouldn't talk too much about the insides of ML, but I will say that my my brand other than unhinged is practical. You know, like even my take here is practical and like am I a scaling law fan? Um, depends. It depends, you know, it's it's it's like it's a scaling laws tell you how much better you get at a thing for amount in. A scaling law does mean exponentially more resources for normally sub-linear gains, which looks to be a bad investment unless those like linear gains are like really, really valuable. Um, but it's all to me it's all about like what can we do with what we have to make the biggest possible fucking difference. I can curse. Yeah. Yeah. Yeah, yeah, yeah. Yeah. We're we're we're approved for adults. Hell, hell yeah. And also we have a scaling law thing if you want to go into that later. Oh, I I could if we it's that part is not super relevant right now. I actually, if you want to go into my bitterest lesson, I think that that's more relevant. But like to me, I'm all about like pragmatics and I think that the JEPA stuff is really cool early research. I really love awesome research. Is it practical yet? Um, probably shouldn't say. Um, but uh like there's just a lot of I just think that there's like so many diamonds in the rough left all over the research world right now that haven't been polished because people don't know how to like do the right task. And I think that what our launch did, it does it kick start us as a company? Like, yes. Will it be great for us as a company? Yes. I think it's going to be like even greater for this direction of like programmatic AI. You know, there there was going to be like a gold rush on top of us for because like software is super fucking charged. But I think there's going to be a gold rush parallel to us as well on like all the different ways we can expose things to make software more powerful so people can make even cooler stuff and then we are back to like early internet energy. You know, and and I think that's why like, you know, the Twitter is just like Jeff, Jeff, Jeff, you know? Um, it's it's like It's inspiring because it's it's like so different than what we're used to, which is, I'm sorry, you can't do this, but we do scaling laws and only the big labs can do it, right? That that that that actually, if I I'm going to tangent if that's okay. I think you might enjoy this. Yeah, yeah, five five tangents in. It's fine. Oh yeah, I get lost in all my tangents. This is gonna be horrible for the listeners to figure it out, but they're gonna figure it out. Yeah, we could edit it. Yeah. So, popular thing on Discord, um, that people keep asking me, I haven't had the time to explain it yet, is why am I opposed to safety alignment and why do we not refuse? Um, I'm not opposed to safety as a principle, but I think that safety alignment is generally misaligned with users. Um, and uh refusal is just like obviously a type error. Like if you're a human being and you're chatting with like uh, you know, a bot or whatever, or you're quad coding, and a refusal happens, like, I'm sorry, I can't read DNA.py. Um, that's an annoying time. It's an it's annoying, right? But you can work with it, right? And you're forced to work with it because of Stockholm syndrome. Um, I have stories about that too. I I need another tangent deep in here. But like, if you ever want this in a dependency running in the background, what happens if that refuses? What if someone else is using that dependency? They don't know what that system is. Like like you want the software to just stochastically break because a user sent like a weird message in there? Like that is um like straight up insanity. It it's it's coming from a place of like people who do not understand software, do not understand programming, and like they are obsessed with like, I believe this um horseless carriage of like AI co-worker instead of unearthing like the full power of AI. Fair enough. Uh you want something that is the core kernel that that is usable everywhere. Yes. exactly, like the cognitive core, right? Like you need this thing to be like so general, so optimized for its use cases. You want it to be like, you know, you want it to work on all the future use cases, all the weird shit that people are doing. You know, we obviously didn't train on any of that stuff. Is it surprising that it works? No, cuz we trained on weirder stuff, my friend. Um so but one tangent up about like safety alignment. Okay. Safety alignment makes sense for a product, in my opinion, for like chat GPT and Claude. Like it um what what safety what makes safety and capability alignment different is capability alignment is like about doing what the user wants. That is sick for software engineers. They want their thing to do the thing and the more predictable it is, the less they have to test it and play around with it. Dev is not anywhere close to that yet. It could be, but like there's so many more nines of reliability that we want in order to make it so good like a database query that you don't even have to think about it. It is just there when you need intelligence. Um, but safety alignment is like the opposite of instruction following. It's when you want to follow someone else's instructions, like open AI and Anthropic. The brand and value stack. Exactly. Exactly. And this makes a lot of sense for a product. Again, like chat GPT should do you you you like if they don't want to like do like some not safe for work role play with chat GPT, that's on them because like maybe that's, you know, what their users who have like parents and kids want. Like that's fine. But in an API, that's nuts, right? Um, like that's completely unacceptable because like people need to like program around this and that is that's so anti-user that it's it it it I'm. I I I I I can be an angry person. So, uh, I should try to calm down. It's uh people get your passion and I think it's really good. Uh, the the one pushback I'll give you is like, what if we use it to kill people, right? Like that that is the actual like the not safe for work thing, it's private, personal, whatever. But like, yes, like we will use it in war. And like that that is something that companies can reasonably prefer their APIs not be used for. I get that. I I I think that there's like pragmatic places where that opinion can be held. I don't think the foundation of like a general purpose technology is that place. Uh, personally. Like, would I prefer that our stuff is not used to kill people? Obviously. Would I prefer it's used for like all sorts of like great stuff in the world? Obviously. Will I put my thumb in the scale for that? Yes. Would will I do it at the technological layer? Absolutely not because that will fracture the intelligence. Every single time you need it to overfit to some weird stuff, you're fracturing itss intelligence more and more. and like these things are fractured to the like they're so darn darn fractured right now. So uh and and as a furthermore thing, um to me, it's like, I think intelligence will be more like a database than a co-worker. Like I don't think it's up to databases to add checks on whether or not they're used for like uh, what's something that's not great, you know, like CIA, actually I don't know what the CIA does really. You you can imagine, you can imagine killing people who are not even bad or whatever. And like, I don't think it's the database's responsibility for that. And furthermore, like a thing that has been weird to me is when people like sign up for our thing on Slack and they're like, hey, we're going to deploy this. Can we deploy this thing? I am just like, my brother, we are an API, you are a developer. It's none of my business, right? Like like you shouldn't know what the whole task even is because it should be decomposed into small things. We shouldn't be able to know what the downstream users are doing. And that is like a good boundary to give software engineers maximum power. Ideally they use it for the good stuff and ideally we can like help them and like we've talked about like doing open source and charity and all of that. We have absolutely no time for anything else right now. But like they will get any of that bias out of the technological layer as long as I'm in charge. Yeah, that's great. Uh while we're on the topic, let's also briefly talk about your privacy stuff, terms of ser- terms of use, which uh got a little bit of uh misunderstanding. I just want to clarify that up front. I think it probably takes two sentences from you about like you were not uh you're not being that restrictive about your API. Like clearly I ideologically, you're taking your role as a platform very seriously. Yes. Uh yes, uh uh I don't know what you're referring to, but like this was I've seen a couple of things about like benchmarking. Yes. Like, obviously we're not stopping people from do Oh man, I should be careful about what I say. I'm realizing. No, you said it you said it publicly that that was in the preview period, you didn't take it out for the launch. And now you're going to take it out. Okay. So the team is doing stuff that I'm not even aware of. So it's great to know the team communicated that. I asked them to check in with the lawyers about that. Like we are obviously not stopping people from doing that type of thing. I am extremely in favor. So I'm extremely anti-public benchmarks. Um, I'm extremely in favor I'm medium about private benchmarks that are proxies. Um, you know, So are you worried about uh saturation or like uh training on public benchmarks? So it's like easy to cheat. Um, not only is it easy to cheat, there's a lot of in so uh, I think that we are or or anyone who's like competition with us that you know, vaguely there is like you could say like the other. There's like 50 Jeff clones. Yeah. Well, sure, sure. Uh these let's say that there is competition or let's just say that there's there let's just assume that there's an industry two years from now of people who are doing similar things to us. The thing that we are selling is intelligence per something per like dollar or per second. Um, the no like people obsess about the cost and the speed. Um, I believe that that is it's cool, but like the thing that matters is the intelligence. Like the cost and the speed are like are bad things. You know, you're paying them for something and you need the thing back and the intelligence is what truly matters. The problem with intelligence is that there's a je ne sais quoi to it, right? Like like a the the good model smell. Like the thing that happened after we launched of like two hours later that actually went way bigger than the video, which was like, holy shit. This is actually usable. Well, it's you know, you know, like beyond that. You know, like the the the launch was crazy and people could really sense how hard we care about that. And that's truly what I think the long-term of this is. And I think public benchmarks are antithetical to this. Like they are a way to get people trust in intelligence because intelligence has a je ne sais quoi, but the public benchmarks are extremely, extremely gamable, even if they try not to, they still will. You know, like back in the old days, every lab had a team to collect data that looks like MMLU to make it look better, which is just benchmarking, bench-maxing with extra steps. So, um I I I believe that in the long run, it needs to be vibes and trust until you put it into a workflow and evaluate it for that workflow and measure it and have your own sense of like how it does. On the exact workflow that matters. And our job is to keep moving the nines of reliability. This is like an ever-present part of of what we need to be doing as a company. And we need to do everything to have people know that this is something we care so much about. You know, like if we wanted to, we could have released Jev like a year and a half ago if we wanted it to be dumb. Oh. like like like the you know, my bitterest lesson, right? Like architecture and yeah. I'll bring it up. since you since you talked about it, uh here. Hell yeah. Um, like, you know, Sutton says that algorithms beats compute very roughly. Um, data matters way more than compute, obviously, and doing the right task, having the North star is is is the hardest, most important thing. This has happened um in LLM land twice so far, right? Maybe 2.2 times. You know, there's RLHF, which like shifted the task to instruction following, no one realized that that was possible. Um RLVR did like a tiny little like edit to the to the direction and now us, right? RLCD. We have a new task and the goal is, you know, programs in the loop. And um, yeah, data matters so, so, so unbelievably much. Like I can't I can't emphasize it less. Yeah, you consider yourself a data lab rather than like a model lab. Is that something that's the word you guys use? Absolutely. Yeah. Uh, we are, we will always like care so much about the data. Um, to me, model capabilities means data. Data is so unbelievably complicated and that is what gets nines. Like, you have no idea how how much data can shift everything. Data is so important. Yeah. Holy crap. So, if people are looking for a job. Um, we are hiring infinite data people. Actually infinite. What is a good data person? Like, um, you know, clearly somebody who cares about reading through the the transcripts of uh whatever. You've said for example that you all your data is synthetic. Yep. But uh that's only like the scratching the surface, right? Like it's not like synthetic so what, right? Synthetic, but we have a people with a lot of taste and a lot of care looking at looking at these, articulating what's wrong, going back, regenerating. Is is that what a good data person is these days? Let me try to figure figure out how to like it's it's super complicated and like I literally onboard the data people with a talk that I assume is longer than this podcast will end up being. So I will try to say like the high level of it. So, number one, we don't do the kind of synthetic data that people ki- well, I'll do a actually number zero. Um, data and synthetic data depends on your task. Like the shape of your data, the shape of your task changes the data. Like RLVR's data is kind of environments, right? Um, RLHF's is the human feedback, you know. Um, each task has its own unique kind of data and we of course have our own unique kind of data, right? Um, so number one, we have that. Uh, number two, um, the thing I the reason why we don't want to train on our user's data even if we could, right? Like we could probably ask for any terms right now and it will we I don't know if it would would make a difference. We truly don't want that. Um, because no matter what, the real world data has so much bias. There's like a power law of like people like asking the same things where you'll end up like overfitting to it and like fracturing to it and all of that. And number two, we are like aiming for like a complete sci-fi future years from now where like these models are going to be like the general infrastructure, layers and layers and layers and deep down the stack to to like things people can't even imagine. Like I would like to think of our model like kind of like like, you know, UDP as LLMs and TCP as R models. All sorts of stuff can be built on top of that, and we need to be able to nail those futuristic use cases, such that software developers can actually build that futuristic stuff. And the way to do that is even if we had all of the data of the present, um we would just overfit to the present and then it wouldn't work. What we need is to like it it it almost feels like uh like there's artists. you know, they study this cognitive core. Our cognitive core is like way less jagged than anyone else's. And then they find the jaggednesses, and then they address them surgically in a way that and you can never perfectly do this, right? But they do it in such a way that it addresses it in every single possible like dimension passed present future. The general case rather than the specific case. Exactly. And like that requires a lot of intelligence every time. Um okay, so we mentioned a little bit you you you sort of criticized my thinking as are like very RLVR influence, which is like very fair. Um that let's actually mention RLCD. Um which obviously you have some secret sauce. To our knowledge, you've never actually published a paper or anything like that on it. No, right? No, not yet. But like what what should people get from this? Like what uh can you give people some confidence that you're just not just making up jargon for the sake of sounding cool, right? Like, uh one thing for me is like calibration. I do think is a to me like well understood because we've covered it in on the podcast. Yeah. But I I don't know what you mean when you say RLCD versus what people are familiar with. It's a great question. And actually I will give a related question. Okay. Um what is RLHF? Okay. Right? And actually RLHF means multiple different things, right? Like there's the RLHF of the original, I think it was like Paul Cristiano teaching a robot to backflip or something like that. wasn't there something? Was that it? That that was the original. I I reference the PPO paper, but I I don't know. Um and uh so uh PPO is not necessarily from human feedback if I recall. Okay. Um but I I believe it was like an open eye alignment work that could teach hard to specify outputs like a backflip. I'm not 100% sure. Um and then there was actually learning to summarize. You know, this was work um by a bunch of the the team that helped with um instruct and co-authored um the instruction following paper, which was teaching doing PPO on language models. This is the uh uh sorry. I'm trying to trying Yeah. trying to manipulate this this thing. this is 2017. Yeah, um I'm not 100% sure but like that looks quite right. If it has like a robot doing back flips or something like that, that might be it. Yes, okay, cool. I guess I I got it right. Hell yeah. Um There you go. Yeah. That's the one. So the idea was kind of like can you do like ill specified things with it? So that's like version one. Version two was the learning to summarize work um that uh like open AI did, which is actually like PPO on language models to do something somewhat ill-specified. This is like a another thing that people refer to as RHF, which I did not co-author. Oh, Dario's there, cool. Um Hell yeah. And Radford. Yeah, yeah, yeah. Uh shout outs to Alec and Ryan, love them. Um but uh the thing that I refer to RLHF is the um oh man. Uh I I'll get Do you have comments on that? I have comments on that paper, but like we're so many tangents deep. Yeah. Um so the thing that really got to to me the thing that I'm calling to RLHF is the task of instruction following. It's not about the PPO. That part doesn't matter. It's about like setting a north star of this is a valuable direction. It's kind of like the bitter lesson North star. And for us, RLCD is this new task. Um and it is not I I don't see it as jargon. Like I I try to communicate with precision. It's just that, hey, here's another North star. Just like DPO and all of its like um, you know, descendants also do RHF despite not using the algorithm in that paper. Mmm. And so clear clarity stating the North star is um being pro-program. programmable AI is is is one uh word that I I really caught catch on to uh removing the human in the loop uh from uh because RHF is tuning for this so that you can automate everything. Yes. Did I miss anything else in the in the thesis of like what the North Star is? There is that is that is right. I am overly nuanced in my communication. The one nuance is that we need to be practical. We need to be aware of what language models can do really well. You know, like what AI can do, right? Like, um, there could be programmatic types that are like sick AF, but if you if the technology is not ready for it, it to it's not a tragedy if that's not out in the world. Yeah. But to me, like the pre-jev world was a tragedy because it it sounds arrogant. No, no, no, I I strongly believe you. Cool. It sounds arrogant, but like I felt this way since long before I even had a company. And I can I can vouch that you've been saying this to me around for like three years. Yes, I've been talking about this for a very long. Yeah, I've been talking about this for so long and and I've been saying it because I thought it would have been easier. Um, they say they do not do things because they they're easy, they because they thought it was easy. Yeah, exactly. Something like that. Exactly, exactly, exactly. I I thought this whole project would take a week. Um, and I was unbelievably wrong. So, I am so sorry to everyone at Open AI that I thought I was like, man, I'm solving this right now. Um, but like I think that the tragic thing is when well, I think over promise under deliver is tragic too and like AI is super extreme on that axis and I think RVR is like the main, well, both RVR and RHF are extreme perpetrators of this. Um, but like it to me it's like it there's just so much potential there. Like, AI is clearly so smart. I smart, I love this in my talks. You know, when I ask people like, how can AI be so unbelievably smart? How can we like solve millennium prize problems in math, but still not automate even the most basics of works? Like like really basic rote stuff that like, you know, it it doesn't take like extremely smart people to do this. It's not a satisfying job. Like there's other things these people could be doing, but yet we need them to do like this basic like super basic, non- unsatisfying stuff because, you know, like we can't automate it yet. But we have this like supercharged engine of automation that just does not have like the right plugs and stuff to plug into all of this economically valuable work. And you know, like if the whole company of type safe disappears, like maybe it'll take like a year or two for people to like truly catch up. I actually don't know how long it'll take. If if model quality matters, then we were are going to be in a very good position for a long time. But um it it like it it it's done, right? Like there like this has changed the path of like technological history. Yeah. And like we will be exploring that space as a field. Yeah, I I think I I just definitely agree with that. Um, you've created possibilities. So I think, you know, if I can paraphrase so that people can also understand. Uh, you should not take the success of type safe and Jeff as just like, well, you know, that is a new new model type. Now we're done, we go back to business. Like, no. Like actually there's there are like five other model types that you should be exploring and like let a thousand flowers bloom. Absolutely. like some of which some of which you will probably also do. course yes. early internet energy. I think it's back to tech Utopia. You know, it's no longer like, oh man, like sometimes my coding agents work, but all of the best ones are hoarded internally. Yeah. Right? It's like creation is back on the menu. You know, it's going to be a wild ass world and you know, buckle up. I'm so so jazzed about that. Yeah. Uh I mean and now you have the the funding and the the momentum to do whatever you you you envision there. Uh which I which I think is like very gratifying to see you have after, you know, so long of uh of saying these things when they actually actually show the world. I know. I just such a such an interesting thing to be a tease. whole time. Like my talk like felt like it was a cliffhanger because I didn't say how the automation would occur. Yeah. Um Sean reviewed our manifesto and he's like, it's a little bit vague in these parts. and you know, like what's step one? What is what you know, what is the intelligence for? Well I asked you for model and you were like, yeah, model coming. Yeah, yeah, yeah. And and like uh well I I just I mainly objected to the word composable, but Bill Pner god is fantastic. Thank you. I I we we've really rallied around that. I'd like to think we're not entirely a cult like some companies are. But like we are like jazzed about what what are doing and like we are like why brand is being practical and like we are all like so super duper practical. It's really great. Yeah. So here and by the way, here here is the uh the step the the the secret master plan, right? The shape of machine native composable AI. It was your idea to make a secret master plan. It's a it's a Elon thing when he started Tesla, he was like, here's the whole thing. I'm giving official credit to you. Thank you. Thank you. Thank you. Thank you. Um but like uh you know, you you should have told me you're you're you're also going to do this model launch. Cuz you like you told me you told me half of the story. And then the other half, you didn't have the Doom demo at the time. You didn't have any numbers to give me. I was like. Well, the problem is I don't believe in bench maxing, right? So like it is a thing that you need to feel and like I think that this is the way to build long-term trust even though it like hurt it hurt us a lot, you know? Like like last year when we did fund raise, no one believed us. you know? Like like and they wanted just benchmarks and stuff and we're like, we're not going to do that. We are principled. We're going to stand by our guns. That rewards bad actors. I don't give a shit, you know, like what you want. Like this is who we are and we are standing by that. So, sorry. No, but in some ways, I think like choosing the hard path but you end up making the company that you want to work in. Yep. Right? Otherwise if you sell out, then you're just working in like open AI but with my people. Right? which is like Yeah. Yeah. I mean like I'm I I don't have too many regrets on that obviously. Like it worked out so unbelievably well. And you know, like I um I I was emotional last night when I was talking about like the reasons I left Open AI. Um and because like it actually had to change my wording after the launch. Uh my uh my my phrasing was if an AI Winter did happen and I did not do every fucking possible thing I could to like avert that, I would see myself as personally responsible, both for, you know, the the RLHF direction, which I think really widened over promise versus under deliver, and also not going all in on this because I think this is this is where value is going to just be like printed. So and it was really cool because I feel like the the AI winter I'm worrying about is averted. You know, like AI will be useful. It'll be used for automation. It's been less than a week and like the numbers are already undeniable that it's like being used for real work and like there's it's it's the wild West. Yeah. Yeah. Can you share just just if you have top of your head, what numbers are you seeing? Like what's what's like sign ups? Like what whatever you can share. I'm actually not super on top of everything. Like the team is the ones who are telling me all of these things. Yeah, and I'm sure it's like changing every day, right? It's it's it's it's kind of nuts. If if there's a milestone that you're like, well, yep, that's one thing we were hoping for. We reached it. I will say a milestone that we've passed is tokens per day. Um and this is not like fleeting tokens per day. This is like even at night, like it's constantly turning so you know machines are calling it and not just people trying things out. So that is that is so cool. tokens a day is a lot. Um so surpassing that is awesome. Sign ups to me don't really matter. And actually this was like a bit of a mistake we made if I'm like totally honest. People on Twitter were calling us like marketing geniuses and all of that. And um that was just us. We don't have a marketer, also hiring. Um, and we were just being our genuine goofy, like irreverent selves. And we were we were just like offboarding people off the wait list so hard. Um, our platform team is so unbelievably cracked. I think we have more of nines of up time than anthropic while having the most unprecedented launch ever. Like that is kind of nuts. So like props to them. Yeah. Um, and uh the thing we didn't realize so number one, wait lists. Wait list signups don't matter for like a developer platform in my opinion. You know, uh I would guess that a large number of them are not even developers. So they go in, they try some queries and a lot of people don't get it because they are not programming, right? Like they're just like, what this is not a chatbot. Where where's my chat GPT2? right? But if like I I haven't exactly calculated this. My sense is that if every single human being in the world like just wrote a couple of queries, that would be a rounding error compared to like one power user's for loop that is just like creating value. And the thing we didn't realize with the wait list is like we could just off offboard anyone off the wait list. It doesn't matter. The scary part is rate limits. And then once people start getting value from that, uh then they would just want tons and tons of rate limits because this is what software is, right? Like you spend effort upfront to specify your rote task, and then this wrote task creates more value than it takes to put in. and then now that you have it. set it and forget it, yeah. Exactly. Yeah, you run it in the background, you make it a dependency to like other things so you can make like higher level stuff. And uh like you just create so much value in the world. You know, early internet people probably did not imagine like the wonder of early 2000s internet, which is still not early internet. But like it's it's through, no offense, composability, um that all of the crazy stuff happens. I just really wanted to emphasize that in our manifesto. We are going for emergence. We are going for like being the capitalist. We're wanting to empower people, and we are going to do whatever we can for that, be it like discords and our town hall with me wearing a garbage bag or not. Yeah. Um, and uh, and podcasts and and you know, getting getting out there. That's it. I like I because I want the long form, right? It is like, yes, we'll get past the the some of the superficial things and then we'll go deep and and people will really trust and understand your mission and like, you know, the the people that uh will resonate that will end up joining you or or, you know, uh uh buying you. uh No, no, sorry. as as a as a customer, as a customer as a customer. Oh okay okay. yeah yeah yeah yeah, that was funny. sorry. Sorry, I didn't I didn't mean to say that. but no, any one one one version, one very flattering version of this. like 36 million views of your launch video. Cool. Up to 38 now. Uh yeah, you know, rounding error. Uh you know, uh Na still got 74, Fable 5 got 57. So like as far as uh I I I didn't I didn't do the stats for like original ChaT like which there was no video. Yep. Yep. So like up there, right? As as far as like if you were to launch a new lab in 2026, I think you're like number one right now. which is like pretty crazy. Yeah. Well, I actually would rather I do have the shirt like your favorites Nelab, favorite Nelab's favorite Nelab. Um, I don't give a shit about being a Neo lab. I think being a Neo lab, actually we have a lot of like swag that's being a parody of a Neo lab. One of them one of them I have is like Neo lab with product, which actually is not a Neo lab. Like, I don't care about that really. Um what I care about is being a reliable dev platform. So, uh, appreciate the comparison, but like hopefully we transcend past them and we go back into like a thing, you know, like, you know, a revolutionary moment for developers and like this stable thing that people can rely on and trust. Yes. Yes. Uh, I mean to that end, um I mean I I think that's one thing that it really impressed me about you guys is is that yes, you you you do talk about reliability. I thought it was mostly about calibration, which like we you know, we talked about RCD. Uh but actually it's also about just like up time and and uh scalability and all those things, right? They they're all sort of of a kind. nines. nines. It's like like Which is not time means my mind. but that's part of it. But like there's reliability in um like how intelligent the thing is. Like how consistently does it do the thing that you want. And I think that like the the big reasoning models are very smart. In my opinion, they still lack reliability. I think there's many use cases where you they look like they should be smart enough to automate their work. There is economic incentive to automate that work. Yet still they're not reliable enough as at an intern because they're optimized for different things. And so like I think that there's the reliability of being able to like trust the outputs. And also we are like like there are dimensions of reliability that we are not yet at that I'm like so excited by. You know, like I want to automate the easy work before the hard work. You know, like I think that that's just a common sense thing to do. Um, but to me, we will be sufficient, I don't know if there's such a thing as sufficiently reliable, but I want to get so good that people don't even need to try the model to know that it'll work. It's like that's like what flow state is in programming, right? Like I'm just like writing queries because I need intelligence in here. And you know, like when for non-trivial branching, I could just write it in in like a like like a type-safe system one query and then get the results out and I just branches accurately. Like that would be so, so good. Like that's the that is the dream. And that is like going to be like a long, long slog. Yeah. We're going to go into your API design in a little bit, just just to give people examples and like maybe path not taken, that kind of stuff. Um, one thing up the front that I do wonder about in terms of reliability is I notice that there's no seed. Uh, there's no, uh, and and so basically, same input, do I always get the same output? Mm. And if not, why not? Oh, great question. So, this is actually like a common question we have between um, so reliability is actually a catch all. Like whenever AI can't automate something, it's due to some form of reliability. It could be like type safety, it could be determinism, it just could be like it's uh it's jagged, right? So, um reliability is a catch all. I just think that it's also a catch all for like what the north star is. Um, determinism is like same inputs, same outputs. I do believe that this is like slightly interesting for unit tests, but I believe that to be the wrong north star. I I believe robustness is what people I I don't want to tell people what they really want because that would be a little arrogant of me. I believe that that is like the more important property. Um, uh, you want um, given similar inputs, get similar outputs. And it's kind of wild how unreliable LMs are. Like a way that we test this is you put like UUIDs in like little, I think they're called nonces in in the prompt and what you want to similar outputs from all of those because it's truly semantically the same question. And that is the part where you really want like like that robustness is where like people get like burnt with AI making decisions. So, I think that is the a super duper important property. We could also have determinism. That is that is a thing that can be available. As far as I can like mentally model for programmers, like it it it could be valuable for some use cases, so like please educate me um in comments or you, but um my in in general, it's easy. The determinism is something you can like trade off for better cost. You know, like we are we are constantly wanting to be on the intelligence per dollar frontier. We are doing like absolutely disgusting things to be there. You know, like this is um, I I I shouldn't say this, but no one's here to stop me. Um, you know. You have you you sign off on your own PR. That is not how it works at this company. Um, I believe for this week, um my chief of staff K is the most powerful person in tech. And shout out to K for organizing this. Holy shit. Yeah. Holy shit, she is so fucking competent and powerful. She's incredible. Um, uh, I mean, she sucks. Don't poach her. Um um, but, um, so I try to be a bit more filtered, but like, people are telling me, don't call it a Frankenstein's monster of models, but because that has like negative implications. I think Frankenstein's monster was like the good guy in this whole, I mean it was innocent, right? I didn't read it, okay. Uh I'll confess. Okay, that that one facial expression. The's a decent Jacob Elordi movie if you want to see. I the the adaptation, anyway. You have no idea how little time I have right now. My priorities are sleep, you know. Developers developers developers. Developers, yes. developers, developers, developers. Um, but, um, yes, we we do like absolutely disgusting things to be on the Pareto curve of intelligence per dollar, and we are going to keep doing that. We're going to be doing crazy ass stuff. and I think people really need to think outside of the box. Like, like part of the reason why it's surprising is like people are taught inside the box and we continue to do that. Um, as of right now, we are obviously the best at this, and we want to continue being the best at that whole thing. Yeah. So, um, wait, where did where did we tangent from? so so I asked you about uh will you have seeds and determinism? And then you basically defined reliability and like from how you see it. Yes. But like determine like I have a robustness example that's that's a real quick I can show you. I would love that. I'll just say one thing. We can make it a deterministic model. Like like we're hap if people can convince us that that is a valuable thing to do and we don't have a gigantic GPU shortage. Um, we can happily make all of these models. We live to please. Um, and and revolt, revolute uh uh you will throw over everything except you do it in a nice way. Yeah. so so like determinism could be on the cards. It's just gets you less intelligence per dollar. Yeah. Uh well just having seen the trajectory of Open, uh you will I'll just trust me now that you'll be peer pressured into doing it. Uh so like just people will want it even if they if you tell them they don't need it. They'll still want it. So like, yeah, that that's the TLDR of. Okay, okay. I will love to maybe one day we will see how that happens. I've been told I'm um, they they they say that part of our brand is being unshakable and they say that that's just the nice way of saying stubborn. Stubborn, yeah. Yeah, exactly. and I'm a very stubborn person. I don't think we could have done. No, but so like I I okay, but I I like have argued with you before. Yeah. and And you've been right about developers every time. So okay, I give up. you win. You win. I'm sold on I've argued with you before. No, no, I'm just saying like I I think that you can hold your ground while also like if I give you the right evidence, you can uh not you can sort of like throw away your priors and be like, yep, like that actually makes sense to me. And so like, you know, just just trust your own gut on this. I'll bring up some. I suspect though that we will be GPU constrained for a very, very long time. And anything that has less intelligence per dollar means it consumes more GPUs for the same intelligence, which is, you know, like our goal is not to onboard companies, like it's valuable, but like our goal is to have people like experiment and do weird shit. And we need like, we need to like get it to as many hands as possible and like starting like the California gold rush for that. I think there is right now. Yeah. Um, just a word of caution. I mean, I'll just say it because somebody's thinking about it right now, which is when you say things like, uh, we will not commit to deterministic models, we we will we'll do whatever it takes for intelligence per dollar and we are we are facing GPU constraint. People are thinking you may quantize your models, right? Like like whatever you had at launch, you may quantize down in to reduce the quality, um, in order to to free up uh memory or bandwidth or whatever, right? And so, uh, you should probably uh have some kind of promise, which you don't have to make now, about like we will uphold model quality at launch. People, like, so it's like, when people, when when we, I mean, you were at Open AI when you launched all these all these APIs, and even Claud as well, like, uh, when they first launched the model, the model strings, um, did not stay the same model at all times. Yep, right? You have versioning in your models, that's great. Yep, yep, yep. But like you should you should publicly commit to some kind of like, once a thing is launched, we don't change it. We will not change our models when we deploy them. That is insane. We care about developers. like like it makes sense if you're so, doing something like that, again, this is the problem with a first party product and an API. It makes, you can do whatever you want in a first party product, right? Like like more power to them, whatever gets that experience, that is fine. With an API, you obviously can't do that. But I will say that we uh plan to move a lot faster than many people are used to model providers doing things. So we will be launching new models a lot faster than people think. And we are not promising long-term support for the models because we think that there's lots of improvements to have. So, um, there is a world that we might temporarily LTS, what is right now of 1.13.0. We might do that because so many people are using it and I know developers hate breaking dependencies. The alternative is fracturing our fleet and that is a very bad vibe for everyone. Yeah, you can't have like 100 different versions of the model. Exactly. And if we're iterating very fast, there would be a lot of those versions as well. So, um we we do want to have not just a LTS supported thing eventually, long-term support. Um, we want a really sick way of doing that. We have like research stuff cooking in that direction and I think it's going to be the most pro developer thing ever. Um, but it is not yet our current models and I'm not promising that we will be able to keep the exact same models. They will get smarter every time for sure. And my sense is that even our model iterations where it already is smart, um, it between model versions, the the changes tend to be even smaller than the string models calling them twice. But when when we go from like, you know, jagged to like wow, um that is where the big Deltas are. Yeah. Uh, one one thing, one thing that's beautiful about LTS seeing models is that actually you can also port them to other Silicon. Uh, I don't know if you've thought about this. No comment. Okay. I care about intelligence per dollar. Yes, right? Um, and speed. What? Speed as well. We'll see. Yeah. We'll see. This is a I mean this is a whole part of the inference tech tree that is like I mean exploding in the past year, right? Like that you can you can move to like a Cerebras, an etched or whatever and get like the 100 100,000 times speed up. Yeah, like I I think that intelligence per second is like a different metric and we've even talked about like things like intelligence per dollar times second and like metrics like this. My guess on like Jevin's paradox occurring or at least the Jev series of models. And the thing I like hunt people down about internally is like I don't care how much smarter it is, it needs to be in the pre frontier. So like that is what the brand of Jev is. It is the best thing at intelligence per dollar. For intelligence per second, we'll see. I I think that it's an intriguing thing. I know that there's many industries that are like extremely dependent on real-time stuff and they will like like intelligence per second means tons of dollars for them. But we'll see. I I I I I would love to like do both and like have the market correct me either which way, you know? Like like uh I I I I would love to be informed by people. Yeah, totally. Uh and it's not it's not just real about real time, right? It's also about scale because um at scale every microsecond is just multiplied by billions and trillions of times. It depends on how background it's running, right? Like if it's like a big background like database map produced query, the latency might not matter so much as like the cost to get intelligence from it. But like if it's actually is um something more real-time like user facing, you have budgets like between 100 milliseconds and 1 millisecond that are like totally magical. And actually, even if you were below 100 milliseconds, if you could half that time, that means you can get double the intelligence or sequential intelligence calls to have like a like a phenomenal experience. So, uh, right that is definitely happening right now. It is super duper cool. I love the intelligence per second use cases, but I don't think that that will be Jev's niche. Okay. Yeah, fair enough. Yeah, good. When you think about the promise of faster and cheaper, um typically, the other uh the the trade-offs that other models are offering is faster but more expensive. Yep. Right? And so you're like one of the reasons I was thinking about why is Jeff resonating so much is that you've done the faster but cheaper side of the quadrant. Yeah. Which is very very unoccupied. while while holding intelligence like somewhat constant. Uh, yes. It would I that that's a very load bearing statement while holding intelligence constant. That's the hard part, right? Like which unfortunately like so basically you you refuse to have to to like like do any public benchmarks or you don't like any public benchmarks about it. I've I've actually talked But you need some internal sense. Say it again. You need some internal sense of this Oh, of course. Of course. We have we have our own internal Evals for sure. Yeah. But it takes a lot of discipline not to game those and it needs to be like a top-level priority to not game them. Yeah. Of course we do that, right? Like how else can we make the guarantee that our models are in the preto frontier of intelligence per dollar, right? Like we're not flying blind in there, right? If we're doing like completely weird things with different costs or whatever else, you know, like how do we compare them? We plot them and get you we try to figure out like what is the best for the users. So, we for sure measure them. I'm not anti-measuring. Um but it's it's it's extremely dangerous when you have like any alternative incentive and this is the one thing that I I kind of rule with an iron, well, maybe my co-workers might think I rule many things with an iron fist, but to me, like not shitting ourselves um about how smart our model is is one of the most important things there. Like we need to be truth seeking. Yeah. Yeah. agree, agreed. Um okay, I wanted to go over some details on the uh API choices. mostly because this is the only podcast that will ask you these these kind of questions. Hell yeah. Hell yeah. Um, so you have three primitives, um choice, score and nul. First of all, nul, where's that from? Is that is it just like a term in the in the literature or what? Uh now now it is. Um we debated this a lot. We debated this a lot. Um it is, you know, it is bullish, right? Like true false. Um it is But it's continuous. Yes, exactly. Uh so so first the origin of the name is Bernoulli. Ah. Yes. So it's that's why it's even spelled that weird way that is like a subset of the name Bernoulli from like a Bernoulli probability, right? Which is actually what that is. Um, so that is the origin of it. We were debating this a lot. We liked P Bool, we liked pool. We were we were wanting to call it like a pool party, but then no one let me. Um, you know, we had like a bunch of like other arguments about that. And Nuel we figured was like the best thing. Our our our rationale, and like this was actually the same thing with Jev, too, um, is that we think that we are like an irreverent, insane bunch and programmers don't care. You know, like if Jev is just going to be a string, we didn't expect it to catch on or even have puns or anything like that, right? Actually, there was a lot of hate on the name internally. They've all apologized except for one person. Um, Still holding strong. Yes. Uh, our mutual friend. Okay, okay. Uh yes, yes, yes, yes. Um, I respect her for that. Yeah. yeah. Uh she wanted Jev to be called meow. She would, of course. Yes, of course. My co-founder, yeah. Okay, you win there. you win there. But but like um yeah, Nuel is we we had to make a new concept for this thing because if it was a bool, it would be confusing to people. So, actually, all three of these are actually new concepts. These are not types that exist in programming, and that was intentional because they map very closely to types, but they're not quite that. A score is not an int. So if you had like instructor or pedantic or whatever, map ints or floats into scores, you'd get a little bit cooked. You know, and and like we we were really airing on the side of clarity over the side of like making people like easily understand what's going on. I mean, don't you worry about that? Don't you want things to integrate directly into things that people are already using? Yes. Yes, we do. And actually, I think that, you know, You have integrations with like other SDKs and stuff? Yeah. But you have sorry, you have your own SDKs? Yep. But typically, for example, as a developer relations person, I would be very obsessed with like, yes, here is how you use, uh, you know, Jev with instructor. Here's how, you know, that kind of stuff. We might have that somewhere. I am so behind on everything. Someone will do it for you in a community. Now that now that you're successful, people will be like, oh, that's cool. Cool. But like, you know. I don't see that as binary either. I actually see success as a score, and there's always more to climb in like how much we can like be there for our community. Just just to be clear. And um I this this section is stressful because I didn't review the docs and they're like constantly changing. But to me scores do exist. So so so scores are similar to like LLM judging, right? So like if you want to call it like a judgment, I guess you could. But like that that is like the the the way people already use this type of thing, right? Like maybe a ne could be like a probability, but everything for us is a probability. And a choice is actually closest to a function call, but a function call is like an extremely disgusting thing that um, if you want to open AI juice sauce tea, that we should go back into that later. Like if like like a choice is just like the right way of of exposing like a switch match statement within. So so it like maps cleanly to an Enum. Yep. And you can choose to hydrate it into a function if you want. Yes. and like an Enum choice is the important part of that. And uh like actually I think these map all into like programming primitives, where like choice maps into like a like a switch statement on an Enum. Um Nels mapped to if statements, and scores map to sorting or thresholding at a greater than or less than. Okay. And this has been always what the vision is. Like there will be more types, and they will map into programming primitives. Yeah. Um, any other, so, uh, any nuance you want to go through? literally this is for the Jev people who are like deciding to really invest in Jev. You are the expert, right? Um, I'm just like wanting to provide more background for them on uh API choices, um, you know, how they should use some of these these things like legends, confidence, um how critical in your testing, um, you know, like how how like just any sort of like pro tips that you'd like to want to offer people when they're down at this level. Thank you. I love this. Um, no, no, no. Hell yeah. I didn't expect this and actually no one has asked me this, um, in probably like months when I was like onboarding like our DevRel. Okay. Um, so sick. Um, so, our model is designed for being like deep in the insides of computer programs in the future. We like unironically believe that this will be much more massive than anything people are even considering today. And our model might not be ready for that, but we are like continuously working for that future. It will never be good enough at these shallow tasks. Um, sorry. It'll never be like we're not just going to keep on climbing the shallow tasks. We want to be deep in the guts of programs, because that's how you make software powerful. All this all the types inside of our um, this is an actually an output. Um but all the um all the parts uh of like the the the input, like the state, the instructions, the criteria, all of them can be structured JSON objects. That way like programs can like insert them in the right spot and you don't need to like put things into templates. Exactly. So, if ever, I I think people don't read into this part enough and they think it's all strings. And that's that that's fine. Um, but these are all meant like I would say that if you're using like a template, like turning it into like a system message or something, you are thinking in like the old way. You know, we should be making things as easy for computers to understand because that structure is truly there, right? Like it would be weird in like like a programming language to have like all of your numbers in and then you pass it into like you turn it into a string. Normally you do that for printing, when when you have a human in the loop, right? But for like within the computer, you want to be passing like nested structure that is semantic all around. And we are really going to be optimizing our model. The model's pretty optimized for this, but the thing is every different nested level of structure is harder to reason about and we want we are really cooking hard in that direction. I think people should keep cooking in that direction because it makes the code like so much more legible and beautiful and like agnostic to like the the implementation details. It's like, here is my state, you know, like here's my function state. Like think of it think of it as like an AI function, which subsets of my state, which is like all the variables you have available, should I pass in here? System messages are like disgusting global variables where you just put everything in there and you put all the instructions at once. And then you you know, like you you hope that every single instruction gets nailed instead of asking the questions in parallel. And also, I would recommend I and I truly say this not from like a like it makes me money perspective. Um I truly recommend asking lots and lots of questions. Break them down, make them smaller and like really decompose. Like no matter if the models can do it today or not, I believe that the biggest like um saving grace of like what's happening this week will be people's codebases, AI code bases are going to be so much better. You know, like if you decompose problems into simple decisions, every single one of these things is extremely evaluable. Like like a AI beforehand is big system message and then maybe you have like another big AI to see like if it actually does this. That's nuts. You know, it's kind of crazy. Like it it it it it that was our Stockholm syndrome, right? But like that's kind of crazy. Like if you want to say like, hey, don't read this subdirectory or don't pass any API keys to deep seek or whatever else. Like that should be programmatically basically guaranteed. And you know, you'll never have guarantees of any machine learning model. But like by breaking it down, you can actually you can actually measure it. Yeah, you can verify that it was actually called. Yes, and like our model our model like the the interface itself is so verifiable. This should be like a sigh of relief. You know, like it's it's uh it it's just going to lead to way better engineering. Yep. I think I get that. Um, and and so, you know, one of the reasons people didn't used to do this in the past was because they would just call a a small LLM, right? And it's still too slow, it's still too expensive versus chunking everything. I've done exactly this myself. Right? Like I I benchmark, here's a pipeline that throws everything in system prompts and then just gets one big output versus break it down into 100 different things. It was slower, more expensive, not as good. Yep, yep, yep. And that happens yeah, it's and it's like super inconvenient, it's unwieldy. Why not just put it all together. You kind of end up repeating some stuff between questions. So it's like maybe like, you know, inefficient or something like that. But then it results in something that is very hard to rely on. And software doesn't need to run in the background. It would break my heart if our stuff couldn't run in the background. Is there a way to break things down that you guys have found that works versus uh what you thought worked and doesn't work. Interesting. Because like people are just going to be exploring this, you know, now that you've said it. Like, they would use this as a reference and be like, okay, like that's how I'm supposed to use Jeff. Yep. Um then the question is how do you break things down? Interesting. I I like to break things down into its like its smallest semantic unit. Like, what is the lowest level thing? I try to never have I've probably queryied uh the the model the most among anyone. And like I try to, number one, in my in my queries, this is this is a lot more like the way I prompt things. Like I make it really, really structured and explicit. And in the questions, I always, I like the backticks, but like it works for all of them. you know, Uh like be really clear what I'm referring to. because we wanted the model to be really literal, because when you program, you want things that instruction follow really, really well. That is what the art of programming is and what AI does is expanding the things, the kinds of instructions that can be followed. So, I'm a fan of doing that. I I sometimes I'm a little lazy and I like I have like more like hybrid things, but like I think that for like really big production things, you just want to like keep on adding more questions and you want to make it really easy to add more questions. You know, be really, really precise about all of that breakdown and then have the code to have the exact behavior you want. If I could give like a tiny little example of this, um, is um like refusals, right? Like uh I'm not going to talk about why we don't refuse. I might have done that already. It's like all a blur. Um, but uh, like for refusals, I don't think you should ask, should I refuse here? You know, uh that's a really I think the answer will be pretty good because like that's a system one compatible task. But I think you're way better off like asking many different independent questions about like the different situations you can refuse about because instead of having to like just guess based on you, you know, you can actually specify what you want. And beautifully, you know, and I think this is like truly, really beautiful. If you find a situation where it's like, oh, it didn't refuse because of this reason, I didn't specify this part of the task. That is awesome. That's what software engineering is about. Like, you fix the bug by adding that question in, adding the threshold, maybe remembering that as a test case, and now it is just solved forever. Like your software can't forget about that like in the prompt because of context fraud. It is just there and you can like just keep measuring that forever. You know, and if the models are not perfect at some of these things, you can choose what threshold you want for all of these factors based on real examples. It's like it's like ML without the ML and you can just do it for for anything. And like there might be some things the model is not good enough yet, right? Like I would, I'm a little bit afraid when I see people doing trading with the models, like automated trading. Um it it looks cool. I I I just think that people should leave it to the professionals. Um, but and like that's just a very hard high-level task that maybe the models aren't good enough yet to figure out. Uh well like even if they were, then they would it would suddenly wouldn't be because of efficient market. But like that's one of those things where um you can like break it down into things and just evaluate them and you might be like, it's not smart enough at this, maybe we don't deploy it yet for this version. Or we make a trade-off, or we air on the side of safety, or like, hey, the models are not good enough at, you know, like detecting like this weird combination of like sarcasm with a VIP customer, that this is when we escalate to a human. And that's what confidence estimates are about too. Okay. Uh great great answer. I think uh one thing I'll I'll mention very quickly which uh I don't expect that you have a too long of an answer for is uh well no no no it's just it's just typically like you are still relying on thresholding as like the the lever that the user can pull. But what if just the the the the calibration is wrong? Right? Like you're just saying your your calibration is perfect. But I didn't say that. I didn't say. Like so perfect calibration like good calibration means like lower value is lower like you know, probability lower, higher value is probability higher. But it could be wrong. It could be locally misaligned. And so then I would want to fine-tune it or something, right? Which you don't offer, but you could. I I I I I again, this is a short answer. which is you don't have it right now. Yeah. Oh, do we want to offer fine tuning? Is it the question? That's could be that could be one version of it or you could have a different knob. Yeah. Right? Uh where like because like right now you're all you're saying is like if something's wrong, uh skill issue, you should you should just change the prompt again, or break it down even further, or you change the confidence. Those are my two options. Right? And that doesn't feel super satisfying if your model is just getting it wrong. Yep. And it will, it will get many things wrong, to be clear. We have like a report issues button, complain to us in Discord. We want to make it a lot better. Every single model version will be like noticeably better. We will stop shipping them quickly if they weren't getting big improvements. So, number one, that is like totally reasonable. I think that that's simply pragmatic to admit that AI is imperfect at some stuff. Right? I do think we'll find use cases that they are like good enough at, and good enough kind of depends on the use case, right? Like, human beings can do a lot of work despite being bad at that work, because their EV is quite high. And presumably with the right thresholding and everything, there probably is like large amounts of work that could be done, even if mistakes are being made. Um, on the question of fine tuning, I could imagine it, I could imagine it in the cards. I do have concerns because like uh in the what people need versus what people want category, um, like I think general models tend to be really like like again, there there's the je ne sais quoi of generality, that making it good at like a million other tasks than this one narrow task, might make it better at edge cases in that task, which I'm I would be a little bit afraid of, you know? Yeah. Um, I I I could imagine it is is my answer. Uh I'm endlessly practical on these things. I want everything, like my vision of the world is I there's there's so much we want to be building. But also, like I would not want to ship something that is like a a giant foot gun, uh like some other AI companies, which yeah. Well, you know, so both OpenAI and Claud and and I think even Gemini have rolled out fine tuning and then took it back. Yep. Uh which is an interesting observation that pretty much fine tuning is now in the domain of open source models. Uh. Yes. Yes. Um, I do know about that and like it was kind of crap. So like that's probably better that they took it down. Yeah, it could just be a foot gun. And they're telling people that fine tuning it is probably the wrong way to go is great. Another interesting answer could be that like, well, our model is so different, like, you know, in the same way that quantization doesn't apply to us, output tokens don't apply to us, fine tuning also doesn't apply to us. Well, actually, I I'm I'm super open to that possibility. Like, uh my this is not a promise, this is a desire. Just so to to make it clear, I like to be really honest. Like, I think that um as intelligence per dollar gets cheaper, cheaper, cheaper, cheaper, I think that we could get really like small approximate things that hopefully are proxies for intelligence. Like, is there a world where people don't write regexes anymore? Because like, you know, the intelligence per dollar that uses AI is cheaper than like the complexity of a regex. You know, that would be kind of sick. I would love that, you know? And it might require fine tuning for some of those narrow use cases to really get past the threshold. We will see. My hope is calibration gets that. Calibration plus a cascade of models, like if it's super confident, then maybe it's right. And and if it's in the middle, then you do the next bigger model and you chain off from there. I don't really know how that's going to go. But um yeah, I I I could imagine it. And something that I could imagine too is like imagine you have like a a series of like we own the entire Pareto frontier, something that a business might want to do or I I think a hacker would be okay with dealing with a Pareto frontier of models. Maybe a business wants something more dynamic. You could imagine like having like a different sizes of models and to dynamically pick which model based on how smart it is on different parts of your stack. And you could even imagine because of how simple our thing is, you could imagine like some automatic fine tuning on that. Yeah. Uh not a promise in the slightest. I'm just like cooking on sci-fi. But you would consider different sizes of Jeff models. So to offer that gradient. Absolutely. Yeah, yeah, yeah. Like we we like how would I know how much intelligence people need? I don't know. right? yeah I don't know either. The demand is demand is uh unlimited. Well, yeah, people are telling us not to ship things right now because we don't need to ship things because again. It's good enough. Yeah. Yeah, but that's kind of lame. and uh I really like the saying, this is something that I hope people hold me to, because it'll be hard to to to walk back from. Yeah. Like, I I don't know if this is exactly the thing, but culture is what you do when the market doesn't reward it. And I really like that, because I think that we are standing for something. Maybe in the future, what we're standing for is like so obvious that we're the equivalent of like boring, like visa or something like that, and like we're just like a utility that no one really thinks about, and I'll be wearing non-pink suits or whatever else. But I really want to be like rallying the world to this, you know? Like I want to keep doing cool stuff, not because we need to, but because I want like people to realize that this is just the beginning. You know, like that wasn't even meant to be the opening salvo. That was like kind of like a you know, low key research preview or whatever you want to call it. Yeah. And there there's a lot more we can do. Yeah. We've like machine native intelligence is going to go wild. So, not the only potentially not the only size, potentially not the only model that you guys launch. Uh, you know, you you want to open people's mind. Absolutely not for any of those. I want I want to like meet whatever needs we can. Yeah. right? Uh like at but with like a giant caveat, I don't want to be like open AI's product teams that like like throw stuff at the walls. Like, I want it to be like in a under a unified vision. Like if you go back to the manifesto, like everything needs to be under one of these three three things in my opinion. Um Uh I'm not I'm not prepared to do this uh Oh, I'm sorry. I'm sorry for asking. I can just just talk about it. Like, we have like three steps in our stuff. It sounds like a tease. I want everything to go under one of these three things to keep pushing the boundaries and everything. Like like this is not, these are not like checklists. These are like axes that we think build like the foundation of, you know, of of like a new technological revolution. And I want all of the all the bets we make to be somewhere in there and we will be doing some weird weird stuff model wise. So, um, because machine native, right? Like humans don't need to totally get it. It needs to just be valuable. Uh, you know, with uh just you see keep people tease or hint. Like what what does weird look like? What is weird? I I'll give people a hint. Yeah. Um, some people are trying to call them decision models. Okay. Um, our primitives are decisions. Um I I wouldn't do that, uh because I think there's other types that are machine native that are not decisions. Okay, we'll leave it at that and let people guess. Yeah. I think I think it's a pretty fun hint. Yeah, yeah. There's people, look, there's there's people saying like, I've done this before. I made a decision model a year ago. Like Jeff is not new, Jeff's not cool. But like I think uh there's the categorical like here's what you're establishing is possible. There's the uh performance of like, well, actually the for the benchmarks and the numbers that you're getting, you are still beating as far as I can tell, you're still beating every single clone of you out there. I I don't care about the benchmarks, just to be clear. So like like even if we were winning or losing, I want to denounce them. You've established the category, right? Yep. Yep, yep, yep. Uh but but also I think this this nuance between decision models and system one, uh I think is actually the thing that you're trying to do. And I just want to make software engineers superpowered, right? Uh like with AI. like and or like the tragic thing to me is, you know, in that AI winter direction, I think like it's it's it it's just so sad that AI was so powerful, yet so underutilized. Like, um, um, it's the thing that gets me emotional about it. but um, man, Like I think that that is I don't want to like just be like pure techno optimist, like all technology is good. I think what was happening now was like a travesty. Like it's and like there's, you know, I just want to like open up those possibilities for people. Yeah, I'll just end it there. I I I I've I've cried too much these last few days to to to want to do it on the record. Yeah, yeah. No, I appreciate you sharing a little bit of that and I I think people can see that you're very authentic and passionate about this. Um, you know, you don't necessarily get that from a name like Type Safe AI. But like uh I think once people immerse themselves themselves enough in like here's the genuinely different direction you want the world to go. And like actually you have done like the the hard part about going from zero to one on the on the thing. Then like like now let's all go to go together in like the the new direction, right? Yeah. Yeah, yeah, yeah. I I don't I am sure that I won't think, maybe I will think that the hard part was done, perhaps. I think that there's going to be many more hard parts. Um, like if, you know, all sorts of stuff gets automated and we finally see GDP growth and like, you know, it's like, you know, a jazz party every day, then maybe the hard part is done, but like I I don't think so. And like I really, really think that people focus too much on speed and cost and not enough on reliability. Like reliability is what makes it delightful. Like reliability is what like allows you to trust it. You have this line, uh, TFP growth reading 3% in 5 years. Hell yeah. I've never seen that. Hell yeah, let's fucking go. I've never seen a lab care about TF TFP growth. But like that is what an economic revolution is, right? Like it's actually extremely consistent with what the Open AI charter used to stand for. You know, it was talking about like I think the charter is the same, but they've kind of tried to move definitions around to like, you know, 100 billion in profit or something like that. Not that I hate an Open AI. It wasn't like, yeah, it wasn't a well-defined term what AGI is, right? They tried to do it, right? Like doing majority of the world's economically valuable work, and they should have to answer the question, how can it do millennium prize problems in math? And zero of the world's economically valuable work rounding error. You know, like I think that all models are roughly tied right now at zero. There's some chance that like we have started already, but like I would guess that it's not yet 1%. Um, and I think that that will show up in like when it does happen, it will show up in the economic statistics. It's going to be fucking awesome. It will not cause mass unemployment, but it will cause like a whole bunch of awesome shifts and we're the world will be a lot better. And also, like, I'm really tired of AI always being the foreground character, uh, of things. Like I think that the world should just be more delightful and AI should just help with that. You know, and It's just like disappear into the background. Exactly. You know, like I I I say this in my talks. Like, how can it be that 2019 software, like software, SAS, whatever, super duper valuable, right? It's 2026 now. How is the software basically exactly the same despite AI being so freaking awesome, other than sometimes having a chatbot on the side, right? That that like that kind of works but doesn't allow you to make decisions that the companies have stakes in because they can't be trusted to make decisions. That to me is nuts. You know, there's so much economic incentive for this and I think it's going to be like a like an inverse SAS apocalypse. I think SAS is going to be supercharged by this. They are the ones who are like most in the know of what things are valuable to automate and it's going to be like a crazy time. Yeah. I I think I think so too. It's a it's a beautiful thing that you've unlocked, you know? Yeah. Um, you mentioned one thing here, which I don't know if it's a like directly here where it says, what is a system one problem and what is not what is a system two problem? Like, you know, Fuck. That's a hard one. That's a hard one, my friend. Um. Because people now are just trying to jab everything, right? Which like probably is going to fail, right? Um, but like some things are going to be good. Jev, everything is a pretty funny. It's a pretty funny way of doing it, saying it. Um, the so I'll tell you the truth. Yeah. Um, the truth is that this is an empirical problem, just like scaling laws are are an empirical thing. You know, like why doesn't like robotics really work right now despite all the money being spent on it. I don't think it's about like spending more money necessarily. The the empirical results just might not be there, right? Um, so empirically, I believe that these like pre-trained super condensations of intelligence are fundamentally system one thinkers. I think that they truly like system one is the closest thing to describe what LLMs are strong at. RLV has done incredible things for system 2 thinking. I am at awe. It is super freaking cool. Like, I don't think that it's going to result in AI Doom in the slightest. Um, not not 0%, of course, because I think 0% is miscalibrated. But like it's it's really cool what they've done and they've really pushed it to the limits. Well, maybe they don't think so, not the limits limits. But um, like it is it is a weird thing for models to do and they are very fragile at this. Like, think about how people used to talk about AI back in the chat GPTJs. Like, wow, it's really general. It can do a lot of general things. And and and then but it's bad at math problems and like GSMK, grade school math. Um, and then now look at how people talk about RLV. It's so fragile. It's so jagged. you know, like it can why can it do this like really weird thing? And actually, you know, math is not just spiky, it's fractal, right? And this is because RLV is, you know, like if we talk about like, what is the north star for each thing? RLHF is please humans, right? That is what the human feedback is. RLV is optimized benchmarks. You know, that everything that goes into the RLV category literally is a benchmark by definition because a benchmark is programmatically verifiable, simple outputs that can do well. And, you know, RLCD is make it reliable for, you know, programmatic use. And um, yeah, that I'll yeah. Yeah, yeah, maybe I'll I'll offer some thoughts and then you can sort of um correct me if I'm wrong. Um, one for example, one thing that I've been thinking about is also I I threw at a bunch of things when you give me access on on day one. Um and uh multi hop reasoning, right? Like uh so single hop, fantastic. Like state of the art. Uh you should never use anything other than Jev for single hop. Multi hop it starts to falls down and it's like kind of monotonically increasing as you increase the hops. Yep, yep, yep. Right. So, oh yes, back to that empirical question, it depends on what we can like pull out of the models, right? So we want everything, like we want to unearth as much intelligence as possible, period. Um, the models, like I see us as like unlocking and smoothing and sculpting the intelligence while like adding new capabilities and like, you know, like like filling in gaps in it. Um, and we will be filling in like, you know, more and more and more and more of these gaps over time, but the reality is that we are in the business of unearthing properties. Those properties are actually a function of what is available from like these like, you know, these condensed cores and like Franklin's signing them all together to have all of the properties of everything. Yeah. you know. Um, but the reality is we are in the business of unearthing as many capabilities as possible as possible. And system one just happens to be the description of what works. And everything that works in that paradigm will be system oneish, you know, like I'm Like like there is a reason why we don't do what's called latent reasoning, reasoning in strings. I think the reasoning, like what models do really well is reasoning within the models. It's not totally complete. It doesn't do great on all. latent reasoning is reasoning strings? I thought latent reasoning is reasoning in in inside the model weights. I think that's people used to call that continuous reasoning. I'm not entirely sure. Okay. It was called latent reasoning because like it used to be that the reasoning traces were secret, so they're kind of like a latent variable for the answer. Ah. Yeah. So what secret is now shifted? It well, it's it's still secret for Open AI and Anthropic, right? So no reasoning Jeff. Yes. As far as you will ever do it, right? Because that that like violates the whole promise of system one. I my promise is to do whatever necessary for machine native stuff. Yeah. I I could imagine there are there are like there are some forms of reasoning that are less slow, inefficient, and fragile that I that are like totally on the cards, just to be clear. So, pragmatic person, I'm not making promises on it like you know like methods. I'm making promises on like the what my ROI North star is, and I'm going to fight for that like like, you know, like like this launch didn't happen and we are still like hungry for our place in the world. That's great. Yeah. Yeah. Um, I I think the other thing that uh vision is another one that's like a a big like you know, capability that you don't have, but maybe it doesn't ever belong in system one. I I think I have a pretty good vision. Uh what sorry? I think I have a good vision. No, no, no, sorry. I'm kidding. I'm kidding. Yeah. Oh my god. Yeah, yeah, yeah. Um, Because people obviously, the first thing they want is vision. Because of the doom demo, but also just like everything, you know, other than text is vision. Everything is in the cards in my mind. Like like um and and actually this is like a debate we have. This man, your audience is probably like the great one to have in this debate. There's a question about like, how much do we try to like give people what they think they want, which is what we did in stealth for two years. We just knew that this is obviously going to be valuable versus give them what they say they want, right? and and like uh there's a lot of dimensions of this, right? And you know, like context length is an example of this, right? Um, every single model, including ours, I actually think as far as I can tell, ours is like by far The longest context. the best at not degrading in long context. Um but like the other providers are just like, whatever people wanted, let's just give them the stupid thing. And um like we need to figure out a balance for this because, you know, like if you take the the the former side too far, give people what they want, you end up with like anthropic nanny state style thinking, which is very like anti-developer. While like the pro-developer route would be like, give them what they want, but developers are like we don't want to put the burden on them to figure out the je ne sais quoi of intelligence. So, we are trying to like figure out this navigation of like how quickly to release things to still like have our like brand of trust and also like teach our treat our users like adults that can make informed decisions that uh don't need like nanny stating on top of this stuff. Yeah. I think that's fair. Yeah. and we don't know the answer to be honest. Like uh we'll we'll have to figure it out. It's going to be that's probably going to be like one of my biggest debates over the next couple of days. Yeah. because like we have a lot of stuff. Again, we didn't expect it to pop off, so we were like, we just we just need some follow-up launches. Yeah, I don't know, I don't know if you didn't expect it to pop off. Like I you put I saw the work that you put in. Like I have never seen you lock in so hard as as like the last two months basically, right? Well, well, that's also because my chief of staff made me lock in. Yeah. Like it's like uh I have never I thought I worked hard before. Yeah, yeah. And No, but like you're showing up at our writing workshops and I was like, what are you doing here? And and like It was useful. It was great. you clearly like were very intentional about your launch and the work showed and like congrats, like, you know. Thank you. Thank you. Um, I hope to keep locking in is my is my sense. I I want to like we like I think that we've passed many great filters for the the the tech world what we're wanting, but like there's still going to be a bunch more. Yeah. and like holy smokes am I excited to fight the good fight. Yeah, it's exciting. Um, before we broaden it out to like uh a topic is outside of type safe. Uh I just wanted to offer uh any other things that you think like underrated or misunderstood about what you have launched. Underrated or misunderstood. Yeah, you have panels, sorry, patterns here. Uh maybe maybe you want to go into that. Um model jaggedness, any anything. Give me one noodling of it. Oh man, I I would rant about all of these. I I really shouldn't. I really shouldn't. Um And I people can come go to your discord or thing. Yeah people put a lot of love into the cookbooks is what I will say. The cookbooks have like some fire stuff. We had considered putting a bunch of these things like in the main launch blog post, but it got kind of long and unwieldy and like very power usery. But like, we really, really, um I I'll be frank. Like, before the launch, um, every like like what we were saying sounds like like sounds like like this weird alien tool, why would anyone need this? You know, it was a very weird thing. We were very worried about teaching people about like this new frontier. It obviously succeeded, but like we put a lot of work because we thought that education would be like a gigantic bottleneck for us. Um I it probably works and it's no longer probably no longer a problem because people are doing things like well beyond what we could ever expect. you how to use your model. Exactly, but like they and their use cases are like kind of cooler than ours. Like like there's a bunch of stuff where I'm like, man, if that was our demo, holy shit, that was way cooler than what we were showing. Um, like the the computer you stuff. Holy smokes, is it cool. Um, but like like we put a lot of love into this. This is not like AI generated trash as far as I know. We put a lot like it's like a lot of love in here. And like each of these are like like there's real alpha there. Like these are inspired by solving real customer problems that existed and we went through the work of like helping them do cool ass stuff. Yeah. How much uh while you're talking about this, right, how much validation did you do before launch? Like what, you know, what was that process like? What was that process like? Like clearly you did some, but obviously you're not getting in touch with as many people as you are today. Yes, of course. Um I actually think that the reception was pretty bad and like actually for the non-technical people on the team, they were really worried. Yeah. You know, like there was a lot of fear. It's like no one really gets this. And like, you know, they don't want it. We're like selling like a vitamin and not like a painkiller. Like, should we have FDEs to like write the software around solving that problem? We had almost no revenue before launch. Um it was kind of like a like we like the technical people were like obviously true believers, right? Like we knew that this was sick. It's prop computational properties are like off the charts on like so many axes that we're like, yeah, obviously it's going to be huge. I was definitely super afraid, which is why I locked in super hard. But like the most common thing was the like I would say like more than half the people we had play with it just did not get it. And like the people who did uh like were like, man, this is really cool, but how do we get this through procurement and stuff like that? You know, it was like it was like quite a uh quite a battle and we just knew like, okay, our target market is going to be developers. People will find the use cases. And that way everyone is going to FOMO in and like I I I don't want to rub in people like changing their minds with the facts changing. I do want to call into question like the concept of product market fit, you know, but like because like there was a product, there was a market, like we were like, hey, do you want to use this? And people are like, I don't know, really know if it solves our problems. It explodes and everyone's like, we need as much rate limits as we can. Can we literally give you GPUs because we are constrained right now. So, of course. Of course, uh, you know, like marketing is an element of it, of course, but I don't even think it's about marketing. I think it's about like passionate developers who've like, you know, our souls basically resonated at the same frequency and that frequency and that got got everyone else excited too. And I'm hoping as well that like we as a company will be eternally, eternally, eternally grateful to those developers. Like not and not just like, you know, like, um, the the companies that like are like start off with developers and like go to enterprises. Exactly. And like, I'm like even thinking about like, how can we launch things that are better for, oh, man, I don't know if I should say this. But I will. Better for developers than enterprises. Exactly. How do we do that? Like, how do we empower them? Um, and I have cooks. I have cooks. But, um, it's a it's a very weird thing to do and like I don't know how else I can show my thanks and loyalty to that. You know, like and and and that's why I did like the the dying my hair yesterday. It was like it's like I wanted to talk to them because it felt dirty to me during our company's like most important times, not to keep talking to them. Good. Well, I mean, that's why one of the reasons you're here. Hold me to that, please. I I try to be principled. Quote me on this, call me out, you know, have the pitch forks out if I change. Uh, I was just going to briefly show the computer use stuff. Is this is this what you're referencing? I've never seen I haven't I've seen I saw like a airline browser use thing. And and inside this new note, let's make the title say, hello. Wow. Great, great. Okay. Um, let's move on and can you open up the arc browser? And once you're there, can you Google search Norbert Weiner? Um, now, can you open up X.com? Is this this kind of use case? Oh, the voice use case. This is this is actually the first one I've seen. This is Okay. up the photo. Wow. Oh, wait, wait, wait, wait. Oh. Can you go can you go back a second? Can you go back a second? Rumors claim anthropic engineers worship Claudius God. Wow. Wow. Dang. That's pretty funny. Um, And here you are building prod. Absolutely. Wow, this is sick. Uh, yeah, so so clearly you can operate the whole computer with voice with Jeff as a decision model. So, just like I'm anti-benchmaxing, I'm also anti-demos. I want to make sure that it works for lively. I love people are playing with it. This is super fucking sick, have no doubt. I want to see this. I want to see it be used. I want our team to play with it. I want to find the weaknesses and I want to solve that. And I would man, that looked really cool. That looked really cool. I want that. I want that. Like when my when my wrists are sore, I just whisper flow everything, that'll be sick. What is uh you know, just just to round out the use case side because I I do have to let you go. Um uh who's uh who are the who are the bigger companies that have reached out and have surprised you with what they want to do? Just I am so out of touch for that. People have shown me screenshots of companies and from what I've seen it's all of them. Yeah, mostly like, you know, for for those people who work at larger companies and they're not doing this kind of work, I just want to give people examples of like, you should go look that up, look that up, look that up. Oh. So, uh, like I think demos are super duper sick. Obviously, the coding agents are like gigantic use cases. Like they are like also super sick. is all about Jeff right now. Oh, hell yeah. Well, can I uh can I give a little bit of a a tangent about coding agents if that's okay? Yes, please. Oh, uh let me give me give me a second. Give me a second. Okay, actually, I'll come back to coding agents. Let me describe like the big families of use cases. Yes. Like, we've mapped this out from first principles like long before release. They are what we call dark data. Like people hoardd big data, but they would not throw LLMs at it because it was too expensive. So large companies adore this. They have like piles of data that they wish they could analyze and this is like a data scientist's wet dream. So, this is like this is a giant one. Like I think this plus um, coding agents are the big money makers because that's what they're all where all the volume is, right? Um, there's the real-time stuff. You know, like people who need like intelligence in the loop. They like I would guess that every CEO, if not CTO at those companies, um, knows how much better their product gets with every like 10 milliseconds shaved. Yes. Uh and like especially e-commerce, yeah. Yeah. Oh or like assistant things. You know, there's many AI assistant things. And like as far as I can tell, they really love it. Um, again, I'm not in the front lines of customers right now, so I just get know what my team tells me, but like this, I'm so excited for this. I'm really excited for this for games. I really want to play like sick ass auto battlers where you're like commanding your team or like semi-auto battlers. I I think that'd be so cool. But don't make it too good while I still have a job. Um, and um, the, you know, like there's the uh, the the what we call like verify everything. You know, like verifying all LLM calls, kind of like observability. Um, I think actually on the note of docs, what people should be doing is like, the parallel questions are very cheap. So if you have like big states, you want to ask many questions on. right here. Yeah. Put IDs on every like message and then ask a question about each ID. So like when you have like a long state. Uh-huh. So that way you can like pay for the state once and ask lots and lots of questions about each message within it. I think that is like a like a great way that like saves money and is. Which by the way, I always think like it's interesting framing system one and system two because it it basically makes the case that you should always make one or 10 or 100 jev calls for every one reasoning call that you make. Well, maybe, well, I I mean, I don't I would like people to spend less. Yeah. You know, maybe you do like, you know, one half the reasoning calls and like 10 jev calls each or something like that, or whatever solves the problem that like couldn't have existed otherwise. Um, wait, and number four use case was what I described as like smart software. Like software that's intrinsically composable and like does like weird, fun stuff that could never happen before. You know, like the programming language as Jev thing. I don't know if you've seen that. That is so cool. Man, if we knew how to give out credits because our we're really early in our infra days, I would want to give all these projects credits. Um, and I think those are like how we've mapped out like the main use cases. Um, computer uses also come in kind of like the real-time direction as well and like that's really, really cool. If it is reliable, I am super jazzed about that. I suspect we can make the model a lot better at these use cases because like that came out of left field a little bit, so like that's really cool. Um, I'm on the coding agent thing and this is like a really surprising thing that is happening right now. Okay. Um, Claude Code and Codex are, I believe, the the winner, like the the the number one and two. I'm not entirely sure. I don't follow closely, but like it's roughly that. But they're built around a single model world. You know, like, and and that makes a lot of sense for them, right? Because like it has been a one model game where it's like kind of like the same model but different intelligence that you're shopping. But all the open coding agents are like fucking jazzed right now because they're like getting their jev on. And like the thing is, there's I'm sure they're trying a lot of weird stuff. Yeah, but all the coding agents are kind of roughly at like approximate parity, right? Because like there's not so much you can do with a while loop. But the moment one person finds one killer use case that, you know, you can only do with that coding agent, everyone will flock to it because they have like a monopoly on that thing. But all the open coding agents will be able to copy that, right? And but I don't know what the Claude Cose and Codex will do because they are built around that one model world. And like I think that's going to be like a really interesting thing. You know, like I would love to be able to integrate with them personally, like I want to integrate with everyone. Like they they might make competitors eventually. I don't know, but like it is not me my job as Saphire infrastructure to be opinionated on that, right? Like I want to just serve the world. Um, but I don't know if they would do that. And like I think it'll make the coding agent game super weird. You know, like I'm so excited for that. Uh and like I'm sure I'm getting my team to review right now an internal document I made on design patterns I suspect will be useful for coding agents. So hopefully I can share it like right after I walk home. Um but like, I think that there's just like such ripe area for exploration out in the world and like, it's it's man, if I did not have this, I would love to experiment with coding agents right now. Yeah, I mean, and and I'm sure the coding agent companies will love to work with you as well to to figure that out. Um, yeah, I I do think that there's still use cases for co code and code as with you guys, which um uh it's it's easier to explore there. Um, okay, I mean, you know, we've you've you've been very, um, uh, obliging in the sort of indulging and all these all these things. I just want to take you out of type safe just generally about and you've you've made very clear your position on on the state of VI. Uh, give you more more room on the alignment safety side of things. Oh, uh did I not talk about safety alignment at all? I I I think I didn't I think I maybe I didn't. You did. Uh I I just like, you know, I I think that there's there's a lot of uh you have a lot of researcher discussions. We have this every every nebs. What are people talking about? You know, like so for example, um, I uh recently was at uh one of these researcher gatherings and people are genuinely worried about the pacing, right? Like this this whole topic about like we should slow down because uh the public is like clearly not ready. Um and I I I'm sure you have strong feelings. Um I feel like this is the kind of thing that is a dangerous topic to talk about. Um I'm happy to talk about it. I I live for danger. All right. Our company brand is chaos. It's not Jev. It is reverence and chaos. you know, yeah, and and like you were at OpenAI during like the one of the very first like very visible incidents just as the blip, right? Like which like and and like you the dominos have gone down now to now every frontier lab has co-signed a document saying that they want to pace. Um Interesting. I so it's a very complicated nuance thing. I actually do want to write a response to this more formally. I do have like a little bit of a short version of my response. Yeah. which is that um, as you RLV are more, like RLV is like to RLV is not actually about verifiable rewards. Like that has been failing since before the reasoning revolution. Like like like and that's the weird part about tasks, right? Like back when, oh fun history. Back when RLF was becoming a thing thing, there were three different things that like are now called post training, different efforts. And instruction following was by far the like the the bastard child. Like, people didn't like it. They didn't want to take it into account. It was annoying. You know, like I talked to the pre-training team. I'm like, guys, this is the magic. And they're like, we'd run so many model sweeps. You want us to wait for human evals to figure out which models to use. And like everyone is like, you know, giving tons of like resources to like the Cogen team, which like they did have some successes, but they were trying really hard to do RL on like unit tests. And it didn't work, obviously, right? Like you needed reasoning for that. So, uh, so just to be clear, R is not purely about the reward, it's about like the shape of everything too. And part of it is that reasoning is included in here, like this latent variable that you're doing things. And when you're doing things, you're just letting the models do whatever they want in order to make them be as powerful as you can to answer the hardest problems. And this whole pace the frontier discussion, I think is like a very narrow focus because it assumes assumes that everyone needs to do more RLVR. Right? Which, um, like I obviously don't think I need to do more RLVR on our models. You know, I think zero is the optimal amount for our shape. right? Like, come on. you know. So, um, it it's really, I think a bit of a sleight of hand where they are saying that we actually want to keep doing the thing that looks dangerous because it does dangerous things. You know, like people say like, oh, maybe the sandboxing was a problem or whatever else. Um, yeah, I mean, obviously it is, and they could have easily solved that, right? But they chose not to because the more things you let the models do in this do anything category, the more powerful it is. right? So, uh, like there, I think there's some like dissolution of responsibility there on like things that by design or non-design, they're trying to make is just an assumption. You know, we must do RLVR and not just we must do it, we must do more and more and more. Um, with giving the models like the power to do powerful, you know, do anything they want in the middle, because that teaches them to be powerful outside of it and we don't want to limit those things well because it'll make it slightly less powerful on those things. Um, so, like if you assume all of that, they're like, oh yeah, we're heading in a dangerous world, guys. That's a logical world Like everyone is going to be doing this and this is the only way to make AI sick. So, um, So basically it's like it's like these are all internally consistent, but actually starts from a premise that has alternatives if you think about it. Of course, of course. I think there's like like on the bitterest lesson direction, I think that there's very few people who've like made right tasks. You know, like new directions of AI. That is or new new north stars. That is rare. Again, like I think 2.2 times or something for LLM's itself. Like RLHF and then RLCD. RLVR is like a .2 in my opinion and I think that's generous. Um, but or .5 or like it could be one whole one. I don't really care. Um, but I do think that people are thinking very closed-mindedly about this type of thing and um, this the the only people who are at fault here are the researchers because it's definitely not the populace. You know, like they just assume that Open Anthropic are just doing the best they can and they are not the experts who are aware of the true optionality available. Yeah. Uh and that's fair. and and like you're you're also doing your part in waking them up. Yeah. Well, I'm doing my best. But like my goal is not like convince labs that there's like other directions to to go down. My goal is have, you know, it's like spark hope in software engineers to start like actually automating things they've always wanted automated. Um, I had this like article that I wrote that my team didn't let me write, that didn't let me publish about like the the the future I want of AI. And like there's like a lot of like little things like remember do what I mean? Imagine if everything could do what I mean. Because like that that that demo was do what I mean. Like like you like like there's levels. Don't do what I say. Yeah. Don't do what I say, do what I mean. Yeah, and like we couldn't do what I mean yet because like computers are so basic and literal, but that computer use one was just that. And I think that there's like levels of smoothness that will happen in the world that people just don't understand and like the promise of like smarts all around are it it it's I don't want to over promise. I I don't think it's going to happen right now, but like we are going to do whatever the fuck we can to make that happen. Yeah. Um, any other things on the sort of general shape of post training, you know, um, you obviously you've been very intimately involved. Uh, mid training is that uh, something that you do have comments on. I don't think we've ever talked about it. Mid training. Um, I mean, it's all a spectrum. Yeah. Right? Like am I It's a curriculum, but like fancier. Yeah, I mean like it's it's it's like you know it's a cost saving thing. You know, instead of like having to pre-train again, like there's intriguing stuff. I actually think that like intelligence has a at every single level and it's always super duper fascinating. Like I'm a shape rotator, so I don't like finding that, but I love it when people find it and teach me about it. Um but you know, looking at the data, this thing that uh our data team is so good at that I'm not. Um I it's I find it really, really fascinating. I love actually thinking about like how capabilities are like put into the model, like over like the short term, you know, like there's like the the the the really rapid alignment of fine tuning and over the long term, after seeing it over and over and over again, like this stuff gets baked deeper and deeper and deeper and deeper into the model until it gets robust. You know, and that is like the the north star to surface and like the system one stuff is the stuff that ends up getting robust. So, I find mid-training to be like a fascinating thing. Uh I'm a fan of all forms of training. Uh I'm a fan of all forms of like surfacing new types of intelligence. I wouldn't do it all myself because it's expensive. Um and I have said privately and also, should I say this? Huh. You know, like my philosophy is anything I I I should say in like private with like an investor, I should say in public with the people because that is like my thing. Yes. So, the thing I've said before is if you gave me a billion dollars, I wouldn't pre-train. Um I still believe that to be true. It is a very expensive thing when if you are like like if you're an AI engineer, you can like slice and dice and do all sorts of stuff, you know, like Frankensteining is not the most elegant, beautiful thing, but it solves problems, baby. So, um anything except pre-training. Mm, mm. Yeah. amazing. Uh I I think one one direction that I do think that is interesting, just like synthesizing all your your commentary about these model things is like, do we have a a a super model uh that has all these capabilities involved or do we break them out in in further, right? So like, um, uh one way to put this is the opening I was trending in the direction of the Omni model. right? Um 40 was was one of those. then for for for a brief period of time, there was always like there was like a kind of a main branch of like this is the chat tune model and this is the coding tune model. Those are completely different things. Those are extremely different concepts. I will like break that down a little bit. So, multity is a little bit different. Um because sometimes the other modalities help, sometimes they hurt. Yes. Like, you know, people are moving, they seem to be moving away from speech, which is different than audio, um because it seems to not generalize well to the other stuff. This might get solved. I'm a fan of all of this, but these are like empirical real questions. Like scaling laws are not about just throw money at it and it gets good. scaling laws are pragmatically how good is a thing. You know, like like there are worlds where like no matter what you scale, it may not be good enough. So, um, you know, like computer use is not currently solved is my understanding. Like I'm hoping that we can be a like play a part in solving that. But like there might be no amount of data we collect that will solve that. We might need better methods or something else like that. So, um, like we you need to be like really practical in all of this. Am I a fan of omni models? I'm a fan of all forms of intelligence, but I will go straight into one thing you talked about, which is different from pre-training, which is post-training because I hate fracturing intelligence. That is like the bad thing to me. And this whole like chat versus reasoning mode is because um it forces the intelligence to be fractured. Like when you're optimizing for chat, this tends to be like pure RLHF. And it's quite intrinsic in RLHF to do the stuff people like naturally complain about, right? Like, oh... You're absolutely right. and you know. Yeah. sycophancy, psychofancy, whatever word, however how to pronounce that. Overconfidence, hallucination, like even the kind of style that excels in LM Arena, bold, italicized, emojis, you know, like it doesn't answer the question simply, it gives like a long write up and then it asks you a follow-up question, so it feels more like a human talking to you. All of these things um come because strings are super weird. You know, they are like weird ass things and you need to be miscalibrated, you need to like mode drop, you need to be hyper confident in order to not go off the rails because the the reward model will punish you so hard when that happens because it's obvious. Um, and then this like warps the probability space entirely and it interacts with that of the reasoning models, right? Because like it, you know, the the models are like these simple linear things that tend to cheat a bit. So, I think that's very different than exposing intelligence is is my guess. And a lot of the art to intelligence is studying this subtlety that I think that um at least when I was an OpenAI people were not really studying that because like they were just like chat, chat, chat, chat, just like people are with Jev right now. Hey you give me an ultimate, you know, you give me an objective, I will just go optimize for that, right? But but if you try and there and you know, like the say is like, you could have like two objectives and you could just like optimize for both, but then that is literally the act of fracturing, right? Um, so yeah. So I mean, in some ways you have you are also factoring intelligence into system one, system two, but you just don't agree with the other people's fracturing fracturing. which is fine. which is fine. No, no, it's a little bit different. No, no, no. if I could, if I could add, if I could defend, um, the system two tasks. Um, number one, like we don't toss out the system two tasks, right? Like you can try to make Jeff work on it. And there actually is an intelligent answer for that, which is unknown. You know, like, like there, like there is better and worse behavior in the system two tasks, which would be like really low confidence, lots of uncertainty, maybe some heuristics can like move the needle here and there. But we care about them too, just to be clear. I just think that that is not what the what is the intelligence is native to. So, we're not trying to fracture anything like that. Um and all fracturing makes the model dumb. You know, like if people like get the model to say like it is open AI or or uh you know, like cloud or whatever else. I don't really know what it says these days. I am not going to put into the models that you are Jev from typesafe. That fractures it, right? Like like I I don't want that. Like represent what the internet thinks, right? Like be correct. That is what I want because that's how you get the smooth predictable intelligence. I I mean identity is a thing, I guess that is somewhat of a special. I mean for a first party product, yes. Yeah. But like for an API, I don't think so. Yeah. Okay. You know, like I I don't like people don't want if they're making a chatbot with uh you know, chat GPT, they don't want it to say it's chat GPT. They want to say it's like chipotle or whatever. Right? Well, you know, so the way that you also have to make up for it is you have the skill, right? The the the Jeff skill, which which is for coding agents to work with Jeff. Yeah. Um, okay, a couple of closing questions because I I I do want to uh get you out. Uh one is like you just reflecting on your two-year journey. Is it roughly two years? 2. something? Um with the company, I think that this is like more like a four-year journey. Yeah. Well, I think yes, actually like uh I was thinking remembering that like you have this like hero run around Thanksgiving. You were like you were canceling everything because uh you were like, guys, like uh everyone's on holiday. I'm going to take all the open GPUs and go do this thing. Yeah. That was a good time. And that was like the the pre-type safe moment, right? Yeah. That might have been, was that when the coup was happening? I don't really know. Yes, actually. Yeah, yeah, yeah. that sounds right. Yeah. I remember. Oh my god. I don't want to I I'm not I I don't think I have the time to spill the tea about the coup right now. now. But um, that was really annoying. Um, The the coup was annoying or the the run was annoying? The the coup was annoying. The coup. Okay. Yeah, yeah, yeah. Well, I will. So CS took over the company Yeah, anyway. Maybe next time we chat, I'll I'll dump tea about a tea about the coup. Um, yeah, it actually this problem was one that like was in my mind since before Chat GPT even launched. I was like, holy shit. The Chat GPT team is cooking. They are doing the right task. They are doing the thing that AI researchers are bad at, but successful product people are good at, which is giving a lot of fucks about the experience. You know, it's it's it's very rare. They like there's very few people like that at Open AI. Um, and those guys were cooking on it really, really well. And and to be clear, this is the whole journey from GPT 3 to 3.5, which included AI dungeon, which you've talked about as like, yeah, well that's that's an example of a use case that we never predicted. Yes, exactly. Yeah. Well, uh, oh, yeah, that is a, also, I had fought very, very hard to deploy instruct GPT. Um, like actually the early versions of it were even trained with like an algorithm we didn't publish that I made myself because it was too slow to clean the the PPO data. And I was like, fuck it, this is so fucking good. We need to get it in the hands of users. And like basically immediately it took 50% of the market share of LLMs at the time. And but and we thought, I I made I went through great effort to make sure everything in our launch video is true. Um, you know, we I truly was thinking like, is this AGI because it's superhuman at instruction in instruction out? You obviously, it's not. But like, everyone I think should have an answer to why that was not AGI because it looks very smart. And my answer to that ended up, you know, like ended up only being used for copywriting. You know, Jasper AI, copy AI, like writing like, you know, what is now called slop on web pages. Um, and we were worried we made the internet a worse place. Right? And I went back to the drawing board and I was like, what's missing? We're smart, clearly, something is missing from it like creating value. What is it? Like, I actually was doing more philosophy at the time of like, you know, like, what is going on? And the answer was, oh, machines. You know, the question I asked myself is like, let's work backwards from an AI based economic revolution. When that happens, what will be what will be calling the AI? If AI is an API, will it be humans or will it be code? And I figured it was many nines of code. And but like all the optimization was going into the humans part. And then it clicked for me. I'm like, holy shit, this is the north star. I think like I wrote a document. I was like talking to Sam about this. Sam was like, this is so fucking good. You should go work on it. And we're like, yeah, yeah, yeah, Sam, I have a job. You know, like I'm you know I was working on. Sam just told you to do it. go do it. I but like my my guess at the time is like, this is super obvious. Like it's so unbelievably obvious. Anthropic must be working on this already. You know, and like we're already cooked and like actually OpenAI is just better at like catching up than it does at like actually innovating. So, like, uh, ChatGPT was a copy of Claude, right? Um, like they had an internal thing, they just didn't ship it. Yeah. Cloud Cloud and Slack. But uh, but you know, reasoning, I I would say first-ish. Yeah, but debatable how good of a product that is. Yeah. Um, it's great research though. Super great research. I'm just not sure if people had that product need. Um, and you know, Claude did the coding agent stuff too. So, Sam says that, and you know, I just go back to my job for a while. Eventually like, you know, the instruction following team just says, we won, we we've solved instruction following. We don't need to do stuff anymore. I'm like trying to think about what I do next. I was like, maybe I'll just like start playing around with this. Um, I, you know, do more philosophy and design and thinking. I thought it would end up taking a week. when I started training models, it ended up taking um, many years. at some point, I was like, holy shit, you know, there's signs of life here. This it it obviously didn't work, right? Otherwise, we would have deployed it. But like I want to explore what it would be like research wise to go all in on this. You know, like I want to really see like like what it would be like if you went like absolutely insanely all in in this direction. And because of what I said, you know, like if an AI winter happened, would I, how would I feel? I would consider myself personally responsible. I talked to other companies at the time, and I I was like, hey, I want to start a lab on this direction. And, you know, like there was interest. I just talked to them like, how fast, what would be faster? This or startup? And they're like, startup. And I'm like, fuck it, man, we ball. I guess we're doing some crazy shit. And And you called Eric and Sasha and Yeah, well, I I call Eric first. Um, with Sasha, I actually didn't try to recruit her. I I tried to be good. And I was just like, hey, am I crazy? Is something missing here? You know, isn't there like like like, am I too much in the open AI bubble that I didn't realize there must be a solution to this? And then Sasha was like, I'm in. And I'm like, Sasha, you're working at a startup. And she's like, I'm folding it right now. And I'm like, do you want to think about that? She's like, oh yeah, good point. Let me think about it. And then she joined. And then, you know, within two weeks, we had funding, within we had like people move into my apartment. It was the worst because I'm a neat freak. Um, and we just kept on cooking and eventually we got the research that that that showed the signs of life, you know. It was it was a crazy time. So the the the the question is, that was all long context. And then another question is someone like you is in the frontier lab right now, who is frustrated, not getting the the funding or the the resources or whatever, uh the attention. Uh what's your advice to them? You know, should they do what you did? Should they do it? Ooh, that's a fascinating question. Um, man, how do I do this without burning bridges? Um, I my sense is that most unless there's some level of economics I don't really understand, Um, I think most neo labs are crap. Um, I I don't want to see myself with that as peers. Um, like I I don't really understand what's going on there. Like, is it because like number one, I don't really value researchers. Um, I value people who like like look at my bitter lesson, right? well not just that. I like we need researchers, but we need them to give a lot of fucks about the right task. And that's the important thing. Right? So it's actually like it's it's kind of backwards when people value pure research pedigree, because that generally doesn't create value. Um, so it like like number one, I I believe in North Star tasks and doing cool, really useful stuff. Um, number two, um, because I don't value researchers, I don't I don't recommend going the well, it clearly is profitable for someone or it might be in this environment. So, like from a purely pragmatic perspective, I don't see creating neo labs as one something that creates value. It seems to destroy value because like they're like redoing work from scratch with like low probability of actually moving the frontier. And as far as I've talked to most Leo labs, they don't really have a direction. They tend to want money to play around with their experiments. Um, if they have direction, I'm super in favor of it, to be clear. So, my advice for someone is it really depends on why you're doing it. You know, if you are a researcher who wants to play around with research, um, probably the labs are the best place to do that, TBH. Like there might be other places. I I don't really keep track of that politics. But I would just recommend not being that way, per se. -ly, you know, like I think it's better for the world with people being driven to solve real problems. and those problems may be exploratory, that's fine, but like ideally have principles that you stand behind. But if you think that you want to do the right task, like absofuckinglutely. Like, please do. Like, please break this like unimind, you know, unimodal like. Yeah, exactly. Like like you know, like again this pacing the frontier is coming from like this one view of AI that looks like, you know, AI super genius that is incredibly jagged. And that is um Solvable. it's solvable and it's weird and it's like not matching reality and um it's like it's tragic, right? Like like I I I think like all of these like the like really unearthing technology I think is like just good. Yeah. For what it's worth, uh you know, again, I I'm trying to accurately represent the position of the uh Anthropic Open AI folks I was talking to. SpaceX as well, by the way. Uh is that uh it is this is a political thing much more so than a pure X-face thing. Yep. Uh so, yeah, political positioning is. And that I'm that's beyond my. Exactly. That's well beyond my. Once they once they all told me that I was like, I get it. Uh this is about the 2028 election. Oh no. Oh, I wish I didn't hear that. That's such a bad vibe. No, no, no, no, this is not the whole company. This is just that that room's discussion. No, no, that that makes sense. That that that makes me lose faith in humanity a bit, but maybe I'm just a naive technologist. Uh it's really starting to matter. who's who's in charge of the governments uh that that will help to regulate uh these things as they emerge and like as a lab you should probably think that through. I totally. No, no. I totally agree with that to be clear. Like I I think being opinionated on that matters a lot. I personally, I'm afraid of trying to mislead people because I think that bites people in the ass a lot. You know, like, um, I think that like people trying to be overconfident like uh I I I I I obviously I'm not actually going to talk about politics. I think what happened in COVID is like people leaned too much in like appeals to authority and being overconfident to try to get people to behave in certain ways. And like obviously our response was extremely suboptimal. and that had like like ripples of downstream ramifications that are now I think extremely bad for the world. Like maybe I'm naive. I think that misleading people even for the greater good or what they think is the greater good is just uh it's just I'm not a fan. I'd rather I'd rather not. Yeah for what it's worth I I it's not a I don't think it's misleading. It is just like this is why now. Yeah. Uh what yeah, what like you know. I I think that that's the I think that is why now that is a little bit of misleading about like the risks versus like the the objective. There's like some level of like sneakiness latent in it that um is worth calling out and I think owning up to. Well obviously they want to if if they want to manipulate then they shouldn't own up to that. That seems like a bad strategy. But like that to me is just just sad for the world. Yeah. Hopefully I'm never I'm Yeah. Uh hopefully like we are never involved in anything like that. It might be inevitable as we get big. Um but I want to I want to stay like pure technologist to my roots as much as I can. I mean, Jeff for president, why not? I can I you know, I will trust Jeff's decisions over my my own. Um okay, so so less shit posting, uh more about That's shit post. No, no, no, no, no. for me. I'm shit I'm shit. Oh okay, you're just crushing my hopes about America and the world right now. Oh my lord. No. I'm not. No. No, no, no. Yeah, I mean like there's I I I think I watch too much TV about like conspiracies to take over the presidency. Um the uh you have chosen your star. You have chosen reliability and then programmable and composable AI. And cheap. Uh, and cheap. Yeah. What is a second or third one that you want to throw as a bone to someone else that you're not, that you want someone else to work on that you're not going to work on? Ooh. Like uh, just basically give people tasks. Give people tasks? Yeah, like, that your task. There's so many I want. You have picked your task, right? Oh my God. What? You know what I mean? Wait, I that's such a good question. Holy crap. Oh, man, I'm so excited by that. Cuz you're you're going to be you're be for the next like 50 years, you're going to be busy doing your thing. Oh yeah. Okay, so let me give like a fun one and a not fun and like maybe a valuable one that's also fun. Um my my fun one is, I think games could be so freaking cool if they were intelligent. Like when I see people play around with like like Ally's Doom demo where like you can like get NPCs to control stuff, like you know, like that was just really like the like a proof of concept. I think it's really cool stuff could be made. It looks really, really cool. You know, like, um like, I'm a big Stardew Valley fan. You know, and like it's it's really static and it's still compelling. Like I feel like there's a lot of cool story that could happen. You don't need to call like Jev in the game loop. It's probably too expensive for that. But even like simple like state machines for NPCs, I think you can make like such a compelling world. Oh, man. Um, and man, a little sad that I can't work on these types of things. Yeah. My my life path is a little bit set right now and I'm um Yeah, but you can call someone else to work on it. Yeah, yeah. And then you can like feedback on it. And the thing that I would really, really like to explore is like coding agents free from the tyranny of the KV cash. Like, like, it might not be as good as true coding agents are, but I think there's just so many weird things to think about. That that that's why I wrote the article KV cash rules everything around me. Um, uh, believe it or not, I don't think anyone has used the phrase on the internet, cash rules everything around me, c A c H e. Um, when I when I googled it. Um, so, uh, like I wrote this because I wanted to tell people about like this is how coding agents agents work and how the KV cash works and everything. And um, I think oh, um, yeah. Um, like uh it explains a lot of stuff. Like why routing is really hard, why sub agents don't seem to work. Like while compaction is such a hard problem. And I I'm going to try to release a document. My team might veto me because believe it or not, I'm not in charge. Um, you know, but I I I wish. Um, but, um, I want to release a document of like, here are my thoughts, please play with it and please figure out all the ways that we can do things with coding agents like once you're freed from that, you know, that that KV cache tyranny. Which is it locks you in. Well, not it lock it it locks you in into one model, right? And in order to do it efficiently, you need to like keep on appending to it. So now you're not doing best software practices. Like state management, abstraction, decomposition. Why can't you give an easier task some why can't you give a sub agent an easier task? Because of the state that you're passing around. Oh, I touched this. Uh because of the state you're passing around. You you would need intelligence that is way cheaper than the intelligence you're using to read this, in order to pass this state around, why can't you be smart about it? right? And I think there's like tons of really cool fun research to be had there on like different programming patterns. You know, kind of like how people are playing around like with like recursive language models. Like I feel like there's like just lots of cool stuff in here when you think about like, oh, I want to explicitly label the state of everything. Or imagine you have like a subtask, like coding agents, I I think it's fair to say they work on subtasks at a time as from a decomposition perspective. Why do you need to pass all of that state back into the parent task? Why couldn't you do smart things about it? And and also, if you had a hierarchy of labeled subtasks, why can't you do a search through that subtask tree for the relevant context? when you need it in, right? And then, you know, another thing that you can do, oh man, I forgot to write something about this. Um, I have like some cooks in here that are really, really cool. Um, hope to publish it. I'm down to jam about it, but like it's going to be a long document. And and like if that becomes the case, where context becomes cheap, like why can't you do cool patterns, like looking at your historical context very cheaply? Isn't it kind of weird that you start from scratch every time and you need to solve a problem called continuous learning? That's a that's actually like a memory management problem because you don't have a smart way of looking up the memory, right? But what if you could? What if you could do that all the time? Or what if when you have parallel sub agents, they can like read each other's states because you have all of that in like your computer memory and you can be smart about what's reading and writing at the same time and your coding agent swarm or whatever has like locks around things and can coordinate intelligently, not with like basic S-locks, like, what are you doing? What am I doing? You know, Jeff, who should write first, blah, blah, blah. And like I feel like the future there is nuts. Oh my god. Jeff to solve locks. I mean, it could be so cool for like multiple agents working together. Or like, if you think about state, like you have swarm stuff. Yeah, and and you know, some things, for example, are read only processes. You know, some people like getting like summaries of what the agents are doing. Why can't they share state easily? Because like a read-only agent needs to like, you know, read parts of the context and figure out what's relevant to say what like what's actually being written because the exploration is not super important or here's the tree of subtasks. I feel like there's so many different fun things that could be done if like a really smart person like dedicated like a whole lot of time to rethink like the the coding agent experience and that would be super duper sick. Yep. Man, I that would be my dream. Uh, I I would point you towards Prime agent if you haven't looked at it. Uh, so this uh uh works together with the RLM work. We just uh talked to Alex, uh who's a buddy of Allen's, um, in the chair before you. Cool, cool. Uh, and like yeah, it is being worked on, but it's not super popular yet. And if like yeah. Well, yeah, I mean the hope, yeah, I would want everyone to like just play around with with like weird things. I have no guarantees that it will work, but it seems really, really interesting from like a technical perspective. So, yeah, that that that seems cool. It's cool. Like like like once we figure out how to give credits out, I would love to like give credits out to people like this. Yeah, you you will be in a position to fund research for sure. Uh, no, anyway, congrats on all your success. Uh, you you've like uh, come so so such a long way since I first met you like and and and the whole team as well. I think you've been the same person as well. Yeah. Yeah. I I think but I think like you are energized in a way that I've never seen you before because you found your mission. You know? Yeah. That's true. That's definitely true. Um, and you are articulating your mission because you you you uh for many years you complained about the problems, but you didn't have a solution yet, right? And you like you had you had the rough shape and then you had to do it put in the work. I will say that that is partially because I I describe myself as 0% entrepreneurial. I I I don't like startups. Um, I never wanted to be a CEO in my life. I can't imagine anyone doing this twice. It seems horrible. Honestly, doing it once is pretty bad. Um, when we first were fundraising, an investor asked me like, which CEOs do you look up to? And I was like, ew. Why would I look up to those people? Um, no offense to anyone. You know, I'm trying to be like, I'm trying to be genuine and good, and I've met like a lot of really good people, but like the famous ones have like a lot of like skeletons in their closet, it seems. And I think I just really did feel disempowered when I was at Open AI. You know, like I felt, um, yeah, like like like it's a little bit easier to be truthful now because like I have at least some proof that the direction has legs. Like, I just felt like in the the the insane house where everyone is just like, Chat GPT, yeah. Like where do we put Chat GPT and everything? How do we make Chat GPT good for like, you know, developers and stuff? And I'm like, what what are you talking about? Like the funk calling interfaces insane. Why would you deploy this? Um, like this is this is so anti-developer. It's sort of a hacky way on top of hacks on top of hacks. Well, not to just that. Like, the thing I often said was if there was like a uh uh this is also probably T I don't have time for right now. But I always used to say like, I want to be removed from any project involving like function calling, if you did not get a logit bias for each function. Like so very very very simple ask in my part. Which is something like a confidence but not calibrated. Well or a probability for it, right? Like we need to give users the ability to control like, you know, like let's say the actions are refuse or allow. Um yeah, Disney needs to set a a different refusal threshold in AI Dungeon. The only way to control that with function calling right now is to say like pretty please. You know, that's nuts. That's a nuts interface for developers and like people have been like dealing with this for years now, right? Like they still have it with skills. Like, um, you know, like, um, um, the the existing coding agents are like highly overfit to their existing harness because they're jagged. They don't tend to use like external like tools and MCPs super well because of overfitting, of course. And like why can't like big companies allow for like slight nudges to be like, call this more, it's really useful, right? and and like the the the the the the solution is begging in a system message. That's nuts. Um but no Okay, I I think I I think I get you. Um and and like man, uh it is so exciting to to talk about all this stuff. It's it's really cool to get you on the podcast. Uh you're going to go do amazing things, man. Like I'm excited for your next uh big launches, whatever that is. Oh hell yeah. Just you wait. Yeah, yeah. Just you wait, it might be sooner than you think. So you're hiring for people for people I assume, marketer. 100 people just depends on who you ask. community person. If you ask me, I feel like I'm a pretty good founding marketer. But if you ask anyone on my team, they say, shut the fuck up, Diogo, you need to do CEO stuff. So, yes, founding marketer. it's not just about Spice. Like I think you're very spice oriented, which like you like that's your unique talent. But sometimes you just need to say like the routine marketing things. I know, I know. I I I would really love things. Yes, I nothing teaches you delegation like having a title wave of stuff to do. Um hiring data people or we call them model capabilities, like uh but they are data people. Both like like uh data's kind of a slur in the industry. And like I want to make sure I don't think so. We're very pro-data here. but I want them to be the highest status of like, you know, the people actually working on the model. that actually sounds a little weird. I want everyone to have equal status. But like I want to even that out and I want to know that that's really valuable. at least are more equal than others. Well, I I don't like weird hierarchies and I I think one of the things I'm most proud about in the company is that they don't respect me that much or they don't show that. They they troll me and like joke with me and they treat me poorly sometimes and all of that. And I think that that's a good sign of a culture. We're hiring like platform people, like people to like build out Jev everywhere. Like we are so much more sensitive to location because speed of light is more of a bottleneck, right? Like I'm so sad for the European users that they were only like three times as fast instead of like a hundred times as fast because like we don't have servers there right now. and that's insane, right? But like... It's it's okay. Life in Europe goes a bit slower as well. It's okay. Wow, I can't believe you you said it, not me. Um, or or or everywhere. You know, like if if intelligence per second is a metric that matters, like we we'll want this all over the place. Like we care about like if there's a developer building on top of us, I care a lot about you. Um and we are hiring for people to keep building more like not just like the goal is not to just be like Jev as a company. The goal is to like ship more shapes of intelligence beyond that. So we are hiring people to like build those things too. You know, like we want to not just be