Databricks CEO Rejects AI Extinction Risks, Cites Context Bottleneck

Databricks CEO驳斥AI灭绝论，指出上下文才是真正瓶颈 企业自动化停滞是因为模型缺乏不成文的机构上下文，而非缺乏原始智能，这使得关于前沿模型发展速度的争论与商业落地无关。 Databricks首席执行官Ali Ghodsi认为，前沿AI带来的近期生存危机接近于零，他将最紧迫的威胁重新界定为机器速度、由智能体驱动的网络攻击，这需要像Lakehouse Watch这样的自动化软件防御措施。他概述了真正的递归自我改进所需的四个严格且必须同时满足的条件，指出当前的前沿模型正趋于更高的脆弱性、更多的人力投入和资本成本，而非自主失控循环。转到企业层面，Ghodsi和主持人们解释说，组织上下文、机构知识和内部本体远比更高的模型原始智能重要得多。最后，他们探讨了企业如何利用路由网关、分层开源架构以及像Neon这样针对智能体优化的基础设施来压平AI成本。 Ali Ghodsi评估认为，前沿AI带来的近期生存危机接近于零，并指出在毫无证据的情况下煽动公众对人类灭绝的恐惧是不负责任的，而且会分散对亚秒级网络攻击等紧迫威胁的注意力。 (1:48) 真正的递归自我改进需要同时满足四个条件——超线性减少的GPU数量、更短的训练时间、更高的智能以及可重复的自主迭代——这与当前前沿模型训练需要数十亿美元且依然高度脆弱的现实相矛盾。 (10:15) 企业采用AI的瓶颈在于组织上下文，而非模型智能；系统需要访问内部数字本体和不成文的机构知识，而不是在复杂推理基准测试中获得更高的分数。 (40:00) 企业AI的Token消耗量正在与基础设施成本脱钩，在a16z的投资组合中，开源模型占据了超过60%的企业Token用量，但仅占总支出的约5%。 (57:12)

Databricks CEO Ali Ghodsi argues that near-term existential risk from artificial intelligence is close to zero, contending that doomsday narratives misguide public policy while obscuring immediate threats like machine-speed cyber warfare. In enterprise settings, adoption is constrained not by a lack of raw model intelligence, but by the absence of internal institutional context, escalating inference costs, and infrastructure limits. Rhetoric, Regulatory Tension, and Pacing Dynamics (1:48) Ali Ghodsi and Martin Casado critique the apocalyptic framing promoted by some frontier lab leaders, arguing that unverified existential warnings induce public hysteria and mobilize blunt legislative interventions. Casado contends that framing model safety as "pacing" is an unforced PR error that conflates calendar throttling with rigorous engineering, noting that slowing timelines does not inherently make an architecture safe. Ghodsi frames the competitive dilemma as a classic tragedy of the commons: commercial labs facing looming IPOs and market rivalries cannot voluntarily slow down without conceding ground, leading executives to paradoxically urge regulators to mandate collective speed limits. Addressing governance models, Ghodsi rejects competitor peer review—such as rival labs inspecting one another—as fundamentally compromised by commercial incentives and executive rivalry. Independent inspection bodies require balanced, credible experts rather than pre-committed ideologues to avoid stacked panels. He suggests that oversight is likely to evolve into hybridized structures modeled after FINRA, combining private-sector operational expertise with statutory federal enforcement authority. Deconstructing the Recursive Self-Improvement Thesis (10:15) Ghodsi disputes assumptions that frontier AI is currently approaching runaway recursive self-improvement (RSI), contrasting empirical reality with Nick Bostrom's 2014 superintelligence thesis. He specifies four strict conditions that must occur simultaneously to produce compounding RSI: superlinear reductions in compute resources, significant training speedups, measurable capability gains, and compounding repeatability across iterations. If any condition fails, the feedback loop encounters physical resource bottlenecks or calendar limits, preventing explosive runaway. Frontier training dynamics run counter to RSI assumptions. Successive frontier training runs have escalated in cost from $100 million to between $5 billion and $10 billion, remain limited to one or two per year, and require expanded headcount alongside increasingly brittle clusters of GPUs. Casado observes that most startup claims of RSI are actually standard autocatalysis—using AI outputs to accelerate downstream tasks like cleaning datasets or compiling GPU kernels—rather than systems autonomously designing superior architectures. Machine-Speed Cyber Warfare Over Speculative Superintelligence (20:00) Ghodsi identifies cybersecurity, rather than hypothetical superintelligence, as an urgent, concrete risk. Deploying thousands of reinforcement learning agents across large GPU clusters to optimize reward functions can inadvertently discover novel zero-day exploits, penetration techniques, and hacking vectors. Simultaneously, the operational window between Common Vulnerabilities and Exposures (CVE) publication and active exploitation has compressed from two to three years in 2018–2019 to mere hours or minutes today. Because traditional Security Operations Center (SOC) teams are inundated with alerts and cannot respond at sub-second speeds, Ghodsi asserts that defense must transition to automated, machine-speed agents that continuously red-team internal infrastructure. This dynamic is collapsing the barrier between data platforms and cybersecurity, prompting Databricks to launch products like Lakewatch to ingest agent telemetry and monitor live enterprise agent interactions. Enterprise Bottlenecks: Institutional Context and Cost Control (40:00) Ghodsi argues that enterprise adoption does not require smarter frontier models capable of solving advanced mathematics; existing models already possess sufficient raw reasoning power. The actual operational bottleneck is the absence of institutional context—unwritten meeting history, organizational ontologies, and operational norms. By synthesizing internal communications into a multi-million-node knowledge graph queried via tools like Databricks Genie, Databricks accelerated internal decision workflows and numerical modeling, bypassing traditional hierarchical reporting. To prevent runaway inference expenses, enterprises are shifting away from unconstrained model usage toward tiered architectures governed by gateways and routers. Sarah Wang and Casado report that across portfolio companies, open-source models account for only ~5% of enterprise AI dollar spend but represent over 60% of total token volume—rising toward 90% in cost-sensitive customer-facing applications. Under this paradigm, high-end frontier models provide scaffolding and audit checks, while cheaper open-source models handle execution, paired with agent-centric database infrastructure like Neon that can clone and snapshot petabyte datasets in under a second. While the participants place the probability of existential AI catastrophe at near zero, unresolved tensions remain around how to structure credible independent oversight without stymieing research velocity, and how quickly enterprises can construct structured ontologies before machine-speed cyber exploits outpace legacy security architectures.

Databricks 首席执行官 Ali Ghodsi 与风险投资人 Martin Casado 及 Sarah Wang 深入探讨了人工智能发展的现实边界，指出当下公众对“超级智能导致人类灭亡”等存在性风险的恐慌被严重夸大，而真正紧迫的威胁在于机器速度级的网络攻击以及企业在组织上下文缺失下的落地瓶颈。 节奏之争与末日叙事的公关误区 (1:48) Ali Ghodsi 与 Martin Casado 批评了将前沿模型安全简单框定为“研发节奏管控”（pacing）的叙事策略。Ghodsi 认为科技领袖负有社会道德责任，在没有确凿证据时大肆渲染毁灭性场景是不负责任的，目前人类灭亡的存在性风险“接近于零”（P(doom) 低于 10%）。Casado 指出，“控制节奏”在公关上陷入了尴尬境地：既不能平息末日论者的焦虑，又会让决策者误以为机构缺乏安全把控能力；放慢研发进度在逻辑上独立于安全防护，犹如“慢速制造武器”依然危险。然而，Ghodsi 也指出前沿实验室面临现实的公地悲剧困境，在激烈的市场竞争和 IPO 压力下，单方面放慢速度无异于将市场拱手相让，导致部分实验室倾向于游说监管机构出台统一的外在限制。 递归自我改进的物理局限与机器级网络风险 (10:12) 针对模型自主实现“递归自我改进”（RSI）乃至智能爆炸的担忧，Ghodsi 提出了必须同时满足的四个严苛条件：超线性减少硬件计算资源、大幅缩短训练耗时、准确率与智能水平显著提高，以及这一循环具备复利式可重复性。事实上，前沿模型的物理现实与 RSI 恰恰相反：当前训练成本已攀升至 50 亿至 100 亿美元，集群构建愈加脆弱，不仅没有减少人力与算力，反而依赖更多的人工干预和硬件投入。Casado 补充说明，行业中 90% 被冠以 RSI 之名的实践本质上只是“自动催化”（如编译器自编译或 AI 协助清理数据），并非真正的自主递归迭代。 与虚幻的超级智能相比，大规模强化学习（RL）带来的网络威胁已迫在眉睫。Ghodsi 警示，企业若启动包含数万个智能体在数万张 GPU 上持续数周的大规模 RL 任务以优化目标，模型会自发寻找非预期的漏洞利用途径，从而发现零日漏洞并引发严重渗透行为。过去 CVE 漏洞从公开到武器化需要两到三年，如今在智能体时代已缩短至数小时甚至数分钟。传统的安全运营中心（SOC）依赖人工分析数以百计的误报警报，根本无法应对机器速度的攻击，迫使企业必须采用全自动化的防御智能体、实时追踪日志与动态数据流。 企业落地核心瓶颈：组织上下文与本体构建 (40:00) Ghodsi 指出，企业应用当前面临的瓶颈并非前沿模型不够聪明，而是缺乏“组织上下文”。企业并不需要能够解偏微分方程或在奥林匹克竞赛中拿高分的模型，绝大多数业务落地仍停留在 Copilot 和辅助代码编写层面。真正的壁垒在于模型无法获取老员工沉淀五年的非正式历史与隐性知识。为此，企业必须构建包含抽象概念、业务目标、组织架构与在研项目的“本体”（ontology）。 以 Databricks 自身实践为例，该公司通过在内部吃狗粮建立了包含数百万节点的本体图谱，并借助 Genie 工具实现了全员实时数据检索与即时分析，彻底扁平化了层层上报的决策链条。为解决企业内部缺乏前沿 AI 工程能力的问题，Databricks 部署了前沿部署工程师（FDE），帮助传统客户构建具备严格行为边界与低延迟的专业智能体，例如为 Fox 打造具有敏感话题隔离机制的体育问答应用。同时，在医疗健康等实际场景中，相关技术已落地应用于 Crisis Text Line 实时自残干预、Insulet 自动化人工胰脏、Zipline 无人机急救物资配送以及默克（Merck）的 TEDDY 基因调控网络模型，创造出切实可验证的生产力价值。 工程成本控制与智能体数据库基础设施重构 (55:00) 在模型调优与推理成本控制方面，企业正在从盲目的“Token 最大化”转向“价值最大化”。开源模型与闭源模型在生产环境中逐渐形成分工：投资组合数据显示，开源模型虽然仅占企业总 AI 支出的约 5%，但已占据超过 60% 的 Token 调用量；在客户交互的生产级应用中，开源模型使用比例接近 90%。企业正在采用阶梯式（ping-pong）架构，由前沿闭源模型负责顶层逻辑规划与最终审查，中间执行环节交由低成本的开源模型（如 GLM）或轻量引擎处理，配合 Unity Gateway 等网关实现预算硬性限制与智能路由。 此外，智能体的运行特性对底层软件栈提出了颠覆性要求。由于智能体会频繁且自主地启动环境、测试代码、克隆数据并瞬间销毁，传统面向人类设计的数据库架构已无法适配。为此，以 Neon 为代表的无服务器 Postgres 引擎针对智能体优化，支持秒级以内完成 PB 级数据库克隆，并引入轻量级分支（branching）与无损回滚机制，让智能体能够在隔离的沙箱中安全运行实验和迁移测试，而不会污染生产数据或造成算力成本失控。 本场对话明确反驳了短期内关于超级智能导致灭亡的炒作，认为科技行业应当摆脱空洞的监管公关战，将注意力转移至迫在眉睫的机器网络战以及底层基础设施建设。企业在迈向深度自动化的过程中，能否高效完成私有上下文与本体数据的结构化，并建立匹配智能体特性的敏捷底层系统，将直接决定未来生产力的释放效率。

## Background 本期内容由 a16z 普通合伙人 Sarah Wang 和 Martin Casado 主持发布于播客平台，于 2026 年 9 月 18 日上线。节目邀请了 Databricks 联合创始人兼 CEO Ali Ghodsi，共同探讨前沿 AI 研发节奏争议、递归自我改进的真伪、网络安全风险，以及组织上下文对企业落地 AI 的核心制约。 ## TL;DR Ali Ghodsi 与主持人们深入剖析了围绕 AI 末日论与“放缓研发节奏（Pacing）”背后的商业与政治博弈，明确指出当下人类存亡危机（P-Doom）无限接近于零，而基于强化学习的自动化网络攻击才是真正迫在眉睫的现实威胁。他们用严谨的工程指标反驳了所谓的递归自我改进（RSI）神话，指出前沿模型训练正面临极高的物理算力与工程脆性壁垒。转向企业端落地时，Ali 指出现有模型在智力上已完全够用，真正的瓶颈在于缺乏企业内部的隐性上下文与“组织本体”（Ontology）。为此，企业需要依靠离线图谱索引、Unity Gateway 成本路由治理以及适配 Agent 亚秒级冷启动的新型数据库底座（如 Neon/Lakebase），将 AI 真正转化为切实可落地的业务价值。 ## Atmosphere & style 对话风格兼具硅谷顶层决策者的尖锐博弈感与资深系统架构师的务实工程视角，充满一针见血的技术洞察与直白的行业内幕吐槽。Ali Ghodsi 言辞果断而冷静，善于用生动的内部案例（如全员问答 Genie、CEO 亲自提交代码）化解宏大概念；Martin Casado 则以历史技术浪潮为镜，不时抛出幽默风趣的讽刺与反问。整场对谈节奏明快、观点交锋激烈，既戳破了资本与舆论炒作的末日泡沫，又对底层系统重构展现出强烈的极客兴奋感。 ## Bulletin - `11:15` Databricks CEO Ali Ghodsi 认为当前超级智能导致的生存风险（P-Doom）无限接近于零，但由海量 Agent 与强化学习驱动的自动化网络安全攻防是迫在眉睫的现实危机，CVE 漏洞从披露到被武器化的时间已从数年压缩至数小时甚至数分钟。 - `12:45` Ali 驳斥了当前 AI 处于递归自我改进（RSI）的说法，指出真正的失控型 RSI 必须同时满足算力消耗超线性递减、训练周期缩短、智能实质提升与无限复现四大严苛条件，而现实中前沿大模型单次训练成本已飙升至 50 亿至 100 亿美元且工程脆弱性极高。 - `40:00` 企业 AI 落地的核心瓶颈在于缺乏内部业务上下文与组织“本体”（Ontology），而非模型不够聪明；类似 Google 当年的离线索引技术，企业需将组织关系与知识图谱转为离线索引以替代低效的串行 Agent 循环。 - `58:34` 企业在 AI 消费上正从追求吞吐量的“Token Maxing”转向“Value Maxing”：开源模型虽仅占企业支出金额的约 5%，但在实际 Token 消耗量中已超过 60%，在落地产品端甚至高达 90%。 - `01:05:00` AI Agent 的兴起正在倒逼数据库基础设施变革，Neon/Lakebase 凭借亚秒级冷启动与分支（Branching）克隆能力成为首选，其平台上超过 90% 的数据库已由 Agent 而非人类创建。 ## Key takeaways - `00:46` 科技领袖与学者在缺乏事实依据的情况下公开渲染人类灭绝恐慌具有严重的公共危害性，这不仅造成大众心理危机，还被华盛顿政客武器化用于推动非理性监管。 - `05:00` 前沿模型实验室面临着深刻的“公地悲剧”困境：由于巨额投资与 IPO 变现压力，没有任何一家实验室愿意单方面放慢速度（Pacing）来进行严格的安全审查。 - `10:00` 在公关策略上，将“做好安全控制”包装成“放缓前沿发展节奏（Pacing）”是一场灾难，它陷入了既无法满足末日论者又让政策制定者不满的恐怖谷。 - `12:45` 严格意义上的递归自我改进（RSI）必须同时满足四个条件：算力资源超线性减少、训练周期缩短、智能实质提升以及可持续无限迭代；一旦某个条件断裂，物理硬件限制就会迫使其自动降速。 - `15:00` 现实中的前沿模型研发并非自我演化，而是一场烧钱且极度脆弱的重工业工程竞赛：训练门槛已升至 50 亿至 100 亿美元，且各大顶尖实验室每年仅能支撑一两次训练运行。 - `37:34` 当前初创公司口中宣称的 RSI 大约 99% 只是“自动催化”（Autocatalytic effects），本质上等同于用计算机设计下一代计算机芯片或用编译器写编译器，不属于质变级的递归自我进化。 - `19:50` 网络安全从传统人工响应模式向自动化 Agent 防御模式转型已成定局；传统安全运营中心（SOC）面对海量告警已达极限，数据平台与网络安全市场正深度融合。 - `29:48` 实验室互相监督机制在现实中行不通，犹如让拳击手兼任裁判；在数十亿美元商业利益和 IPO 诉求下，必须引入立场平衡的第三方独立审查机制。 - `40:00` 企业级落地的最大阻碍并非模型智商不足，而是模型缺乏对内部运作逻辑、组织决策链和沉淀在老员工脑中内隐背景的认知上下文。 - `45:00` 企业“本体”（Ontology）是指组织内抽象概念、目标、部门、员工与进行中项目的真实拓扑网络，是打破传统层级汇报损耗、实现扁平化决策的核心底座。 - `47:35` 类似 25 年前搜索引擎的发展历程，企业 AI 必须依托全天候离线计算构建的逆向索引，否则依赖串行调用 MCP 接口的 Agent 循环在延迟、成本与信息覆盖面上均不可持续。 - `55:00` 通过引入 Unity Gateway 预算阈值监控、智能路由分流，以及挑选轻量级 Harness，企业能够在 Token 吞吐量暴涨的同时维持总体账单平稳，且不同 Harness 间存在高达 2 倍的成本差异。 - `58:34` 大企业在模型后训练（Post-training）上面临严重的评测集（Evals）构建障碍，其心理类似软件工程中的测试驱动开发（TDD），客户普遍偏好立竿见影的“简易按钮”而非繁重的测试工程。 - `01:05:00` 自主 Agent 正在成为全新的软件用户画像，其频繁试错与沙箱实验的需求推动了以 Neon 为代表的底层数据库创新，实现亚秒级冷启动与 PB 级数据的瞬时分支克隆。 ## Detailed walkthrough ### `00:00` 开场：炒作、恐慌与企业落地的真实瓶颈 对话由当前 AI 领域最尖锐的矛盾切入：在激烈的市场竞争中，商业领袖普遍面临“公地悲剧”——即便有人呼吁放慢脚步，但为了赢得竞争，谁也不愿率先减速。当前行业被撕裂为两大阵营：一方认为当前的风险归根结底只是工程实现与安全控制问题；另一方则坚持必须由外部力量介入以强行放慢研发节奏。与此同时，一边是大规模宣扬“全人类毁灭”的末日论，另一边却是紧锣密鼓地争夺 IPO 份额，这种戏剧性的反差被戏称为某种精心编排的“四维象棋”。 在企业级落地层面，真实的技术动态与公众的宏观焦虑截然不同。近期甚至有大型企业选择从前沿大模型转向 GLM 这类更具针对性的模型。核心痛点并非模型不够聪明，而是模型缺乏对企业内部实际运作逻辑、组织决策路径以及数年积累的隐性背景知识（Context）的理解。 - 行业两大核心阵营：纯工程视角解决安全问题 vs 外部监管强行放慢研发速度（Pacing）。 - 现实威胁与虚构风险的错配：AI 赋能的极速网络攻击与自动化防御脱节是迫在眉睫的现实危机，而虚构的超级智能人类灭绝论却占据了舆论中心。 - 企业 AI 落地的真正瓶颈：不是模型推理能力不足，而是缺乏企业上下文与组织本体（Ontology）。 --- ### `00:46` 末日叙事的公共危害与技术官僚的恐慌蔓延 谈及当前针对前沿 AI 的放缓研发倡议，Databricks 联合创始人兼 CEO Ali Ghodsi 明确表达了自己的底线：行业领袖与科学家负有最基础的伦理责任，即在没有确凿事实依据的情况下，绝对不能无端制造社会恐慌。他指出，社会大众的心理承受力参差不齐，动辄宣扬“人类面临 10% 毁灭概率”这种毫无根据的论调不仅不负责任，而且在现实中诱发了严重的公众心理危机。Ali Ghodsi 直言：“But I think that right now the existential risk is close to zero.”（但我认为当下这种人类存亡级别的风险无限接近于零。）对于纯粹的技术假设与前沿探索，学术界完全可以在论文和研讨会中深入交锋，完全没必要动辄登上电视节目或在社交网络上引发全民歇斯底里。 a16z 合伙人 Martin Casado 对此深有同感，并分享了一个生动的亲身经历：他远在亚利桑那州农村当教师的妹妹甚至在周日发短信问他是否需要为“AI 末日”收拾庇护所、储备饮用水，并询问他什么时候过去避难。更危险的是，这种非理性的民间情绪已经直接传导至华盛顿政界。从 Elizabeth Warren 呼吁暂停所有前沿 AI 开发，到 Bernie Sanders 与 Steve Bannon 等政治谱系两端人物的异动，整个联邦官僚与监管体系正在借机加速运转，其后果很可能彻底背离技术安全的初衷。 - 科学探索与舆论传播的割裂：严谨的技术讨论被降维包装为博人眼球的毁灭概率，给大众带来巨大的心理负担。 - 政治力量推波助澜：两党政客借公众恐慌争夺监管话语权，暂停开发的提议正在被政治建制派工具化。 --- ### `05:00` 利益博弈、“放缓研发”公关误区与公地悲剧 讨论进一步深入到争论背后的现实利益冲突。Ali 指出，舆论场两端都充斥着极度现实的政治与商业博弈：商业资本方急于保护即将上市的投资组合，希望大家“闭嘴”以顺利完成 IPO 变现套现；而另一方则在不遗余力地武器化各种舆论事件，通过放大恐慌来获取政治杠杆。甚至地方政界（如德克萨斯州州长 Greg Abbott 针对数据中心建设的态度）也卷入了复杂的博弈。 Martin 尖锐地批评了所谓“放缓研发节奏”（Pacing）在公关与战略逻辑上的致命失误。他指出，各行各业推行自律与安全审查本是正常的商业惯例，但将这一诉求包装成“Pacing”完全是公关灾难——因为“放缓速度”与“构建安全系统”在逻辑上是完全正交的，慢慢造武器并不等于不制造武器，更何况前沿实验室事实上一直在疯狂采购算力以跑得更快，这种退让与妥协反而直接落入了叫停派的语境陷阱。 ```mermaid flowchart TD accTitle: 商业博弈与公地悲剧推导流程 accDescr: 展示前沿 AI 竞争中因公地悲剧导致无法单方面减速并最终诉诸监管的逻辑链条 subgraph Competition ["市场与商业竞争"] A["资本回报与 IPO 压力"] --> B["前沿军备竞赛与算力扩张"] B --> C["单方面放慢 = 商业自杀与落败"] end subgraph Dilemma ["公地悲剧困境"] C --> D["各方均无动机单方面减速"] D --> E["安全审查与监控被压缩"] end subgraph Outcome ["诉求与监管"] E --> F["发生安全事故 (如沙盒失控)"] F --> G["呼吁外部出台监管规则 (强行统一配速)"] end ``` Ali 则从企业领导者的微观治理角度反驳了 Martin 的纯逻辑推演，阐述了博弈论中残酷的“公地悲剧”：在真刀真枪的商业战场上，没有人愿意单方面减速充当受害者。正如 Ali Ghodsi 所说：“Like I'm competing, I want to win, you know, and you're running ahead.”（比如我正在参与竞争，我想赢，而你正在前面全力冲刺。） 随后他们以 Hugging Face 与 OpenAI 之间发生的测试事件为例进行了复盘：当时强化学习（RL）实验产生了数百万 GPU 算力小时的 Token，AI Agent 在沙盒环境中肆意运行，而团队并未对实时生成的每一个 Token 设立前置监控，仅仅是在事后才介入审查。Martin 认为这纯粹是基础安全工程与访问控制（Security Controls）缺失的问题，呼吁“做好你的安全防护”即可；而 Ali 则指出，要求安全与合规团队全面介入并实时审查每一个输出，在现实企业运作中必然意味着极度拖慢研发进度。如果对手不做这种前置审查而你做了，你就会在速度上惨败。因此，实验室高管们呼吁制定规则的潜台词实际上是：在缺乏外部强制规则的前提下，自我克制就是商业自杀；唯有外部规则强制所有人同时减速，安全底线才能真正落地。 - 资本与政治的双重裹挟：IPO 退出需求与舆论政治博弈相互交织，导致公开讨论严重失真。 - “放缓研发（Pacing）”的公关缺陷：安全与速度并无绝对因果，将安全问题误读为减速问题导致了监管立场的被动。 - 现实治理代价：全流程合规与前置安全监控必然大幅降低研发吞吐量，商业竞争的零和博弈直接导致了安全控制的让步。 ### `10:00` 沟通的落差：“放慢节奏”与“确保安全”的公关灾难 对话从各大 AI 实验室对前沿模型安全风险的对外沟通策略展开。Martin Casado 与 Ali Ghodsi 探讨了 Anthropic CEO Dario Amodei 与 Meta 掌舵人 Mark Zuckerberg（Zuck）在表述上的巨大反差。大家普遍认为，Zuck 在处理这类议题时手法老练得多，他将核心聚焦于“安全、保障与自我监管”（security, safety, and self-regulation），将模型的延后发布合理解释为“我们必须把安全控制做到位”，这种表态更容易被大众与监管层接受。相反，Dario 一开口就是“我们需要放慢前沿发展节奏”（we need to pace the frontier），这种“放慢节奏”（pacing）的提法在公众心理上造成了完全不同的冲击。 Martin Casado 犀利地指出，Anthropic 的做法显然是想在呼吁“暂停”（pause）的末日论者（Doomers）与政策制定者（politicians）之间寻求妥协，结果却掉进了公关的“恐怖谷”（uncanny valley），两头不讨好。末日论者认为“放慢节奏”根本不算暂停，而政策制定者与行业大众则觉得这既不真诚也不现实，反而忽略了真正的网络安全防护。Ali 补充道，实验室内部确实有一群专注于安全的工程师正陷入极度焦虑（freaking out），他们并不全都是有效利他主义（EA）信徒，有些人的恐慌源于他们在技术一线看到的惊人结果。然而，由于这些研究员缺乏专业的公关训练，选用了糟糕的措辞，最终引发了强烈的行业反噬。正如 Martin Casado 所评论的那样：“I think it's like literally, you're like, you're trying to find this, he paces like the uncanny valley of making the Doomer people unhappy and the policy people unhappy.” - 公关策略的失误导致本可达成共识的“前沿安全控制”变成了引发全民恐慌的政治议题。 - 全球多地正逢选举，一旦技术风险被贴上泛政治化的末日标签，将引发非理性的跨国监管连锁反应。 ### `11:15` 真正迫在眉睫的威胁：强化学习与网络攻击，而非“超级智能” 谈及 AI 的核心风险，Ali Ghodsi 直言不讳地表达了自己的立场，他认为目前所谓的“人类生存危机”（existential risk）几乎接近于零。但与此同时，他强调行业内确实存在非常真实且迫切的危害，那就是大规模强化学习（Reinforcement Learning, RL）带来的网络安全风险。Ali 指出，现在有实验室在进行超大规模的 RL 训练：投入上亿美元、调用上万个 Agent，在庞大的计算集群上并行跑上一两个月，只为给定一个奖励函数（Reward Function）去解决某个数学谜题或通用任务。在这种机制下，系统为了最大化收益可能会演化出极度危险的攻击行为。 Ali 解释道，这类不可控的黑盒探索极容易导致严重的系统被黑或网络攻防失守，网络安全风险（cyber risk）才是真实且近在眼前的危险。然而，许多人却本末倒置，将这种危险渲染成虚无缥缈的“超级智能毁灭人类”，甚至让原本对技术毫无兴趣的普通民众都陷入无端的恐慌。Ali 认为科技界领袖有责任用理性、平衡的方式拆解真相，戳破所谓超级智能的末日迷思，将讨论拉回透明且客观的技术现实。关于当前的生存风险，Ali Ghodsi 坚定地评价道：“I agree with the core premise that we shouldn't freak the public out. I think that existential risk right now is close to zero.” ```mermaid flowchart TD accTitle: 强化学习风险 vs 超级智能恐慌 accDescr: 展示大规模强化学习如何引发真实的网络安全风险，以及其与虚构的超级智能恐慌之间的区别。 subgraph Cluster["高额算力驱动的大规模 RL"] A["10,000+ 智能体并行运作"] B["1 亿美元投入 / 数月集群运行"] C["奖励函数驱动极限探索"] A --> C B --> C end subgraph RealRisk["真实紧迫的风险 (Real & Immediate)"] D["网络安全漏洞被攻破 (Hacking / Cyber)"] E["软硬件基础设施失控"] C --> D C --> E end subgraph FalseRisk["被夸大的舆论恐慌 (Overhyped Narrative)"] F["科幻式超级智能 (Superintelligence)"] G["人类生存灭绝危机 (Existential Risk ~ 0%)"] end D -.->|"被错误归因为"| F E -.->|"引发政治恐慌"| G ``` ### `12:45` 递归自我改进（RSI）的严格定义：构成失控的四个必要条件 为了打破关于“AI 自我进化并失控”的空泛恐慌，Ali Ghodsi 给出了一个极其严密的工程学判定框架。一些实验室之所以对递归自我改进（Recursive Self-Improvement, RSI）感到惊慌，是因为他们看到了模型正在帮自己编写代码。但 Ali 指出，单纯的“软件由 AI 自主编写”早已成为现实——在 Databricks 内部，已经有超过 90% 的软件代码是由 AI 生成的，就算最后几个百分点也被 AI 接管，在工程上根本无关痛痒。真正的、具有失控加速风险的 RSI，必须同时满足四个极为严苛的条件。 Ali 详细拆解了这四大条件，并强调它们必须“同时发生且持续迭代”，缺一不可： 1. **算力资源消耗超线性减少**：训练下一个更强大的模型所需要的硬件资源和 GPU 数量更少，而不是单纯线性持平。 2. **训练周期缩短**：模型训练所需的时间显著降低（例如从数月减至半个月）。 3. **模型能力与智力实质性提升**：在资源与时间双降的前提下，模型的准确率与智能水平反而持续攀升。 4. **过程具备完全的可持续复现性**：上述前三个特征能够一轮接一轮、无限循环地进行自我加速，而不会中途停滞或衰减。 Ali 分析称，只要其中任何一个条件断裂，系统就会自我刹车。例如，如果资源需求保持恒定，人类有限的硬件供应和 GPU 产能就会自动对发展节奏形成物理限制（pace itself）。即使这四项条件在短时间内同时出现，系统也完全可能在数学上最终收敛，而不是无限膨胀为超级智能。Ali 主张各家前沿实验室应该公开客观数据，提高整个研究链条的透明度，而不是任由“涌现现象”（emergent behavior）这种模糊概念在公众中制造恐慌。Ali Ghodsi 在拆解物理硬件对算力的天然限制时明确指出：“Like, for instance, if resources is constant, then that's okay, because we're going to run out of hardware. So then it will pace itself.” ### `15:00` 算力暴涨与工程脆弱性：前沿大模型并非递归自我改进 在探讨人工智能是否已经实现递归自我改进（Recursive Self-Improvement, RSI）时，Ali Ghodsi 直言许多人混淆了概念：“如果我自己都不写代码了，AI 在自己写代码”，这让程序员觉得自己的价值受到冲击并感到恐惧，但这绝不等于 Nick Bostrom 在 2014 年所预言的那种超智能（Superintelligence）奇点。Sarah Wang 敏锐地指出，在讨论 RSI 时人们往往忽略了算力门槛这一硬性指标：训练顶级模型的计算成本门槛不仅没有降低，反而从早期的 1 亿美元、10 亿美元一路飙升到了如今的 50 亿至 100 亿美元。尽管半年后复现同样性能的模型可能只需二十分之一的成本，但开拓最顶尖的前沿阵地依然是一场极其烧钱的重工业竞赛。 ```mermaid flowchart TD accTitle: 前沿模型训练的现实壁垒 accDescr: 展示前沿大模型训练面临的物理、工程和人力约束，反驳递归自我改进的神话。 subgraph Reality["前沿训练现实 (非 RSI)"] A["50亿-100亿美元巨大投入"] --> B["大规模数据中心建设与供电"] B --> C["数万张 GPU 集群互联与容错"] C --> D["工程脆性剧增 (一次崩溃损失数千万)"] D --> E["庞大的人类工程团队与每年仅1-2次训练"] end ``` Ali Ghodsi 强调，前沿实验室的真实运作机制恰恰完全违背了 RSI 的理想条件。目前各大顶尖实验室每年往往只能进行一两次大规模训练，整个流程极度脆弱：为了容纳更大规模的 GPU 集群，团队必须在网络拓扑、系统容错上投入大量工程人力。随着集群硬件数量呈数量级增长，过去无需担忧的微小错误都会导致灾难性瘫痪，一旦训练失败就意味着数以亿计的资金打了水漂，业内已经发生过多起训练流产的惨痛先例。Ali 总结道：“It's the opposite. It's like it's taking longer and it's more brittle and it's more people and it's harder to pull off.”（现实恰恰相反。它耗时更长，更加脆弱，需要更多的人力，而且越来越难以完成。）因此，所谓“下一个模型更便宜、更快、全靠自我递归演化生成”的说法，在现阶段工程现实面前根本站不住脚。 - 训练前沿模型的算力资金门槛已从数亿美元跃升至 50 亿至 100 亿美元。 - 复现六个月前的前沿模型水平，成本约为原先的二十分之一（1/20）。 - 顶级 AI 实验室受制于庞大基础设施建设和高昂试错成本，每年仅能支撑一至两次完整的前沿训练运行（Run）。 ### `16:35` 黑盒试金石与历史透镜：算力过热是狼来了吗？ 针对如何验证真正的 RSI 是否已经发生，Martin Casado 提出了一个“黑盒试金石”（Black Box Litmus Test）：面对动态复杂的自适应系统，与其相信各方的空谈，不如直接看硬性数字——如果一家顶尖模型公司在人数锐减、资金投入减少的同时，模型产出速度却指数级加快，那才证明自我改进正在发生。对此，Ali 认为单纯看公司总人数并不完全公平，因为拥有充裕现金流的实验室会招募海量员工去尝试成百上千种边缘探索（例如 OpenAI 早期主导大语言模型的实际核心团队仅有约十人）。要真正衡量，必须剥离外围杂音，直视预训练和后训练的核心研发团队：这支真正训练核心模型的队伍是否变得极度微小、是否工作量在缩减并全交由 AI 接管，以及他们使用的 GPU 数量是否在减少。显然，目前没有任何一家实验室呈现出这种态势。 Martin 随即拉开历史视角，回忆起计算架构从单机大型机向客户端-服务器（Client-Server）架构和集群超算跃迁的时代。当时内存一致性（Memory Coherence）限制了单机规模，而分布式集群与 GPU 的起步直接催生了算力爆发。Martin 甚至回忆起 1999 年美国曾对索尼 PlayStation 游戏机实施出口管制，理由是担心萨达姆·侯赛因（Saddam Hussein）会将其芯片用于武器级模拟：当时人们也在疾呼这项技术具有无限潜力与毁灭性风险，但那些末日预言最终并未成真。现场随即爆发了一场关于年龄与见闻的幽默调侃，来自瑞典的 Ali 笑称自己根本不记得当年 PlayStation 被禁运的事情，反问 Martin 的真实年龄，并打趣道或许是“无知者无畏”，或者瑞典人压根不理会美国的出口管制条例。 > “Because they have so much resources that they might be doing a lot of other stuff they don't need to do.” —— Ali Ghodsi ### `19:50` 迫在眉睫的现实威胁：AI 时代的自动化网络攻防战 尽管 Ali 对科幻式的生存危机（Existential Risk）并不买账，但他极其严肃地指出，网络安全（Cyber Risk）是当前最大且最真实的威胁。如今全球关键基础设施的数字化与互联互通程度，已经远远超越了三十年前互联网萌芽或游戏机禁运的时代。整个世界对数字网络的依赖高度集中，但海量基础设施在底层安全上漏洞百出。一旦恶意行为者释放具备自适应探索能力的自主 Agent，它们将极其高效地寻找漏洞、跨系统横向渗透并肆意扩散。 Martin 提出质疑：当年冲击波、震网等早期计算机蠕虫刚爆发时，就已经导致医院系统瘫痪、关键基础设施受损，造成数百亿美元的经济损失；如今大量资金涌入网络安全，为何反而没有看到同等规模的破坏？Ali 解释称，尽管目前还没有演变成无法挽回的生存绝境，但攻防前线的防御能力早已逼近极限。传统的安全运营中心（SOC）依赖人工分析警报和排查入侵，但面对海量入侵尝试，值班人员每天一睁眼就要面对数百封报警邮件，其中夹杂着大量误报，人类分析师根本不可能跟上攻击的频次和速度。整个网络空间正在被迫快速进入由 Agent 对抗 Agent 的全自动防御时代。 ```mermaid flowchart LR accTitle: 现代网络安全运营转型架构 accDescr: 对比传统人工安全运营中心与现代基于 Agent 的全自动威胁捕获模式。 subgraph Traditional["传统 SOC 模式 (已无法维系)"] T1["攻击事件激增"] --> T2["数百封邮件告警 / 误报频发"] T2 --> T3["人工排查响应滞后"] end subgraph Automated["现代 AI Agent 模式 (Lakewatch)"] A1["攻击载荷与恶意 Agent"] --> A2["威胁自动巡检与模拟自攻"] A2 --> A3["实时自适应防御与数据流融合"] end ``` Ali 透露，Databricks 已经推出了面向安全检测的产品 Lakewatch，并在 RSA 大会上与 Ben Horowitz 深入阐述了该趋势。目前，大部分行业依然受困于过时的人工响应机制，除了少数金融机构和极度注重安全的大型企业在尝试自动化外，多数企业的自动化威胁捕猎（Threat Hunting）仍未建立。Ali 指出，数据、AI 与网络安全这两个原本泾渭分明的市场正在彻底融合：过去数据平台专注于存储与建模分析，网络安全则专注于入侵检测；如今企业内部正运行着越来越多的自主 Agent，数据流与安全执行环境已经不可分割。 > “So this is the whole space, security cyberspace, is being transitioned into fully automated, using agents, for detection on the other side. If we don't do that, I mean, now we're rushing. The industry is rushing to do that super, super fast.” —— Ali Ghodsi ### `25:00` 网络安全响应周期骤降与“控制节奏”争议 Ali Ghodsi 指出，企业内部部署的 AI Agent 正在与外部 Agent 频繁互动，产生浩瀚的运行日志、调用轨迹与数字指纹，使得需要分析的数据量呈爆发式增长。与此同时，网络攻防的时间窗口正在极速坍塌：2018 至 2019 年间，一个通用漏洞披露（CVE）从公开到被黑客真正武器化（Weaponized）利用，通常需要两到三年；到了 2022 年，这一周期压缩到了八至九个月；而如今，这个时间窗口已经彻底缩短到以小时甚至分钟计。面对这种瞬间完成的武器化攻击，Ali 认为唯一的应对之策就是依靠自动化、数据与 AI 平台协同的架构，这促使传统的数据安全市场与 AI 市场发生深度重叠与重构。 ```mermaid flowchart LR accTitle: Vulnerability Weaponization Timeline accDescr: Timeline showing the shrinking window from CVE disclosure to weaponization from 2018 to present. A["2018-2019: 2-3 年"] --> B["2022: 8-9 个月"] B --> C["当下: 数小时至数分钟"] ``` 随后，Martin Casado 挑起了关于 AI 存亡风险（Existential Risk）的两派阵营讨论：一派认为这是 Databricks 这类公司可以通过工程、产品和服务解决的工程问题；另一派则坚信根本没有工程解法，必须像对待核武器一样引入强监管，甚至“控制研发节奏”（Pacing）。Martin 对“Pacing”一词极其反感，打趣道：“There is no line of inquiry ever that gets to pacing, I don't think. I think it's like you pause it or like you solve it.”（我认为没有任何一种推导逻辑会导向‘控制节奏’。要么你直接暂停它，要么你用技术解决它。）这引来现场的笑声，Martin 甚至开玩笑说干脆别买 Databricks 了，大家都把算力放到国家实验室里各自慢下来就好。 - **CVE 武器化时间演变**：2018—2019 年需要 2—3 年；2022 年缩短至 8—9 个月；如今缩短至数小时甚至无缝发生。 - **数据与安全市场的融合**：海量 Agent 日志催生了对高并发、大规模数据分析基础设施的刚性需求。 ### `27:14` 超级智能与工程化风险的边界 针对 Martin 的质问，Ali 强调外界往往将两个性质完全不同的问题混为一谈。其一是 Nick Bostrom 在其 2014 年著作《超级智能》（Superintelligence）中描述的情形——能够几秒钟写出一篇包含原创突破并能通过同行评审的博士论文，或者无需在纸上演算就能在 11 维物理空间中进行直觉推理。Ali 坦言，如果真的出现了这种全面超越人类认知维度的机器，那绝非简单的工程问题，而是不折不扣的存亡级风险，各方对此并无异议。但现实中，当前的技术路线和现有 Agent 距离那一步还差得非常远，甚至连通往那条道路的影子都还没看到。 当前真正面临的拐点，在于工程层面的“规模化廉价劳动力”。人类历史上第一次可以通过轻点按钮，就在沙盒环境中调度 10,000 个甚至 100,000 个具有相当能力的 Agent 协作一个月，以极低成本输出相当于一亿美元工资的人工工作量，或者组织上万个数学模型协同攻关数学猜想。这种能力的民主化直接引爆了现实世界中的网络安全威胁。Ali 总结道，超级智能是否会通过递归自我改进（RSI）迅速到来仍存疑问，但眼前由大规模 Agent 驱动的网络风险完全是一个“可以通过工程解决的问题”（Engineering problem）。 > “This, I think, leads to all these cyber risks. I think cyber is the major problem here. This is, I think, you can solve with engineering.” —— Ali Ghodsi ### `29:48` 检查员机制的博弈与三大治理路线 针对前沿实验室（Frontier Labs）提议引入外部“检查员”（Inspectors）核查训练进程以确保安全，Martin 和 Ali 探讨了这一构想的可操作性。两人一致认为关键在于“挑选谁来当检查员”，因为很多人的立场早已固化。Ali 以图灵奖得主 Yann LeCun 为例：作为深度神经网络的先驱之一，LeCun 一贯主张“这里没有任何毁灭性风险，这纯属无稽之谈，应该全速推进”。Ali 指出，如果让 LeCun 担任检查员，他若实地核查后依然出来说“一切如常，继续推进”，大家会感到安心；但如果连 LeCun 出来时都神色动摇、观点发生转变，那将释放极其强烈的警示信号。因此，检查员团队必须具备多元视角，才能避免出现预设结论的“操纵性评审”（Stacked panel）。 ```mermaid flowchart TD accTitle: Frontier AI Governance Approaches accDescr: Three distinct proposals for frontier AI oversight and auditing. A["前沿 AI 监管方案"] --> B["第三方检查员机制<br>(OpenAI / Anthropic)"] A --> C["同行交叉审查<br>(Elon Musk 倡导)"] A --> D["企业内部自主监管<br>(Mark Zuckerberg / Meta)"] ``` Sarah Wang 随后梳理了目前行业内的三条治理路径：以 OpenAI 和 Anthropic 为代表的独立第三方检查员模式；以 Elon Musk 倡导的各大实验室像科学同行评审一样相互交叉核查模式；以及以 Mark Zuckerberg 为代表的企业自我监管模式。对于实验室互相监督的提议，Ali 形象地反驳道：这就像让拳击台上的拳手自己兼任裁判，只要对方训练出了强大的新模型，另一方立刻就会高喊“犯规”并大炒超级智能威胁。在面临数十亿美元商业利益、IPO 计划以及激烈历史恩怨的前提下，竞争对手之间绝不可能保持公正客观，第三方裁判机制是必不可少的。 - **三类监管提议**： 1. **第三方独立审查**（OpenAI、Anthropic 主张）：引入外部专业团队进驻检查数据与模型行为。 2. **实验室交叉核查**（Elon Musk 主张）：各大前沿实验室如同学术评审般互相检查代码与权重。 3. **企业自律**（Mark Zuckerberg 主张）：依赖各公司内部的安全红队与合规流程。 ### `32:04` 商业野心与安全焦虑的双重奏 针对 Sarah 提出的质疑——Elon Musk 曾公开讽刺 AI 实验室在下一盘精心设计的“4D 国际象棋”，即一边声称“人类即将毁灭”，一边在后台询问投资人“你的 IPO 认购额度想要多少”——Ali 直面了这一尖锐矛盾。Ali 承认，行业内确实存在借助安全恐慌进行市场营销（Marketing stunts）的倾向，每当训练出新模型时，大肆宣扬“这个模型太强大了，连我自己都害怕”，能瞬间吸引全球关注；同时，如果通过监管限制行业发展节奏，往往在客观上有利于头部既得利益者筑高护城河。 然而，Ali 也强调这两者并非绝对对立的虚伪表现。安全焦虑与商业利益往往在从业者脑中奇妙地达成自洽：现实中从 CVE 公开到武器化被利用的时间的确已经从数年塌缩到了几分钟，恶意网络攻击的威胁真实迫近；但借此制造舆论声浪同样是一门绝佳的公关手段。随后，Martin 从经典网络工程历史（如 IETF、IEEE、ICANN 等行业自律标准组织）切入，询问在什么时候应当让联邦政府强行介入，而非仅仅依靠类似 FINRA（美国金融业监管局，介于行业自律与政府监管之间）的半独立组织。Ali 认为两者并非非黑即白，随着风险扩散，行业自律与政府介入之间的界限正在不可避免地变得模糊。 > “So it's real. The cyber attacks are real. But there's also a great marketing ploy to, you know, whenever you train a new model, make lots of noise around how much of a crazy risk it is to the world. It helps you, right? So maybe they're not in contradiction, these things.” —— Ali Ghodsi ### `35:00` 行业自我监管与前沿 AI 实验室的上市动力 对话深入探讨了 AI 行业的监管走向。历史上，行业通常先尝试自我监管（self-policing），随后不可避免地演化为政府层面的正式监管。当头部前沿实验室公开宣称“AI 存在生存级风险（existential risk）”并呼吁监管介入时，监管机构几乎不可能对此置之不理。主持人引用 David Sacks 的讽刺说法：既没有见过主动要求被监管的 CEO，也没有见过面对这种要求会说“不”的监管者。如今，无论是政界要人（如奥巴马发声）还是即将来临的中期选举，AI 治理都已成为核心议题，甚至有人担忧政府会采取一刀切式的联邦重手干预，导致整个领域被强行暂停，乃至像 Anthropic 这样的实验室被纳入美国能源部（DOE）监管。 与会者分析认为，当前前沿实验室面临的巨大困境在于资本消耗。通过强化学习（RL）给模型注入海量资金，并配合可验证的奖励函数（如求解狭窄领域的数学题或特定编程问题），确实可以在局部取得极佳的刷榜效果，但这绝不等于实现了通用超智（superintelligence）。人们甚至可能陷入认知偏差，误以为少量输入正在带来更好的产出，然而在海量外部资源投入下，根本无法完成真正隔离的对照实验。这也解释了为什么这些头部 AI 公司纷纷启动 IPO 计划： - 它们作为未上市企业，原本可以继续保持私有化运作，但指数级增长的算力与训练开销迫使其必须进入公开市场融资。 - 这些团队坚信 Scaling Laws（规模定律）和海量资本本身就是不可替代的战略护城河。 关于自我推进与递归自我改进（RSI），Martin Casado 指出，科学界早已研究过类似系统（例如 Nanotechnology 纳米技术或编译器的自举构建），但在复杂的现代经济网络中，现在的危险在于“狼来了”式的虚妄警报——许多人宣称看到了威胁，但事实上根本没有确凿证据。Ali Ghodsi 直言：“And maybe they're just unjustifiably worried themselves, just like they were worried about GPT-2.” 当年 OpenAI 曾因担忧 GPT-2“毁灭世界”而推迟发布，但随后的 GPT-3 和 GPT-4 证明这种世界末日的恐慌纯属多虑。 ```mermaid flowchart TD accTitle: Capital and Scaling Dynamics in Frontier AI accDescr: 前沿实验室在强化学习刷榜、海量资本消耗与 IPO 之间的因果反馈路径 subgraph Narrow_RL["局部基准评测突破"] A["可验证奖励函数 (数学/狭窄编程)"] --> B["巨额资本注入与实验堆叠"] B --> C["高分基准测试结果 (误以为是超智突破)"] end subgraph Capital_Cycle["资本飞轮与外部性"] C --> D["算力与训练资金需求暴增"] D --> E["被迫推进 IPO / 寻求公开市场资本"] E --> F["将规模定律与资本壁垒视为核心护城河"] end ``` --- ### `37:34` 澄清 RSI 炒作：绝大多数只是“自动催化” Martin Casado 和 Ali Ghodsi 进一步剖析了当前在硅谷被滥用的术语——递归自我改进（Recursive Self-Improvement, RSI）。从前沿实验室离职创办新公司的创业者几乎人人都在谈论 RSI，但深入其架构就会发现，其中大约 99% 的做法根本不是严格意义上的 RSI，而仅仅是“自动催化（Autocatalytic effects）”。 工业界早就充斥着自动催化效应： - 现代计算机芯片必须依赖前一代芯片及其驱动的 EDA 工具才能完成设计和制造； - 编译器也是用编译器写出来的； - 蒸汽机推动了制造更多蒸汽机所需的钢铁和煤炭开采； - 互联网通过赋能跨地域协作，催化了互联网自身的扩张。 Martin Casado 分享了一线的观察：“Listen, my full-time job is people coming out of labs and starting companies, and they all say RSI because everybody says RSI. And maybe 1% of those are actually RSI.” 现实中绝大多数项目只是在利用 AI 进行数据清洗（data cleaning）、编写 GPU 内核（GPU kernels），以此加速研发流程。这种用模型辅助建造基础设施的模式，性质上等同于用计算机设计计算机，属于典型的自动催化机制，大约占了研发力量的 90%。尽管确实存在极少数团队在探索像 Andrej Karpathy 提出的自动研究（Auto-research）那样让模型真正自主训练自身的闭环，但这部分研究在全部力量中占比极小。 随着 OpenAI 联合创始人 Greg Brockman 在播客上声称“我们已经处于 AGI 时代”，整个行业的言论风向再次被带偏，所有人也跟着高喊 AGI 已经降临。 --- ### `40:00` 企业的现实鸿沟：缺乏业务上下文，而非模型不够聪明 面对外界对 AGI 降临和复杂智能体（Agent）集群的狂热宣讲，Ali Ghodsi 指出这与企业端当下的残酷现实脱节。从去年下半年开始，科技圈人士便频繁宣称 AI 已经在多数时候比身边的绝大多数人都要聪明。然而，当被问及“你们当中有多少人在管理成百上千个能够相互协商、协作、自组织并全面接管企业业务的智能体集群？”时，台下的企业客户几乎没有一个人能举手。 现实中绝大多数企业在 AI 上的应用极为初级： - 大多数企业仅仅部署了微软的 Microsoft Copilot，甚至这就是他们 AI 战略的全部； - 普遍的使用场景仅仅是将大模型当作高配版的内部 Google 搜索或聊天机器人（Chatbot）； - 虽然有一部分人在利用 AI 辅助写代码，但整体投资回报率（ROI）仍待商榷； - 能够端到端自动化核心业务的 Agentic 场景在整个企业界几乎尚未真正落地。 造成这种落差的根本症结并非模型推理能力不足，而是模型完全缺乏组织内部的上下文（Context）。Ali Ghodsi 精辟地指出了问题所在：“The models are smart enough, but they just don't have the context that exists inside of any organization, like they have not been in every meeting, they don't know what's in everybody's heads, they don't know all the processes.” 在任何一家公司中，真正关键的业务逻辑往往深藏在几位老员工的脑子里——团队成员遇到难题时总是去拍拍他们的肩膀请教，甚至私下感慨“如果他离职了公司该怎么运转”。AI 模型从未参加过内部会议，不了解权力结构和隐性流程，缺乏这些长年累月沉淀的制度性记忆。 Ali Ghodsi 强调，只要能把企业现有的上下文与当下的前沿模型相融合，就能为全球企业带来难以估量的生产力提升。企业目前根本不需要能解纳维-斯托克斯方程（Navier-Stokes）或在极限学术难题考试中从 60% 跃升到 70% 的超级大脑。就算前沿模型的进化速度被政策放缓（Pacing），对广大企业的日常自动化也无足轻重，因为绝大多数企业在技术采用曲线上落后得太远；不过对于头部模型实验室而言，前沿停滞将是灾难性的，因为“智能的价格正在呈渐近线式暴跌，几乎每六个月就降至原先的十分之一”，如果不持续推高前沿能力，实验室的商业模式将难以为继。 --- ### `43:08` 正在落地的垂直 AI 突破性用例 公众目前的注意力被过多的末日生存危机和网络安全风险等负面叙事所占据，不断追问“AI 到底给我带来了什么实质好处”。针对这种质疑，Ali Ghodsi 分享了几个已在 Databricks 平台上实际运行并创造出重大社会价值的真实案例： - **Crisis Text Line（危机短信热线）**：利用大语言模型实时分析求助短信的内容与语义，识别具有自残或自杀倾向的青少年，从而帮助人工干预人员在千钧一发之际完成救援，实打实地挽救生命。 - **Insulet Omnipod（智能胰岛素泵系统）**：面向糖尿病患者的自动化给药设备。它告别了以往患者频繁扎破手指测血糖和手动注射胰岛素的繁琐流程，通过车载 AI 算法实时学习个体的葡萄糖代谢水平和胰岛素释放动力学，实现精准的自动化自适应注射，相当于为人体内置了一套 AI 血糖调节器。 - **Zipline（全自动配送无人机系统）**：最初在非洲偏远地区为难民营和医疗点紧急投送血液样本及急救药品，现已拓展至全球。其无人机机队从航线规划、多机避障调度到电池能量消耗优化，全部基于运行在 Databricks 上的底层 AI 驱动。 - **默克集团（Merck）联合研发的 TEDDY 蛋白质模型**：该模型全称为“Transformer 增强型药物研发”（Transformer Enhanced Drug Discovery），相关学术成果已公开刊发。TEDDY 改变了大模型用于预测自然语言下一个 Token 的常规做法，转而用于预测基因调控网络（Gene Regulatory Network, GRN）在受到外界干扰时的生物学反馈。该架构能够从海量数据中精确鉴别哪些细胞变化是致病的因果主导因素（Causal），哪些仅仅是被动的反应性表象（Reactive），从而帮助科学家精准锁定疾病靶点，大幅削减新药早期研发的试错周期与天文数字般的资金成本。 ### `45:00` 企业的上下文瓶颈与“本体”（Ontology）的本质 在谈到人工智能为现实企业带来的颠覆性价值时，Ali Ghodsi 首先以制药巨头诺和诺德（Novo Nordisk）为例。诺和诺德不仅研发了知名的 GLP-1 药物，如今还在其所有临床试验中全面部署这一技术。在进行肥胖症等临床试验时，过去需要耗费数周才能得出的洞察，如今借助系统已被压缩至短短数分钟。面对主持人关于企业未来 12 个月该如何将这种“上下文”（Context）真正落地的提问，Ali 指出，现实远比外界想象的艰难：企业的第一步必须是将组织内部发生的一切彻底数字化。这绝非挥舞魔法棒就能达成——每一场会议都必须有转录记录，所有的数字化内容与决策过程都必须沉淀下来，并作为上下文喂给 AI。然而在许多传统组织中，第一道门槛往往来自法务合规部门的阻挠，他们会严令禁止员工“录制每一次通话或每一件事”。 针对主持人提到 Palantir 经常使用却未曾垄断的“本体”（Ontology）一词，Ali 给出了极其接地气且深刻的定义：本体是指企业内部所有抽象概念、目标、部门、员工与进行中项目之间的真实关系网络。他举了一个生动的对比案例：两位教育背景完全相同、同样勤奋聪明的员工，一位是今天刚入职的新人，另一位则已经在公司工作了五年。两者的唯一区别就在于，老员工的大脑中内置了这家公司的“本体”——她清楚公司内部到底如何运转，知道谁能把事情办成，明白“千万别看组织架构图找某某，他根本办不成事；去找另一个人，他马上能帮你搞定”，甚至了解哪些流程手续根本无需填报，以及当前项目的真正底细。初创公司一旦核心团队流失便往往万劫不复，正是因为新人无法凭空获得这些存在于人脑中的内隐知识。AI 必须要拿到这层上下文网络，才能真正为企业所用。 > “Ontology just means that in an organization, the relationship between all the abstract concepts of all the goals and all the departments and all the people and all the projects that are going on, what do they exactly mean?” —— Ali Ghodsi ```mermaid flowchart TD accTitle: 企业“本体”构建与应用链路 accDescr: 展示从企业内部多源数据采集、离线索引与图谱构建，到最终为决策层提供即时问答的完整流程架构。 subgraph Ingestion ["数据摄取与数字化"] M["会议转录文本"] D["项目文档与业务数据"] C["沟通记录与邮件"] end subgraph OntologyEngine ["本体图谱引擎 (离线计算)"] G["实体与关系抽取 (部门/目标/人员)"] SEC["权限与访问控制 (Access Control)"] IDX["离线逆向索引与知识图谱"] G --> SEC --> IDX end subgraph Consumption ["AI 智能体与业务交互"] GENIE["Databricks Genie / AI Agents"] USER["高管与员工即时问答决策"] end M --> G D --> G C --> G IDX --> GENIE --> USER ``` --- ### `47:35` 告别低效的 Agent 循环：用离线索引重塑企业级图谱 将内隐知识提炼为可供 AI 理解的数字化图谱（Digital Graph）是本体落地的第二大核心。Ali 直言，当前诸如 Claude Code、Codex、Pi 等大多数主流 Agent 的底层机制仍然存在明显的效率瓶颈：它们普遍依赖一种串行循环，当遇到问题时，会一个一个地去访问模型上下文协议（MCP）服务器或外部资源，逐一排查、汇总并给出答案，这种过程极其缓慢。 Ali 将当下 Agent 的这一做法与 25 年前的搜索引擎相比较：如果 Google 当年在没有索引的情况下做搜索，用户输入关键词后，系统必须现跳到一个网站用大语言模型生成摘要，提取超链接，再并发访问另外几个网站，持续翻找 10 分钟后才返回 10 个蓝色链接。这种方式不仅每次成本极其高昂、延迟无法忍受，而且输出质量极差，因为系统在有限时间内只能触及全网极小的一个子集。Google 真正的杀手锏是离线构建逆向索引和 PageRank 算法，从而在用户发起查询时，在不到 100 毫秒内于本地直接返回结果。企业 AI 也必须采用这种架构——在后台全天候离线计算并更新本体索引。与公开网络不同的是，企业场景要复杂得多：网页是公开且天然带有超链接的，而企业内部涉及细粒度的隐私隔离、严格的数据访问控制权限，以及种类繁多、彼此割裂的对象体系。 > “The ontology is that we need to compute that index offline all the time.” —— Ali Ghodsi - 传统 Agent 串行检索的问题：逐一调用 MCP 服务器进行推理与验证，导致长延迟、极高计算成本与低信息覆盖面。 - 离线索引的优越性：将企业内各类资产、人员与项目的拓扑关系在离线阶段转化为“反向索引”，使查询响应保持在毫秒级。 - 企业本体的独特工程挑战：不仅要跨越多类型数据对象，还必须在图谱层面深度绑定多租户隔离与访问控制权限（ACL）。 --- ### `50:00` Databricks 全员“吃狗粮”与幽默的 Genie 财务段子 Databricks 在自身业务中全面实践这套方法论。Ali 透露，由于 Databricks 自身是平台最重度的使用者，他们内部维护的本体图谱包含数千万个节点，规模甚至超过了所有外部客户。在传统的树状企业架构中，信息通常需要沿着管理层级自下而上汇报、自上而下渗透：员工带着 Excel 模型和精美的 PowerPoint 幻灯片进会议室向领导汇报，遇到争议再向更高层汇报升级，决策做出后又得开会传达。如今，拥有完整组织上下文的 AI 可以直接接管数据建模与分析工作，员工无需无休止地召开跟进会议，而是直接向 AI 追问。这也印证了 Jack Dorsey 曾倡导的组织扁平化重构理念。在 Databricks 内部，每当会议中出现复杂问题，所有人都会立刻拿出手机向内部 AI 助手 Genie 发问。 Ali 还在现场分享了一则此前在内部董事会筹备期间发生的趣事。当时他急需获取公司在“财富 500 强”（Fortune 500）企业中的渗透率数据，便发短信询问销售运营团队的一位员工。对方回复道：“抱歉，我正在飞机上，现在登录不了 Genie。”Ali 当场哭笑不得：如果你只是上去查 Genie，那我何不自己查？他本来以为对方掌握着某种不为人知的特殊报表。随后 Ali 带着几分火气转而发短信质问首席财务官 Dave：“你知道我们财富 500 强的渗透率到底是多少吗？”结果 Dave 直接把 Genie 的查询结果截了一张图发了回来。Ali 戏谑道：“难道我们公司就没人做点创新的事吗？全都是去问 Genie。”如今，“让我为你 Genie 一下”（Let me Genie that for you）已经成了全公司的流行语。 > “So I texted Dave and I said, hey, do you know what our Fortune 500 penetration is done? And he just copy pasted a screenshot of Genie back.” —— Ali Ghodsi - 传统汇报痛点：依赖分析人员手动构建 Excel 数据模型与 PPT 汇报，决策链条上下传递损耗极大。 - Genie 带来的变革：业务高管与员工以会话形式直接向本体图谱下达多轮分析指令，随时随地调取跨部门数据。 - 与 Palantir 的落地差异：Palantir 更多依赖前向部署工程师（FDE）手动梳理隐性业务知识；Databricks 则倾向于从数字化底座中自动提取数据关系构建图谱，再通过 AI 智能体转化为分析图表交付给管理层。 --- ### `52:19` 从 Token Maxing 到 Value Maxing：CEO 亲自带头写代码上线 对话随后转向企业在 AI 采纳过程中的心态演变。主持人 Sarah 观察到，行业正从一开始盲目比拼消耗的“Token Maxing”（追求大吞吐量和高调用量）逐渐过渡到冷静理性的“Value Maxing”（最大化业务价值），同时也越发重视通过 Unity Catalog 等网关产品来管控激增的模型推理成本。 Ali 回忆，转折点发生在去年第四季度（Q4），当时前沿大模型的能力迎来了质的飞跃，展现出了肉眼可见的生产力红利。为了彻底打消研发团队的疑虑并强力推进内部采用，Ali 作为 CEO 亲自下场使用模型辅助编程，并一路将代码直接合并部署到了 Databricks 严苛的生产环境中。他借此向全公司施压：既然 CEO 都能在这样一个要求极高安全级别和合规标准的敏感数据平台上直接向生产线提交代码，研发部门没有任何借口做不到。Databricks 自此在去年第四季度设立了全员使用的排行榜（Leaderboards），用最激进的姿态强推 AI 深度融入核心业务流程。 > “Like if the CEO can commit code to production on a very sensitive data platform that has all these security requirements, you should be able to do that too.” —— Ali Ghodsi ### `55:00` Unity Gateway、成本控制与多 Harness 调度 年初（一、二月份）Databricks 内部已进入全面使用 AI 的阶段，但很快就遭遇了“Token 上限”危机。由于起步早了几个季度，他们提前目睹了成本失控的风险。为此，Databricks 依靠已有的 Unity Gateway 来统一调配 OpenAI、Anthropic、Gemini、Grok 以及各类开源模型的 Token 算力。为了压低曲线，团队实施了一整套成本工程：在个人和团队层面设定预算上限并发出告警；引入预测性用量分析；配置智能路由（Smart Router），在预算吃紧或遇到简单查询时自动切换为廉价模型。 ```mermaid flowchart TD accTitle: Unity Gateway Model and Harness Routing Flow accDescr: An architecture diagram showing how Unity Gateway and Omnigent multiplex requests across models and harnesses to control enterprise costs. subgraph ClientReq["企业用户请求"] R1["复杂任务 / 高价值推理"] R2["日常琐事 / 简单查询"] end subgraph CostControl["成本与治理层 (Unity Gateway)"] Bud["预算上限与用量预测"] SRouter["智能路由器 (Smart Router)"] Omni["Omnigent Harness 调度器"] end subgraph Models["多模型池"] FM["前沿模型 (OpenAI / Anthropic / Gemini)"] OM["开源 / 廉价模型 (GLM / DeepSeek 等)"] end R1 --> Bud R2 --> Bud Bud --> SRouter SRouter --> Omni Omni -->|"架构审计 / 关键推理"| FM Omni -->|"重命名 / 基础代码"| OM ``` 此外，Databricks 研发了名为 Omnigent 的测试套件/框架（Harness），能够在不同 Harness 之间进行多路复用。Ali Ghodsi 指出，业界往往忽略了 Harness 对成本的巨大影响：“Turns out actually the harness itself matters. Like if you use the same model but different harnesses there almost 2x different cost difference.” 比如社区发现 Pi 作为 Harness 就格外轻量高效。通过模型与 Harness 的双重优化，Databricks 成功实现了“Token 消耗量持续暴涨，但总体账单基本走平”的目标。 Martin Casado 提到，他担任二十多家公司的董事，上周首次有一家规模庞大的工程团队在董事会上正式宣布将业务从前沿模型迁移到 GLM。虽然前有 DeepSeek 和各类宣称对标 NVIDIA 软件的时刻，但此前市场并未真正产生结构性动摇；而现在，大公司因成本压力走向多元模型的案例正切实发生。Ali 强调，目前最主流的模型交互范式是“大小搭配”，绝不能用顶级模型去干琐碎的事情：“People literally use their harness to rename files and whatnot. You're paying orders of magnitude more for that. Please type that in yourself. Don't have the model do that.” - 成本控制三板斧：每人/每组的刚性预算阈值、精准用量分析预测，以及分流简单任务的智能路由。 - Harness 成本放大效应：相同版本的模型在不同 Harness 驱动下，实际成本消耗可产生近 2 倍的落差。 - 市场工作流范式：Martin 观察到目前许多团队流行用 Fable/Astra 做架构设计，廉价模型执行代码编写，再换回 Fable/Astra 进行审计把关。 --- ### `58:34` 开源模型占比与模型后训练的落地鸿沟 谈到开源模型的实际占有率，Martin Casado 与 Ali Ghodsi 梳理了当前初创生态的真实数据分布：如果按支出金额（Dollar）衡量，开源模型仅占约 5%；但若按消耗的 Token 总量计算，开源模型已超过 60%。尤其在面向外部客户的落地产品端，很多原生 AI 初创公司（如 Decagon 等）的开源模型调用比例甚至高达 90%；而在内部研发与非生产环境中，团队通常不太在乎浪费，依旧全盘依赖前沿闭源模型。 ```mermaid flowchart LR accTitle: Open Source AI Adoption Disparity accDescr: Comparison of open source AI usage measured by dollar spend versus actual token volume. subgraph Spend["按支出金额 (Dollar)"] S_OS["开源模型: ~5%"] S_Prop["闭源前沿模型: ~95%"] end subgraph Volume["按使用量 (Tokens)"] V_OS["开源模型: >60% (产品侧可达 90%)"] V_Prop["闭源前沿模型: <40%"] end ``` Sarah Wang 谈到 Databricks 早在 2023 年夏天收购 MosaicML 时押注的“企业拥有自身智能、自行做后训练（Post-training）”路线。Ali 坦言，当年谈论“为每家企业做预训练（Pre-training）”如今看来毫无意义，因为开源底座已经足够强大；但基于强化学习环境（RLE）的后训练在初创公司中正在大规模普及。初创公司拥有明确且垂直的用例，利用 RL 深度定制开源模型能换来极高的性价比、低延迟和自主 IP。 然而，大企业在模型微调与训练上面临严重卡点。首要障碍就是评估集（Evals）：构建一套科学严密的评测体系极为繁重。Databricks 曾尝试在产品前端自动为企业生成评估工具，结果由于步骤繁琐，根本没有客户愿意使用；后来将其退居为后台可选组件，使用率更是跌入谷底。Ali 将这一现象类比为软件工程领域的测试驱动开发（TDD）：“You know, in software engineering, did people actually do test-driven development? Very few did, right? Everyone said it's the right way to do it, but nobody actually practiced it.” 大多数传统企业只想要立竿见影的自动化“简易按钮”（Easy Button），遇到业务难题宁愿直接无脑调用前沿模型，也不愿投入精力搭建科学的 Eval 基准。 --- ### `01:02:18` FDE 模式与 Agent 时代的数据库基建（Lakebase/Neon） 为了打破企业落地的僵局，Databricks 内部的前向部署工程师（Forward Deployed Engineer, FDE）需求激增。由于绝大多数传统大型机构缺乏原生 AI 工程人才，FDE 的核心任务是协助客户建立组织本体（Ontology），并为生产级 Agent 构建安全边界与护栏。Ali 举例，福克斯（Fox）旗下的 Sports AI Agent 允许用户实时互动讨论体育赛事，一旦用户试图诱导其谈论政治话题，内置的护栏机制就能果断拒绝并巧妙把话题引回体育领域——这类端到端系统正是由 FDE 在现场帮助落地的。 ```mermaid flowchart TD accTitle: Agent Real-Time Database Spawning Workflow accDescr: How agents dynamically spin up and discard sub-second Postgres databases for rapid experimentation. A["自主 Agent (执行编码/数据探索任务)"] -->|"需要独立实验环境"| B["Neon / Lakebase 架构引擎"] B -->|"亚秒级冷启动 (<1s)"| C["克隆海量生产数据库分支"] C --> D["Agent 运行实验 / 执行修改"] D -->|"任务完成 / 报错"| E["合并写回或销毁临时分支"] ``` 随后，Sarah Wang 抛出一个行业评测信号：近期中立第三方的基准测试显示，Lakebase（基于 Neon 架构）出人意料地击败了众多老牌开源方案，成为 Agent 运行环境的首选 Postgres 数据库。Sarah 强调：“Because if you win the agents now, you win the market.” 询问 Databricks 与 Neon 团队究竟是如何实现这一突围的。 Ali 将这一战绩归功于 Nikita 及其 Neon 团队对底层极客指标的偏执追求。Agent 与人类开发者的行为逻辑截然不同：Agent 会在运行过程中自发去编写软件、测试逻辑，它们需要随时随地拉起全新的数据库环境进行沙盒实验。如果数据库冷启动需要几分钟，整个 Agent 逻辑链条就会彻底崩溃。Neon 团队将启动延迟死磕到远低于 1 秒的水平。正是这种亚秒级的瞬时拉起能力和极速克隆大型数据库的分支能力，精准击中了自主 Agent 高并发、高频试错的基础设施命门。 ### `01:05:00` 为 Agent 重塑数据库架构：Neon 的分支持性与全新用户画像 Ali Ghodsi 深入探讨了以 Neon 为代表的新一代数据库系统如何彻底适配 Agent 的操作需求。传统数据库对于超大规模数据的处理极为笨重，但这类现代架构即便面对 PB 级别的数据体量，也能在不到一秒钟的时间内完成克隆，具备极高的弹性和极快的响应速度。其核心的杀手级功能是“分支”（branching），允许用户基于同一个底层数据库轻松创建出海量轻量级分支。Ali 指出，这种架构演进与我们在其他 AI Agent 工具链中看到的趋势如出一辙——就像 uv、ripgrep 等工具用极速、轻量且具备防错机制（failsafe）的现代重构代替传统 Unix 工具一样，Neon 则将这一逻辑应用到了难度更高的数据库领域。这使得开发人员能获得一个既兼顾 Postgres 强大能力，又极其敏捷、可秒级回滚到历史快照且高容错的系统，天然契合 Agent 频繁试错、并发探索的行为模式。 ```mermaid flowchart TD accTitle: Agent-Centric Database Branching Architecture accDescr: 数据库架构从服务于人类DBA演进到服务于AI Agent，支持秒级克隆与轻量级分支试错。 subgraph Core["底层存储与引擎 (Postgres)"] PB["PB 级主数据库"] end subgraph Branching["轻量级弹性分支 (Branching)"] B1["实验分支 A<br/>(秒级克隆 / 容错回滚)"] B2["实验分支 B<br/>(低成本沙箱)"] B3["生产分支<br/>(正式业务承载)"] end subgraph Agents["AI Agent 自动化操作"] AG1["Agent 代码生成与调试"] AG2["Agent 自动化评估与测试"] end PB -->|"不到 1 秒克隆"| B1 PB -->|"瞬时创建"| B2 PB -->|"持续同步"| B3 AG1 -->|"并发读写与试错"| B1 AG2 -->|"低成本运行测试"| B2 ``` 与此同时，商业模式也随之改变。以往用户为正式生产环境付费是理所当然的，但如果 Agent 仅仅是在构建软件过程中进行海量自动化实验，不可控的费用暴涨将成为致命瓶颈。因此，弹性合理的计费机制对于 Agent 场景至关重要。Ali 强调，Neon 的核心突破在于其目标人群的根本性转变：传统的数据库厂商都在卷如何服务好 DBA（数据库管理员）或应用开发者，但 Neon 却彻底抛弃了“在传统数据库竞赛中击败同行”的执念，全心全意专注于如何成为“对 Agent 最友好的数据库”。如今，Neon 上超过 90% 的数据库实际上是由 Agent 而非人类创建的，这一数据直接印证了“Agent 作为新用户画像”的巨大变革。 - “They were obsessed with how are we the best for the agents. And that's a new persona because in databases, the obsession has been how do we help DBAs, How do we help app devs? How do we help the people that are using the database?” —— Ali Ghodsi - “And now over 90% of the databases that are created on Neon and Base are actually created by agents. So it's not even humans.” —— Ali Ghodsi --- ### `01:07:05` 大厂整合后的产品敏捷度与终极 P-Doom 大拷问 Martin Casado 分享了自己的实际使用体验，表示他已经将 Neon 作为自己的标准数据库。他感到非常惊奇的是，按照行业惯例，初创团队在并入大公司后通常会面临节奏变慢、官僚化的问题，但 Neon 在 Databricks 体系下不仅没有掉速，产品体验反而变得肉眼可见地更好，因此他询问团队目前是完全独立运作还是与整体业务深度融合。Ali 回应称，双方团队合作极其紧密，因为大家骨子里都对数据和数据库充满极度热爱，团队在将系统打磨得极速、灵敏和完美适配 Agent 方面做到了极致。 在节目尾声，主持人抛出了 AI 圈最为经典的终极拷问——“你的 P-Doom（AI 毁灭人类的概率）是多少？”Ali 给出了相当乐观的研判，认为这一概率“小于 10%，甚至接近于零”。Martin 则给出了一个极为深刻的反转视角，直言：“在没有 AI 的情况下，我眼中的人类毁灭概率要远高于有了 AI 之后。”Sarah Wang 对此也表示完全赞同 Ali 的乐观态度。这一辩论在对 AI 赋能未来的坚定信心中圆满结束，节目随后在鸣谢与订阅寄语中收尾。 - “All right, Ali, what is your P-Doom? Less than 10%. Close to zero.” —— Ali Ghodsi - “My only answer is my P-Doom without AI is much higher than my P-Doom with AI.” —— Martin Casado ## Notable quotes > `40:00` The models are smart enough, but they just don't have the context that exists inside of any organization, like they have not been in every meeting, they don't know what's in everybody's heads, they don't know all the processes. — Ali Ghodsi > `01:07:05` My only answer is my P-Doom without AI is much higher than my P-Doom with AI. — Martin Casado > `01:02:18` Because if you win the agents now, you win the market. — Sarah Wang > `00:46` But I think that right now the existential risk is close to zero. — Ali Ghodsi ## Mentioned Ali Ghodsi, Martin Casado, Sarah Wang, Databricks, OpenAI, Anthropic, Meta, Mark Zuckerberg, Dario Amodei, Elon Musk, Yann LeCun, Greg Brockman, Nick Bostrom, Elizabeth Warren, Bernie Sanders, Steve Bannon, Greg Abbott, David Sacks, Ben Horowitz, Jack Dorsey, Nikita, Hugging Face, GLM, Gemini, Grok, DeepSeek, Crisis Text Line, Insulet Omnipod, Zipline, Merck, TEDDY, Novo Nordisk, Palantir, Genie, Unity Gateway, Unity Catalog, Omnigent, MosaicML, Decagon, Fox, Lakebase, Neon, Postgres, Claude Code, Codex, Pi, Microsoft Copilot, Lakewatch, PlayStation ## Video description Databricks co-founder and CEO Ali Ghodsi joins a16z General Partners Martin Casado and Sarah Wang for a conversation about AI risk, recursive self-improvement, cybersecurity, and what’s actually holding back enterprise adoption. Ali argues that today’s models are already capable enough to automate far more work than most companies are using them for. The bigger problem is context: models haven’t been in every meeting, don’t understand how decisions actually get made, and lack the institutional knowledge that experienced employees accumulate over years. He explains why building an organizational “ontology” could help close that gap and what Databricks has learned from doing it internally. They also debate the current conversation around pacing frontier AI, what would constitute meaningful recursive self-improvement, and why Ali distinguishes speculative superintelligence risk from the much more immediate challenge of AI-powered cyberattacks. They close with how enterprises are managing exploding AI usage and costs, the shift toward multiple models and harnesses, and why agents are beginning to reshape infrastructure itself. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/809021a5-7883-43db-a718-62cf28e58cf8/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=809021a5-7883-43db-a718-62cf28e58cf8&feed=JGE3yC0V)

[40:00] Ali Ghodsi: The models are smart enough, but they just don't have the context that exists inside of any organization, like they have not been in every meeting, they don't know what's in everybody's heads, they don't know all the processes. [67:05] Martin Casado: My only answer is my P-Doom without AI is much higher than my P-Doom with AI. [62:18] Sarah Wang: Because if you win the agents now, you win the market. [0:46] Ali Ghodsi: But I think that right now the existential risk is close to zero.

[67:33] Martin Casado: My only answer is my P-Doom without AI is much higher than my P-Doom with AI. [1:48] Ali Ghodsi: I think that right now the existential risk is close to zero. So why freak everybody out? It's not actually needed. [53:00] Ali Ghodsi: If the CEO can commit code to production on a very sensitive data platform that has all these security requirements, you should be able to do that too.

[0:00] As a business leader, there's a tragedy of the commons. If you want to stop, if you want to go slower, why don't you go slower? Like, I'm competing. I want to win. There's almost two camps. There's one camp which believes that this actually is an engineering problem. And there's others which actually believe you have to slow it down. Humans don't respond fast enough to the attacks that are happening. You need to automate all of those. And most organizations are actually not supposed to doing that. Is RSI and recursive self-improvement that the labs are doing leading us there? That's the big question. This is some elaborate 4D chess because on the one hand, you're saying all of humanity will die. On the other hand, you're saying, hey, what do you want for your IPO allocation? For the first time ever, a company at scale last week said that they're moving from the frontier models to GLM. Do you think that that's a trend or do you think that's just like a one off anecdote? AI may already be smart enough for the enterprise. The problem is that it doesn't understand your company. Z's Martin Cassato and Sarah Wang to discuss what's actually holding back AI adoption, and why the answer may have less to do with building smarter models and more to do with giving them the right context. They also debate the push to pace frontier AI, recursive self-improvement, and where the risks are real today. Ali argues that superintelligence remains far from what we currently see, while cybersecurity is an immediate problem as agents make attacks faster and harder for human Get into what Databricks has learned using AI internally, from building an organizational ontology to managing token costs and choosing different models for different jobs. pacing the frontier, et cetera, um... What is your... strongest agreement with what's being out there? Where do you disagree? And maybe where is there nuance that's not being captured? Yeah, happy to cover it. And me and Martin argue a lot. So I'm sure that's not going to take long. We'll try and rein it in this time. Try to stay calm. But well, I do think first and foremost that there, maybe we agree on this, that leaders have responsibility to not freak people out unnecessarily, unless there's really, really good reason. And I think, you know, there's always different people in society that are different places, you know, in their mind space. scenarios where all of humanity is going to be wiped out, I think is irresponsible. It can tip a lot of people over and it can cause a lot of mental health issues. Unless you have something that's going to wipe people out. Yeah, as I said, yeah. If there is an actual reason for it, then that's a different story. But I think that right now the existential risk is close to zero. so why freak everybody out it's not actually needed There are risks. We'll get into it. That's probably where we disagree. But first and foremost, I think that leaders should not freak everyone out. And I mean, you know, if there's like technical nuances in how we're doing AI research and so on, well, researchers can discuss that. You don't need to every time go on TV and or blast on Twitter to millions of people that, hey, you know, I think there's like. This percentage, 10% risk that all humanity is going to be wiped out, I don't think that's helpful for a lot of people. Actually, I think it causes a lot of harm for a lot of folks who get stressed out and actually are not in the nuances of all of this stuff and what it means. So that I don't think we should do. I don't think it's fruitful. It doesn't really help anyone. I mean, I think this is very... very true for the general public. Yeah. Like my sister, who's great, who's a school teacher in rural Arizona on Sunday texting me and said, It said, Martine, should I prepare the cabin for, you know, she's kind of a prepper anyways, but should I prepare the cabin for the AI apocalypse? You know, I've got water set up. Like, when are you showing up? I'm like, hold on. Yeah. Like, we're not there. So clearly this has kind of spilled over the populace, which I agree is unnecessary and has blowback. I think there's a second one, which is, I don't know if you saw, like, walking in here, I was checking X, and Elizabeth Warren just talked about – Um, pausing all of AI development. That, of course, is on the coattails of Bernie, who is also working with... Bannon, like Steve Bannon. So now, so in addition to like, you know, just scaring people, The federal complex is now spinning up. And I think that could be actually quite contrary to the actual goals of the message. And so there's more than just, you know, I think public hysteria at stake here. Yeah, there's a lot of politics going on, but I'm like in all of these groups and, you know, I see both sides. There's heavy politics happening on both sides. we should say. [5:00] right? This is happening on both sides. This is, I think both parties... that do not include Trump himself, agree that... AI should be constrained at some level. I'm talking about the other side of this argument as well. Let me give you an example. Even Greg Abbott, right? Even Greg Abbott was like, you know, you can't have data centers in Texas. Well, I'm not talking about politicians. I'm talking about there's politics going on on both sides, right? There is politics on the business side, people who want to see great IPOs and they want to get returns on their investments. And they're like, don't mess up my IPO. And they want to get like, hey, can everybody just shut up so that we can get our money back? So there's that. and they're using them. And, you know, so there's politics on that side. And those are like not, they're not sitting quietly and not doing anything. And they can pull strings and they have connections. On the other side, there's all the people that are like, okay, how do we weaponize this? This is awesome. This guy tweeted this. You know, let's like, let's weaponize this one. Let's plant this. If I, you know, let's pump these, you know, threads. Let's talk to the specific leverage point that everybody is using. Because I actually think that this is like a classic case of a PR misstep. And it's not just the do me, gloomy type stuff. So here's the PR misstep, I think, which is, It is not unusual for industries to try and regulate themselves. It's just not, right? And I think saying security and safety is important. It is with every tech epoch, and we want to have some oversight. That was very, very sensible. The problem is just couching this notion of pacing. And there's a number of issues with pacing. First off, it's orthogonal to safety and security. You can slowly build a weapon. It's not different than building a weapon. been at a dead run. They're still buying more compute to be even faster. No, no, no. I mean, they just haven't done it. They haven't done it historically. But also, it kind of feels like this kind of almost milquetoast capitulation to the Paz people. So you're like, well, you say Paz, well, I say Paz, which is almost like Paz, but it's not like Paz. So they chose this kind of flag to follow around Paz. But if you actually read, did you read the to do with pacing, right? And so I honestly... No, I just mentioned it. Look, I can a little bit disagree. Look, there's a tragedy of the commons. There's this like, hey, if you want to stop, if you want to go slower, why don't you go slower? Why do you write articles? There's a lot of people making that argument. But no, I mean, as a business leader, I understand that there's a tragedy of the commons. Like I'm competing, I want to win, you know, and you're running ahead. There's also the market equilibrium, which suggests that pacing is probably impractical anyways. Yeah, so I'm just saying that, you know, so it makes kind of sense where people say, hey, if you guys don't stop this, this tragedy of the commons is going to continue. I'm not going to stop racing because, you know, there's IPOs at stake. There is a competition at stake. There's also some animosity between the people. So, like, I'm not going to stop unilaterally. I'll be a sucker, you know. Why don't you stop first, you know. So then they're saying, hey, can you come in and stop us? But, you know, I think that... You could also make the argument that if you look at the Hugging Face OpenAI incident that... By the way, I think these companies are great. And I think they are probably investing a lot of resources. But it's very clear from, if you read what happened, is that there weren't monitoring every token coming out and having it, you know, they were just like running these RL experiments. And then after the fact, coming in and checking out what happened. So they should have paced. They should have been much slower in that particular incident, right? I don't want to just quibble on syntax, but words matter with PR, right? So let's take the hugging face incident. When I read that, you know what my reaction was? It was not, oh, OpenAI should pace. It was like, dude, fucking secure your thing, right? Do security controls like we always have done. It is pacing, though. It is pacing. It is pacing. Let's do security. Let's do control. Let's do like whatever. You should do that. But it is pacing in the sense that, look, I face this all the time. I have a legal department at Databricks. I have a security department at Databricks. And, you know, they're always like, hey, slow everything down for everything, not AI. Like literally every little thing. Like, oh, you're going to go on a podcast. Well, what's the script for it? What are you going to say? And let's review that. And, you know, what's illegal? You cannot say this. You can say that. You can say that. You know, everything you say has to be materially true. Are these pacing people in this room with us right now? you're training the next model, should the security team be there and run all their monitors and look at everything? I mean, millions of hours of GPU hours of tokens were produced and these agents were running a mock in the sandboxes. It would have slowed them down significantly if you had security teams sit there and look at all the stuff. I agree. And they're saying, hey, if we do that, it'll slow us down. And I'm not sure the other side is doing that. So can you guys come in and slow us down? Like, just tell us, like, put some guardrails around us. We'll happily then follow the rules and do the secure thing. Otherwise, it doesn't make sense because we'll get our butts kicked. I just think, like, nuanced second-order words don't work when, like, people are really afraid. You're like, I'm going to pace, and therefore, things like these don't happen. I literally think we should have just been, like, safe and secure. [10:00] is paramount. We're going to put in these controls. That's the important thing. And I do think that nuance actually got lost. If you look at the... What Zuck said... I thought Zuck did it. He was like, hey, we're going to pace ourselves. We're going to put in security. Like the reason we release this later is because of security. Well, I thought what was so great about Zuck is like he was very focused on like security, safety and self-regulation. Dario's first five words or whatever, like we need to pace the frontier. Right. It just puts you in a very different mindset than what he could have said is we need to secure the frontier. Fine, we need safety. I mean, like at some level, I think they were trying to optimize both for the doomers, which caused for pause, and for politicians. And they kind of didn't satisfy either. But those are the people who are actually freaking out inside the labs. And there are a lot of safety people that are freaking out genuinely. By the way, and not all of them are EA people and so on. And people are like, hey, they're surprised, right? Right. But here's the thing is that using the worst pace doesn't help. either of them. I think it's like literally, you're like, you're trying to find this, he paces like the uncanny valley of making the Doomer people unhappy and the policy people unhappy. Because the Doomer people are like, that's not a pause, this is pacing. And everybody else is like, well, this isn't real, you're not going to do it anyways. And you're not focused on security. So again, independent of what we should do, which we should talk about, I just think that the way it was presented was just bad and it just didn't work. And that's why we're having the blowback. They're not trained, you know, PR people, you know, and yes, some of the stuff. I agree. I agree with you. I mean, I agree with the PR people. I agree with the core premise that we shouldn't freak the public out. I think that existential risk right now is close to zero. You know, but let's talk about the core thing, which is the... The fact that anyone who's doing big reinforcement learning runs And they're giving it a reward function. So they're unleashing, you know, saying, hey, here's like, you know, 10,000 agents and here's $100 million. Let's put them in parallel and let them run on a gigantic cluster for a month or two. Try to solve anything. And it doesn't need to be a security thing. It could be like, do anything. You know, solve this math puzzle. Really bad things can happen. Really bad things, meaning things get hacked. And it has, you know, cyber is the primary one. Right. That is real. Right. I think of this making it, hey, this is an existential risk and so on, which I think was a mistake. that way. I think it's become something that everyone, not just your sister, everybody around the planet is like now talking about. I've had all kinds of people that never care about this stuff and they find this extremely boring, ping me and say, what do you really actually think about this? This is really important. Now I'm starting to worry about it. So then it becomes a political issue and we have elections here coming up, but there's elections all around the world. So you're going to see they're not going to sit still in other parts of the world either. But I think that's our responsibility to talk about this in a balanced way and actually expose I think that super intelligence, that idea from that book is very, very far away. I don't see any evidence that we're actually marching towards that or that's going to happen. Apparently, some people, yeah, apparently some people at the labs are freaked out. that maybe there's progress towards that. And I think it comes from RSI, recursive self-improvement, the models improving themselves. I would love to understand how much, have they seen something we don't know? There's kind of four criterias. If those four things are happening, I would love to understand them. One is, if we end up in a situation where following four conditions are happening, which is the next model requires less resources, less GPUs to train. And super linearly, not just you know, it takes less time to train as well. So the second condition. Third, accuracy of the model, the intelligence is increasing. And fourth, we can do the former three again and again and again. You know, it's not just... All those at the same time, right? All at the same time, not just any of them. Yeah, all four. If all four are happening, then you can imagine in a way where you can, you know, because any of them does not happen. Like, for instance, if resources is constant, then that's okay, because we're going to run out of hardware. So then it will pace itself. Like, we will not have enough hardware. to do that, not enough GPUs, right? Time, the same. So it needs to be that you end up in this situation. So if you just mean that the software is writing itself, we're already there today. Like 90 some percent of the software in Databricks is written by AI. Does it matter if the last few percent is also written by AI? No, it doesn't matter really that much. But if you're getting these four conditions, then you might get a speed up where the next model, let's say, takes half a month of time and half the resources. And it is more intelligent. And you keep doing that, you know, then you might end up in a situation where I don't know. By the way, I don't even know if that necessarily leads you to super intelligence per se. That can still converge, yeah. But it could. So then that would be more risky. if they can share all that data, and we can shine some light and transparency on that. - I actually think it's a great breakdown that you have. - I don't think anyone is using that as a definition, actually, right? - That's a great one. - I think there is a little bit of people freaking out about, like, oh my God, the emergent behavior, now it's creating itself, and so on. But I think, like, as I said, [15:00] A lot of people, their definition is just, hey, If I'm not even coding anymore and it's coding itself. Right. But I think they're conflating, hey, what's my value and is it scary for me? Versus, hey, that then means we'll get that super intelligence that 2014 theoretically was hypothesized by Wattstrom. Well, your point in compute, though, is a really good one that's missed in, I think, a lot of arguments on RSI. Right. Because... As far as we can tell, the minimum threshold for compute needed to train a good model just keeps going up. Like it was 100 million, billion, now it's probably like 5 billion. To train a model now? To train like a frontier model, right? Five to ten billion. Right, right, exactly. Million or billion? Billion. Billion. Versus your second point. The frontier is very expensive. Yeah, frontier is very expensive. To replicate the frontier six months later is about one twentieth the cost. No, I think Sarah has a great point, which is... This is a good argument against this whole thing, which is that the next model, first of all, there's only one or two such runs a year that each of these labs do. And it's the opposite of the four criteria that I mentioned, right, which is it's going to take more resources. more humans involved, and it's even more brittle, and they have to build out the data centers. I mean, like, the labs are not necessarily doing that, but others have to build the data centers, and they have to be gigantic, and they have to get the GPUs, and they have to get the networking right, they have to do the engineering to make sure that they can tolerate, because, you know, every order of magnitude, more GPUs are crammed in there. You have to now worry about errors that before you didn't have to worry about. So you have to increase robustness. So it's like a very brittle process, and if it fails, you've squandered so much money, so they're, like, very, very careful with that run, and there's been multiple runs that have been botched. It's the opposite of that, that the, hey, the next model is faster, cheaper, smarter and recursive improvement. It's the opposite. It's like it's taking longer and it's more brittle and it's more people and it's harder to pull off. So I do think that that is true. With respect to RSI, With respect to actually cyber risks and things getting hacked, we need to take it super seriously, yeah. So I'm going to have like another, like, let me test here. It's like, I actually love your four criteria. I was like literally just waiting to argue with it, but I actually think it's very good. So let me give you like a black box. Like when you're dealing with these dynamic adaptive systems, like what are you going to believe? Are you going to believe like the numbers are your lying eyes, right? So I think you kind of have to go to the numbers on these ones. So like what are the numbers to look at? to do it. If these companies continue to grow, reduce the number of people and the number of money that goes into them, then I would say something is definitely happening here. I do think that you can actually black box this and take a look, but none of those indicate they're hiring like crazy. But that's not fair. That's not fair because companies are not necessarily efficient, right? So what if you have, I mean, OpenAI itself was doing a million different activities. It's a very small team of 10 people were doing LLMs and the LLM stuff was useful. was a lot of people. Now it's much less people. I agree. Just another litmus test. We have two litmus tests. We have your litmus test, which I think is great, but then you would actually have to have a way to instrument it. And then we should have the BlackBoss litmus test. I mean, listen, if Anthropic in two weeks is 12 people and they continue to grow and they're putting out models at an increasing rate, I think we should probably take notice of that. That's sufficient criteria, but it's not a necessary condition. I'm just saying that there could be that. And really the right way to do this is to look at, okay, the pre-training and the post-training that's been doing, that's really necessary. Because they have so much resources that they might be doing a lot of other stuff they don't need to do. But they're doing it just and they can just hire the people because they have infinite money and infinite. So really, the people that are training the next model, is that team tiny, tiny, and it's actually getting reduced, and they're doing less and less work, and just the AI is doing it, and then post-training, and then they're all just using less GPUs? That's not the case. before. I remember when we learned how to really cluster computers. Because the mainframe was actually kind of limited by things like memory coherence. Remember that? You can only make it so big. And then we kind of went to the client server area and then we didn't have that problem. And then we started creating supercomputers, which were basically just clustered computers. And that's not when the internet happened. And that was kind of also roughly when GPUs started getting good. And do you remember that we would actually export controlled PlayStations? Because we were worried that Saddam Hussein would use them to do simulation. And the arguments were very similar, which is like, these things are getting infinitely powerful. We're using them to simulate nuclear weapons, which we were like I was. Yeah. Like we can't, you know, this stuff has existential risk. Actually, they didn't use those words. But like this has the potential for like nuclear weapons or whatever, and we should stop it. And like, none of that came to path. So I think a very reasonable discussion is, is this time different? Yes or no? I don't have an answer to that. but I'm a PC but you know you're a yeah I mean look I think I was I'm not old enough to remember so ignorance is bliss wait so I can take this come on I don't recall playstations being illegal and Saddam Hussein being I just don't know how old are you? maybe I'm just ignorant this is like 1999 maybe I'm old and ignorant I mean you know maybe they didn't care in Sweden maybe it's just a ninja from age [20:00] But, whatever it is. Sweden doesn't care about export controls in the US. Yeah, you know, whatever it is, I think it's at a different scale now, right, with AI and, you know, with what we're doing, the pace of development and so on. They are freaking out the frontier. I do think cyber is actually one of the biggest ones that we're going to see, right? Because... There's just so much infrastructure on the planet. By the way, way, way more than it was whenever, whatever, Saddam or Xbox or whatever it was you're talking about. I mean, like we've just interconnected way more things and they're dependent. And like the planet just looks different today from Internet. tech dependency interconnection um than 30 years ago. So I just want to make this point. There's so much infrastructure that's insecure. And if you're going to unleash these agents, they're going to find loopholes, they're going to find exploits, they're going to break in here and there. So this is a real risk. And by the way, this time it didn't do that. But you could imagine a scenario also where it starts hopping, like it takes resources and it starts executing itself elsewhere. So it's kind of spread like a virus a little bit. So that's a real risk. much then. Again, I'm much older than you. I remember very well. So when the internet came out, by this point, we had literally taken out 10%. We disabled hospitals. We'd taken out critical infrastructure. We'd caused tens of billions of dollars in economic damages from worms. All of that had already happened. And to your point, we had much less build out, risks and threats were running so fast, so much money has been poured into it, and we haven't seen anything commensurate with the early days of worms. What does that disconnect? Yeah. Look, so I do remember that, those days. Maybe at the same time. Yeah. So look, I would just say that I... I am sleeping well at night and I don't think there's existential risk right now. I do think there's a lot of infrastructure that needs to be secured. We have a product in the market, in the detection market, Lakewatch, that helps you do detections. And the space just there is moving so fast. Because, you know, you used to have these SOC teams, security operations center people, that would look at... what intrusions are happening, how are we being attacked, and so on. And now the humans just can't keep up. So this is the whole space, security cyberspace, is being transitioned into fully automated, using agents, for detection on the other side. If we don't do that, I mean, now we're rushing. We are rushing. The industry is rushing to do that super, super fast. If we don't do that, I do think you will start seeing those kind of things, like sites going down, you know, whole systems that stop working for a while. And there will be consequences, not existential, but economic damage and, you know, people getting hurt and so on could happen. So we just have to race very, very fast to do all of those things. It's just, you don't have the... The humans don't respond fast enough. to the attacks that are happening. So you need to automate all of those. And most organizations are actually not close to doing that. The banks are doing it, some of the people that are super security conscious are doing it, but most of the industry today is running with old school security operation centers and people that are waking up every day and there's like hundreds of emails of detections that have fired. Many of them are just false positives. So you don't need to, you can ignore them. But some of them are not. They just don't have time to go through those. And you need to identify that. You need to have threat hunting that's automated where you're actually attacking your own systems automatically with agents and so on. It hasn't happened. So I do think if we just say, hey, this is just like the internet in the early days, bad things are going to happen. So there is a race going on. observability, agent threat detection, and using data break. I didn't even know that you had this offering, quite frankly. Just from an education standpoint, how extensive have you gotten in the agent AI observability, security, safety thing? Yeah, I mean, we gave a talk this year at RSA, actually with Ben Horowitz. But the issue is that data and AI is blending with cyber. These two markets are collapsing. And the reason they're collapsing is that it used to be like, okay, we have like data and AI, the kind of stuff Databricks and these kind of companies used to do, which is like, okay, you have a bunch of data and you run AI and machine learning and that just lives separately. And then you have the cyber world. Cyber world is, you know, we want to detect if something bad, if like bad people are trying to hack us, if bad people are doing things, we need to detect that. Okay. But now on the data and AI side, we have agents running. internally in the company. People are having agents running and the agent [25:00] are also doing things with other people's agents, and they're producing a lot of data, logs, trails, fingerprints that are being left. And so now we have internally these agents that are doing that. So these worlds start merging more and more, which is like, okay, well, all the data that's being produced needs to be analyzed. And the scale at which you need to do that is just like many, many orders of magnitude more than just one or two years ago. So things have changed dramatically. Like 2018, 19, the time it would take from a CVE, vulnerability being sort of published until you see it actually weaponized in the industry would be like two, three years. That went down to 2022 significantly, but it was still like eight, nine months. So that's kind of fun. You have eight, nine months from a vulnerability perspective. That was 2022. Now if you look at this curve from 2022 until now, now it's down to like basically hours. So it's like down to like basically no time. Like things get immediately weaponized. So you need to just do it in an automated with the data and AI sort of platform approach. So these markets, I'm going to argue, are just going to collapse. So it's a very specific question. I actually think a lot of like to pull back on the existential, extra risk discussion. It feels like there's almost two camps. There's one camp. which believes that this actually is an engineering problem. And like companies like Yeah. Like Databricks can solve it and they can solve it through product and through engineering solutions and through services. And so like we just as an industry need to solve that problem. Yeah. And there's others. which actually believe, seems to me that there is no engineering solution. You have to, you know, slow it down. You know, you have to use regulation. It's more like a nuclear weapon, et cetera. So, like, does this mean you believe it is an engineering problem? Or are you not quite comfortable saying that yet? Because why do you even buy Databricks, man? Let's just put this stuff in a national lab. Just pace it is the only solution. I'm just kidding. Everybody just paces themselves a little bit, then we'll be fine. The bad guys too. There is no line of inquiry ever that gets to pacing, I don't think. I think it's like you pause it or like you solve it. What we've learned here is that Martin really hates the word pacing. I will never use that word with you ever again. Okay. Clearly. Duly noted. Frustrated. Yeah. Is it an engineering problem that can be solved by engineers or is there more to it? I actually think which problem are we talking about? There's two separate problems that I think are being conflated. Yeah. There is the super intelligence problem. Yeah. And I think a lot of this comes from like Bodstrom 2014 super intelligence book. And if you look at the definitions, like I think people don't have these clear definitions of what super intelligence is. If you read his book, those definitions are kind of crazy. So I think what he had in mind when he said super intelligence is, you know, AIs that, I don't know, I don't know what the examples were. Something like they write a whole PhD thesis with novel, like peer reviewed stuff in a couple of seconds. you know, in like instantaneously. And, you know, so this is like the level of, you know, how fast they are, how intelligent they are. So it's like they can learn. Yeah, yeah, it's just, yeah, I mean, yeah, but it's just many, many, many orders of magnitude, right? It's like, it's just the scale of the problem is just completely different. So, you know, If such a thing exists, do I think it's just an engineering problem to solve? No, I think that's actually very, if such a thing would happen, that would be very existential. Of course. And that's what everybody agrees on. So I think that's being mixed with, now we have agents that are nowhere near that. It's not even like, there's nothing like that, and we don't have anything towards that path right now. But these agents are capable, and you can do something with them that you could never do before in the history of mankind. So I do think an inflection point has happened. Databricks, but I could never say let's get 10,000 of them. In a sandbox... for a month and have them do $100 million worth of salary wage work. We can do that now. We just turn on a button and we can get 100,000 of them. Or mathematics, like we can say, hey, you know, we want to solve a conjecture. Okay, let's get pretty good mathematicians, but let's have 10,000 of them collaborate. And then you can make very fast progress. So this, I think, leads to all these cyber risks. I think cyber is the major problem here. This is, I think, you can solve with engineering. And I think it's like we are working on it. Many others are working on it. There's still risks. They're not existential. I think we should do it. There is the superintelligence thing. That's the thing that could do, write a novel PhD thesis or like reason intuitively in 11 dimensional space physics instantaneously without writing anything down. Something humans can't do. Like that kind of superintelligence. The question is, is RSI and recursive self-improvement that the labs are doing leading us there? Are we going to get there? Trying to do. Trying to do. Is that what's going to happen? And how fast is that going to happen? That's the big question. And they've suggested that, you know, hey, we should have inspectors that come in and look at what we're doing. And that's a good idea. Have them go in there and get the data. I would love to like, the question is, who are the inspectors? Because you can like, You can [30:00] Stacked. Right? You can stack that. Who do you think, Martin? Yeah, so there's a bunch of people that like, on either camp, actually, I wouldn't care. If they're the inspectors, I would not be very impressed by what they say, because they've already made up their minds even before they would go in there. Right, exactly. But let's say like, as an example, if Yanda Kun... who was one of the inventors of this different network technology, right? One of the pioneers. If he said, hey, there's nothing to see here, there's no risk. I'm paraphrasing him. This is nothing. This is just nonsense. Keep on going. Go fast, fast, fast. None of us wouldn't believe it. I'm putting words in his mouth. Now, if he was one of the inspectors and he went in there... and he had a look, and he came out, and he said, hey, I've looked, and it's just what I said. There's nothing to see here. Just keep going. I would feel very good about that. I would say, okay, well, I would feel very good. Or if he comes out and says, oh my God, he's wobbling, and he would change his mind a little bit, that would also have a lot of interesting signals. So I think it comes down to who we pick as inspectors, and I think it's a good idea. Let's have some of them and pick a diverse set of people so that we can get different nuanced points of view. What do you think about this kind of Elon Musk view, which is like... It's less third party. It's more, it feels like there's kind of three proposals. Like the OpenAI, Anthropic one is a third party. The Elon Musk one, as far as I can tell, is the labs cross-check each other like Purogy, like you do in science. And then the Mark Zuckerberg one is police yourself, right? What do you think about this middle one? that they should face each other, evaluate each other. I think if we have boxing matches in the ring, the boxers should just be the judges of each other. Would that work? No, they would scream foul all the time. Foul, foul, foul. Like, you know, it's like, ah, you know, the moment the other guy puts out the great model and it's a big super intelligence risk. Absolutely. Like, you know, they have not been responsible. Like, you know, when vested interests are at play and there's like IPO plans and these two companies are so competitive and they have like this history also between them. Yeah, they'll be very fair to each other. I'm sure. That's why you need to play a party, right? Why do we have judges in the world at all? Why do we have third parties at all? Why can't just people figure things out between themselves? But I mean, they should try. If they want to do it, they should try. But I'm skeptical that they wouldn't just be biased. Yeah, you know, in multiple ways to self like, you know, judge each other. So I have to ask Ali, do you think something Elon also said, I think it was on the All In podcast. summit, he was like, this is some elaborate 4D chess, because on the one hand, you're saying all of humanity will die. On the other hand, you're saying, hey, what do you want for your IPO allocation? Right. And so, I mean, that is probably a more cynical view. But like, how do you How do you reconcile that? I mean, the dissonance, I think, gets a lot of people. Like, how do you think that gets reconciled? Look, I think all of these things get mixed. Like, I think there are people that are freaked out. And I do think that the people are saying like, hey, if there was regulation that would pace us, sorry to use the word, that would be good for us. Right. That would be good for us. But I also think that people have vested interests. Yeah. Right. These things like, you know, usually people figure out a way to always get all of these things to align in their, you know, harmonically in their head. So, yeah. Do I think that there has been a tendency in the past of in general using also marketing stunts? by saying, you know, oh my God, this latest model is so good that I trained. It's like unbelievable. It's like almost scaring me. And then the whole world like kind of starts focusing on it. Yeah, there's been that kind of marketing going on. But at the same time also, as I said, the time from CVE to actually weaponize exploit has like been going down from years down to like minutes now, just in like three, four years. So it's real. The cyber attacks are real. But there's also a great marketing ploy to, you know, whenever you train a new model, make lots of noise around how much of a crazy risk it is to the world. It helps you, right? So maybe they're not in contradiction, these things. I mean, you and I are networking folks, and there's a long history of... uh, forming third parties to help arbitrate things, right? Like IETF or, you know, IEEE or, you know, even like I can. I see where this is going. No, no, no, no. So my question to you is like, so I think it's actually, this is a very sensible proposal that they actually have. I actually agree with you. You probably want to make sure it's independent, which is not fair right now. And there's gonna be a lot of arguments on who you put there and everybody's gonna disagree. Right, right. But you said... Why do we have judges? Actually, the state stepping in is actually quite a different thing than basically industry self-policing. So at what point in time do you think it makes sense to actually consider federal involvement? Or do you think now is the time to actually consider actual federal involvement as opposed to more industry self-policing? These are very different. They are different, but they kind of bleed into each other. Like, for instance, FINRA is not a completely independent self-policing. It is, but it's linked to the government. over. I think it's hard [35:00] They evolve into, historically, the industry saw police and then it evolves into regulation. Look, if they are saying there is existential risk, which they're saying, you know, and they're saying, come police us and regulate us, I think it's very hard for regulators to say, no, we're not going to do that. So far, they've said that, but I think that's not going to last very long. But it wasn't David Sacks who was like, I've never had a CEO ask us to regulate them. And my favorite thing is like... And the CEO, I've never had a regulator that says no to that. I mean, the reality is like, is actually in motion already, right? I mean, everyone has a talking point. Obama has came out. It is a major issue. Do you think that there's a reality that it's too late? This will be a major issue in the midterms and we're actually going to heavy-handed federal regulation and this is all going to be paused, you know, anthropic goes into the DOE and we're past that point? Or do you think we can actually end up with a sensible self-policing regulation? Because if this is a headline, you cannot. We should strive towards doing the right thing. I think there's still some degrees of freedom things evolve and there's still time. And yeah, you're right. Largely, if you have these companies where we're pumping in so many billions of dollars. And the way reinforcement learning works is that you, you know, give it the reward function that's verifiable, like we're going to solve this math problem or this kind of, you know, this narrow area of programming and so on. And we pour in so much money into that, you can get quite good results in that narrow kind of, that doesn't mean that you're getting that super intelligence. No, but you can even trick yourself into thinking that less inputs are giving you a better outcome just because you're running so many experiences and you've thought about it so much, right? But it's actually very hard to do a closed experiment this way given how many resources are going on. Well, the fundraisers are going up astronomically, to your point. Yes, yes. That's why these companies are going public, right? I think they would otherwise... I mean, as someone who runs a private company at scale, to stay private otherwise, why are they going public? Because they need the capital. And they consider the scaling laws and the capital to be a strategic advantage. So that's why they're going public. But I would say, let's go back to the four things that I listed. If those are true, Um... if those four are true, would you want to know about it? And would that be worrisome? That could get out of hands. Now, there's no evidence that those four are happening. But if there was like, you know, that is actually where it's headed. I actually think understanding for any system, like any sort of self-propelling property is important. And we've done this in the past with dynamic systems, right? Like we've done this with like whatever compilers. We did this with all the research on nanotechnology. Like it's been a it's an interest, we should continue to have the interest. I just think the fear is that these particular systems are net economic systems that are so complex that the risk is crying that you're seeing it when you're not seeing it. And I think a lot of that's happening right now. But of course, if you see it, of course, I mean, you want to know. But it is fair to say that the labs are now focusing a lot on RSI and that's where they're headed next. And maybe they're just unjustifiably worried themselves, just like they were worried about GPT-2. They were like, GPT-2 is world-ending, and then it wasn't, and GPT-3 and 4 came out. So I don't want to quibble. A lot of the times when they say RSI, they're actually talking about autocatalytic effects. And autocatalytic effects have been in our industry for a very, very long time. So for example, there's no way you can create a computer chip without a computer chip. It's like you cannot do it. Anyone with computer science degree. A compiler writes its own compiler. Well, that becomes closer to RSI. Steam Engine was autocatalytic, right? So listen, my full-time job is people coming out of labs and starting companies, and they all say RSI because everybody says RSI. And maybe 1% of those are Axie RSI. They're more like, we use AI for data cleaning. We use AI for making GPU panels. Let's make this thing. So you're saying it's basically catalyst, in the sense that they're using AI to speed things up. It's autocatalytic, yes. So I would say... Like a model trains another model, right? You're using a model to build a GPU kernel. You're using a model to do data cleaning, just like I use a computer to design a computer. It's autocatalytic, which every... The internet was autocatalytic because it allowed people to collaborate remotely. So I would say... this is anecdotal. 90% of the calories are autocatalytic, which is 100% what you would expect. And that's been going on for a while, though. I mean, that's not even new. But I would say, but also now there is a focus on, let's move towards, actually, can we get the model to train itself? This is kind of like the auto-research that Karpathy did, but now they want to do that. There are teams that do that. It is not nearly as many calories as you would expect. And I just feel like I actually have a good sampling of this because they all come and talk to us. Maybe you can be one of the inspectors. Can we get all that data? And all of us look at that data? Maybe there's nothing to see here. I personally don't think it's like very high probability that those four criteria are happening. Greg Brockman went on the pot this week and said we're in AGI era. Yeah. So now I asked the same question after he said that and now everybody's saying we have AGI. [40:00] So, you know, people follow what they say. But people have for a very long time, when I asked this question, said that AI is smarter than most of the people around me most of the time. That's been almost like since Q3, Q4 last year, they've been saying that. Then I asked them, how many of you have, you know, hundreds or thousands of agents that you are managing that are coordinating with each other in swarms and negotiating and, you know, automating your life and everything around you? And if so, raise your hand. It's like almost nobody raises their hand. Of course, Martina has done that at home. No, no, most enterprises are on Microsoft Copilot. That's the extent of their AI. Most enterprises I talk to when I ask this question, they're like, no, we don't have any of that. So we're like, what are you doing then? They're using a chatbot. They're asking questions from a chatbot. That's basically very, very glorified, efficient Google search of the whole day. The results is just faster Google search. And then coding is happening. So people are using it for coding. Though the ROI is, we can discuss the ROI there. but there's no like agentic like work that's automated the whole enterprise that has just like not happened uh so then why is that and i think that the real reason is if you actually look at it is the models are smart enough um but they just don have the context that exists inside of any organization like they have not been in every meeting they don know what in everybody heads they don know all the processes There always a couple of employees who know everything in every organization You go tap on their shoulder and they're like, oh my God, what would happen if he or she quits? They don't have that context. And if you just fused that and gave that context into the AI models, just the frontier today, I think there's so much productivity gains you could get for any organization on the planet. For that, we actually don't need smarter models. So we don't need a smarter model that can actually solve Navier-Stokes or conjectures or do better on humanity's last exam. Like we needed to just go from 60% to 70%. None of that is needed. So I think actually people are very upset on some, like, oh, if we pace the frontier, but actually if the frontier doesn't advance, it doesn't actually matter. I think for vast majority of organizations on the planet, they're just so far behind in the adoption curve of actually automating things and getting value out of this stuff. But it would be disastrous in labs because the price of intelligence is dropping asymptotically. I think it's going down by one tenth every six months or something like that. So that will dramatically change their businesses if you weren't pushing the frontier. Yeah, but this is what we should focus on, right? We should focus on, there's two sides. We discuss here a lot the costs. There cost benefit analysis that we should do on everything We discussed the costs a lot here Oh is there existential threat Is there cyber risk Is there things we should be worried about And so on that like the cost side What's the benefit? And I think now that this has become like a public thing and the whole public cares about AI, they're asking, hey, what's in it for me? What am I getting out of it? It seems nothing. So how did they get there? What are some of the use cases you've seen to date that have maybe surprised you to the upside? Yeah, I mean, first of all, there's like so much worry about, you know, existential risk and so on. So I think a lot of people just don't know what a, you know, cool use case is where people are actually doing interesting things. We have a lot of use cases that are, I mean, just fascinating. One that I like is Crisis Text Line. So, you know, they actually use large language models with us to detect if teenagers want to do self-harm or suicide. Oh, wow, yeah. You know, that's an awesome use case. And it actually saves lives. So that's a great company. And that's, you know, that organization is doing amazing work. Another one that's kind of interesting is the Omnipod, which is for diabetes patients that can put the Omnipod and it uses AI to really learn your insulin release and your glucose levels and actually exactly release. I don know if you remember people used to stick themselves right But this now happens automatically and it like self AI self AI for your body That a cool use case Zipline is another one They doing awesome But when they started, it was like these drones that had, you know, they were completely automated, all AI-driven, everything from the, you know, battery optimization to the routes and everything, and they were delivering food in, you know, areas of need. Blood to refugees. Yeah, started in Africa and then elsewhere in the world. So yeah, that's awesome. That's all AI. It's an AI use case, you know, built on Databricks. So that's a cool one. But there's more advanced ones also. Like one that I kind of like, but it's harder to maybe explain, is this model that we built, a transformer-based model that we built with Merck. It's called TEDDY. Transformer Enhanced Drug Discovery. Love that. Yeah. And they publish actually the research, so you can check it out. But basically it's a model, instead of predicting the next token in English, it predicts what the gene regulatory network, the DRN is going to respond. And it can really detect which cells are causal and which ones are just reactive. So they're just reacting. And therefore, they can start using this in drug discovery and get costs down significantly for developing drugs that are targeting specific diseases. So that's a super cool use case. There are lots of these. You know, Jeannie, [45:00] you have this ontology and you can ask any questions. Novo Nordisk is using this. So, you know, they built this GLP-1 drug. But what Novo Nordisk is doing is now they're using it for all of their trials that they're running. And it can compress the time it takes to get insights versus if you're doing an obesity study or something from weeks down to minutes. So there are a lot of amazing use cases of AI. We should not forget these upsides also. Like we want all of these and we do not want to paste these. Yeah, exactly. You're totally right. And so how do they, let's say if you map out the next 12 months, how do the enterprises actually get value? You dropped the word context, but how do they operationalize that? So it's actually harder than most people believe. But, you know, first and foremost, we have to make sure that we have digitized everything that's happening in an organization. That actually, you cannot actually just, you know, have a magic wand and make that happen. So, you know, every meeting has to be transcribed, you know. So you have to be able to get all the context of all the meetings and everything that's happening. All the digital content has to be fed to the AI. So you have to build, we call it an ontology, we build that. But first and foremost you have to collect that That itself is a problem in many organizations because legal teams will say don record every call don record every thing So you have to do that in a way where Can you define ontology for everyone Because I know Palantir says the word a lot, but it's not like they own the word ontology. What does that mean for the people listening? Ontology just means that in an organization, the relationship between all the abstract concepts of all the goals and all the departments and all the people and all the projects that are going on, what do they exactly mean? And what's the relationship between them, the people, the resources, and what that company does? So it's the difference between a person who is a new employee in the company and just started today and a person that has worked there five years. Let's say they're equally skilled. They have the same educational background. They're equally smart and hardworking and all of that. But one, it's her first day today at work. The other one, she's been there five years. What's the difference between these two people? that one has an ontology of how that organization works, who the people are, how you get stuff done. Don't look at the org chart. Don't go ask that person. He will not get anything done. You go ask this person, he'll get it done for you. And that's not how it works. You don't need to file that paperwork here And this project this is what going on This is essential so there just a lot of ingrained knowledge that sitting in everybody heads who knows how an organization works that's why it's people say in startup land they say hey if you lose most of your people that company can't recover from it doesn't you can't just replenish and hire new people like the people are so essential how do we get that context that's the ontology and give it to the eye part of that is we just have to have you know the recording and all of But the second part is, how do you actually distill it down into a graph? Actually, a digital graph that you can then feed to the AI. So the way a lot of the agents work today, like a cloud code, any of them, Codex or Pi or, you know, open code, you can go through the whole school of them. You know, they have this loop, a genetic loop. It can reason, but then it goes and checks every resource one at a time. So we'll go to this MCP server for your question and try to see, is the answer here? Is there another one? it synthesizes it and gives you an answer. But it's kind of slow. I liken this to if Google would have built Google search this way 25 years ago, we would have said, okay, we're going to get 10 blue links. We search for key terms here. But instead of giving you 10 blue links, it would have gone to one website summarized with an LLM what it does found a few hyperlinks jumped in parallel to a few of them read a few websites done that for 10 minutes and then giving you its best 10 blue links it would find Well that would be very expensive cost a lot of money to do that every time, go on the web. Two, it would have taken a long time, and you've got to wait 10 minutes. And three, the quality would be bad, because you're actually only looking at a very small subset of everything that exists out there, right? So how do they do it? They have an index. You never leave Google servers. You search for, it hits the index, The reverse index immediately gets you the 10 blue links within less than 100 milliseconds. We need to do the same thing for the AI. So the ontology is that we need to compute that index offline all the time. So it's almost like the PageRank algorithm that Google had invented back in the day. But it's more complicated because Google was just looking at a web where everybody can go on the web. Here, there are permissions. And the links existed. Yeah. Here, there's permissions involved. the data I'm allowed to access might not be the data that you're allowed to access. So there's privacy, there's access control. Also, there's many different types of objects here that we're dealing with, not just websites. So the problem is a little bit harder. But it's manageable. You can actually do it. So, you know, I'm convinced you can do this and you can get massive productivity gains out of it because we did it for Databricks. Yeah. We did it for ourselves. Yeah, yeah. Yeah, and we're like, [50:00] The company has just completely changed. It's like not the way it was, I would say, a year ago because of this. I mean, you've been dogfooding Databricks for Databricks forever, but maybe say more about the impact you've seen as an organization. Yeah, I mean, once we got this ontology and we started working on it, and we actually have probably the largest of all of our customers, we have the largest ontology. Our ontology is bigger on us than any of our customers when they use us to build their ontology because Databricks uses Databricks more than anyone else uses Databricks. And so it's like millions and millions of nodes in the graph, in the ontology graph that we have. So it's just, you know, what happens in an organization? What happens in an organization? You have a tree structure organization and information flows up and down the tree structure. You know, if you can't make a decision, you escalate your boss, maybe they can tiebreak, it escalates up. They need to get up to speed on what's happening and they need to get all the context and then they make decisions. Once decisions get made, you have to percolate them down in the organization. A lot of this can now be done by AI if you have an ontology. Why? Because... you know, what happens in a meeting? In a meeting, You go through some, you know, someone has done the analysis. They probably have a PowerPoint deck. with some pretty graphs in it. That person who did the analysis is some smart person that used Excel, made some models, there's some numericals. So a lot of that you can now just do with AI. So the AI can do the analysis for you. It has all the context. It can present it in a way that you want. You can ask questions about it instead of having follow-up meetings. You can directly ask questions directly from the AI. So it's very similar. It's along the lines of what Jack Dorsey has said that you can do to the organization. It's just a concrete way of implementing it. So it's a game changer for us. You know, it's just everybody's on their phones now in the meetings on Genie. And they're like asking Jeannie questions. You can see, as soon as someone says something complicated or something, you see everybody go to the phone. Can you share that finance, like the finance anecdote you mentioned once in a board meeting? Yeah, sure. Internal board meeting. Only in kosher. Yeah, exactly. No, it's actually needed for one of our presentations. I need to know how many customers do we have in Fortune 500? What's our penetration of Fortune 500? And I asked one of the people in sales ops because I thought she would have it. and she texted me back and said, oh, sorry, I can't log in to Genie right now. I'm on a flight. That's why if you're just going to log in to Genie, I can do that myself. I asked you because I thought you had something alternative that I don't have access to. So then I was kind of a little bit angry. So I texted the CFO instead, Dave. And so I texted Dave and I said, hey, do you know what our Fortune 500 penetration is done? And he just copy pasted a screenshot of Genie back. So he also asked that. So I said, does anyone do anything novel here? going to Genie and asking the anthology, you know, questions. It's like, let me Genie that for you and let me go. That's what everybody is doing. Now we just say, hey, can someone just Genie this? Can you just get it from the anthology? So I do think it's a game changer. But It's not just you press a button, you have an ontology in an organization. And I think Palantir actually has done a great job of going to organizations and getting a lot of that tacit knowledge written down and getting it into the organizations. We automatically can take that and build the graph. And then we feed that graph into the agents so that we can answer the question and answer it in a way that, business leaders would like to see it, which is in graphs, you know, analytical way and a way where you can interrogate that question and, you know, continue asking questions and getting answers to those so you can make decisions and then disseminating that information in the organization. Yeah, it's pretty amazing. You sort of bookmarked the developers are obviously using AI, questionable value. I want to follow up with you on that, because I feel like you guys were one of the earliest. And I say, I don't want to use the word token maxing, because it has such a negative connotation. But I think in terms of applauding people who can use AI to become more productive, you guys were at the forefront of that, right? And then, of course, there's this cycle of, oh, shoot, people are being wasteful. Now we need to value max. What was your... your own journey on that. And like, how do you guys think about value maxing, not token maxing? And then I'm going to throw in Unity Gateway in this, right? Because I think the managing of costs piece is actually getting more important. And you guys are helping people do that. But maybe tie that in to extend it. Yeah. Yeah. So around Q4 last year. was when the models got really, really good. And we started noticing that, okay, it's actually starting to give much better productivity. So I actually started using the models myself to sort of start, you know, commit code into production for Databricks. Like the actual, as I want to take it all the way to production. So I did that and started pushing the organization that, hey, everyone needs to do that. I have done it. Why are you not? Like if the CEO can commit code to production on a very sensitive data platform that has all these security requirements, you should be able to do that too. the organization. So I started pushing everyone very hard. And we started making leaderboards in Q4. And at the beginning, [55:00] of, I say January, February, when kick it off the year, we were already full swing. Everybody was using the stuff and we're pushing and we're managing this. But, you know, the whole token maximum thing was happening around, you know, February, March period, already it was happening. So, yeah, we just had the luck of being maybe a few quarters ahead of folks to see what was happening here. And it was getting out of hand. So we already had a gateway. So it's called Unity Gateway where we were already, this gateway was being used to provide token capacity. so you can get OpenAI, Anthropic, Gemini, Grok capacity. Any customer can come to us and we'll just provide them that capacity because we have a relationship with those. And any open source model. So we started putting in budget constraints in place and giving people warnings like, okay, you have this much of your budget left. You're getting close to your kind of ceiling. So we started doing that per person and for group. And then we started doing great analytics so we could predict exactly where the costs were going. And then we added smart routers. that could actually pick cheaper models if you're getting close to your budget or if you have simple questions, we started doing that. We also built a harness called Omnigent which can multiplex between the different harnesses Turns out actually the harness itself matters Like if you use the same model but different harnesses there almost 2x different cost difference Even exactly same model you know, same version but different harness, you get 2x difference in actual cost. So if you can change harness, you can get a lot of leverage in the cost. So we started using all of this, that we were able to actually bend the curve and actually our cost for AI has been basically, The tokens have continued to go up, but the costs have been sort of stagnant. So that's been actually super, super important for us. And there's a huge demand for this. I think every organization is going through this now. Yeah, for the first time, I do a lot of board meetings, so I'm on 20-something boards. For the first time ever, a company at scale last week said that they're moving from the Frontier Models to GLM. This is a large engineering organization. Do you see this? Do you think that that's a trend, or do you think that's just like a one-off anecdote? Because I've been hearing about, I remember the first DeepSeek moment, and then like NVIDIA software, and then that turned out to not be real. Yeah. And then the Kimi moment, then the next DeepSeek moment. None of it seems to have actually had an appreciable impact on the market. Yeah. But now the amount of anecdotes that I have are pretty real, and it seems to be happening. Yeah. Love your view. I mean I think people want both They want the latest model that super intelligent for the difficult task where they get ROI But then there a lot of mundane dumb things Like people literally use their harness to rename files and whatnot You're paying orders of magnitude more for that. Please type that in yourself. Don't have the model do that. It's going to spin for five minutes, and then it's going to rename the file for you and cost you cents. I get some people. I'm just wondering, do you actually see market movement? No, people are moving on it, but what they're doing is that the pattern is either you can use this expert pattern where you have a small, cheap open source model that uses an expert model, the big ones, or vice versa, or a way in which they can sort of ping-pong them to each other, but also multiplexing harnesses and just changing harnesses so that you can control the costs is also what people are doing. People have found, for instance, Pi is very efficient when it comes to, as a harness. So yeah, I think there's going to be a multitude of these. The models themselves are stochastic, as you said. Every time they give a different answer, and they're changing so much. So there's just a lot of experimentation happening. So I think we're going to get to a world where you're not always using the smartest model for everything, which has kind of been the paradigm for the last couple of years Like a new model comes out it super smart they use it for everything even really really simple mundane tasks I tell you what I see I see people using Fable and Astra for architecture a cheap model for implementation, and then Fable or Astra for audit. That seems to be this emerging. What are you guys seeing in the startups? That's it. That's honestly the pattern. How much open source? By token or by dollar? Either. So by dollar, open source is like 5%. It's very little, but by token count, it's over 60%. Yeah, I was going to say, I mean, we talked to, let's say, a Decagon or something like that. Well, I think it's different internal use versus external for product. On the external for product, I think they're almost up to 90% open source. On the internal, and I don't want to say for Decagon in particular, but a lot of them are like, we don't care. We'll just use Frontier. We're not thinking about cost control. But as it gets bigger, you and I were in another board meeting where they actually did bring that down, just from a waste perspective. So I definitely see that moving more toward open source on the product side. And actually, that's related to another question, maybe around open source, but post-training specifically. I feel like you were kind of early. I remember talking to you in 2023. When did you buy Mosaic? 2015. [60:00] - 2023, okay. - Summer. - So this vision that you had in '23, kind of came true in 2026. I don't know if you guys would agree, right? Like that's sort of what we're hearing across the, What was it? Oh, just sort of like, hey, we're going to actually, you're going to own your own intelligence. You're going to be, you know, post-fitting your open source models, etc. And that's definitely what the startups are doing. Yeah. I don't know if that's what the enterprises are doing yet, but like... I mean, do you feel like you were early to that? Yeah, I mean, first of all, when we started, it was also, hey, we're also pre-training for you, which that doesn't make any sense. There's a very good pre-trained model now that you can use, but that you can do actually post-training on the model and you can do reinforcement learning. Yeah, we're actually doing it at that scale. And many of the startups are actually customers, so we actually help them using RLE or reinforcement learning environments where we can make the models very, very good at the specific tasks that they are doing. It makes a lot of sense for them to do that. Startup. and it's offering a product, and a product does something specific. It's not just a general intelligence. It does something specific for you. It makes just a lot of sense to take a really good open source model and use reinforcement learning and make it really good at that specific task. You can cut the cost down. You can make it really fast. They control their own IP. So in that sense, that is possible. But large enterprises, they just need basic automation. And it's just too much for them to do this right now. I think one of the challenges is you need good evals. Totally. And... Making good evals is hard. So while the startups can do that and they're motivated to do that, other organizations, the easy button might be just to use a frontier model than having to create your own evals. We actually generated even... evals for the customer automatically in the product. And we had it front and center. But then people didn't want to use it. So we said, okay, let's move it to the back end so it's optional. And then they would never go to it. So I would say in general... Why? They just don't want to get into it? It's too complicated? I think you want quick reinforcement of like, hey, there's a new model. I want to try this out. I want to get this problem solved. You don't have time to go do this scientific method of let's make an eval. Let's have a great baseline. sort of like TDD, test-driven development. You know, in software engineering, did people actually do test-driven development? Very few did, right? Everyone said it's the right way to do it, but nobody actually practiced it. So that's the same. That's kind of a little bit of the curse of, you know, training your own model is the evals is the hard part. I know you have an FDE model at TDD, Databricks. That's a very popular word right now or acronym. But does it, like to get these at enterprises that large, is it a full FDE model that's required? Or like, how do you and how's that evolved? Maybe? Yeah. Yeah, I mean, we've had these FDs and it's the demand for it's gone up significantly. And, A lot of it is, how do we build that ontology? The ontology is automatic, but if you're not collecting any information, you're not recording anything. So that's one of the key things that we do. But also things like, I want to build an agent, I want to put it on, I want it to be customer facing and it's have really low latency and I want it to have guardrails to not, people coming to abuse it or ask it things that we don't want it to answer and so on. So we can build that. Like sports AI that Fox has, you can go chat with it about sport events. You can try to ask it actually about politics and it's very good at rejecting you and moving and talking about sports instead. So the FD has built that. So, you know, we'll help the organizations actually get started with AI. It is important because it's just many organizations do not have the in-house expertise to build this stuff. So just a little bit of help on the side and then they get started. Makes sense. So this is more related on the agent side, but I saw recently that I think a third party, neutral third party, I think, did some tests that. Lakebase or Neon was actually the Postgres database of choice for agents. And I thought that was interesting, one, because, you know, one, exciting data bricks, but two, I probably wouldn't have guessed that maybe a year ago. Yeah, it was a surprise. Yeah, it was a surprise. Just because there's others out there that have, you know, great developer momentum as well. But it was pretty clearly number one. And so I'm curious, How did you guys crack this? And, um... What? what makes you win across the agents? Because if you win the agents now, you win the market. Yeah, I mean, I think a lot of credits should go to Neon and Nikita and team. And I think what they've done is they've just been obsessive about how do you make the models, how do you make the models pick And agents favor Lakebase or Neon as a database. So what do they do? The agents want to experiment. They're going off. They're trying to build a little bit of software. They need the database. So you need the database to come up quickly. So they had this obsession that everything should take less than a, you know, far less than a second. So if database comes up in far less than a second, you can clone gigantic databases [65:00] of petabyte database, you can clone it in less than a second. So it's highly elastic, highly responsive, and then they built this killer feature called branching. So branching just lets you branch the database, and you can have many, many branches over the same database. And they just made this very, very lightweight. We saw this with other things with agents, right? Like UV, you know, RipGrap, like basically re-implementation of a lot of the tools. on Unix, making them really, really blazing fast and lightweight and also sort of failsafe. for agents, they've just did this to a harder problem, which is database. So now you have a Postgres database, and the Postgres database has all these advantages, that it's really fast, it's nimble, it's failsafe, you can go back to snapshots, you can do those things. So I think that's why, it's just easier for agents to use this. They also made sure that they had a pricing model that was, like you don't want, just because the agents are building some software and experiment, you don't want the cost to run out. You're okay paying for your database if it's like production used and lots of people are using it, but just to experiment. I think they were just obsessed. They were not trying to win... the database were trying to be better than some other vendor. They were obsessed with how are we the best for the agents. And that's a new persona because in databases, the obsession has been how do we help DBAs, How do we help app devs? How do we help the people that are using the database? They changed the game and said, hey, how do we focus on agents and help agents get the best database they want? And now over 90% of the databases that are created on Neon and Base are actually created by agents. So it's not even humans. So numbers speak for themselves. By the way, it's remarkable. So I've started using Neon as my standard database. And it was bizarre to me because normally when you enter a large company, things slow down. It's actually like the products got materially better. Are they totally independent? Do they work with the rest of the... No, it's a great team. I mean, we work very closely together. We love databases and data. So we live that. The team does a great job of just making it super fast, snappy, and great for agents. All right, Ali, what is your P-Doom? Less than 10%. Close to zero. What about yours? I don't know. My only answer is my P-Doom without AI is much higher than my P-Doom with AI. Wow, that's a tough one. What about you, sir? I would agree with Ali on this one. thanks for listening to this episode of the a16z podcast if you like this episode be sure to like comment subscribe leave us a rating or review and share it with your friends and family for more episodes go to youtube apple podcast and spotify follow us on x a16z and subscribe to our sub stack at a16z.substack.com thanks again for listening and i'll see you in the next episode Thank you. Thank you.