Cerebras, Black Forest Labs Expand Global AI Buildout Beyond US

Cerebras与Black Forest Labs拓展美国以外的全球AI建设 递归推理模型的飞速演进将数代技术进步压缩至数天之内，推动全球AI格局向本土开源基础设施和多模态物理自动化转变。 本期内容探讨了全球实体AI数据中心与硬件前所未有的建设规模，将向先进推理模型的转变定性为从简单文本预测器迈向自主多智能体工作流的根本变革。行业领袖探讨了开源模型如何缩小与闭源系统的差距，从而推动企业与国家数据主权。此外，对话还重点介绍了潜扩散与多模态架构的突破，这些突破正通过顶尖影视合作重塑创意媒体，并进一步拓展至实体机器人领域。 全球 AI 算力中心扩建规模堪比历史级巨型工程，单体耗电已超中型城市；Cerebras 积压订单高达 250 亿美元，芯片厂商正全力交付已被锁定的既有算力需求。 (0:00) AI 交互模式正由提示词工程跃升为具备自主反思能力的复杂推理系统，而 Cerebras 晶圆级芯片通过架构优化突破了传统摩尔定律限制，为长时间自主思考提供算力支撑。 (7:15) 受严格合规约束的金融与医疗行业对数据泄露和“智能主权”高度警惕，促使企业加速采用本地化部署与开源多模型架构，减少对单一闭源前沿模型的依赖。 (17:14) Black Forest Labs 联合创始人 Robin Rombach 透露正与著名导演马丁·斯科塞斯（Martin Scorsese）合作，利用 Flux 模型作为电影前期构思的视觉探索载体，实现可控的人机协同创作。 (46:32) 多模态生成式视频架构正在与物理 AI（Physical AI）深度融合，通过理解真实世界物理规律，同一个生成与感知底座可直接作为实体机器人的“大脑”进行物理动作预测。 (53:19)

Executives from Cerebras and Black Forest Labs argue that the artificial intelligence sector is transitioning from raw model scale to specialized hardware deployment, open-source adoption, and multimodal physical world models. Infrastructure Scale and Changing Compute Economics (0:00) Cerebras CEO Andrew Feldman frames the current AI buildout as a global mobilization that requires data centers shaped like football fields with energy needs rivaling midsize cities. The geographic expansion spans North America, Europe, and emerging hubs such as Kazakhstan, Tajikistan, Georgia, and Armenia. Feldman notes that hardware providers are attempting to meet previously booked demand rather than speculative capacity, leading to a reported $25 billion backlog for Cerebras. Enterprises are simultaneously adapting their compute consumption from broad token usage to tiered deployments. Feldman suggests companies are shifting toward hybrid strategies that assign routine data-manipulation workflows to open-source models, reserving expensive closed models like OpenAI, Anthropic's Claude, or Google's Gemini for high-value reasoning. He adds that Cerebras's wafer-scale architecture aims to outpace traditional Moore's law scaling over the next 18 months through specialized inference optimization rather than standard fab node shrinks. Open-Source Models, Enterprise Sovereignty, and Cybersecurity (17:14) Regulated sectors such as healthcare and finance are increasingly prioritizing on-premise, domestic open-source architectures to satisfy HIPAA and FINRA compliance while mitigating data-leakage risks. Feldman reports that Cerebras deploys both open and closed options—including Qwen, GLM, Kimi, and bespoke models for entities like GlaxoSmithKline and G42—while Nvidia promotes its own open-source models to retain neutral infrastructure positioning. Rapid enterprise model deployment also reveals significant security risks. Feldman cites a test run by Palo Alto Networks CEO Nikesh Arora, where a frontier reasoning model identified unmapped software vulnerabilities in less than an hour, prompting a six-week operational pause to issue patches. While regulatory coordination faces domestic political polarization, the speakers advocate staged red-teaming periods with security agencies like the NSA before frontier models undergo broad public release. Recursive Reasoning and Accelerated Iteration Cycles (7:15) The hosts and Feldman describe a shift in AI agent capabilities from rigid prompt engineering to internal deliberation and recursive self-correction. In autonomous agent testing using models like GLM 52, the system demonstrated internal strategic debate over information gathering across Reddit, Hacker News, and Instagram. When run on low-latency hardware, these recursive feedback loops compress extensive deliberative processes into single-day operational windows. Feldman likens this accelerated software iteration to the reproductive cycle of fruit flies, which enables thousands of generations of evolutionary testing in days rather than biological generational scales of 15 to 20 years. Proponents forecast that these compressed feedback loops will systematically dismantle institutional inertia across science and education, though human leadership and cross-functional coordination remain organizational friction points. Multimodal Diffusion in Creative Production and Robotics (40:52) Black Forest Labs co-founder and CEO Robin Rombach describes the evolution of latent diffusion—which compresses raw sensory data into lower-dimensional representations—into multimodal frameworks that train on images and audio while predicting actions. In film production, Black Forest Labs has partnered with director Martin Scorsese to use its Flux models as iterative pre-production storyboarding tools rather than automated replacements for filmmakers. Lower down the budget spectrum, dynamic AI backdrops have cut indie production costs, illustrated by an indie film starring Gal Gadot that completed shooting for $30 million instead of an estimated $150 million physical-set budget. Rombach explains that the visual and physical representations learned from massive video datasets allow multimodal generative video architectures to operate as foundation models for physical robots. After broad perceptual pre-training, factory robotic hardware requires only a few hours of task-specific fine-tuning, moving toward the goal of direct in-context natural language prompting for physical navigation and manipulation. While the participants maintain an optimistic outlook regarding compute expansion, abundant creative output, and accelerated scientific discovery, significant near-term friction remains in enterprise systems integration, severe cybersecurity vulnerabilities, and the technical challenge of translating digital world models reliably into physical robotics.

本期节目探讨了人工智能基础设施建设的空前规模与生成式多模态技术的落地变革。Cerebras 首席执行官 Andrew Feldman 强调全球算力正经历由推理模型驱动的超级周期，企业端正转向开源与分层部署策略；Black Forest Labs 首席执行官 Robin Rombach 则指出，潜在扩散模型正从协助影视制作深化为控制实体机器人的世界行动模型。 基础设施扩张与算力分层配置 (0:00) Feldman 认为当前全球 AI 基础设施建设在人力与资本调动上堪比万里长城或金字塔等历史级奇迹。新建数据中心单体面积已达到足球场规模，其电力需求超过中型城市，未来数年的能耗预计将超过此前计算产业50年的总和；这一扩张正从北美、欧洲迅速蔓延至哈萨克斯坦等新兴节点。与以往技术周期不同，供应商目前在竭力满足已签约的存量需求。单 Cerebras 一家便拥有250亿美元的积压订单，微软与 AWS 等超大规模云厂商同样面临算力供应不足的压力。 面对这种供应紧缺，早期员工无节制消耗算力的“代币狂热（token maxing）”模式正在演变。Feldman 借用 Costco 超市购物的成熟过程比喻企业 AI 战略：用户不再像早期那样漫无目的地采购，而是转向系统化分工。企业开始将大批量日常、常规的数据处理任务交给低成本的开源模型（如同驾驶经济的“面包车”），仅将极具挑战性的高价值难题保留给昂贵的前沿模型（如同昂贵的“法拉利”），在成本与实用性之间达成平衡。 推理范式跃迁与安全治理边界 (7:15) AI 交互正从早期的精密提示词工程转向深层意图理解。当系统获得充足算力支持时，推理模型能够持续进行长达一天以上的内部反思与多渠道策略推演，将数周的人工推断压缩至数小时内完成。为了支持海量内部代币的快速生成，Cerebras 依托晶圆级芯片架构摆脱了依赖晶体管微缩的传统摩尔定律制约，并预期在未来18个月实现远超两倍的性能提升。Feldman 甚至主张，按照数十年前的图灵测试标准，人类事实上已经达成通用人工智能（AGI），技术演进正进入由递归自反驱动的“循环最大化（loop maxing）”。 然而，模型能力的飞跃也带来了网络安全风险。Palo Alto Networks 首席执行官曾透露，测试前沿推理模型时，AI 在一小时内即发现其企业软件中未记录的关键漏洞，迫使公司暂停常规业务六周进行紧急修复。Feldman 认为不可预见的大规模数据泄露是统计学上的必然事件，主张借鉴美国 FDA 对药品的审核机制，建立由国家安全局等政府机构主导的分阶段红队演练与准入测试，以在全面公开发布前筑牢关键基础设施的防御壁垒。 开源生态崛起与企业数据主权 (17:14) 企业运营中绝大部分流程属于常规的数据流转与执行，并不需要顶尖的数学理论突破。与此同时，受制于 HIPAA 和 FINRA 等严格合规要求的医疗与金融机构，对数据泄露以及外部模型厂商的意图深感担忧，这促使企业强烈要求采用本地化、国内可控的开源方案，以保障自身的“智能主权”。 开源与闭源前沿模型之间的性能差距正在快速收窄。Cerebras 透露其系统目前广泛运行 Qwen、GLM、Kimi 等开源体系，同时为葛兰素史克（GlaxoSmithKline）、G42 与 MBZUAI 等机构部署定制化专有模型。云厂商为降低对单一芯片供应商的依赖而自研芯片，英伟达亦积极分发自身开源模型以保持中立，多模型共存的多元生态已成为不可逆转的趋势。 潜在扩散模型与影视工业变革 (40:52) Robin Rombach 带领的 Black Forest Labs 推出了代表性图像模型 Flux。作为潜在扩散（latent diffusion）算法的联合发明人，Rombach 解释该技术通过类似于 JPEG 或 MP3 的数学压缩方式，将高维原始视听数据映射到低维特征空间进行高效训练。针对生成式 AI 类似“随机老虎机”且缺乏可控性的痛点，团队通过引入多图语义摄入与多层操作控制，赋予创作者精细调整视觉元素的能力。 在影视工业落地方面，该团队正与导演马丁·斯科塞斯（Martin Scorsese）合作，利用模型将导演脑海中难以用语言精确表达的构思具象化，作为类似传统概念分镜的高效头脑风暴工具。这种工作流正在重构制作成本结构：初创公司大幅压缩了宣传片预算；女演员 Gal Gadot 主演的独立电影借助 AI 动态生成全部背景环境，取代了实体布景与传统绿幕，使制作成本从预估的1.5亿美元降至3000万美元。针对大型制片厂的 IP 资产保护，Rombach 主张在公共工具端设置防侵权围栏，同时通过模型授权变现粉丝社区的二次创作生态。 多模态生成向物理具身智能收敛 (53:19) 多模态视频生成的演进正超越虚拟屏幕，向物理行动预测（action prediction）领域深度拓展。Rombach 指出，海量视频预训练使神经网络具备了对现实世界物理规律与交互动态的直觉理解；因此，用于生成高质量视听内容的同一套核心架构，可直接作为实体机器人的认知“大脑”。 这种世界模型的交汇改变了机器人学习方式。当前的工厂级硬件只需利用基础模型的丰富视觉理解，辅以数小时的特定物理任务微调即可胜任工作，其长远目标则是实现全语音即时提示驱动的机器人交互。与人类代际繁衍的缓慢周期相比，软件系统的迭代被比作生命周期极短的果蝇（Drosophila），在压缩的时间窗口内完成上千代演化，打破了传统制度与学科范式的代际停滞。 尽管两位受访者对技术攻克疾病、推动教育均等化与降低创作门槛抱有高度乐观态度，但核心瓶颈正逐渐从模型层转移至现实物理与社会层面：超大规模数据中心的扩张受到能源供应与电网基础设施的物理约束，而组织管理摩擦、网络安全漏洞以及地缘政治对立，依然是决定 AI 能否安全变现其生产力潜力的关键不确定因素。

## Background 本期节目由知名商业科技播客《All-In Podcast》于 2026 年 7 月 27 日发布在 YouTube 平台。页面简介显示，本期访谈深度对话了 Cerebras 首席执行官 Andrew Feldman 与 Black Forest Labs 联合创始人兼首席执行官 Robin Rombach，核心探讨了规模堪比城市的超大规模数据中心扩建、突破摩尔定律的深度推理架构、开源生态与主权智能的崛起，以及生成式视频在机器人动作预测和好莱坞顶级 IP 创作中的前沿应用。 ## TL;DR 上半场 Cerebras 首席执行官 Andrew Feldman 解析了当前 AI 算力供需脱节下的 250 亿美元订单积压，阐释了晶圆级专用芯片如何打破摩尔定律以支撑万亿级推理 Token 消耗，并直言依照传统定义人类已然步入 AGI 时代。随后，对话探讨了递归学习自迭代（Loop Maxing）所带来的指数级跃迁，指出 AI 正以超越果蝇演化的速度压缩人类知识的代际传承。下半场 Black Forest Labs 首席执行官 Robin Rombach 剖析了潜在扩散模型向多模态世界模型以及具身机器人动作预测的演进路径。最后，访谈结合马丁·斯科塞斯（Martin Scorsese）的实测体验，深入展示了 AI 如何重塑好莱坞前期分镜与低成本虚拟制片，并构想了兼顾版权保护与粉丝二创自由的未来交互式娱乐范式。 ## Atmosphere & style 整场对话节奏紧凑明快、信息密度极高，洋溢着硬核工程理性与极客浪漫主义交织的澎湃张力。主持人 Jason Calacanis 凭借敏锐的商业直觉与生动的隐喻穿针引线，嘉宾在现场从随身展示研发成本数亿美元的晶圆级芯片，到生动复盘好莱坞电影大师坐在身旁测试 AI 提示词的幕后细节，极具画面感与现场感。交流不仅充斥着对黑天鹅网络风险与行业资本浪费的坦率审视，更贯穿着对攻克绝症与终极物质富足的坚定乐观，真实再现了身处技术范式大重构震中时那种激动人心的前沿眩晕感。 ## Bulletin - [03:08](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=188s) Cerebras 首席执行官 Andrew Feldman 透露，公司在芯片投产前客户即已排队锁定产能，当前积压订单高达 250 亿美元，凸显算力供需严重倒挂。 - [25:10](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1510s) 大模型在网络安全上构成即时挑战，Palo Alto Networks 曾因 AI 在一小时内挖出自研软件中数十个未知高危漏洞，而被迫全员停工六周修复。 - [46:29](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2789s) 好莱坞传奇导演马丁·斯科塞斯（Martin Scorsese）与 Black Forest Labs 合作，将生成式 AI 视作克服自然语言“损耗沟通”的视觉画笔，加速影视概念探索。 - [51:14](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3074s) 盖尔·加朵主演的比特币题材新片借助生成式 AI 背景取代传统实体搭景与绿幕，将拍摄预算从 1.5 亿美元大幅压缩至 3000 万美元。 ## Key takeaways - [03:08](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=188s) AI 超大规模基础设施建设已超越以往科技周期的投机性逻辑，当前完全是在拼命追赶昨日未被满足的存量需求。 - [04:22](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=262s) 企业对算力 Token 的使用正从前期的“极大主义狂飙”转向精细化治理，依托系统思维在核心工程与轻量端侧模型间分流任务。 - [07:15](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=435s) 交互范式已告别“提示词调优”时代，转向由密集 Token 驱动的多 Agent 自主推演与交叉辩论的深层推理模型。 - [10:23](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=623s) Cerebras 打造的晶圆级芯片旨在打破传统通用处理器的摩尔定律演进节奏，为需要长思考周期的复杂推理任务带来跃迁式提速。 - [12:46](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=766s) 超大规模云厂商自研芯片不仅是为英伟达压价，更核心的诉求是吸取 x86 与 GPU 时代的历史教训，摆脱对单一供应商的致命依赖。 - [16:09](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=969s) 开源模型在日常任务中已堪比“买菜车”，并因金融医疗等监管合规与主权智能诉求，成为企业保障数据隐私的刚需路径。 - [20:49](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1249s) 前沿模型发布应留出 2 至 3 周红蓝对抗窗口供关键基础设施打补丁，而芯片超快推理能抵消安全护栏带来的高延迟损耗。 - [27:34](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1654s) 按照几十年前图灵测试与科幻作家的智力评测标尺，人类实际上已经全面迈入了 AGI 时代。 - [30:43](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1843s) AI 模型的递归反馈循环（Loop Maxing）彻底打破了人类知识传承的生物代际限制，展现出指数级能力飞跃。 - [40:52](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2452s) 潜在扩散模型（Latent Diffusion）的底层核心是神经压缩，该架构打通了多模态感知并自然延伸出机器人的动作预测能力。 - [44:49](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2689s) 生成式 AI 正从不可控的“老虎机”黑盒演进为融合多张参考图、文本与连续视频的多层精细控制创作工具。 - [50:27](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3027s) 生成式 AI 在影视工业的最佳定位是“人在回路”的灵感放大器，通过将数月手绘分镜压缩为数分钟多方案并行验证来重塑前期制作。 - [54:23](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3263s) 具身智能世界动作模型依托海量互联网视频习得通用物理常识，仅需极少量真实硬件动作数据微调即可完成实体操作落地。 - [58:02](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3482s) 未来的流媒体娱乐将演化为官方受控 IP 与粉丝二创并存的即时交互式创作空间，重现并放大同人文化的创造力。 - [01:02:02](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3722s) 视觉生成领军企业正通过开源底座赋能掌控力、专有模型提供顶级质感，并在大规模训练、系统工程与物理 AI 定制上大规模扩张团队。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=0s) 超级智能竞赛与史上最大规模的基础设施建设 对话开篇聚焦于当前全球正如火如荼进行的“超级智能（Superintelligence）竞赛”。Cerebras 创始人兼首席执行官 Andrew Feldman 再次做客节目，探讨其公司在专用推理芯片领域所开创的路径以及刚刚完成的成功 IPO。主持人 Jason 回忆起双方在达沃斯论坛及 Liquidity 峰会上的交流，随后将话题引向当前 AI 基础设施建设的空前规模。Jason 指出，人类历史上几乎从未见过如此巨大的动员规模——动用的资本、耗费的时间以及倾注其中的顶尖智慧，甚至可以与古代的中国万里长城、埃及金字塔，或者二战时期的战时工业动员相提并论。 作为这场基础设施浪潮的核心硬件供应方，Andrew Feldman 直观地感受到数据中心物理形态的蜕变。他指出，人们谈论芯片或软件时，往往只联想到微小的硅片或机箱，但现实中的支撑体系已经演化成宏伟的物理实体。在未来数年内，新建数据中心所消耗的电量预计将超过地球过去 50 年的总和；单个建筑物的体量堪比数座足球场，输入电力甚至超越一座中型城市。这种疯狂扩建不仅席卷了美国和加拿大，更拓展至北欧、法国巴黎、中东，乃至哈萨克斯坦、塔吉克斯坦、格鲁吉亚和亚美尼亚等曾经绝少处于计算中心的边缘国家。全球几乎每一个国家与美国的每一个州，都唯恐在这场算力竞赛中错失身位。 ```mermaid flowchart TD accTitle: AI Infrastructure Demand and Supply Chain accDescr: 展现从前沿 AI 巨头需求到算力短缺及全球超大规模数据中心建设的链条。 subgraph Buyers["前沿大模型与算力买家"] A["OpenAI / Anthropic"] B["xAI / Google / 微软 / AWS"] end subgraph Bottleneck["供需严重脱节"] C["未满足的存量需求<br/>(非投机性建设)"] D["Cerebras 达 250 亿美元订单积压"] end subgraph Physical["超级物理基础设施建设"] E["占地如足球场大<br/>用电量堪比中型城市"] F["跨国扩张: 美加、北欧、中东、中亚各国"] end Buyers --> C C --> D D --> Physical ``` --- ### [03:08](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=188s) 供需极度失衡与“追赶昨日需求”的 250 亿美元订单积压 当谈及谁在消化这些庞大算力时，OpenAI、Anthropic、xAI、Google、微软和 AWS 等前沿厂商展现出了永无止境的胃口。Andrew Feldman 指出，眼下的技术周期与以往科技炒作有着本质不同：在过去的周期中，企业往往遵循“建好设施，客户自来（If you build it, they will come）”的投机性逻辑；而如今，所有厂商是在拼命“追赶昨日的存量需求”。实际市场需求已经大幅超越了全球建造数据中心并将硬件填满机房的物理交付极限。 以 Cerebras 为例，公司甚至在芯片彻底完成生产前，客户就已经提前下单锁定产能，目前的订单积压（Backlog）高达 250 亿美元。不仅是芯片设计公司，各大云服务巨头与顶级模型实验室也都在竭尽全力扩建数据中心，其核心焦虑并非产能闲置，而是“如何避免排队的客户因等待过久而流失”。 - **前所未有的订单积压**：Cerebras 坐拥 250 亿美元的未交付订单，反映了算力供给与市场爆炸性需求之间的巨大鸿沟。 - **反常的市场逻辑**：并非提前预支未来的虚假繁荣，而是各家机构手握已确认的客户预算，却受困于物理电力和芯片制造周期的严重滞后。 --- ### [04:22](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=262s) 从“Token 大跃进”到企业理性治理与系统思维 面对近来行业热议的“Token 极大主义（Token Maxing）”以及“算力狂飙是否真在创造价值”的争论，Andrew 提出了客观审慎的判断。他承认市场充斥着海量的试验与试错，这与当年 AWS 云计算刚兴起时的情形如出一辙——当时为了绕开效率低下的企业内部 IT 审批流程，各部门工程师纷纷刷个人信用卡开通云服务，其中创造了巨大价值，但也伴随着大量毫无意义的浪费。 Andrew 用 1988 年加州帕洛阿尔托刚开业的 Costco 仓储超市做类比：起初，人们像逛传统超市 Safeway 一样挨个货架扫视，结果推车里堆满了四件单价 22 美元但毫无用处的东西；等到消费者变得成熟理性后，他们直奔货柜深处拿一只烤鸡，或者买齐孩子生日派对所需的 18 个纸杯蛋糕，整个购物路径极具战略性。企业对 Token 的使用正在经历同样的演化：从最初不计成本的“无限开放”，转变为精打细算的商业化治理——对高价值、超高产出的核心工程团队放开算力，而对一般性任务分流至开源模型或成本更低的端侧小模型。这种转变催生了一批深谙“系统思维（Systems Thinking）”的新型人才；许多创业者与投资人甚至被 AI 主动反问其根本目标与需求文档，被迫重新审视自己的架构设计。 --- ### [07:15](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=435s) 从猜谜式 Prompt 到意图感知与深层推理能力的爆发 在短短 24 个月内，AI 交互范式发生了翻天覆地的跃迁。Andrew 回忆起二十年前一位顶尖计算机科学家的名言：“计算机其实非常愚蠢，它们只会严格执行你下达的指令。”早期的提示词工程（Prompt Engineering）完全印证了这一点——措辞的微调往往会导致输出结果天差地别，逼得人人都在研究“提示词调优技巧”。而如今像 OpenAI 乃至更先进的模型，已经进化到能够直接洞悉用户的潜在意图。即使用户输入简陋，系统也能主动提供超出预期的方案（例如在需要图表时主动生成折线与柱状结合的双轴图表），彻底告别了“提示词操纵师（Prompt Whisperer）”时代。 Jason 分享了自己当天的实测经历：他体验了基于去中心化算力网络 Bittensor 部署的最新 ZAI GLM-52 模型的自主 Agent 系统。在被赋予“每小时找出全球未被察觉的新兴趋势”的任务后，模型在未展开的推理折叠链中展开了自我辩论——它权衡了 Hacker News、Reddit 等极客社区的信息源价值，随后又主动推演趋势往往先在 Instagram 等社交媒体发酵的过程。这正是典型的“推理模型（Reasoning Model）”运作轨迹：在折叠状态下，模型依靠底层庞大的 Token 消耗进行自我质疑、推演与逻辑闭环。Jason 感叹，对于习惯了以往简单问答的普通用户而言，一旦见识到近乎“无限 Token 即无限推理（Unlimited tokens means unlimited reasoning）”的威力，整个认知都会被彻底颠覆。 ### [10:23](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=623s) 推理与代币消耗：晶圆级芯片的算力优势 主持人与 Cerebras 首席执行官安德鲁·费尔德曼（Andrew Feldman）探讨了前沿人工智能向“推理模型”演进所带来的算力范式转变。当模型需要思考 24 到 48 小时来解决难题时，会展现出令人惊叹的效果；而若借助 Cerebras 的加速硬件实现 15 倍的提速，用户在 24 小时内便能收获相当于数周乃至数月的思考深度。主持人感慨道，早在早期阶段，Ilya Sutskever 和 Sam Altman 就预见到了这种演化方向并率先洞察行业趋势。当被问及“下一步是什么”时，Altman 给出的回答正是“推理（Reasoning）”——不仅理解人类意图，更要规划策略、多 Agent 线程协同校验，彻底告别了“预测下一个词”或“总结 PDF”的初级阶段。 费尔德曼指出，推理的本质就是极其密集且耗费计算资源的推断过程，核心在于模型在输出最终答案前会在内部消耗巨量的 Token 进行自主推演与验证。为了应对这种吞吐需求，Cerebras 打造了极速算力硬件。现场，费尔德曼特意展示了一块随身携带的晶圆级芯片（WSE）。他打趣说，一块研发投入高达 5 亿美元的芯片自然要片刻不离身，甚至此前在达沃斯论坛上也随身携带；他的妻子更是调侃他就像一个在 8 岁生日得到越野泥地摩托车的小孩，晚上连睡觉都想把它放在卧室里。这款初代的晶圆级巨型芯片在硬件历史上占据着独特地位，正是专门针对此类大规模深度推理负载所定制。 ```mermaid flowchart LR accTitle: 推理任务与芯片加速协同架构 accDescr: 展示复杂推理任务中内部Token密集生成，经由Cerebras芯片加速，最终产出策略的完整流程。 subgraph S1["复杂任务输入"] A["意图理解"] --> B["策略规划"] end subgraph S2["内部推理与Token生成"] C["多Agent并发线程"] --> D["交叉校验与试错"] end subgraph S3["硬件级加速执行"] E["Cerebras 晶圆级芯片"] --> F["极速产出高质量解"] end B --> C D --> E ``` --- ### [12:46](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=766s) 突破摩尔定律与 AI 时代的 CEO 挑战 针对推理硬件的演进节奏，主持人询问 Cerebras 内部是否存在类似于传统摩尔定律的翻倍规划。费尔德曼明确表示，在 Cerebras 之前，通用计算处理器产业基本严格遵循着“每 18 个月性能翻一番”的传统摩尔定律，而 Cerebras 的全新晶圆级芯片架构打破了这一定律的节奏，开辟了全新的增长曲线。在未来 18 个月内，其性能提升将远超 2 倍。他分析指出，对于像 GPU 这样已有二十多年历史的老化架构而言，架构层面的红利早已消耗殆尽，必须严重依赖台积电等先进制程节点的物理缩圈；而全新架构仍处在生命周期早期，有大量空间去根据实际工作负载特征进行软硬件协同优化，从而斩获超额的跃迁式性能提升。 面对行业内高达 250 亿美元规模的算力需求浪潮，CEO 的治理挑战异常艰巨：团队必须以极高的速度部署硬件、招募顶级工程人才、制定产品路线图，同时还要跟上像 OpenAI 这类史上发展最快、要求最严苛的客户的节奏。针对当下 OpenAI、亚马逊等云巨头自研芯片的现象，费尔德曼回应称，当前市场需求极其旺盛，没有任何可用的芯片会被闲置。超大规模云厂商自研芯片并非仅仅是为了向英伟达施压争取定价折扣，更关键的教训在于摆脱技术依赖：正如 x86 时代云厂商受制于英特尔、GPU 时代芯片制造商受制于少数超大规模买家一样，各家都在试图掌控自身的命运闭环。 --- ### [16:09](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=969s) 开源模型的崛起与主权智能的合规考量 随着对单一硬件与闭源模型的依赖引发行业反思，开源生态在推理能力上迅速逼近闭源巨头。主持人提到自己在实际工作流中运用开源工具结合智能路由（Smart Routing）时发现，成本远低于旗舰模型，但输出质量在许多任务上已难辨优劣。费尔德曼借用“买菜车与跑车”打了个生动比方：日常去杂货店无需开法拉利，哪怕车后座掉了饼干屑也无所谓的“家用微型客货车（Minivan）”才符合常态。现实商业环境中，大量工作只是将数据从 Workday 转移到表格中，根本不需要“奥数金牌级别”的尖端模型，坚实可靠的开源模型完全能够胜任日常任务。 在企业实际落地中，数据隐私、反泄露与“主权智能（Sovereign Intelligence）”正在成为监管严苛行业（如金融、医疗受 HIPAA 与 FINRA 规管）的硬性要求，迫使越来越多机构选择本地化或开源方案。费尔德曼强调，美国亟须打造更多本土开源模型，以防全球市场在开源领域形成二选一的被动局面。在硬件层面上，Cerebras 始终保持中立姿态，同时支持开源与闭源生态的多元化运转： - **开源模型生态**：稳定支持包括智谱 GLM、月之暗面 Kimi 以及阿里通义千问（Qwen）等主流开源体系，同时呼吁美国国内提供除 OpenAI OSS 12B 外更丰富的开源选择。 - **闭源前沿模型**：为 OpenAI 的闭源旗舰大模型提供算力承载。 - **垂直行业模型**：运行制药巨头葛兰素史克（GSK）针对生物计算与 DNA 研究所自主研发的行业专用模型。 - **国家级主权 AI**：为阿联酋战略合作伙伴 G42 及穆罕默德·本·扎耶德人工智能大学（MBZUAI）运行并支持其自主研发的模型，满足国家级数据主权与在地化合规诉求。 ### [20:49](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1249s) AI 安全红线、政府监管与网络战威胁 对话探讨了前沿大模型发布所带来的现实威胁，尤其是在网络安全与网络战领域的潜在风险。主持人提及 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）具有极强的沟通表达感染力，并询问将此类极具潜在破坏力的模型采取渐进式排期发布（scheduled rollout）是否明智，以及这些模型是否真的构成了重大网络安全隐患。 Cerebras 首席执行官安德鲁·费尔德曼（Andrew Feldman）对此表示，政府要求分阶段推进发布完全合理。强大的新技术如同高效能药物，虽然没有人希望看到像 FDA 那样动辄历时七年、充斥繁琐官僚文件的审批泥潭，但至少需要留出时间进行内部红蓝对抗（Red Teaming）。他指出，国家级基础设施和国家安全局（NSA）等部门需要两到三周的窗口期来检验关键系统，并修补被模型挖掘出来的明显安全漏洞。 然而，在当前的政治极化氛围下，理性的安全考量容易被党派情绪劫持。不论是唐纳德·特朗普（Donald Trump）、亚历山大·奥卡西奥-科尔特斯（AOC）还是任何中间立场的决策者主导这一过程，过度党派化都会严重阻碍清晰的理性思考。安德鲁认为，政府基层公务人员实际上非常努力地想要跟上这项技术的脚步。健康的竞争格局至关重要——达里奥与萨姆·奥尔特曼（Sam Altman）的全力竞争不仅极大地推动了技术突破，也倒逼谷歌（Google）和亚马逊（Amazon）等科技巨头全面提速。行业既需要这种蓬勃创新的冲劲，不应重蹈某些地区“监管先行”扼杀活力的覆辙，同时也必须认识到目前尚未有现成的“安全规则手册”。各大实验室正在边前行边摸索构建“安全护栏”（guardrails）。而更快的芯片推理性能则能显著降低由安全护栏带来的高延迟痛感。 - 网络攻防需要合理的缓冲窗口：给政府机构及关键基础设施两到三周时间进行补丁修复并验证防御有效性是合情合理的诉求。 - 安全护栏会增加推理延迟，但超快硬件架构能够缓解这种性能损耗，让安全检查在毫秒级内完成，提升用户体验。 ```mermaid flowchart TD accTitle: AI模型分阶段发布与安全审查闭环 accDescr: 展示前沿大模型在公开发布前经过政府红蓝对抗、企业补丁修复以及硬件加速支持的流转机制。 subgraph Evaluation ["安全评估与验证"] A["前沿大模型研发完成"] --> B["政府与安全机构红队测试 (Red Teaming)"] B --> C{"发现高危系统漏洞?"} end subgraph Defense ["防御与加固"] C -- "是" --> D["预留2-3周修补期: 关键基础设施与企业系统打补丁"] C -- "否" --> E["进入渐进式发布流程"] D --> E end subgraph Deployment ["部署与护栏优化"] E --> F["部署安全护栏 (Guardrails)"] F --> G["超快推理芯片加速: 抵消护栏引发的高延迟"] end ``` --- ### [25:10](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1510s) 黑客攻击、未知漏洞与必然发生的“黑天鹅” 随着模型推理能力的跃升，大模型在挖掘代码隐患方面展现出了惊人的破坏力。主持人提到自己曾与 Palo Alto Networks 首席执行官尼克什·阿罗拉（Nikesh Arora）交流，当他们将最新模型对准自家的网络安全软件运行时，AI 在短短一个小时内就找出了数十个此前完全未知、未公开的严重漏洞（critical opens）。这一结果直接迫使整个工程团队停下手中所有正常业务，连续花了整整六周时间进行紧急打补丁。 安德鲁将当下的网络风险比作沃伦·巴菲特（Warren Buffett）关于再保险行业的名言：灾难必定会降临，只是具体时间未知。必须提前储备资金与心理预期，因为无论是龙卷风、大地震，还是前所未有的大规模数据泄漏与网络崩溃，都是统计学上的必然事件。我们身处一个由“未知的未知”（unknown unknowns）主导的黑天鹅世界，唯有在心理上提前做好最坏的防御准备，并设计好危机发生时的应急响应机制，才能应对突发的大规模安全事件。 - Palo Alto Networks 曾因 AI 发现未披露漏洞而全员紧急停摆六周执行代码补丁。 - 软件发布的常规分级测试（内部测试、安全测试、公测）在 AI 时代需要以更高烈度的安全红队形式常态化推进。 --- ### [27:34](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1654s) 推理能力的跃迁与我们所身处的 AGI 时代 面对复杂的未来，具备高级推理能力的 AI 正在从根本上改变人机交互的方式。主持人分享了自己的 Prompt 使用技巧：在指令末尾总是明确要求 AI “检查你自己的工作”，指出遗漏了哪些尚未考虑到的盲区，并在任务执行过程中主动向用户反向提问。这种反思机制使模型的产出质量呈现质的飞跃。 安德鲁回忆起过去十年在融资过程中遇到最聪明的一个提问：“在我们刚才谈论的内容之外，你听过别人问的最聪明的问题是什么？”这种思维方式体现出深度的好奇心与认知谦逊。如今，人们可以通过向 AI 询问“一个博士级专家或国际数学奥林匹克金牌得主会提出什么问题”，来瞬间拓宽自己的认知边界，触及原本“不知道自己不知道”的领域。 谈及通用人工智能（AGI）和超级智能（Superintelligence）的发展阶段，两位对话者达成共识：**事实上我们已经步入了 AGI 时代**。如果按照 20 年前乃至图灵测试（Turing Test）设立时的技术定义，如今的模型早就已经轻松将其碾碎并远超预期。早在几十年前，科幻小说家们所能设想出的所有智力极限测试，现代 AI 系统已经全部解答完毕；倘若让他们看到眼前的技术现状，甚至会陷入“已无问题可问”的境地。安德鲁感慨道，正因如此，行业应当尽早倾听那些起初听起来极端前卫、犹如身处边缘的声音——正如伊利亚·苏茨克维（Ilya Sutskever）在八至十年前便极力呼吁对 AI 安全给予高度重视，事实证明他的远见是完全正确的。 ### [30:43](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1843s) 递归循环、超级智能与“循环最大化”（Loop Maxing） 对话在探讨技术成本与进化的语境中展开。主持人提到，正如当年埃隆·马斯克（Elon Musk）谈及制造火箭并将发射载具成本降至近乎为零时，人们的第一反应是“怎么可能？”，但如今事实就摆在眼前，这也是当下成为一名技术专家最令人兴奋的原因。更关键的是，如今人们不仅在构建各种 AI 工具，而且这些工具正开始在一种递归循环（recursive loop）中反哺并构建自身。主持人分享了一个趣事：他在做趋势分析时，算法不断抓取“looping”（循环）和“maxing”（最大化）这类词汇，甚至自动为他造出了一个热词——“loop maxing”（循环最大化）；令人惊讶的是，随后人们真的人人都在谈论“loop maxing”，AI 仿佛提前预判了人类会发明并流行这一词汇。 对此，Cerebras 首席执行官 Andrew Feldman 深入阐释了递归学习与通往超级智能之路。他指出，早在五六年前，萨姆·奥尔特曼（Sam Altman）、伊利亚·苏茨克韦尔（Ilya Sutskever），以及后来的达里奥·阿莫代伊（Dario Amodei）和德米斯·哈萨比斯（Demis Hassabis）等人就已经洞察到：强大的递归收益（recursive gains）呈指数级爆发。其核心逻辑在于，当你向模型提出问题，从反馈中学习并让其再次迭代，答案便会随着更多信息的注入而大幅跃升；多次迭代带来的不是微小的渐进改良，而是数量级的质变。这种指数级收益的曲线斜率极为陡峭，当前业界才刚刚揭开其冰山一角。眼下最迷人的核心知识命题在于：如果持续投入算力，这种指数级上升究竟会在何处停下，还是会永无止境地持续向右上方攀升？ ```mermaid flowchart LR accTitle: AI 递归优化与推理循环机制 accDescr: 展示 AI 从输入问题到反馈自迭代的循环过程及指数级收益 subgraph IterationLoop ["递归学习闭环 (Loop Maxing)"] A["输入复杂问题"] --> B["生成初始解答"] B --> C["从结果中提取新信息与反馈"] C --> D["模型再次自我迭代/推理"] D -->|大幅提升答案质量| B end D --> E["指数级能力增长 (Superintelligence)"] ``` --- ### [33:20](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2000s) 人类瓶颈与石匠代际传承：从果蝇进化看 AI 迭代速度 当 AI 不断解决复杂算力与数学难题时，Andrew Feldman 提出了一个深刻的转折：未来的瓶颈或许不再是纯粹的智力难题，而是“人”的问题。现实中运转一家公司也是如此，大量困扰领导者的往往不是高深的技术课题，而是如何组织团队协同合作。管理者常常要扮演拿着“WD-40 润滑剂”四处化解人际摩擦、激励团队的角色。未来世界模型（World Models）若能通过观察人类行为，或许会为我们提供前所未有的行为洞察。主持人打趣畅想，未来当技术跳出屏幕走入现实，人们或许不再只是要求它重写一个 Salesforce 软件，而是在内华达或得克萨斯的数百英亩土地上派出一千台 Optimus 人形机器人，直接平地建造一座凡尔赛宫。 这种宏大工程引发了关于知识累积与代际演进的讨论。建造凡尔赛宫或宏伟教堂耗时数十年乃至上百年，往往由三到四代石匠家族接力完成，技艺在父子与叔侄之间传承创新。Andrew Feldman 指出，人类学习的致命弱点在于它受限于生物代际——人类像大象或其他大型哺乳动物一样，每隔 15 到 20 年才更替一代；但在遗传学中，科学家选择研究果蝇（Drosophila），是因为果蝇一天能繁衍两代，几千代的演化信息可以迅速在 DNA 中编码呈现。而 AI 的核心突破正在于将人类数代人的知识积累浓缩到极致，它在极短时间内完成了数千代人经验的演化迭代： - **范式转移的生物学本质**：主持人回忆起攻读心理学学位时教授的名言——“科学范式本身不会死，死的是坚守旧范式的人”（Paradigms don't die, people do）。无论是弗洛伊德（Freud）、斯金纳（Skinner）还是荣格（Jung），他们的学说往往需要等待几十年、直至门徒退出历史舞台，新范式才能诞生。 - **代际鸿沟的终结**：AI 的递归学习彻底压缩了“代际间隔”（intergeneration gap），让演进速率达到了连达尔文（Darwin）都会为之叹服的地步。 --- ### [37:44](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2264s) 乐观主义账本：消除癌症、孔子式因材施教与终极富足 面对社会对 AI 抢夺饭碗和技术失控的恐慌情绪，主持人对 Andrew Feldman 严谨而充满乐观主义的构建理念表示赞赏。Andrew 直言，技术变迁必然伴随着局部的经济阵痛，正如当年汽车取代马车时，造马车和钉马掌的工匠遭遇了生计冲击；但在评估技术时，必须建立一张清晰的“损益账本”（T-chart of pros and cons）。AI 带来的潜在收益是人类历史上前所未有的：人类有极大的希望借助这项技术，让我们的后代及其所爱之人彻底告别因癌症死亡的阴影。 除了攻克绝症，AI 还将彻底重塑延续千年的教育模式。千百年来人类早就明白一对一个性化辅导的效果远超大班灌输——亚里士多德曾作为亚历山大大帝的导师，苏格拉底也曾悉心启发门徒——但由于现实资源受限，人类过去一千年里只能采用工厂流水线式的折中教育，面向中等水平教学。如今，为每个孩子量身打造专属 AI 教学 Agent 成为可能，它可以精准契合每个孩子的学习习惯与心智节奏进行因材施教。主持人总结道，未来的账本上写满的是无限能源、无限热量、无限知识、优质教育与廉价住房等“终极富足”（massive abundance）。双方在约定半年后再次复盘交流后，圆满结束了该部分访谈。随后节目接入了 Nasdaq 的合作伙伴赞助鸣谢，介绍其连接全球 135 多个交易市场与监管机构的金融科技底层能力。 ### [40:52](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2452s) 从潜在扩散模型到机器人：Black Forest Labs 的技术演进 主持人 Jason Calacanis 介绍了本节的对话嘉宾——Black Forest Labs（黑森林实验室）联合创始人兼首席执行官罗宾·龙巴赫（Robin Rombach）。在开场寒暄中，Jason 误以为“黑森林”是一座城市，Robin 笑着纠正那其实是一片山脉，正是自己从小长大的地方；如今 Black Forest Labs 的办公室分布在黑森林地区的弗赖堡（Freiburg）以及旧金山两地，Robin 也在这两座城市之间奔波。Robin 曾任职于开发出 Stable Diffusion 的团队，如今 Black Forest Labs 则以开源模型 FLUX 及其闭源旗舰模型闻名于世。当被问及公司的商业定位与终极愿景时，Robin 回顾了技术源头：他和联合创始人早在慕尼黑攻读博士学位期间，就发明了潜在扩散模型（Latent Diffusion）。这项算法构成了当今几乎所有主流图像生成、视频生成乃至物理世界 AI（Physical AI）模型的底层基石。 潜在扩散模型的核心哲学在于“数据压缩”。自然界的数据（如高分辨率图像、连续视频和音频）体量庞大，潜在扩散首先将其压缩至一个高效率的低维潜空间（Latent Space）表征中，然后再利用 Transformer 等骨架模型进行训练与去噪生成。Robin 将其形象地比作神经网络版的“JPEG”或“MP3”压缩编码机制。正是基于这一突破，他们先是打造了风靡全球的 Stable Diffusion，进而在 Black Forest Labs 孵化出更新一代的生成式世界模型。当前团队正在研发的是全方位的多模态视觉模型，不仅同时在图像和音频数据上进行预训练，更跨入了一个全新的范式——引入“动作预测”（Action Prediction）。这意味着同一个模型不仅能生成音视频，还能理解物理实体交互，未来可直接部署在物理世界的机器人上。 ```mermaid flowchart TD accTitle: 从潜在扩散到物理AI的技术演化路径 accDescr: 展示了从高维多模态数据压缩到潜在空间，再经Transformer扩展到内容生成与动作预测的流程 subgraph Compress ["数据表示层"] A["高维真实世界数据<br/>(图像 / 视频 / 音频)"] --> B["神经压缩算法<br/>(类似 JPEG / MP3 潜空间编码)"] end subgraph CoreModel ["模型基底层"] B --> C["潜在扩散模型 (Latent Diffusion)<br/>+ Transformer 架构"] end subgraph Capabilities ["能力拓展与输出层"] C --> D["直觉视觉智能<br/>(理解物理交互与世界常识)"] D --> E["多模态生成<br/>(图像 / 视频 / 音频)"] D --> F["物理 AI 与具身智能<br/>(动作预测 Action Prediction / 机器人)"] end ``` 针对将生成模型应用于机器人的推论，Jason 赞叹道，如果模型能够逼真地生成真实世界的视频，就代表它在底层已经内化并理解了现实世界的物理规律与物体运动。Robin 对此深表认同，并提出了关于智能形态的深刻见解：他认为人工智能存在两种互补的维度——一种是基于感知与物理规律的“直觉智能”（Intuitive Intelligence），另一种则是深层的“逻辑推理层”（Deep Reasoning Layer）。完整的通用智能必须由二者结合而成。Black Forest Labs 选择从直觉智能切入，而图像正是最理想的跳板，其算力门槛远低于视频；如今随着视频预训练的深入，模型自然习得了现实世界的物理交互机制，使得同一套模型自然延伸出机器人动作预测能力。 - 团队技术源起：Robin Rombach 及其联合创始人在德国慕尼黑大学读博期间发明了 Latent Diffusion（潜在扩散算法）。 - 双重智能架构：直觉智能（对物理世界感官常识与交互规律的直觉把握）与深层逻辑推理相结合，构成完整的通用人工智能。 - 终极目标：构建统一的多模态世界模型，打通图像生成、长视频、音频以及针对真实物理机器人的动作预测。 --- ### [44:49](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2689s) 摆脱“老虎机”困境：生成式模型的多模态控制演化 Jason 指出了当下生成式 AI 面临的最大痛点与批评——它往往像一个不可预测的“老虎机”（Slot Machine）。创作者在提示词框中输入一段指令，系统便随机吐出一种结果；但创作者无法洞悉内部运行机制，一旦需要微调特定画风、局部色彩或视觉美学，传统模型往往束手无策。对此，Jason 询问技术上将如何彻底破解这一黑盒难题，赋予创作者精细的控制权。 Robin 回应称，解决这一局限的核心逻辑在于“为用户和开发者暴露尽可能多的操控层”（Manipulation Layers）。在技术发展初期，图像模型主要局限于单一的“文本到图像”（Text-to-Image）流程；随后迅速拓展为“文本+图像到图像”（Text + Image-to-Image），让用户能够在现有图像（无论是真实拍摄还是 AI 生成的）基础上输入提示词进行定向修改、局部编辑与风格迭代。接着，模型能力进一步进化为可同时接收多张参考图像与复杂文本，在语义层面上将它们融合并生成全新的内容。如今，这套由粗到细的控制演变正被完整迁移到视频生成领域。更具突破性的是，当所有这些模态都被统一整合进同一个模型的输入与输出端时，创作者将获得对镜头语言、时空连续性与多元素组合前所未有的精准调控能力。 - 控制维度演进路径：文本生成图像（单一文本） $\rightarrow$ 图像微调迭代（文本+单张图像） $\rightarrow$ 复杂语义融合（文本+多张图像） $\rightarrow$ 连续视频控制。 - 消除“老虎机”随机性：通过多层次的参数介入、多图像引导和多模态交互，让生成过程从不可控的随机抽取转变为可精细调整的创作工具。 --- ### [46:29](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2789s) 携手马丁·斯科塞斯：AI 作为导演视觉想象的放大器 对话转入重磅话题：Black Forest Labs 宣布与好莱坞传奇导演马丁·斯科塞斯（Martin Scorsese）达成合作。Jason 坦言，当今市面上的 AI 视频大都是 5 到 10 秒的碎片化短片，拼凑在一起更像是廉价低劣的“AI 泔水”（AI Slop），充斥着令人不适的“恐怖谷效应”，根本算不上真正的电影。Jason 追问，AI 究竟还要多久才能制作出一部具有专业运镜、声音调度、足以让斯科塞斯感到骄傲的完整长片？ Robin 对此表达了截然不同的务实哲学：“我们认为 AI 模型是一种表达媒介（Medium），我们绝不会去规定它应当被怎样使用，更不会去教马丁·斯科塞斯这样的殿堂级大师如何拍电影。”Robin 兴奋地分享了自己与偶像共处一室的经历——作为底层核心算法的研究者，能坐在斯科塞斯身边，看着他亲自测试并探索自己研发的模型，那种震撼无以言表。在实际测试中，斯科塞斯脑海中勾勒出了一处可能用于新片的场景——一个位于东欧某处的偏僻村庄。斯科塞斯用极其细腻的语言描述其构想，系统生成输出后，他们再基于视觉反馈持续交互微调。最终斯科塞斯感慨，人类的自然语言其实是一种“充满损耗的沟通媒介”（Lossy Communication Medium），不同的人对同一段文字会有截然不同的脑补；而图像与视频则蕴含着极其丰富的高维信号。AI 的最大价值，在于能把导演脑海中清晰的心理意象（Mental Picture）瞬间具象化为一系列视觉画面，从而极大地降低沟通阻力。 Robin 强调，让单一 AI 视频模型端到端自动吐出一整部完整电影，或许根本不是技术追求的终极形态。虽然将视频模型接入自主智能体（Agentic Workflow）来跑通长视频非常有探索趣味，但真正激荡出艺术火花的时刻，永远来自于“人在回路”（Human-in-the-loop）——由人类创作者在迭代过程中把 AI 当作纯粹的视觉画笔与思考媒介。 > “语言归根结底是一种充满损耗的沟通媒介……而视觉信息是如此丰富，图像或视频中蕴藏着庞大的信号量。能够把你脑海中的心理图景提取出来，通过图像将其视觉化地表达与传达，这正是这项技术所带来的美妙之处。” —— 罗宾·龙巴赫 - 斯科塞斯实操案例：在闭门测试中，斯科塞斯将脑海中关于“东欧村庄”的构想输入模型，通过图生图与多次迭代将潜意识中的镜头氛围快速具象化。 - 拒绝纯自动化“AI 泔水”：真正的颠覆不在于一键自动生成全片，而在于顶级创作者借助多模态反馈机制，精准锚定镜头语言与艺术表达。 --- ### [50:27](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3027s) 并行化头脑风暴：好莱坞分镜传统的 AI 现代化 针对斯科塞斯的使用体验，Jason 提炼出了一个极具商业价值的定义：AI 是影视前期构思与制作阶段的巨大飞跃，它能够实现“头脑风暴的并行化”（Paralyze your brainstorming / Parallelize your brainstorming）。这种前期的视觉探索完美契合了好莱坞历史上最悠久且核心的创作传统——电影分镜脚本（Storyboards）。 Jason 举例指出，影史诸多巨匠早已将视觉分镜视为创作核心：执导《异形》与《角斗士》的雷德利·斯科特（Ridley Scott）向来以亲自手绘极为详尽的分镜稿而闻名；史蒂文·斯皮尔伯格（Steven Spielberg）在构思《夺宝奇兵》时同样极度依赖手绘草图；而打造《星球大战》宇宙的乔治·卢卡斯（George Lucas），更是长期雇佣庞大的全职顶尖画师与物理模型团队，通过大量的缩微模型与分镜来验证庞大的世界观构想。如今，AI 为创作者提供了一种前所未有的超高效分镜与视觉概念探索手段。然而，影视工业的成熟范式与颠覆性初创公司的野心之间依然存在博弈，初创团队永远在探索如何从更底层彻底重塑内容制作的每一个环节。 - 影视分镜大师传统： - 雷德利·斯科特（Ridley Scott）：坚持亲手为《异形》《角斗士》绘制极其严谨的分镜。 - 史蒂文·斯皮尔伯格（Steven Spielberg）：在《夺宝奇兵》等经典作品创作中极度依赖草图与可视化原型。 - 乔治·卢卡斯（George Lucas）：在《星球大战》制作期间雇佣庞大团队常驻制作实体缩微模型与大量分镜。 - 工业落地优势：AI 模型并非取代导演，而是将数月昂贵的传统手绘分镜与场景探索周期，缩短为数分钟内多方案并行验证的高效工作流。 ### [51:14](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3074s) 初创公司发布视频革命与好莱坞虚拟制作的演进 在商业与叙事的最前沿，生成式视频技术正在重塑传统的生产成本结构。Jason Calacanis 指出，过去初创公司在融到 1000 万美元后，往往会豪掷 10 万到 25 万美元聘请制作团队打磨一段产品发布视频（Launch Video）。而如今，由他投资的许多团队只需花费一两周时间，配合一名导演借助 Flux 等生成式模型，便能以极低成本制作出节奏紧凑、引人入胜的 30 至 90 秒产品演示片。Robin Rombach 证实了这一趋势，并强调看到不同的产品和发布内容生长在同一个底层基础模型之上，正是这种技术最强大、最迷人之处。 叙事工具的下沉不仅影响了创业圈，更直接渗透到了高成本的好莱坞电影制作中。Jason 分享了他在尤里·米尔纳（Yuri Milner）举办的突破奖（Breakthrough Prize）活动上与著名影星盖尔·加朵（Gal Gadot）的交流：加朵近期参演了一部关于比特币的电影，全片未采用传统绿幕，演员们直接在摄影棚（Sound Stage）内无实物表演，所有背景画面均由生成式 AI 在后期或实时生成。这一技术让一部原本需要 1.5 亿美元搭建实体置景的宏大电影，仅用 3000 万美元预算便成功落地制作；若没有生成式 AI，这部电影甚至根本无法立项开拍。 Robin Rombach 承认，高端电影工业是技术验证中最严苛的场景之一。回顾这项技术的发展路径，当他多年前攻读博士学位时，模型仅仅能够生成 64×64 像素的模糊网格；而如今，模型已经能生成高分辨率、长达数分钟的多输入视频。他指出，尽管技术仍在快速迭代，但目前影视制作的核心依然需要保持“人机协同”（Human-in-the-loop）的工作流，赋予创作者精细控制的能力，而不是单纯期待一键成片。 - 传统初创公司发布视频制作预算：10 万至 25 万美元（占 1000 万美元融资额的相当比例）。 - 盖尔·加朵参演的比特币题材电影：凭借生成式背景取代实体搭景与绿幕，将预算从 1.5 亿美元削减至 3000 万美元，使项目得以顺利立项。 - 技术飞跃对比：从几年前博士阶段的 64×64 像素图像，跨越到目前的高清、多模态多输入连续视频生成。 --- ### [54:23](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3263s) 从多模态生成到具身智能：世界模型与机器人大脑 ```mermaid flowchart TD accTitle: 多模态生成模型在具身智能中的演进路径 accDescr: 展示从互联网大规模多模态数据到机器人在物理世界和数字世界行动决策的完整训练与部署流。 subgraph Perception["视觉理解与感知层"] A["大规模互联网视频与图像语料"] --> B["多模态世界模型 (World Action Models)"] B --> C["物理世界动态、因果关系与视觉感知"] end subgraph FineTuning["动作映射与微调层"] C --> D["各硬件平台专属动作表征"] E["数小时真实场景动作微调数据"] --> D end subgraph Execution["智能行动与泛化层"] D --> F["工厂装配与通用机器人操作"] D --> G["数字世界自动化交互 (Computer Use)"] H["未来愿景: 纯上下文即时 Prompt 驱动"] -.-> D end ``` Robin 提出，多模态生成模型最令人震撼的特质在于其通用性——驱动生成电影画面与驱动实体机器人操作底层的神经网络架构本无二致。当前学术界与产业界热议的“世界模型”（World Models）和“世界动作模型”（World Action Models），本质上都是建立在视觉输入理解基础上的预测系统。为了在现实中预测合理的下一步动作，模型必须首先具备高维感知力，深刻理解所视场景的物理规律、因果演化与几何空间，正如生成逼真画面必须理解光影和物质互动一样。这种同构性使得多模态视觉生成系统天然成为了机器人具身智能的大脑。 针对 Jason 提出的疑问——机器人究竟是依靠佩戴手套与第一视角眼镜采集高保真物理数据（如倒水时液体的飞溅、水珠凝结使杯壁湿滑等触觉细节），还是直接从 YouTube 海量视频中自学物理规律——Robin 给出了务实的工程解释。目前的范式是：基础模型通过浩瀚的视觉数据获得通用的物理感知与空间认知能力，随后面对具体硬件（工厂中机械臂迥异的关节角度与动作表征规范）时，仅需少量的“数小时级”特定动作微调数据，即可实现高精度的任务落地。长远来看，行业终极目标是像对待大语言模型那样，无需额外微调，仅凭上下文提示词（In-context Prompting）直接对机器人下达指令（例如“去拿起那杯橙汁”），但目前这仍处于前沿科研阶段。 - 世界动作模型逻辑：感知（Perception）与预测（Prediction）互为表里，动作决策本质是对物理环境下一帧合理状态的仿真推理。 - 机器人模型部署工作流：先依赖海量通用视觉感知预训练，再辅以少量小时级的硬件动作微调（Fine-tuning Data）。 - 具身智能与数字智能的延伸：同类视觉理解逻辑正同时被推向物理工厂装配与纯数字环境下的电脑自主操作（Computer Use）。 --- ### [58:02](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3482s) IP 保护、开源主权与流媒体交互娱乐的未来 当讨论转向内容主权与版权资产（IP）时，Jason 提及了迪士尼庞大的星战、漫威资产库，以及 OpenAI 曾与各方探索授权的背景。对于握有顶级 IP 的传媒巨头，Robin 建议的核心在于赋能“未曾存在之物”的创造，而非机械复刻。针对版权保护，Black Forest Labs 在公开部署的工具矩阵中设置了严格的边界，禁止公网模型肆意侵权生成特定受控 IP；同时，他们积极与主流版权方建立官方合作，探索基于开源模型或专属定制私有模型的联合研发路径，例如此前宣布的马丁·斯科塞斯（Martin Scorsese）相关合作。 展望两三年后的消费端场景，流媒体形态或将发生质的突变。随着模型推理速度的指数级提升，未来的视频消费将从单向播放演进为“实时可交互”的新形态。例如在未来的 Disney+ 平台上，用户可能不再只是被动观看一集剧集，而是能够利用集成的高性能生成工具，在受控的官方 IP 宇宙内进行高互动性的即时内容创作。 Jason 将这种演进同互联网早期的同人文化（Fanfiction）与影迷自制短片（Fan Films）进行了深刻的类比。当年乔治·卢卡斯（George Lucas）对待星战社群展现了极大的包容，只要影迷不进行商业变现，便公开支持粉丝穿上绝地武士长袍自导自演，官方甚至下场发布光剑特效制作指南与正版音效包。如今，生成式 AI 将同人创作的技术门槛彻底击碎，让每个普通粉丝都能以好莱坞工业级的质感，亲手编织属于自己的银河史诗。 - Black Forest Labs 的 IP 商业策略： - 公共平台部署：严格设置 IP 防护栏，禁止未授权知名角色的无界生成。 - 企业级深度合作：基于开源或私有底座，为顶级版权持有者量身打造专属工具包与生成模型。 - 流媒体体验展望：实时推理速度提升将推动 Disney+ 等平台从单向流媒体视频，演变为用户参与的交互式创作空间。 - 同人文化与卢卡斯精神：AI 视频工具将继承星战同人电影（Fan Films）的开放精神，在保护官方版权的同时，极度释放全球粉丝的无界创造力。 ### [01:00:57](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3657s) 生成式 AI 赋能粉丝创作与 IP 商业化新范式 在探讨粉丝文化与 IP 生态时，播客讨论了从过去的道具制作教学（比如如何制作逼真的光剑音效），演变为当下由生成式 AI 主导的粉丝二创热潮。如今，许多创作者正在利用 AI 工具，将《星球大战》（Star Wars）宇宙中尚未被官方讲述的衍生故事具象化呈现出来。这类作品在 YouTube 上迅速走红，其中名为《Star Wars Stories Untold》的频道尤为突出，单条视频观看量动辄达到数百万次。主持人指出，这正是娱乐产业未来的真正走向：让用户或粉丝群体支付授权费、软件订阅费或基于生成产出的费用，自由运用经典角色与故事世界去尽情发挥创造力，而 Black Forest Labs 等领先的视觉生成模型团队正处于赋能这一庞大市场的独特位置。 Robin Rombach 对此表示完全赞同。他认为，如果能够建立一套既能保障 IP 版权方利益、又能满足创作者深度定制需求的商业与合作模型，将会带来颠覆性的变革。Robin 分享了自身的切身体会：平时在阅读书籍或观看电影时，脑海中总会冒出无数种“如果情节这样发展会怎样”或“这里换一种呈现方式会不会更好”的想法。而强大的生成式模型最令人兴奋之处，正是让每个人都能轻松将脑海中的奇思妙想转化为高品质的视觉画面。 - **粉丝二创案例**：YouTube 上的《Star Wars Stories Untold》频道利用 AI 工具重构星战宇宙故事，每条视频吸引数百万次播放。 - **IP 商业化模式探讨**：结合版权许可、软件租赁或按产出计费的混合模式，既保护版权方利益，又赋能粉丝进行超高度个性化的叙事创作。 ### [01:02:02](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3722s) Black Forest Labs 的团队扩张、人才画像与未来展望 随着业务的迅猛增长与新一轮大额融资的完成，Black Forest Labs 的团队规模刚刚突破了 100 人大关。公司目前在德国以及美国旧金山（San Francisco）两地设有办公室，并正积极招募顶尖技术人才。当被问及正在寻找什么样的人才画像与技能背景时，Robin Rombach 详细列出了几个核心招聘方向：首先是具备大规模模型训练经验的 AI 研究员，尤其是精通扩散模型（Diffusion Models）和流匹配（Flow Matching）前沿训练技术的学者；其次是面向客户的工程技术人才，负责与外部客户及顶级 IP 版权方联合研发定制化的物理世界 AI（Physical AI）解决方案；此外，团队还亟需拥有超大规模算力基础设施管理经验的系统工程师，专注于保障海量训练集群平稳运行，并极致提升模型浮点运算利用率（MFU, Model Flops Utilization）；最后，公司也在吸纳致力于将前沿技术落地普及的产品与部署人才。 主持人总结道，Black Forest Labs 的未来前景极为广阔。尤其是在开源与专有模型的双轮驱动下，大型企业既希望借助开源生态获得掌控力，又高度依赖团队专有的前沿模型来获得更精致细腻的视觉表现。随着双方互道感谢与祝福，本期对话在全速前行（“Going all in”）的欢快氛围中正式画下句号。 ```mermaid flowchart TD accTitle: Black Forest Labs 核心招聘方向与技术栈 accDescr: 展示 Black Forest Labs 正在招募的关键技术人才类型及其负责的核心业务板块。 subgraph Research["前沿算法研究"] R1["大规模模型训练研究员"] R2["扩散模型与 Flow Matching 专家"] end subgraph SolutionEng["客户定制与落地"] E1["客户解决方案工程师"] E2["物理 AI 与 IP 定制模型联合开发"] end subgraph Infra["底层算力与系统"] I1["大规模算力基础设施工程师"] I2["集群训练调优与 MFU 吞吐最大化"] end Research --> TechStack["Black Forest Labs 核心生成平台"] SolutionEng --> TechStack Infra --> TechStack ``` - **团队规模与分布**：团队人数已突破 100 人，持续在德国总部与旧金山办公室开放核心岗位招聘。 - **关键技术招聘领域**： - **核心前沿研究**：大规模模型预训练、扩散机制、流匹配（Flow Matching）。 - **定制与工程应用**：Physical AI 开发、商业 IP 版权方联合定制模型交付。 - **系统与算力基建**：大规模分布式训练基础设施搭建、故障监控与 MFU 效率最大化调优。 ## Notable quotes > [03:08](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=188s) 在过去的周期中，企业往往遵循“建好设施，客户自来（If you build it, they will come）”的投机性逻辑；而如今，所有厂商是在拼命“追赶昨日的存量需求”。 — Andrew Feldman > [07:15](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=435s) 计算机其实非常愚蠢，它们只会严格执行你下达的指令。 — Andrew Feldman > [07:15](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=435s) 无限 Token 即无限推理（Unlimited tokens means unlimited reasoning）。 — Jason Calacanis > [25:10](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1510s) 灾难必定会降临，只是具体时间未知。必须提前储备资金与心理预期，因为无论是龙卷风、大地震，还是前所未有的大规模数据泄漏与网络崩溃，都是统计学上的必然事件。 — Andrew Feldman > [27:34](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1654s) 在我们刚才谈论的内容之外，你听过别人问的最聪明的问题是什么？ — Andrew Feldman > [27:34](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=1654s) 事实上我们已经步入了 AGI 时代。如果按照 20 年前乃至图灵测试设立时的技术定义，如今的模型早就已经轻松将其碾碎并远超预期。 — Andrew Feldman > [33:20](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2000s) 科学范式本身不会死，死的是坚守旧范式的人（Paradigms don't die, people do）。 — Jason Calacanis > [46:29](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2789s) 我们认为 AI 模型是一种表达媒介（Medium），我们绝不会去规定它应当被怎样使用，更不会去教马丁·斯科塞斯这样的殿堂级大师如何拍电影。 — Robin Rombach > [46:29](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=2789s) 语言归根结底是一种充满损耗的沟通媒介……而视觉信息是如此丰富，图像或视频中蕴藏着庞大的信号量。能够把你脑海中的心理图景提取出来，通过图像将其视觉化地表达与传达，这正是这项技术所带来的美妙之处。 — Robin Rombach > [50:27](https://www.youtube.com/watch?v=Y7p4rUCdqi0&t=3027s) 头脑风暴的并行化（Parallelize your brainstorming）。 — Jason Calacanis ## Mentioned Andrew Feldman, Robin Rombach, Jason Calacanis, Sam Altman, Ilya Sutskever, Dario Amodei, Demis Hassabis, Elon Musk, Nikesh Arora, Warren Buffett, Martin Scorsese, Ridley Scott, Steven Spielberg, George Lucas, Gal Gadot, Yuri Milner, Cerebras, Black Forest Labs, OpenAI, Anthropic, xAI, Google, 微软 (Microsoft), AWS, 亚马逊 (Amazon), 台积电 (TSMC), Palo Alto Networks, 迪士尼 (Disney), Nasdaq, AppLovin, 智谱 AI (GLM), 月之暗面 (Moonshot AI / Kimi), 通义千问 (Qwen), 葛兰素史克 (GSK), G42, 穆罕默德·本·扎耶德人工智能大学 (MBZUAI), FLUX, Stable Diffusion, 潜在扩散模型 (Latent Diffusion), 晶圆级芯片 (WSE), Bittensor, Optimus, Disney+, 《星球大战》 (Star Wars), 《异形》 (Alien), 《角斗士》 (Gladiator), 《夺宝奇兵》 (Indiana Jones), Star Wars Stories Untold ## Video description (0:00) The AI Buildout: Data Centers Bigger Than Cities (Andrew Feldman) (1:50) Reasoning, Inference, and Breaking Moore's Law (16:28) Open Source, AI Sovereignty, and the Road to AGI (40:54) The Innovation Behind Generative Video (Robin Rombach) (47:31) Martin Scorsese, Robots, and the Future of Hollywood IP Thanks to our partners for making this possible! AppLovin Ads - AppLovin's AI advertising platform reaches over a billion daily active users across mobile games. Full-screen video ads with a 35-second median watch time. Advertisers are profitably spending hundreds of thousands of dollars a day and advertiser access is still in closed beta. The window is open at applovin.com/ALLIN Nasdaq - Industries, capital, and intelligence are converging into a single, interconnected system, and the infrastructure behind it needs to evolve just as quickly. Nasdaq was built for this moment: Powering more than 135 marketplaces and regulators globally and connecting capital to the companies shaping the future. As the innovation economy accelerates, connectivity becomes the critical asset. Nasdaq is the technology platform that makes it possible, and scalable. Learn more at Nasdaq.com. Follow Andrew: https://x.com/andrewdfeldman Follow Robin: https://x.com/robrombach Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@allin Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg #allin #tech #news [Source](https://www.youtube.com/watch?v=Y7p4rUCdqi0)

[3:08] Andrew Feldman: 在过去的周期中，企业往往遵循“建好设施，客户自来（If you build it, they will come）”的投机性逻辑；而如今，所有厂商是在拼命“追赶昨日的存量需求”。 [7:15] Andrew Feldman: 计算机其实非常愚蠢，它们只会严格执行你下达的指令。 [7:15] Jason Calacanis: 无限 Token 即无限推理（Unlimited tokens means unlimited reasoning）。 [25:10] Andrew Feldman: 灾难必定会降临，只是具体时间未知。必须提前储备资金与心理预期，因为无论是龙卷风、大地震，还是前所未有的大规模数据泄漏与网络崩溃，都是统计学上的必然事件。 [27:34] Andrew Feldman: 在我们刚才谈论的内容之外，你听过别人问的最聪明的问题是什么？ [27:34] Andrew Feldman: 事实上我们已经步入了 AGI 时代。如果按照 20 年前乃至图灵测试设立时的技术定义，如今的模型早就已经轻松将其碾碎并远超预期。 [33:20] Jason Calacanis: 科学范式本身不会死，死的是坚守旧范式的人（Paradigms don't die, people do）。 [46:29] Robin Rombach: 我们认为 AI 模型是一种表达媒介（Medium），我们绝不会去规定它应当被怎样使用，更不会去教马丁·斯科塞斯这样的殿堂级大师如何拍电影。 [46:29] Robin Rombach: 语言归根结底是一种充满损耗的沟通媒介……而视觉信息是如此丰富，图像或视频中蕴藏着庞大的信号量。能够把你脑海中的心理图景提取出来，通过图像将其视觉化地表达与传达，这正是这项技术所带来的美妙之处。 [50:27] Jason Calacanis: 头脑风暴的并行化（Parallelize your brainstorming）。

[0:00] We've never seen a build out like this since, you know, the Great Wall of China, right? The pyramids, [3:09] Andrew Feldman: The irony is... they're trying to capture yesterday's demand, [3:09] Andrew Feldman: People used to shop Costco like they shop Safeway. They go down every aisle... and that's a horrible way to shop Costco because you end up with four things you didn't need and each was $22. And as people got more sort of accustomed to it, you go to the back, you get the chicken, 18 cupcakes for the kids' birthday party, bang, you are strategic. And it's exactly the same. [20:06] Andrew Feldman: The polarization hurts a great deal. It hurts clear thinking... and both sides are going to do some dumb things and some really smart things. But in fact, what I found is that the people in the government are trying really hard. [36:50] paradigms don't die. People do. [35:50] Darwin would be in awe of this pace of evolution, [46:32] Robin Rombach: We don't want to tell anyone, especially not someone like Martin Scorsese, how he is supposed to use these models, [58:00] Robin Rombach: For myself, when I read a book or watched a movie, I had so many ideas how it could be done differently or this could have happened... it’s so nice that you can actually enable people to visualize these ideas.

[0:00] We are in the race for super [0:01] intelligence and uh Andrew Feldman is [0:04] back uh and obviously CEO and founder of [0:08] uh Cerebras doing inference chips [0:11] pioneered the space had a successful [0:13] IPO. We've talked about this a couple of [0:16] times. We got to see each other in [0:17] January at Davos. IPO happens. Uh the [0:20] boys and I got to sit with you recently. [0:22] >> That was fun [0:23] >> at liquidity. [0:24] >> That was really that was really fun. [0:25] >> Had a great discussion with the boys. I [0:27] wanted to deep dive with you about a [0:30] couple of topics. The first one is the [0:33] buildout of AI. We've never seen a build [0:37] out like this since, you know, the Great [0:40] Wall of China, [0:41] >> right? Who knows? [0:42] >> The pyramids. I mean, it feels like the [0:45] amount of capital, time, and intelligent [0:48] people on the planet dedicating [0:50] themselves to the buildout of something. [0:54] Um I I can't think of anything in our [0:56] lifetimes with per perhaps uh you know [0:58] before our lifetimes the war effort [1:00] >> right [1:01] >> this is a mobilization and a scale that [1:04] we read about we hear about but you're [1:07] actually doing it you have customers who [1:09] are building data centers and you're a [1:11] key piece of that [1:12] >> all [1:15] app started with an $8 domain and no VC [1:18] funding and became one of the largest ad [1:20] platforms in the world now that same [1:22] engine powers Appleven ads for [1:24] e-commerce. Your ads run inside mobile [1:26] games, reaching over a billion people [1:29] with full screen, distraction-free [1:31] attention. The platform finds buyers and [1:33] optimizes for profit. You set the [1:35] target, it does the rest. One cookware [1:37] brand went from 4 million to $16 [1:39] million, turned profitable, and is on [1:41] pace for 80 million this year. Visit [1:43] apploven.com/allin [1:45] to launch your first campaign today. [1:49] >> Maybe you could just enlighten us in [1:51] 2026. What is Cerebras doing and what is [1:55] happening with this buildout out in [1:58] Texas? These are some gigantic gigantic [2:01] efforts. The the size and scope of what [2:04] is being built, the physical size and [2:07] scope. Usually when we talk about [2:09] software or we talk about hardware, [2:11] we're talking about chips or boxes and [2:12] and they don't have the same sort of [2:14] physical enormity. Right. [2:16] >> Right. And what we're talking about now [2:18] are data centers that are in the next [2:21] several years going to use more power [2:23] than the previous 50 years on Earth [2:26] took. Wow. Right. We're talking about [2:28] individual buildings the size of [2:31] football fields that have more power [2:32] coming into them than midsize cities. [2:36] And they're being built they're being [2:38] built across the US. They're being built [2:40] in Canada. They're being built [2:42] throughout the Nordics. are being built [2:43] here in Paris and throughout France, in [2:45] Europe, in the Middle East in nations [2:48] that sort of weren't front and center in [2:50] anybody's mind previously. You know, [2:52] Kazakhstan, Tajjikstan are building out [2:55] Georgia building out data centers of [2:56] size. Armenia, everybody's sort of [3:00] focused. [3:01] >> Every country [3:03] >> um and every state obviously in America [3:05] feels they need to participate in this [3:08] >> and the people who are buying the [3:10] capacity, the open AI, anthropics, [3:12] SpaceX, [3:13] SpaceX AI, [3:16] uh the Googles, they are insatiable [3:20] right now. [3:20] >> Yeah. [3:22] and they're building how many years out [3:26] when you talk to them? [3:28] They were ordering chips from Cerebras [3:31] before you were finished with the chips. [3:32] They're putting orders in ahead of time. [3:34] The irony is unlike many sort of [3:38] exciting times in technology. They're [3:40] trying to capture yesterday's demand, [3:42] right? The demand is way outstripping [3:45] our ability to build data centers and to [3:47] fill them with hardware. [3:49] >> All right? And so, you know, we have a [3:51] $25 billion backlog, [3:54] >> $25 billion backlog [3:56] >> and we we are not alone in that that [3:59] that OpenAI uh Anthropic, you go through [4:02] this list of of uh Google wants more [4:06] data centers, Microsoft wants more data [4:07] centers, AWS wants more data centers, [4:09] right? All of these players are not [4:12] chasing sort of if you build it, they [4:14] will come. They're chasing the demand is [4:16] booked, [4:17] >> right? How do we keep them from leaving? [4:19] Right. And and that that's extremely [4:21] unusual. [4:22] >> It's very unusual. And now we have [4:25] people who are, you know, we have a term [4:28] for a token maxing. [4:29] >> Yeah. [4:30] >> And there's a great debate. Is this [4:32] actually creating value? I'm curious [4:36] where you stand. You know, is it even [4:39] possible that this much demand could be [4:41] created if value did not exist? There is [4:44] clearly massive value happening. [4:46] >> Yeah. But there's also massive [4:48] experimentation. [4:49] >> Oh, for sure. You know, I [4:52] you know what I I I liken this to when [4:54] we first started with uh AWS [4:58] and it was so good to get around your [5:00] own IT organization, [5:01] >> right? [5:02] >> That you told every engineer, yeah, go [5:03] ahead, put on your credit card kite and [5:05] sign up. [5:05] >> Yeah. [5:06] >> Right. And a lot of it was really useful [5:08] and some of it was like, God, I wish we [5:10] didn't do that. [5:11] >> Yeah. And so for sure there's [5:13] experimentation, but it doesn't mean [5:16] that the net value isn't enormous. It it [5:19] means some of it is going to go nowhere. [5:22] >> And you know, it was the same. I [5:24] remember when Costco opened up in in in [5:27] in the Palo Alto area in 1988. And [5:30] people used to shop Costco like they [5:32] shop Safeway. They go down every aisle. [5:34] >> Yes. [5:34] >> And that's a horrible way to shop Costco [5:36] because you end up with four things you [5:38] didn't need and each was $22. [5:40] >> Right. And as people got more sort of [5:43] accustomed to it, you go to the back, [5:45] you get the chicken, [5:46] >> 18 cupcakes for the kids' birthday [5:48] party, bang, you are strategic. And it's [5:50] exactly the same. [5:52] >> I think at first people opened up and [5:54] said everybody as much tokens as you [5:56] want. [5:56] >> And the in enterprises, there's no open [5:59] loop. We don't give sort of any resource [6:03] unconstrained to people. And now we're [6:06] jumping on saying, "Whoa, all right. [6:07] These guys should have as much as they [6:09] need. They're enormously productive over [6:11] here. We can use maybe an open- source [6:13] model, maybe a cheaper model over here. [6:16] And now we're sort of running like a [6:18] business and we're really seeing a [6:22] certain type of person emerge who knows [6:25] how to deploy this technology. Systems [6:27] thinking, [6:28] >> yeah, [6:30] >> which developers kind of have innately [6:32] CEOs tend to be great strategists and [6:35] understand systems. Uh but this the [6:39] intelligence is getting so much better [6:41] every step along the way that I'm [6:44] watching individuals typically startup [6:46] founders but also venture capitalists [6:48] and associates who work at my venture [6:49] firm. They start playing with the tool [6:52] and then the tool starts playing with [6:54] them. They start to go, "Oh, I haven't [6:58] clearly defined what my goal is. I don't [7:01] understand what a system is. I don't [7:03] under I've never heard about making a [7:05] requirements document." And the [7:07] software's like, "Do you have a [7:08] requirements document? What's your [7:09] goal?" The AI starts telling people, [7:11] "You're token maxing and you need to get [7:13] a little more focused here." [7:15] >> One of one of my colleagues 20 years [7:17] ago, a really smart smart computer [7:20] scientist said, "Computers really dumb. [7:23] They do exactly what you tell them." [7:25] >> Yeah. [7:26] >> And at first, prompting was like that, [7:28] right? You modified your prompt a little [7:30] bit and it changed the answer [7:32] >> dramatically. [7:33] >> Dramatically. And increasingly it's [7:36] understanding what your intent was. [7:38] >> Right. Right. And if you if you have a [7:40] chance to to to play with Fable or or 56 [7:43] from from OpenAI, increasingly what you [7:47] don't have to get the prompt just right. [7:49] You don't have to be a prompt whisperer. [7:51] Instead, you ask it and it says, "Well, [7:53] here are some things and and by the way, [7:56] maybe you wanted the chart to to to go [7:58] two ways. You wanted a line and a bar." [8:00] And it's like, "Well, that's exactly [8:01] what I wanted. I didn't ask for it, but [8:03] that is better. And and so it's it's [8:07] understanding intent and that's a huge [8:09] leap which if we were sitting here two [8:12] years ago, the idea [8:14] >> we would never have been able to predict [8:16] in a short 24 months that it would go [8:19] from being a great summarizer researcher [8:22] of web results, [8:23] >> right, [8:24] >> to actually understanding your intent [8:25] and then providing a solution and [8:28] abstracting it all from you. [8:29] >> That's right. That's right. [8:30] >> Which is a very weird thing. I don't [8:32] know if you've played with the Hermes [8:33] agent yet. [8:34] >> Yeah. [8:35] >> Have you played with it yet? [8:36] >> I mean, I asked it just this morning. [8:40] Uh, and I I was given a secret bit [8:43] tensor project that has the new um ZAI's [8:48] uh model 52 [8:50] >> and they gave me [8:51] >> GLM 52. Yeah. [8:52] >> GLM 52. So, somebody in that Bit Tensor, [8:55] I think you understand Bit Tensor, [8:56] you've heard of it, the distributed [8:59] >> um crypto project. And so they have all [9:02] this extra capacity. I was a whisperer [9:04] told me probably some capacity in China [9:06] that has free energy. Okay, fine. So [9:09] they gave me unlimited capacity. So I [9:11] started having to do some really crazy [9:12] jobs where I was saying like every hour [9:15] I want you to tell me what the trends in [9:16] the world are that nobody else has [9:19] identified yet. And you can do whatever [9:20] you want to do that. But my goal is to [9:22] be the smartest trend hunter in the [9:24] world. And I watched what it was doing [9:26] in the background and it started [9:27] debating itself [9:28] >> on where it should find the things. He [9:30] said, "Well, we should probably go to [9:31] Hacker News and Reddit." And then I was [9:32] like, "Yeah, but there's also social [9:34] media and trends tend to manifest on [9:36] Instagram." [9:36] >> That's a reasoning model. You were [9:38] watching a reasoning model work out. [9:40] >> Yeah. Isn't that interesting? I mean, [9:43] that's amazing. And and [9:45] it was collapsed. So, as a civilian, [9:49] >> right, who doesn't hit the uncolapsed [9:52] moment, and if you were using chatbt 3.5 [9:55] or you were using 4.8, whatever it was [9:57] and you haven't used this new level of [10:00] reasoning and inference and unlimited [10:03] compute essentially, [10:04] >> right? [10:05] >> It opened my eyes just this morning of [10:07] what a world [10:09] >> of unlimited tokens might look like, [10:11] >> right? [10:12] >> Cuz unlimited tokens, I believe, means [10:14] unlimited reasoning. It does. [10:17] What does that mean? [10:19] Yeah, it's uh I mean if you run these [10:23] for 25 or 48 hours, you get amazing [10:26] things now and what what if by using [10:29] Cerebrus we were 15 times faster and [10:31] then you ran it for 24 hours, [10:34] >> right? A and you got weeks or months [10:37] worth of thinking. [10:38] >> Yeah. And I mean it it is uh it is [10:41] extraordinary. And I I think one of the [10:43] things is people like Ilia and Sam in [10:46] the early days were saying this was [10:48] coming. [10:48] >> Right. [10:49] >> Right. And I think when you look back [10:51] you say to yourself, "Holy crap." Those [10:55] guys saw it. [10:56] >> Yeah. They were They could see around [10:57] the corner. [10:58] >> That's right. And the rest of us could [10:59] see like what? I'm not sure. It's when [11:02] we had Sam on Allin uh at one point he [11:06] uh and he said, "Oh, you know, I'd love [11:08] to come on at some point." I said, [11:09] "Sure, come on." And he he was talking [11:10] about it. He said, [11:12] >> you know, I said, "What's next?" He [11:13] said, "Reasoning." I said, "Unpack that. [11:15] What does it mean?" [11:16] Well, [11:17] >> understanding what your intent was just [11:19] as you're saying and then figuring out a [11:21] strategy and then maybe talking to other [11:24] agents and other threads about like is [11:26] this the right thing to do and vetting [11:27] each other's work and I'm like wow we [11:29] have come a long way from guess the next [11:31] word [11:31] >> right right fill the sentence in you [11:34] know summarize this PDF [11:36] >> now Cerebrus is at the center of this [11:38] because this reasoning is inference [11:41] >> this reasoning is inference and it's [11:42] computationally intensive [11:44] >> right [11:44] >> right and So fast compute makes this [11:48] sort of work fast and sort of tractable. [11:51] It doesn't it by taking a huge [11:54] amount of time to get a good answer. [11:56] >> And so it's exactly the fact that that [11:58] that this reasoning consumes a huge [12:01] amount of tokens internally [12:03] >> that allows a blisteringly fast machine [12:06] like ours. And I I brought one because I [12:09] I'm never far without you know when one [12:10] costs half a billion to make you you [12:12] bring it everywhere with you. Uh we we [12:14] were um we were tossing this back and [12:16] forth at Davos. [12:18] >> Uh what's the model number of this one? [12:21] Th [12:21] >> this was in the first eight or 10. Um [12:23] >> so this has a special place. [12:25] >> This has a special place. I mean my wife [12:27] says it's like I'm a kid with a a dirt [12:28] bike for his 8th birthday. It was in his [12:30] bedroom at night. I I carry him with me. [12:32] >> I I mean when you have um you know your [12:35] your next party at the house, I highly [12:37] recommend just a little or something or [12:40] I [12:40] >> I think it would be like a great fit. It [12:42] would be a great bit if you had some. [12:44] >> That's right. Um, but what we're looking [12:46] at here is the ability to do that [12:49] reasoning at scale. And [12:53] what is Moore's law for inference and [12:57] for cerebrus? Do you have something [12:59] internally you discuss as [13:02] we're going to double this every x time [13:06] period? So all chips prior to us in the [13:10] processor world followed Moore's law. [13:12] >> Got it. [13:13] >> And we broke [13:14] >> doubling every 18 months. [13:15] >> Doubling about every 18 months. [13:17] >> Got it. [13:18] >> Um and we crushed it with this chip [13:20] >> and we've carved out a whole new [13:22] trajectory. [13:24] And uh my view is in the next 18 months [13:28] we'll be way over 2x [13:30] >> interesting. And so, uh, I I think that, [13:34] uh, early in an architecture, you have [13:36] room to to do much better than what was [13:39] traditionally Morris law. Now, if you've [13:41] got a 20-year-old architecture like the [13:43] GPU, it's much harder, [13:44] >> right? [13:45] >> You you have to rely on things like [13:47] smaller geometry, [13:48] >> right? Going to the next fab node. But [13:51] in a newer architecture, you have a huge [13:54] amount of room still to to to learn [13:57] about the the work that that is being [13:59] presented and make optimizations that [14:02] that give you huge gains. H how do you [14:05] run the company? Like just being the CEO [14:07] now in the age of AI, um you have $25 [14:12] billion in demand. You have to you have [14:16] to deploy at an just an incredible [14:18] blistering pace. You have to hire [14:21] people. You have to create a roadmap. I [14:23] don't mean to give you a panic attack [14:25] here. [14:25] >> You have to keep up with somebody like [14:27] OpenAI who's moving so unbelievably [14:29] quickly. Yes. [14:30] >> Right. And they're they're competitive. [14:31] You got to keep up. [14:32] >> Right. [14:33] >> Right. Your hardware, your software, [14:35] your deployments have to keep up with [14:37] some of the fastest moving organizations [14:39] in history. [14:40] >> They're demanding customers. [14:42] >> They are not they're not pushovers for [14:45] sure. [14:46] >> Yeah. A and also potentially competitors [14:49] down the road. [14:50] >> I look I I think there is so much demand [14:53] right now that that that [14:56] there is no silicon that will go unused, [14:59] >> right? [14:59] >> But why isn't OpenAI releasing Jalapeno? [15:02] Why is Amazon making their own chips? [15:05] You see this reoccurring trend? Is it a [15:09] way to let you know to let Jensen and [15:12] Nvidia know, hey, we can do this, too. [15:14] So we we need good pricing. Is it is it [15:16] a little bit of a flex that way or is [15:19] that the future that they're going to be [15:21] in your business? [15:22] >> No, I I I think nobody likes being [15:24] dependent [15:26] >> and I I think some of the lessons [15:28] learned by the the hyperscalers of the [15:31] x86 world is they were dependent on [15:34] Intel. [15:36] And uh some of the lessons learned by uh [15:40] the GPU makers was they were dependent [15:43] on a small number of hyperscalers. [15:45] >> Yeah. [15:46] >> And they wanted more customers and so [15:48] they set about to to help fund these [15:51] neoclouds. [15:53] >> And so I I think mostly it's about uh an [15:56] opportunity to control at least an [15:58] important part of your destiny. [15:59] >> Got it. [16:00] >> And I I think that's a very reasonable [16:02] thing. I think you don't have to sort of [16:04] make the fastest chip. You you just [16:06] can't be entirely dependent on other [16:08] people's chips. [16:09] >> And that dependency has become a hot [16:11] topic. Not sure if you caught the [16:12] episodes over the last two weeks, but [16:14] we've been talking over the last year [16:17] about open source. I've been championing [16:19] that a lot just because I was early into [16:22] Open Claw and quickly started using [16:24] Kimmy and was like, "Wait a second. I'm [16:27] blowing out my claw tokens, but this [16:30] Kimmy, I can't tell the difference. And [16:32] then we started smart routing it and [16:34] suddenly this open source started to [16:37] figure out reasoning and the gap has [16:41] >> suddenly closed this year. [16:43] >> Well, I I you know, you you don't want [16:46] to take your your Ferrari to the grocery [16:48] store, [16:49] >> right? You you you there there are times [16:51] you want to drive your fun car. [16:52] >> Yeah. [16:53] >> Right. And there are times you want to [16:54] throw the kids in and and don't worry if [16:56] they're Cheerios on the floor. [16:58] >> Minivan time, [16:58] >> right? There's minivan time. And and and [17:00] I think that as the sort of [17:03] sophistication of the user grows, right, [17:06] you're going to have hard problems and [17:08] those are going to be frontier model [17:09] problems. They're going to be open AI [17:11] problems. They're going to be anthropic [17:12] problems. Going to be Gemini problems. [17:14] And behind that, they're going to be a [17:16] lot of ordinary problems. Right? I mean, [17:17] if you think about a company, you know [17:19] how much time is spent cutting things [17:21] out of workday and getting it in a [17:23] different cell for Yeah. [17:26] >> Right. Think about [17:27] >> the cutting and pasting economy is real. [17:29] >> That's right. And and this doesn't need, [17:32] right, gold medal math. No. [17:34] >> What what this needs is sort of rock [17:37] solid open-source capabilities. Yeah. [17:40] And if you think about what I mean, [17:42] well, we've been thinking a lot about it [17:43] in DNA, but a huge amount of DNA, all [17:46] right, is not invention, [17:50] >> right? And you you may not need sort of [17:54] the most sophisticated agents for this. [17:56] And another [17:59] card that's turned over recently is some [18:02] folks uh maybe have concerns with the [18:06] ambition of the frontier models and [18:07] maybe sharing their data data leakage [18:11] and sovereignty of intelligence and [18:14] they're saying hey our company is going [18:17] to choose maybe we're in a regulated [18:19] industry finance healthcare HIPPA you [18:22] know FINRA all kinds of different [18:23] regulations [18:25] >> we need to have this on [18:27] >> prem [18:29] domestically and we'd liken an [18:31] open-source version where we have uh a [18:34] little bit more control. Yeah. And I I [18:36] think [18:37] >> are you seeing that now? [18:38] >> We are seeing that for sure. And I I [18:39] think OpenAI made a good call releasing [18:41] OSS 12B some months back. That was a [18:44] good open-source model. Um but I think [18:47] in the US we need more domestic open [18:48] source models. We need to give the world [18:50] a choice, [18:51] >> right? Right. If they want to run open [18:53] source right now, it's OSS 12B or [18:56] Chinese models. [18:57] >> Nvidia has some in [18:58] >> Nvidia has seen the same opportunity to [19:01] push open- source models. I I I think [19:05] giving them more power might might be [19:07] sort of [19:07] >> Well, I was about to that was you cut me [19:09] off at the past like my understanding [19:12] was Jensen was like, "Hey, we we don't [19:14] even want to talk about these open [19:15] source models we have because our [19:17] customers, [19:18] >> right? [19:19] >> We're now going to be competing with [19:21] Sam, Daario, Elon, [19:23] >> uh Sergey, like do we want to be in that [19:26] position, [19:27] >> right? [19:28] >> So, but we do need some more champions [19:30] here and it's open source so people can [19:31] fork it." Um, but that puts you in a [19:34] more neutral position. [19:35] >> That's right. We we we run today, we run [19:38] GLM, we run Kimmy, we run the Quen set [19:41] of models and we run OpenAI's models, [19:43] the closed source ones. We run models [19:47] for say Galaxos Smith Klein, which they [19:49] wrote and developed. Um, we run models [19:52] for uh our partner uh in the UAE uh G42 [19:56] and MBZ UAI. Yeah. Um that are are are [19:59] their models that they designed. So we [20:02] we have a a a wide variety. [20:05] >> So sovereignty is a trend. [20:06] >> Sovereignty is a trend. And I think uh [20:09] the the government's actions with regard [20:12] to fable and uh uh 56 um where they [20:18] said, "Oh, whoa, let's think and then we [20:21] can act." [20:22] >> Um I I think sort of particularly here [20:24] in Europe was a bit of a wakeup call. [20:27] And when you saw this going down, [20:30] there's a layer of partisanship in our [20:32] country right now. It's pretty fervent. [20:35] Daario is pretty explicitly, you know, [20:38] not part of this administration. They [20:40] they've been very adversarial. Both [20:41] sides have been have admitted that [20:43] they're starting to work it out now. So, [20:44] it's hard, I think, for us not being in [20:47] the room with these parties to [20:49] understand what's partisanship, what's [20:50] games here. But do you believe that what [20:54] they released was truly dangerous for [20:58] cyber warfare, for cyber attacks, and [21:01] that if you were to rate Daario's [21:04] not communication, because he's a very [21:06] effrovescent communicator, um I think is [21:09] a diplomatic way to say it, um but to [21:14] have a scheduled rolled out release, [21:16] >> right? We'll put aside the government's [21:18] control of it, but do you think that is [21:21] is a wise thing for us to do at this [21:22] point? And do you think it's there was [21:24] actually a major threat there? [21:26] >> So what's interesting is I hadn't seen [21:28] it before, [21:30] >> right? And I I think you know if we just [21:33] step back and say is it reasonable I [21:37] don't know whe whether this was the [21:39] right time but at a time [21:41] >> that that that a model is sufficiently [21:46] uh creative in its thinking that it [21:48] poses a meaningful threat for the [21:50] government to say we'd like you to roll [21:52] it out in steps. [21:53] >> Yeah. [21:53] >> Now this doesn't seem unreasonable to [21:55] me. [21:55] >> Not at all. [21:56] >> Right. I mean we we do this with [21:57] powerful pharmaceuticals. [21:59] >> Right. We we we'd like I mean we're [22:02] certainly not encouraging seven years of [22:03] trial and the amount of paperwork and [22:05] all the garbage that has acred to the [22:07] FDA, [22:08] >> but with a powerful new technology, it [22:11] certainly doesn't seem unreasonable to [22:13] say, "Hey guys, um let's at least do [22:17] some red teaming at the government so we [22:20] know our defenses can block this." [22:23] >> Yeah. Have we checked Have we checked [22:24] >> the infrastructure of the country like [22:26] >> of the NSA? Have we checked the [22:28] infrastructure up? Right. And can you [22:30] give us two or 3 weeks to patch any [22:32] obvious holes that are found? This [22:34] doesn't seem to be [22:36] >> an unreasonable thing for the government [22:38] to ask. [22:39] >> Right. We But we in this very polarized [22:42] time [22:44] >> put on top of it, well, oh my god, it's [22:46] President Trump doing it. And then you [22:47] have to think, well, what if it was [22:48] President OA AOC or President anybody in [22:52] between the the two extremes? [22:53] >> I I think the polarization hurts a great [22:56] deal. It hurts clear thinking, [22:58] >> right? It hurts clear thinking and and [23:00] and both sides are going to do some dumb [23:02] things and some really smart things. [23:04] >> Right. [23:04] >> Right. And in fact, what I found is that [23:07] the people in the government are are [23:09] trying really hard. [23:10] >> Um [23:11] >> the rank and file, [23:12] >> the rank and file are trying really hard [23:14] and this is moving fast. And I I think [23:18] that that an ability to to set aside [23:20] some some of the the polarization and [23:23] say h how do we do this in a reasonable [23:26] manner? I mean we we want Dario and Sam [23:29] competing like crazy. [23:32] It's awesome, right? It's good for the [23:34] technology. It's good for it's good for [23:37] entrepreneurs to see even with thousands [23:39] of people this is what what you can [23:41] continue to achieve. Right. [23:43] >> Right. This is Google in the ass made [23:45] them get sharper. Amazon got everybody [23:48] got better because of that. We want that [23:50] and we certainly don't want to become [23:52] sort of a region where the first thing [23:54] we want to do is regulate it. [23:55] >> Right. [23:55] >> Right. But as it gets more powerful [23:59] and the industry really should do a [24:01] better job of regulating itself perhaps [24:03] and and it did seem like they were [24:05] starting that process but then the [24:08] communication was lacking maybe. [24:10] >> Yeah. Uh, you know, I I think not only [24:14] are they racing hard, but they're [24:16] inventing this as they go too. [24:18] >> Yeah. Right. There's not a playbook. No. [24:21] >> Right. They're inventing the we we say, [24:22] "Oh, just put on guardrails." Well, they [24:24] have to design the guard rails. [24:25] >> Sure. [24:26] >> Right. They the guardrails have an [24:27] impact. Um, you know, one of the things [24:30] that fast does is it makes the guard [24:31] rails less painful. And so that that we [24:34] we discovered that in the last six [24:36] weeks. Yeah. is that the very guard [24:39] rails can add time and make it feel [24:42] slower and so fast ships like ours can [24:45] really help that. But so they're racing [24:49] against competition. They're racing [24:50] against their own sense of greatness. [24:52] >> Yeah. [24:53] >> Right. Which is maybe even the biggest [24:55] driver here. A and I think that they're [24:58] in earnest trying to think about how I [25:02] do the right thing. A and all of those [25:04] are are are mixed in this bucket. and [25:07] and sometimes you're on one side rather [25:09] than the other. And [25:10] >> yeah, and as you're saying, this is a [25:13] first time, right? When we when 3.5 came [25:16] out, it wasn't like when we're using [25:17] chatbt 2.5, 3.5, it was taking down [25:20] networks, but in talking to Nicash uh [25:23] from PaloAlto Networks, I asked him [25:26] like, "Hey, well, how would you grade [25:28] this?" And he said, "We put it against [25:30] our software and we found bugs we were [25:33] not aware of and [25:34] >> it killed them." [25:35] >> Yeah. He said we had to stop everything [25:37] we're doing and do patches for 6 weeks, [25:39] >> right? And and that's when you know, [25:41] right? I mean, Nick leads, you know, [25:44] maybe the leading security software [25:46] firm, right? [25:47] >> And when it finds in an hour, [25:51] right, tens of critical opens, you're [25:54] like, whoa, this is a powerful tool [25:57] >> and we need to think and and maybe you [25:59] you show it to a a group first, right? [26:02] May maybe you I I don't know what the [26:04] right thing is but [26:05] >> yeah I mean red teaming and uh we've [26:07] always had just when you were releasing [26:10] the new version of um an operating [26:13] system you know when you have your [26:14] iPhone you can say I want to be part of [26:16] the beta. [26:16] >> That's right. you know, right? And [26:18] there's like two other baiters that you [26:19] don't even get the chance to opt into as [26:21] consumers. And those ones are for [26:23] security. Those ones are for, you know, [26:26] making sure you don't lose your data or [26:27] data. [26:28] >> That's right. Disappear or leak or [26:30] >> corruption. Any any number of these [26:33] things. [26:33] >> I think we can also know [26:36] that that there will be a massive data [26:38] leak. [26:39] >> Of course, [26:39] >> we we know this. Yeah. [26:40] >> Right. And it's like uh Warren Buffett [26:43] talked about the reinsurance industry [26:44] that you know something bad's going to [26:46] happen. You don't know when. [26:48] >> Yeah. [26:48] >> But you got to save up for it, right? [26:50] You put money away for reinsurance, [26:53] >> but there will be a tornado. There will [26:55] be a massive earthquake. I mean, we we [26:57] know this and we can do our best to [27:00] plan, but there'll be a massive breach [27:03] and they there'll be and we we have to [27:05] steal ourselves in advance and we have [27:07] to think about it and think about the [27:08] right response at the time and sort of [27:10] prepare ourselves for a future that is [27:13] in specific unknown. But in general, [27:15] we're pretty sure it's something's going [27:16] to happen. [27:17] >> Something will happen and [27:21] >> yeah, it's typically a black swan, [27:23] right? By definition, it's going to be [27:25] something we didn't consider or a [27:26] question we didn't know to ask, [27:28] >> right? But but even knowing that there's [27:30] some unknown unknowns is a useful place [27:33] to start. [27:34] >> Yeah. What are we not asking ourselves? [27:36] >> That's right. [27:36] >> With reasoning, the AI is going to be [27:38] able to tell us, "Hey, schmuck humans." [27:40] >> That's right. By the way, here's what [27:42] you're not thinking about. This is now [27:44] my closing sentence when I do my [27:46] prompting is I need you to make me a [27:48] prompt that will help me do this trend [27:51] scouting for an example. And then I [27:53] always say at the end um please check [27:56] your work [27:57] >> right and uh then tell me what I haven't [28:01] considered in terms of my goals and uh [28:04] give me ask me some questions every time [28:06] you run the job and that has changed [28:08] everything because it's like I checked [28:09] my work by the way this was incorrect [28:11] right [28:12] >> and uh I'm wondering hey uh would you [28:14] like me to also do this and some of the [28:16] tools like perplexity do that [28:18] automatically they give you your next [28:19] three prompts but if you give it [28:20] explicit instructions [28:22] My lord is it um good at that. [28:24] >> So, you know, over the course of the [28:27] last 10 years as I was raising money, I [28:29] I thought one of the smarter questions I [28:31] got at the end of a conversation where [28:33] someone asked, "What was the smartest [28:35] question you heard that that wasn't [28:37] covered by what I asked?" [28:38] >> It's incredible. right now. Now, that's [28:40] somebody who's curious and thinking and [28:42] humble and and trying to [28:44] >> sort of use this to get a picture of of [28:48] the space and and to the extent that you [28:50] can ask the AI that [28:51] >> and and and that it can sort of broaden [28:54] your your view, you know, maybe what [28:56] question should I have asked [28:58] >> to to be an expert in this? What what [29:00] what would a a PhD level uh uh [29:04] questioner ask of this or a gold medal [29:06] math? I mean I I think those are the [29:09] sort of questions that you you you know [29:12] you don't even know how to ask [29:14] >> which you know if you start thinking [29:16] about AGI and super intelligence [29:20] um you know they're just definitions but [29:22] they're important definitions I think to [29:25] kind of keep in mind because they're way [29:27] points. [29:27] >> That's right. And AGI I think I suspect [29:32] you'll agree with me that we've hit it. [29:35] we just haven't exactly deployed it [29:38] fully. We we we have artificial general [29:40] intelligence now. It feels like when [29:41] we're talking about these reasoning [29:42] moments and you know the the the ability [29:46] for it to be as smart as any human but [29:48] >> let's talk about [29:49] >> by any definition we had 20 years ago [29:51] we've hit it. [29:52] >> Yes. [29:52] >> Right. I mean if you think about oh [29:54] there's a touring test blew it away. [29:56] Yes. [29:56] >> I mean, you think about that that any [29:57] period of time sort of 10, 15, 20, 30, [30:01] 40, 50 years ago, we we we any [30:04] definition we would have previously put [30:06] forward, right? [30:07] >> We've blown past it. And so, which goes [30:09] back to our previous point of like, do [30:10] we know the questions to ask 20 years [30:12] ago, science fiction authors, you know, [30:16] had their say and we answered all their [30:18] questions, [30:19] >> right? [30:20] >> If they were to look at this today, [30:21] they'd be like, "Well, I'm out of [30:23] >> I'm out of questions. [30:23] >> I'm out of questions. Sorry." And that's [30:25] where sort of the the sort of listening [30:29] to people who we who sound sometimes [30:31] like they're on the fringe, [30:33] >> right? When when when Ilia was talking [30:36] eight or 10 years ago about the need for [30:38] safety and then and you're like, "What?" [30:41] And dead right. [30:43] >> Yeah. [30:43] >> Right. When when when Elon was talking [30:45] about building rockets and driving the [30:47] cost to to to near zero of of of a [30:50] launch vehicle, you're like, "What?" And [30:52] there it is. And now you can see and [30:55] that's I think that's why it's really [30:57] fun to be a technologist now. [31:00] >> Well, and with these tools specifically, [31:02] you know, we're talking about building [31:04] all these tools and then the tools are [31:07] starting to build themselves in this [31:08] recursive loop. [31:09] >> That's right. [31:11] >> We're kind of just starting to see [31:14] people apply loops. Uh, in fact, loop [31:17] maxing became when I was doing my trend, [31:19] when I did my trend thing, it kept [31:21] picking up loop looping and it kept [31:23] picking up the maxing stuff and it [31:25] created a buzz word for me, loop maxing, [31:27] right? And then it [31:29] >> magically people started talking about [31:30] loop maxing and I was like, "Wow, this [31:32] is really weird." It anticipated that [31:33] this would other humans would come up [31:35] with this word. But talk a little bit [31:37] about recursive and then the road to [31:39] super intelligence. And do you have a [31:41] way Andrew that you think about super [31:43] intelligence and what [31:45] it will mean for humanity and how we [31:48] will define it and how we'll experience [31:51] it? Yeah. [31:52] >> I I I think let's begin on on on loop [31:54] maxing or sort of recursive learning. I [31:57] I I think um [32:02] I think what what what Sam and Ilia and [32:04] then later Daario and and and Dana [32:07] Dennis saw [32:09] um [32:12] six years ago or five years ago was that [32:15] um [32:16] powerful recursive gains [32:19] are are exponential, [32:22] >> right? you get better, you do it again. [32:23] And if you continue to get gain, the the [32:26] the the slope of that curve is so steep. [32:30] >> Yeah. [32:31] >> And that um we're just beginning to see [32:34] that now. [32:35] >> You ask it a question, you learn from [32:37] the results, you ask it to do it again, [32:39] it the results get better and more [32:41] information is added. Your answer gets [32:43] better. You ask it to do again, it [32:45] covers more material. And the these sort [32:47] of loops are producing sort of not a [32:51] little bit better answers but vastly [32:53] better answers. [32:54] >> Yeah. A [32:55] >> and that is enormously powerful because [32:58] we don't quite know where it ends, [33:00] >> right? [33:01] >> You keep throwing compute at it. I mean, [33:03] how much better does the answer get? [33:04] >> You know, we we run out of tokens or our [33:06] budget or or or but but holy cow. I [33:10] mean, when does the exponential stop or [33:12] does the answer keep going up and up and [33:14] up to the right? [33:15] >> Yeah. And that's sort of an enormously [33:17] interesting intellectual question right [33:19] now. [33:20] >> Yeah. Like when do we run out of [33:23] problems to solve and [33:25] >> Well, that's right. And and when are are [33:27] the the problems [33:29] no longer sort of intellectual problems [33:32] and they're now people problems? [33:34] >> Yeah. [33:35] >> Right. How to organize people to to get [33:37] done what the AI asked for. [33:39] >> Right. Right. I mean, as you know, in [33:41] running your company, a lot of your [33:43] problems aren't hard intellectual [33:44] problems. They're people working [33:46] together problems. [33:47] >> Yeah. [33:47] >> Right. And you [33:49] >> motivation. [33:49] >> Motivation. You spend a lot of time as a [33:51] leader spraying WD40 on your team. [33:53] >> Right. [33:54] >> Right. It just so so friction is reduced [33:57] and [33:58] >> um how do we learn about those from AI? [34:03] >> How do we get behavioral insight from [34:06] from AI? And and I think that's some of [34:08] the things the world models are going to [34:09] bring us as they begin to watch human [34:11] behavior. [34:13] >> Yeah. We didn't even get to that. This [34:14] is going to be for another interview. [34:16] But um when these things jump off the [34:19] screens, [34:20] >> right, and they're in the real world and [34:23] the recursiveness starts, not trying to [34:25] solve math problems and right, you know, [34:27] humanity's most difficult ones, but hey, [34:30] you know, there's an incredible world [34:32] out here and here's the palace of [34:33] Versailles, right? You're just like now [34:35] we're like make me a new version of [34:38] Salesforce and we're like hey you know [34:39] what I'd like a palace of Versailles [34:41] I've got a hundred acres somewhere out [34:43] Texas or Nevada I'll just send a [34:46] thousand optimistes out there make me [34:48] the Palace of Versailles [34:49] >> right [34:50] >> sounds fantastical but the Palace of [34:52] Versailles would seem fantastical to [34:54] people who lived a thousand years before [34:56] it [34:57] >> and it was fantastical I think to the [34:58] people who built it. [35:00] >> Yeah. [35:00] >> Right. Even to the builders, I I think [35:02] they were awed [35:04] >> at it as they built it. [35:05] >> Yeah. They're compounding. They're [35:07] compounding recursive learning. [35:08] >> That's right. A [35:09] >> and generations. We talked about you had [35:11] a really such a great insight of in [35:13] building this place, you had generations [35:15] of masons. [35:16] >> Yeah. I I think in in in all these large [35:19] projects um often there were families uh [35:23] who were specialists who you and uh you [35:26] you apprenticed under your father, your [35:27] uncle and when you had a project that [35:29] took 50 or 70 or 100 years, you might [35:32] have three or four generations of the [35:34] same family, right? The same stonemason [35:37] family working on the same structure [35:39] >> and passing on passing on the learnings, [35:42] >> new innovations, [35:43] >> right? which is what we've modeled with [35:47] this new [35:48] >> That's right. [35:48] >> models and and what you're building in [35:50] the infrastructure. [35:52] >> It's pretty incredible when you think [35:54] about it. [35:54] >> Especially when we're sitting here and [35:57] the and that's what I mean I I think the [36:00] the problem with human learning is um it [36:05] often moves it at the pace of a [36:07] generation. uh and like uh elephants and [36:12] other large mammals, we don't have [36:13] generations but every 15 or 20 years. [36:17] And if you want to move really quickly [36:18] across generations, [36:20] you want them happening more like [36:22] Drosophili, like fruitfly. You want two [36:24] a day. [36:24] >> Yeah. [36:25] >> Right. Then and you see that in [36:27] genetics, that's why we study them in [36:28] genetics because learning encoded in the [36:31] DNA, you can study over thousands of [36:33] generations. [36:34] >> And I I think that what we're getting is [36:36] that equivalent in AI. We're getting [36:38] sort of learning so quickly over the [36:41] equivalent of thousands of generations. [36:43] >> Yeah. Darwin would be [36:46] >> in awe of this uh pace evolution. [36:50] >> That's exactly right. You think about it [36:52] as there was I remember when I was [36:54] getting my psychology degree and they [36:56] were teaching us about paradigms and I [36:58] was like trying to understand how the [37:00] paradigms shifted and the professor said [37:03] to me uh Jason uh what you have to [37:06] understand is paradigms don't die. [37:09] >> They don't. [37:09] >> People do. [37:10] >> That's right. [37:11] >> And that's how Freud [37:13] >> he and Thomas Freud and Skinner and [37:17] Young like it took them dying. [37:19] >> That's right. It took generation [37:22] question it and that was 20 years [37:24] sometimes 40 years as their students [37:28] maintained positions of leadership until [37:30] someone said maybe we could do it [37:32] differently and I I think what you're [37:34] seeing is this iteration is a shortening [37:37] of the the the intergeneration gap [37:41] and the learning is so fast. [37:44] It's uh always so great to talk to you [37:46] because uh one it's just intellectually [37:50] um uh so your your approach to it is so [37:54] intellectually rigorous but also um with [37:57] so much poom in the world. I I feel so [37:59] good that you're such an optimist about [38:02] this technology and you're building it [38:04] >> with such thoughtfulness and it I think [38:06] for people who are hearing these horror [38:08] stories about AI and job loss and [38:10] everything, they need to understand [38:12] there are people like yourself who are [38:13] building this in an incredibly [38:14] thoughtful way and this is going to be a [38:17] net benefit for humanity that just is [38:20] unimaginable. Yeah, [38:21] >> we have a shot with this technology. So [38:26] not our children nor anyone they know [38:28] dies of cancer. [38:30] >> All right. I mean say it like that. [38:32] There will be some dislocation in the [38:33] economy. Sure. There will be there was [38:35] dislocation when cars came and and and [38:38] it was a bad deal to to to be a guy who [38:40] shed horses, right? Or built carriages. [38:43] >> But you got to also against that, you [38:45] know, make your tea of the cons and the [38:47] pros. [38:48] >> Yeah. [38:48] >> Right. There's a shot that our children, [38:51] none of them nor their people they love [38:53] will die cancer. And that that's one [38:56] >> thing that we can work on wi with this [38:58] technology and we will have great [39:00] purchase on [39:01] >> and and I think you begin listing those [39:03] and then it's a more thoughtful [39:05] discussion. [39:06] >> Yeah. Unlimited energy, unlimited [39:08] calories, unlimited knowledge, unlimited [39:10] education, unlimited housing [39:12] >> and how we do it. We imagine imagine [39:15] sort of we know how to teach children [39:17] and we don't do it right. Aristotle was [39:20] a tutor to Alexander the Great. Socrates [39:22] was his tutor. We know that if you give [39:24] a child a tutor and the tutor modifies [39:27] the teaching for the child, they learn [39:29] better. [39:30] >> That's not how we do teaching classes. [39:32] Factory farming. [39:33] >> That's right. We we teach to some sort [39:34] of mid-level. Imagine if we beat built [39:37] agents that that taught children for [39:40] their way of learning. [39:42] >> Right. [39:42] >> Right. And here's a way we we've been [39:44] doing it the same way for a thousand [39:45] years. And during that entire time, we [39:47] knew how to do it better and we chose [39:48] not to. [39:49] >> And and here's a way we can do it. Put [39:51] that on the pro side. And so as long as [39:53] we're sort of thoughtfully and fairly [39:55] writing the good and the bad, I I think [39:57] it'll come out. [39:58] >> You got to get out there, Andrew, and [39:59] keep communicating your version of the [40:01] world because some people see around the [40:04] corner and they get a little nervous and [40:06] >> Okay, fair enough. But I think the [40:09] ledger, as you describe it, [40:11] >> is heavily weighted towards abundance. [40:15] >> I think it'll create abundance for sure. [40:17] >> Massive abundance. [40:18] >> Andrew, pleasure. Always a pleasure to [40:19] see you. I'll see you in six months for [40:21] our checkup. That'll be great. [40:24] >> Industries, capital, and intelligence [40:26] are converging into a single [40:28] interconnected system, and the [40:29] infrastructure behind it needs to evolve [40:31] just as quickly. NASDAQ was built for [40:33] this moment, powering more than 135 [40:36] marketplaces and regulators globally and [40:38] connecting capital to companies shaping [40:40] the future. As the innovation economy [40:41] accelerates, connectivity becomes the [40:43] critical asset. NASDAQ is the leading [40:45] technology platform that makes it [40:47] possible and scalable. Learn more at [40:49] nasdaq.com. [40:52] Robin Rombach is uh the co-founder and [40:56] CEO of Black Forest Labs. You are uh [40:59] based in Germany in Black Forest which [41:01] is a city [41:02] >> in Germany. [41:03] >> A mountain range actually. [41:04] >> A mountain range. [41:05] >> Yes. [41:06] >> Where you grew up. [41:07] >> Where I grew up. Yes. [41:08] >> And uh you are working on open- source [41:11] image and video models. You worked at [41:14] stable diffusion for a little bit. [41:15] >> That's correct. [41:16] >> Cut your teeth on that. and uh you're [41:18] known for the open source model flux and [41:22] maybe also for some closed source [41:24] models. Tell us about the business of [41:27] Black Forest Labs. What is the business [41:29] and what is the goal? [41:30] >> 100%. Uh one uh quick addition, we are [41:34] based in uh the Black Forest. It's a [41:36] town called Frybook and in San [41:37] Francisco. We have like [41:38] >> in San Francisco of course. Yeah. [41:40] >> Um we [41:40] >> you're splitting your time or [41:42] >> uh I'm splitting my time to a certain [41:45] degree. Um we um yeah started a company [41:49] 2 years ago. Um me and my co-owners as [41:52] you said like we've worked on stable [41:54] diffusion in the past. Before that we [41:56] invented like an algorithm called latent [41:58] diffusion which is basically like the [42:00] fundamental algorithm behind all of like [42:02] generative models that are being [42:03] deployed for image generation, video [42:06] generation, even like physical AI now. [42:08] >> Yeah. Um it basically makes use of this [42:10] principle that uh you can compress [42:12] natural data such as images, such as [42:15] video, such as audio into a much more [42:18] like efficient representation and then [42:20] train a transformer model on that. And [42:23] um I mean this is the stuff why like you [42:26] know like JPEG uh MP3 and all of that [42:30] works. And we basically translated that [42:31] into like a neural uh algorithm uh a few [42:35] years ago when we were still like PhD [42:36] students in uh in Munich actually. And [42:40] uh then build on like on top of that we [42:42] build stable diffusion and then um on [42:45] top of that uh yeah uh [42:49] the generative models that we are [42:50] developing today and of course like the [42:52] technology has advanced um but we are [42:55] now tackling I would say models that are [42:59] really made for understanding like the [43:01] whole world around us multimodal visual [43:04] models uh pre-trained on images audio [43:08] data at the same time and we are now [43:10] like entering a new paradigm which is [43:13] combining that with uh something that's [43:15] called action prediction such that you [43:17] can actually use the same model to make [43:19] images to make videos to make audio and [43:22] to predict actions which means you can [43:24] ultimately deploy it on a robot in the [43:26] real world. [43:27] >> Wow. So from the image to the video, the [43:32] audio and then eventually the real world [43:34] with robotics and a real world model [43:38] because if you can make the image you [43:40] and you can train the model that means [43:43] by default you understand the world. In [43:45] order to make a video of the world you [43:48] have to understand the world. Yeah. [43:50] >> And the objects in [43:51] >> I I think that's Yeah. I think that's [43:53] like a a really good like way to think [43:55] about it. Yeah, it's like it's like an [43:56] intuitive way uh to [43:59] interact with the world, right? Like I I [44:02] would say there's like these like [44:03] complimentary forms of intelligence [44:05] ultimately. There's like intuitive [44:06] intelligence and then there's like a [44:08] deep reasoning layer. Now ultimately you [44:10] need for like a kind of like complete [44:12] form you need both um and you need them [44:15] to interact and I think like we've been [44:17] approaching it more from like the [44:18] intuitive side um images is like a very [44:21] natural way to approach this whole field [44:23] because it's not as computationally [44:25] intensive as let's say video right but [44:28] now yeah I think like we're combining it [44:30] it's converging into like a a multimodal [44:33] model and yeah we see like exactly like [44:36] pre-training on videos gives like [44:37] implicit understanding of the physics of [44:40] interactions with the real world and [44:42] then you can get stuff like action [44:43] prediction like robotics out of the same [44:45] model. And with these models and the [44:49] training [44:51] there kind of um been a limitation in [44:55] creating videos and creating images [44:57] where the criticism of generative AI is [45:01] it's a bit of a slot machine. I give a [45:03] prompt, it gives me something back. But [45:06] how did it come up with that the [45:08] training data, but you know, maybe I [45:12] want a different style? Maybe I want u a [45:15] different color. Maybe I want a [45:17] different uh you know, aesthetic. [45:20] >> Yep. [45:21] >> Has that how does that problem get [45:24] solved? And do you actually understand [45:27] what's happening [45:29] when the image is being made under the [45:31] hood? Yeah. [45:33] Yeah, I I think [45:35] like ultimately it's about like exposing [45:38] as many like manipulation layers as [45:40] possible to like I don't know like a [45:43] user or developer that builds on top of [45:45] this model, right? And um I think like [45:48] we've seen that in the past with like in [45:50] the past uh image models they basically [45:53] started from simple text to image [45:54] systems, right? Then they've expanded [45:56] into a text plus image to image systems [45:59] which means you could suddenly take an [46:01] image like a real image or generated [46:04] image um and iterate on that based on a [46:06] text font like edit it um modify it [46:09] right and then this expanded into taking [46:12] multiple images um and a text font and [46:15] combining them in a in a semantic way [46:17] and producing new content. And the same [46:20] principle now applies to video. And I [46:22] think now it becomes actually even more [46:23] interesting when like all of these like [46:25] modalities are actually combined inputs [46:27] and outputs of the same model. [46:29] >> So let's talk about video. [46:32] There's an announcement that you're [46:33] working with the greatest director of [46:35] all time or living director Martin [46:36] Scorsesei. We'll talk about that in a [46:39] second. Yeah. [46:39] >> Fantastic. Uh but in a movie uh this [46:43] promise of being able to make a movie in [46:47] which the camera angle [46:50] uh the sound could be [46:55] something that a Martin Scorsesei would [46:57] be proud to release to his fans. How [47:01] close are we? And maybe tell us a little [47:03] bit about this partnership, the [47:05] technology being able to make an actual [47:09] movie like Good Fellas or a scene from [47:11] Good Fellas, uh, versus where it is [47:14] today where you can make interesting [47:17] five or 10 second clips and then maybe [47:20] people struggle making 10 of them and [47:22] then they use some post- editing [47:24] software to put them together, but you [47:26] immediately understand this is not that. [47:30] It's not a movie. It's AI slop. It's [47:35] cluji. It's doesn't pass the uncanny [47:37] valley. Well, I think it's important and [47:41] that's at least like the view that we [47:43] have is that um these AI models, they [47:45] are a medium, right? They we don't want [47:48] to set like any way of how they are [47:49] supposed to be used. We don't want to [47:51] tell anyone especially not someone like [47:53] Martin Kazi how how is he supposed to [47:54] use these models? Like he is one of the [47:56] like greatest filmmakers ever. Uh it was [47:59] insane sitting in the same room with him [48:01] multiple times and actually him seeing [48:03] like exploring our models like as like [48:05] one of the like core researchers behind [48:07] it was like just an insane feeling. [48:09] Right. And at the same time I'm also [48:11] like a big fan. Um [48:12] >> so you sat in a room with Marty Scarces [48:15] and showed him your tools. [48:16] >> Exactly. Yeah. [48:18] >> And what was his reaction? What what did [48:20] he key off of? What was the thing that [48:23] he found most inspiring or interesting? [48:26] Okay. Um I think it was really this idea [48:28] of like he has clearly um a a vision in [48:33] his head of like a scene or a scenery [48:37] where like maybe a new movie um will be [48:40] shot and he's trying to explore that and [48:43] kind of like um we we we basically [48:46] looked at the scenery of like a village [48:47] in Eastern Europe somewhere and he was [48:49] describing it. we saw some outputs, we [48:51] iterated on the outputs. Um, and [48:54] ultimately I think and that's what he [48:55] said in the end is like the like getting [48:58] like the mental picture of something out [49:00] of your head and communicating it in a [49:02] visual way by making like these images [49:06] um or the series of images um is [49:09] something yeah that just makes it like [49:11] easier to communicate and con convey [49:13] like an idea of like what is actually in [49:14] your head and I think that's like one of [49:16] the like very interesting and powerful [49:18] ways to use this technology and I think [49:20] ultimately [49:21] >> is to get the inspiration [49:23] to get the vision out of his head onto [49:26] an image. [49:27] >> Yeah. I mean like language ultimately is [49:29] like a little bit of like a lossy um [49:32] communication medium, right? [49:34] >> Yeah. [49:34] >> Um it's also interpreted in different [49:36] ways, but then visual information is so [49:38] rich. So rich uh like an image or video, [49:41] there's so much signal in it and it's [49:43] just like another way of communicating. [49:45] And I think that's like one of the [49:46] beautiful things that this technology [49:48] ultimately enables. And I think like to [49:50] your question of making like full movies [49:53] with [49:55] I don't know like a video uh generation [49:57] model for example, I'm not sure if that [50:00] is like the ultimate goal. Maybe it's [50:02] like interesting to plug this into like [50:04] some kind of aic workflow and make like [50:06] a very long video. And I think that's [50:07] really cool to explore, but I think [50:09] ultimately like the real interesting use [50:12] cases they come when you have like a [50:13] human in the loop who iterates and uses [50:15] it as a medium. And I think this is this [50:17] is at least like a perspective that that [50:19] I take um that makes it interesting and [50:22] that this is most often when the most [50:24] interesting outputs um arrive or are [50:26] actually being made. [50:27] >> The brainstorming production level is so [50:30] obviously a huge win for [50:32] >> you can paralyze your brainstorming [50:34] basically. [50:35] >> Yeah. And yeah, I like that. Paralyze [50:37] your brainstorming and and they have an [50:39] analogy for this. They do storyboards [50:41] and some of the great directors Ridley [50:43] Scott of Aliens and Gladiator was known [50:46] for making his own. I also believe [50:48] Spielberg was also like to sketch [50:50] Raiders of the Lost Arc and some of [50:52] these. George Lucas was known for [50:54] collaborating with many amazing artists. [50:57] Um even making miniatures and making [51:00] storyboards for the Star Wars franchise. [51:02] He had those people on full-time helping [51:04] him with that. So that's the obvious [51:07] place to start. But if we look at [51:09] startups, startups uh always want to try [51:12] to figure out how to do something [51:14] cheaply. And people used to make a [51:16] launch video for their startup for, you [51:19] know, $100,000, $250,000. So they take [51:22] their $10 million venture raise and [51:24] spend $250,000 on a launch video. I've [51:27] seen with a lot of the startups I'm [51:28] investing in now, they'll just spend a [51:30] week or two working with um you know, a [51:34] director to make a launch video. You've [51:37] probably seen this trend. Yeah. And I'm [51:39] sure people use Flux and some of your [51:40] models for this. Uh have you seen this? [51:43] >> Yeah, of course. Yeah. [51:44] >> Yeah. What's your take on that? Because [51:46] that feels like the early stage of [51:49] storytelling. You're trying to [51:51] communicate a product or service in a [51:53] fun, engaging, punchy 30-second, [51:57] 90second way. Yeah. [51:58] >> I mean, like again, like I think we [52:01] support these like exploration based on [52:04] these tools, right? And I think [52:05] ultimately it's great to see like all [52:07] different kind of like [52:08] >> I don't know like launch videos products [52:11] being built on top of like the same kind [52:13] of like base model or the same [52:15] technology. Um and I think that's what's [52:18] making it so interesting and also so [52:20] powerful. [52:21] >> Yeah. [52:22] And what else are people using the [52:24] technology for? I understand there's a [52:26] Bitcoin movie coming out. instead of [52:28] using a green screen in this Bitcoin [52:29] movie. Um, I was talking to Galado, you [52:33] know, the woman who played the actress [52:34] who played Wonder Woman. [52:36] >> Oh, Galado, of course. She I was talking [52:38] to her at an event and she was telling [52:40] me um it was the Breakthrough Prize uh [52:44] Yuri Milner's event and she was telling [52:46] me she just did a Bitcoin movie and they [52:48] did it on a sound stage without green [52:50] screens [52:51] >> but all the actors just worked in like a [52:54] sound stage [52:55] >> and then all of the scenery behind them [52:57] was being done by Generative AI. That's [53:00] a real movie. That's a $30 million [53:02] budget movie. She said it would have [53:03] cost $150 million if they had to build [53:06] sets and the film would have never been [53:08] green lit. Are you starting to see [53:10] people use that in production? Not just [53:12] in the back end and the ideation phase, [53:16] but actually in production yet with your [53:18] tools. [53:18] >> Yeah. [53:19] >> Um, yeah, we see some use cases like [53:21] that in production. I think like [53:23] high-end film production is kind of like [53:24] the one of the like most demanding use [53:26] cases. And I I think I'm glad that it's [53:28] being explored, but I also [53:30] >> really want to um [53:33] like it's I think it's important to see [53:34] that this technology is like on a [53:36] trajectory and it's improving. It's [53:37] improving rapidly. I don't know if I [53:39] look back at like where we started like [53:41] a few years ago when I was doing my PhD [53:43] in this field like the only thing that [53:45] you could do was like images of 64x 64 [53:48] pixels. Now you can do like multi-inut [53:52] videos, right, at like a high [53:53] resolution, but it's like it's not going [53:55] to stop there. It's going to it's [53:56] continue to improve and I think like [53:58] then it's going to unlock like even more [54:00] of these like high-end use cases. But I [54:01] think the main thing [54:03] >> before we get to that, yeah, [54:04] >> hard to predict. I think hard to predict [54:06] and I think ultimately [54:07] >> couple of years [54:09] >> ultimately I think you still want to [54:11] have like the tool that enables like [54:13] this human in the loop kind of [54:14] >> of course. Yeah. [54:15] >> Um [54:17] production workflow, right? But I think [54:20] when I look at [54:23] multimodal generative models um as a [54:25] whole, I think what really excites me is [54:28] you can use the same kind of um AI model [54:33] to make a movie and deploy that as a [54:35] brain on a robot. Ah, and I think this [54:38] is like this is so interesting. Um and I [54:42] don't know like there's like some [54:43] thoughts around trying that [54:46] in the digital world, right? which would [54:48] be for example computer use remains to [54:52] be seen if that is actually something [54:53] that works or not but I think like the [54:54] technology is so powerful and so [54:56] versatile and it's just just moving into [55:00] that and all the all the talk around [55:02] like world models world action models [55:04] all of that it's basically all the same [55:06] and I think that's what's making it so [55:08] interesting and what I find like most [55:09] exciting [55:10] >> so do you believe that the technology [55:13] will be used to analyze or primarily to [55:16] analyze is real world like here's a [55:19] video of somebody you know making a [55:22] sandwich. Now we have the robot study it [55:25] and make the sandwich or do you think [55:27] there'll be a lot of synthetic data made [55:30] that then the robots will just study the [55:32] synthetic or they're going to just in [55:33] some way innately know based on all this [55:36] massive amounts of training data. Um I [55:40] think it's a combination of prediction [55:42] right and p prediction and is a way of [55:47] you can think about it as simulation as [55:49] generation it's um predicting actions [55:52] which is you have to understand the [55:54] input the visual inputs in order to [55:56] actually predict a reasonable next [55:58] action um and it's about perception it's [56:01] like you can only do that if you [56:02] understand if you perceive the content [56:05] you then you can only I don't know like [56:07] transform it into a new piece of content [56:09] or predict an action or describe what [56:11] you actually see in that um scene. And [56:13] the and the combination of all of that [56:15] is I think um [56:18] yeah is I think what's what's driving [56:20] it. There's not a single one of them. [56:21] It's a combination of these thoughts. [56:24] >> And what's the best way to get that [56:25] training data? Do you need to have [56:28] people put on glasses, get a firsterson [56:30] perspective, have them put on gloves so [56:33] you have that, [56:34] >> you know, fidelity of understanding, [56:37] hey, this glass is moving. I'm pouring [56:39] this glass. I'm putting ice into it, you [56:41] know, and and here's how that's works [56:43] and the splashing and the condensation [56:45] water so I can pick it up and not drop [56:48] it because it's wet on the outside. Or [56:50] is it going to be just, hey, take the [56:52] corpus of YouTube videos and the robots [56:55] know exactly what to do because they'll [56:57] find a thousand videos of people pouring [56:59] drinks. [57:01] >> I mean, ultimately, I think you would [57:03] want to go to a place where you could [57:04] like prompt a robot in context, right? [57:07] as you can do with like a language model [57:08] basically just tell it hey go and I [57:11] don't know pick up this glass with the I [57:13] don't know orange juice or whatever it [57:14] is yeah exactly [57:17] >> we're not there yet um but I think this [57:20] this is like one of the goals and I [57:21] think like how these models are deployed [57:23] currently is there's like a lot of like [57:25] different hardware different robots that [57:28] are running in factories um that all [57:30] have like some different kind of action [57:32] representation that you need to kind of [57:35] tune the models towards right So in [57:37] practice um what you do is you have like [57:40] all this like visual understanding in [57:42] the models um and then you need only a [57:45] very little bit of like a few hours of [57:47] um fine-tuning data to adjust the model [57:51] on that specific task and I think the [57:53] goal would be to kind of move away from [57:55] that uh towards like as much in context [57:58] as possible but it is a little bit of a [58:00] research problem. Uh, I I think that [58:02] >> open source is kind of having a moment [58:06] right now. We've been discussing it on [58:07] the podcast a whole bunch recently and [58:11] people are also talking about [58:12] sovereignty. You have companies that own [58:14] incredible IP libraries. I mentioned [58:16] Star Wars before. Disney owns an [58:19] incredible library. What should your [58:21] advice, what would your advice be to a [58:23] company like Disney? Should they take [58:25] your open source software, train their [58:27] own models or work with you to train [58:29] their own models to control it and then [58:32] hey this is our IP? They've already made [58:35] a point of working with chat GBT and [58:37] saying hey you you can and cannot use [58:39] certain characters. In fact, OpenAI had [58:41] a relationship with them that's for Sora [58:43] that's no longer happening but they [58:46] officially licensed on the output some [58:48] characters. So how do you think about [58:49] those major IP holders? What's your [58:51] advice to them? Are you in discussions [58:53] with them? We know about the Martin [58:54] Scorsesei [58:56] or tour deal, but how do you think about [58:58] content libraries? [58:59] >> I think it is um look, I think like the [59:02] most interesting use cases of this like [59:04] if you think about like content creation [59:06] um is in generating something making [59:10] something that hasn't been there before, [59:12] right? Like that that's a fundamental [59:13] like interesting aspect of this [59:15] technology. And then I think like yeah [59:17] when it comes to IP what we implement [59:19] for example on like our public facing [59:21] tools is you cannot generate certain IP [59:24] with these models right and I think [59:25] that's something that is a a sensible [59:27] approach and then yes we do work with [59:30] certain IP uh holders to develop models [59:35] together with them um some of them based [59:37] on our open source models some of them [59:39] based on like our more powerful [59:40] proprietary models but I think that is [59:42] like a very like attractive [59:44] >> value there what do you think that will [59:47] look like for consumers in another [59:48] couple of years? What would potentially [59:51] happen when you open up Disney Plus? [59:53] >> I mean, that's a good question. I'm not [59:55] I'm I'm not in Disney, right? So, it's [59:56] up up to them to decide that. But, I [59:58] think we want to enable them to build [59:59] all kinds of stuff that they that they [60:01] that they envision. And I think we can [60:04] support them. We can support like other [60:05] companies in that space to [60:07] >> I don't know integrate the technology in [60:08] the best possible way. I think like one [60:10] of the very interesting angles of it is [60:11] that it is like it's becoming much [60:13] faster. It's becoming more interactive. [60:15] I can imagine like a whole bunch of like [60:17] very interesting interactive content [60:19] creation tools that you could host on [60:21] Disney Plus Plus or elsewhere. [60:23] >> I think the most interesting thing I've [60:25] seen in this regard is uh fan films, [60:28] >> right? So there's a category before [60:31] generative AI fanfiction. People would [60:33] write their own Star Wars story. [60:36] Then there came fan films where people [60:40] would dress up as Jedi Knights and [60:41] record their own films. And George Lucas [60:44] said, "As long as you're not doing it [60:46] commercially, you're not selling it, I [60:47] give you permission to go make Jedi [60:50] movies." And they even released how to, [60:53] you know, how-tos on how to make a [60:54] lightsaber or, you know, sound files of [60:57] like how to make the lightsaber sound. [61:00] Now, people are taking the stories that [61:04] haven't been told from the Star Wars [61:06] universe [61:07] and they're recreating them using AI. [61:11] And for the fans, they're becoming quite [61:13] popular on YouTube. Uh Star Wars Stories [61:16] Untold is, I think, the biggest one. [61:18] It's getting millions of views per video [61:20] already. And I think that's really the [61:23] future is letting the customer base pay [61:26] a licensing fee or pay a fee, uh maybe [61:29] rent software or maybe based on the [61:31] output and let them be creative with the [61:33] characters, let them make their own [61:35] stories, and you could be in a unique [61:37] position to empower that. [61:38] >> No, 100%. I think like if you find like [61:40] a model that works for like the um IP [61:42] owners uh but then also can enable like [61:45] these super like creative customization [61:47] use cases I think that's great. Yeah. I [61:48] mean like like I mean like for myself [61:51] like I when I read a book or whatever [61:53] like watched the movie I had like so [61:54] many like ideas how it could be done [61:56] differently or this could have happened [61:58] right this is like so nice that you can [62:00] actually enable people to visualize [62:01] these ideas. [62:02] >> Yeah it's uh going to be incredible [62:05] continued success with it. You have an [62:07] office in San Francisco. You're hiring [62:09] people? Yeah, [62:10] >> we do. Yeah, we just [62:10] >> we raised a bunch of money. [62:12] >> We raised a bunch of money. Uh we just [62:14] crossed 100 people. We're hiring in [62:16] Germany and in San Francisco. [62:17] >> Fantastic. Who are you looking for? [62:19] What's the right type of person, the [62:20] right type of scale? [62:21] >> Yeah. Um on the one hand, we're always [62:24] looking for [62:26] researchers who have experience in large [62:28] scale model training. Um experience in [62:31] diffusion model training, flow matching [62:32] training. We're looking for engineers [62:35] who want to be working with the [62:37] customers to you know develop these like [62:39] customized um physical AI solutions or [62:43] for example with like a IP owner like [62:46] develop these models jointly with them. [62:48] Um we are looking for engineers who have [62:50] experience in just like large scale [62:53] compute infra managing that um and [62:57] making sure that the training runs runs [62:58] smoothly that we maximize our MFU and [63:02] all that um and we're looking for people [63:04] who have interest in [63:08] you know like uh getting the technology [63:10] out there [63:11] >> in the hands of people [63:12] >> the the the for deployment of this [63:14] there's just so many great ideas and so [63:16] many great partners for you I think [63:18] you're going to with the open source [63:19] specifically. You know, it seems like [63:22] the corporates really want to have some [63:24] additional level of control, but they [63:26] also need the frontier models or your [63:27] proprietary ones for some of those [63:29] refined features. So, I think you have a [63:31] very bright [63:31] >> future. [63:33] All right. Continued success. Thank you [63:34] so much for doing the show. Pleasure. [63:36] I'm going all in. [63:52] I'm going all in.