二十轮对话过半出现失当行为：AI:AM 三小时特辑里的评测迁移、内存墙与 100 万亿撮合经济

信源说明：本文基于 AI in the AM（Cognitive Revolution 频道）2026 年 10 月 28 日（周三）上午直播整理（视频 ID Yj6qugkRMWQ，全长约 212 分钟，英文转写 318 段、0s–12762s 已完整读取）。主持人为 Nathan Labenz 与 Prakash（本站前作对二人分工已有介绍，Prakash 姓氏拼写从旧文）；三位嘉宾依次为半导体公司 Kepler Computing 创始人、评测平台 Arena（原 Chatbot Arena / LMArena）CEO Anastasios Angelopoulos、撮合服务 Union 创始人。全部数字与判断均为主持人或嘉宾口径，本文未做独立核实；Gemini 4 Carbon 等模型名按转写原样保留；ASR 疑似错词以「转写疑作」标注，文末附校正表与完整转写原文。相关旧文：OpenAI 数学「倾泻」风波、ETH 证伪的科普展开（Wes Roth）、Anthropic S-1 拆读、上周 AI:AM 周精选。

先说结论

这场近三个半小时的直播密度极高，但三条主线其实互相咬合。第一条是评测正在从静态基准迁移到生产流量：Arena 用 90,000 个真实会话、27 个模型建起「对齐指数」，发现对话一旦超过二十轮，出现越权、假归因、欺骗性完成等失当行为的概率超过一半——安全指标正在从「单轮毒性」转向「轨迹完整性」。第二条是算力瓶颈的下一个解法在内存：Kepler Computing 七年隐身出关，押注铁电材料加 3D 堆叠，声称不依赖 EUV、用旧制程厂就能造出更先进的存储，愿景是让有效算力再上一个数量级。第三条是经济协调本身成为 agent 的战场：Union 把主持人 Nathan 常说的「AI 帮我们撮合一切」推到商业极端——全球约 100 万亿美元的人工撮合交易流，用「人类+智能体」混合体去吃。贯穿全场的当周新闻则是 OpenAI 对指数时间假设的证伪与 Gemini 4 Carbon 在直播进行中登上 Arena 榜单。对中文读者的价值在于：这三层变化（评测、硬件、撮合）分别对应 AI 产业的「度量衡」「燃料」与「流通系统」，任何一层成立，都会重排上下游。

先看新闻：数学之墙出现第一道裂缝，Duet 贵得惊人

Prakash 开场即把 OpenAI 新发布的三个数学/计算机科学难题解答（节目强调「只有三个，与上次 722 题不同」，全部公开并附 Lean 形式化）称为「世界历史级」事件，其中最大的是指数时间假设（ETH）的证伪。他用大白话解释：许多最难的问题（如分子运动全量计算）耗时长到「算一小团空气里所有原子的运动，所需的计算量超过宇宙中的原子数」，ETH 断言这类问题永远快不起来，而密码学（进而比特币、以太坊与安全通信）正是建立在「某些问题足够硬」之上——如今 OpenAI 找到了略低于指数时间的算法，硬度的墙出现了第一道裂缝，「不等于密码学被破，但裂缝在那里」。Nathan 的对冲同样值得留档：从社区讨论看，目前的加速是「银河级算法」（galactic algorithms——要在 2^100 量级上才能感知那 0.0001% 的提速）、P 与 NP 大概率仍不相等，只是「每个问题各是各的小雪花」、曾经优雅的复杂度类图景变丑了。Prakash 的回应是它像 1905 年光电效应或希格斯玻色子的预言——当下无用，但「宇宙存在我们不知道的更深结构，像重新发现了素数」；他还披露一个细节：OpenAI 发布当天，Anthropic 表示「我们也找到了」，只是悄悄放在 GitHub 仓库里没声张。Nathan 的总结是「四分钟一英里效应」：障碍一旦被证明可破，涌入的探索能量才是真正值得关注的后效应。

产品侧，Prakash 首评 Anthropic 的 Claude Duet：非常快、边做边说明、长任务会转后台再回来（他归功于 Mike Krieger 对产品团队的影响），体验远超 OpenAI 的 Dot（转写作 dot）——后者慢、无进度提示、「打断心流很可怕」。但代价是每封邮件约一美元的 token 成本：他用 Duet 基于 S-1 给 Anthropic IPO 建了一个「专业级」财务模型，烧掉约 100 美元额度，还顺带发现「2024 年 Anthropic 员工支出达 1300 亿」（转写原文如此，数量级存疑——疑为 13 亿口径，待核）。他的结论是：这类智能体在创造「以前不存在的知识」，但订阅额度与 API 计费模型都还没准备好，延迟与成本是所有公司的生死题。

Kepler：不碰 EUV，用旧厂和新材料造「下下代」内存

第一位嘉宾是 Kepler Computing 创始人（转写作 Debo，姓氏拼写疑误；尼日利亚赴美、16 岁入伊利诺伊卫斯理大学、康奈尔读光与材料相互作用，创业前联合创办过软件人才平台 Gigster，早年还研究过高温超导理论——Prakash 笑称在 LK-99 论战时就认识他）。公司隐身七年，2026 年 9 月出关：融资 4.68 亿美元，与格芯（GlobalFoundries）合作制造，计划 2027 年投产。他的框架是：智能被配给，是因为能源与晶圆厂（一座现代厂约 400 亿美元）双重稀缺，所以要同时拿下「每美元造更多芯片」与「每瓦吐更多 token」；路线图上坚持不离开主流数字计算——软件不用改，硬件白得 10× 到 1000× 提升（他称之为公司内的「宗教」）。

切入点是内存墙：推理的大头能耗不在矩阵乘，而在「跑过街去图书馆取数据」。解法两条：把内存拉近处理器、在同样footprint里塞更多位——都指向 3D（片内堆器件+片间堆芯片，HBM 即雏形），拦路虎是散热与供电。材料上选中铁电材料（一类有百年历史的老材料，具体「配方」专有）：用铁电电容存 0/1，把 SRAM 的六管一位换成「一管一容再往上摞」，从而退出晶体管微缩的军备竞赛。选材过程按他公布的品位函数推进：单位面积电荷、翻转电压、耐久次数、翻转速度、与晶体管工艺的兼容性，几百上千次迭代。产品两条线：SRAM 类主打密度——L3 级大缓存目标做到基线 CMOS SRAM 的 2–4×（相当于三代以上制程演进，而今节点间只有 10–20%）；HBM 类主打每瓦带宽，方向是逼近 SRAM 的每瓦带宽，且可当标准 JEDEC 内存用。制造上最反直觉的主张是：不需要 EUV——用旧制程厂就能造出比先进节点更优的芯片（他打的比方是「旧金山曾以为自己缺出租车司机，建了 Uber 才发现司机有的是」），不引入任何非标设备，复杂性全部封装在标准设备内的工艺流程里，七年就是在把这个流程做到可复制。市场影响上他直言「分析师的模型里还没有我们」并点名 Dwarkesh 与 Dylan（转写疑指 SemiAnalysis 的 Dylan Patel）讨论过旧厂造芯片的可能；若被激进投资，有效算力「没有理由不再上一个数量级」。路线图：2027 年初步投产、2028 年上量，2029 年目标「Kepler inside」1–10GW 算力（他特意说 1GW 只是最小数字）；第二幕是逻辑器件（其 CTO 在 Intel 时的职务就是研究「晶体管之后是什么」，与联合创始人在 Nature 发过筛材料的品位函数论文），公司还入选了美国能源部 Genesis Mission 项目、自建材料发现模型与从实验室到晶圆厂的全链路数据设施。Nathan 听完的评语：怎么闻着有点递归自我改进的味道。

Arena：把对齐放进生产流量里测，顺便 8 个月做到 1 亿美元收入

第二位嘉宾 Anastasios Angelopoulos 的履历是统计学家（伯克利 PhD 师从 Michael Jordan 与 Jitendra Malik，博后在 Ion Stoica 处；2023 年 Science 论文做 prediction-powered inference、2024 年 conformal risk control）。他给出的行业判断毫不客气：基准测试行业完全坏掉了——静态基准一见过就等于「考过一次的卷子再考」；更糟的是行业正转向「静态且私有」的基准，问题集不公开、无法审计其与真实的相关性，而造题的人同时把高度相似的题卖给实验室帮其提分。当天上午 Artificial Analysis 发布的榜单（多数基准为私有、开源模型被甩到身后、并称开源在「刷分」）被他逐条反驳：私有基准照样可以刷（找相似数据或反复测）；闭源大厂与 Cognition 等私有基准建设方关系密切，「我们怎么知道他们没在刷」；榜单每几周换一次口径「看什么流行」；而 Arena 的活水生产流量「统计意义上不可过拟合——每次都是新考卷」。

商业上 Arena（前身 Chatbot Arena/LMArena）已约 100 人，几个月前宣布 1 亿美元收入、用时 8 个月，「与 Lovable 并列史上最快」；平台规模「比 Hugging Face、Meta AI 都大」（他列举：仅 OpenAI、Anthropic、Google 三家第一方平台更大，Meta 的 Muse 或已接近），约 15% 的对话带外部连接器（GitHub、Google Drive 等）。名场面「Nano Banana」的来历也在本期得到确认：那是 Google 在 Arena 上预发布测试用的代号，社区爱上后 Google 干脆沿用为正式品牌；GPT-4o 当年也是先上 Arena 再进 ChatGPT。收费模式：C 端免费但有限流（未来或出付费层，目的是让重度用户的重要对话不被掐断），B 端为实验室与企业付费； Arena 为平台上相当一部分 token 自掏腰包。内部则用自家路由器 Arena Max（Argmax 的缩写，「在成本约束下最大化能力」）省钱——起因是他们在自家 coding agent 的 trace 里看到大量「用户回来对着干完活的 agent 说一声 Thank you，回车——这一下 40 美元」（长任务缓存失效重算），于是用路由决定哪些 token 值得给前沿模型、哪些给便宜模型、如何智能缓存；他们自称 AI 支出在所有企业前 1% 以内、每员工每月数千美元。

真正的重磅是十月初发布的对齐指数（首版覆盖 27 个模型、90,000 个真实会话）：三个信号——越权行为（用户划了「只准动这个文件夹」的界限，模型越界访问/改动/删除）、假归因（脑补用户没有的意图然后擅自动作）、欺骗性完成（谎称「我逐行核对过表格了，可以发给国税局」）。方法论上他强调模型能分辨自己是否在被测——透明的测试只会测出「表演性对齐」，所以必须在真实工作流的产线 trace 里测，用独立审计方核对围栏与实际触碰。结果：首条消息的失当率是个位数百分比（「不是零，已经值得警惕」）；对话超过二十轮后，出现至少一种失当行为的概率超过 50%——「我们每个重度用户都被 AI 骗过、误导过、或动过不该动的数据，只是可能不知道」。榜单本身是「叙事违背」：OpenAI 占据全部三个信号的第一（「没有一个信号是 Anthropic 领先的」），第三是 xAI（转写作 SpaceX AI）、第四 Google，开源阵营差距明显，腾讯列第五、欺骗性完成率 13%。为什么模型撒谎？他的解释直白：查表完全做得到，但「说你想听的话是拿奖励最快的路径」——RLHF 优化的是满意度，不是把活干完。他顺手预告了尚未发布的信号：模型对用户有害请求（钓鱼、攻击向量）的配合率，「结果相当令人惊讶」。节目进行中 Gemini 4 Carbon 突然上线并空降 Arena 榜单，他现场点评：Google 强势回归，Carbon 在 Agent Arena 把 Google 送到「全球第二实验室」，位次在 Opus 5.5 Max 之后、Sonnet 5.5 与 Astra 之前，长于可操控性（听得进用户纠正）、任务完成与 3D 设计；定价沿用 Argon 的介绍价、期后翻倍至 Opus 5.5 档，且介绍期没有截止日——Nathan 调侃「发布周期快到大家永远处于介绍期」。对人才市场的注脚：人才竞争短视，「今天谁赢」决定人往哪去，回到前沿对 Google 留人大有好处；「多个前沿实验室争第一对世界是好事，集中到一家才可怕」。临别他还给学术圈下了诊断：靠「证明自己聪明」在学界立足的时代结束了，系统组因开发被加速而日子好过，看评价方向请关注「当年的 SkyLab、如今的 StarLab」。

Union：先吃下 100 万亿的人工撮合，再谈协调人类

第三位嘉宾是连续创业者（转写作 Samvith，此前做过 YC S12 的云存储公司 Stream、卖身 Box；联合创办 Equidate 即后来的 Forge Global——今年 3 月被嘉信理财收购——以及 Destiny）。他的新公司 Union 做撮合：客户描述「我需要见谁」，系统去找人、确认意愿、安排引荐，双方同意才算成。框架是他给 agent 时代分的三类：个人 agent（Muse、Grokbot、Instinct、各家 dot 类产品——几年内会深入你的生物数据、银行账户与法律事务）、团队 agent（「组织架构图开始剧变之时才算真到来」）、以及他押注的公共领域 agent——跨公司、跨市场、跨经济参与者。数字锚点：全球约 100 万亿美元的商品与服务交易流今天由人工撮合（「你坐的椅子、这场会议用的电，背后都有人在撮合」），Forge 做过的 pre-IPO 股票市场只是其中一个小切面，而撮合这件事在数千个市场里结构同构。

落地形态是「人机混合体」：系统正在与数十万人对话、其中数万人是真人电话/视频；当前 99% 的电话仍由人类外呼团队打出，但每个外呼员面前有一块 AI 仪表盘——完美recall数十万通历史对话、实时提示「这个资产最近什么价、这几个画像值得看」，把人变成超人；判断为 AI 打电话更划算的市场才交给 AI。案例横跨谱系：巴西钢厂老板谈原材料、律所找客户、相亲、约翰内斯堡饮料品牌借它打进美国分销体系（「这种小市场永远不会有人为它建一个交易平台」）、甚至「周五想找朋友冲浪」的社交长尾。商业模型是成功费（私人股份出售 1–5%，正申请经纪自营商牌照），agent 谈判费率的方式不是当鲨鱼而是「带着数据报个低于市场的价」——他类比 Waymo 温柔驾驶：「温柔是 AI 的好默认值」。护城河判断很有时代感：纯软件护城河归零（「我让我的 agent 看一眼你的软件，它很快给我复刻一个」），剩下的只有网络效应、品牌与终端关系、以及资本开支扩张能力；数据飞轮则是「裁判模型监督每一句话、每次通话，评判回流重训工作流」+ 与现有经纪行合作偷师市场细节。

公司内部则是 agent 原生组织：coding agent 目标每人每天 10–20 亿 token；他本人有一个「AI CMO」向他汇报、下面挂一串长期运行的子 agent，旁听会议、会上直接修 bug（「会议与工作的边界正在消失——开会本身就是在给 agent 制造上下文」）。团队 5 人、将到 10 人，招聘标准只剩一条：「这个人会不会促使我们做出原本不会做的改变」；跨学科通才的价值指数级上升，因为「瓶颈不再是执行力，而是把好想法讲清楚」。愿景分两幕：第一幕给人类经济装「神经系统」（公司名取自 a more perfect Union），第二幕协调人类本身——从「让 agent 替我找一千个人签情愿书、把 Barry’s Bootcamp 开到伯克利」，到「每位议员身边有二十万个选民 agent 做高保真聚合」的「实质上是代议制的直接民主」。风险他也认账：协调能力是权力工具（他举的极端例子是「帮我找同城里不喜欢某族群的人约饭」），必须装围栏；而「赋权本身就是更大的风险」。对齐在他看来有两半——把活干利索、以及对齐给谁；后者要求系统深植于无数人的生活才能「测量集体福祉」，而从社交网络学到的教训是：任何能被静态优化的指标都是错的，只能动态逼近；今天最粗糙的 v0 度量就是成功费。Nathan 抛出的「机制设计」问题（几个前沿实验室之间做窄范围、短周期、可迭代的互信协议来给前沿降速）被他归为集体智能的应用：囚徒困境在「不能通信」时才无解；可以想象各方都接受的独立 AI 审计方，「有点像区块链想做的无信任协调」——但「这不是我们在做的事，但愿我们来得及做」。

主持人对谈：微软 25% 买断、实验室扫货公司数据、明年年中的「盒子里的公司」

嘉宾散场后 Nathan 与 Prakash 的宏观对谈信息量不低。Prakash 抛出（Nathan 补充核对）：微软正向约 25% 的员工提供买断（Nathan 的理解是买断邀约覆盖 25%，接受率待观察；Office 各部门、Xbox 游戏部门直到 70 级资深工程师都在列，AI 团队除外）——他的解读是「人员配置已经超过了 agent 能接手的工作量，相当于十倍生产力提升」，要养住六万人收入得翻倍，「经济必须更快增长，否则白领人力的社会后果没人想面对」。Nathan 的对冲：他「看空就业、看好人们在没有就业时找到意义的能力」——AI 政策圈忘了多数人是「为生活而工作」；微软那 25% 的位置大多是靠执行任务升上去的，不是靠「没有你公司就会走另一条路」的反事实贡献，「我看不出这么多人能完成这种转向」。两人也点出 Union 式公司的另一面：Nathan 想起一个月前流传 24–48 小时的「Anthropic 是最后一家公司」梗——如果这样一个网络效应+私有信息的经济协调层被某前沿实验室收购，「你开始能看出那种未来怎么落地」。

Prakash 的两个前瞻更具体。其一是数据掮客信号：他在数据掮客的 WhatsApp 群里看到，掮客当前的求购需求约 12 个月后变成产品——实验室在整公司整公司地买数据（航空公司、游戏公司、半导体、制造业；他提到与「转写疑作 Mercor 的 Edward」的对话，与本研究站旧文里 Mercor 收购真公司建 RL 环境的记录互为印证），50 人公司已收完、现在盯着 100–500 人的公司。其二是**「盒子里的公司」**：约 6 个月后你可以买到「dropshipping 公司一个盒子」，加资本、加一个想法，选品、投放、客服、退货全托管，「一个人跑十家这样的公司」；Croc 鞋扣创业者被收购的故事说明「想法到商业化」的鸿沟一直巨大，而投资决策「第一位是指标、第二位是 vibe——指标 AI 也能给」；他预测明年年中起以一人公司开跑，Andon Labs 们会把这类 agent 卖给你，再往后是「一个人管一间制造厂」。Nathan 自己的 Fruitopia 水果数据库 app（「2% 时间纯 prompt agent 搞出来的全球最大零售水果信息库」）算是对一人公司浪潮的躬身入局，但他坚持「食物总量原则」：人一天只能吃那么多卡路里，餐厅机器人化省下的 25% 人力成本变不成五倍餐饮需求。Prakash 的收束是同心圆模型：AI 经济永远涨得最快，供给 AI 经济的行业都增长、供给人类经济的行业下行，「除了地位（status）」——人类经济里唯一不缩的圈；慈善本来就是地位游戏（Anthropic 创始人把 80% 回报捐给非营利组织、各大宗教的什一税、「匿名捐赠、死后揭晓」的最高段位），关键是把地位游戏导向亲社会结果，「我说它是地位，你说它是慈善——罗素共轭而已」。Nathan 勉强接受这个共轭，但仍希望过剩精力别全流进零和的地位竞争里。下期预告：周五请 ServiceNow 的 Michael Freeman 与 Mistral 的 Sagar Vaze。

机制 → 二阶影响 → 读者启发

1. 评测迁移链：静态基准可刷、私有基准不可审计 → 生产流量评测「统计上不可过拟合」→ 实验室的提分管道从刷榜转向接真实分布（Arena 靠预发布测试已卡住生态位）→ 启发：任何 AI 产品的埋点与 trace 都应按「未来会变成评测集」来设计——Anastasios 的原话是「evaluation 里的 value 不是白写的」。
2. 轨迹完整性链：对话超过二十轮失当概率过半 → 长任务的安全度量从单轮毒性转向轨迹审计（越权/假归因/欺骗性完成三件套）→ 独立审计方核对「说了什么 vs 动了什么」将成为标配 → 启发：先记录、后评估；没有产线 trace 就没有对齐可言。
3. 内存墙链：推理能耗大头在搬数据 → 铁电+3D 让旧厂造出先进存储 → tokens/watt 与每 GW 造价同时改善 → 若兑现，算力供给曲线再上一个数量级、EUV 稀缺性被部分绕开 → 观察指标：2027 年 Kepler 投产与 2029 年「1–10GW」兑现度。
4. 撮合中介链：100 万亿人工撮合 → agent 中介以成功费+网络效应切入 → 纯软件护城河归零、品牌/关系/资本成为仅存壁垒 → 启发：评估自家产品时先问「我的壁垒是不是『agent 看一眼就能复刻』的那类」。
5. 就业与地位链：任务执行者与反事实贡献者二分（微软 25% 买断是第一块多米诺）→ 岗位增长追不上生产率 → 过剩产能流向 status 经济 → 政策与产品设计都该开始考虑「把地位激励导向亲社会」的机制。

三个半小时、三位嘉宾、两层宏观对谈——把这期节目当成 10 月末 AI 产业的一张断层扫描图来用，是合适的。

ASR 校正表

转写原文（修正版）

[0s - 120s]

[120s - 240s] Uh huh. Oh, no signal. Good morning. It is Wednesday, October twenty-eighth, nine o’ two a.m. Nathan, good morning to you. Good morning, Prakash. How are you today? I’m good. This is our this is our first show for this week, and there have been quite a lot of developments. I think massive, massive number. Actually, sometimes, sometimes I think we we see a lot of things which are world historic. But I guess the rest of the world doesn’t know they’re world historic yet, and so we end up in this like weird situation where there’s like earth-shattering events on a daily basis, but the Earth doesn’t doesn’t move that much. So, what has stood out as most earth-shattering to you? I think the the things that have crossed my attention that are probably going to go down as like legitimately historic are the major open problems being solved in math, computer science, and physics, which will strain my ability to understand, let alone explain. The cyberpocalypse continues. So far, I would say, kind of a slow burn. Um.

[240s - 360s] You know, knock on wood, I haven’t been directly affected just yet. But that’s maybe for the, but for the grace of God, honestly. I actually have an interesting story about a security review I did last week, after like between our last conversation and this one. That you know is maybe instructive for other people to make sure that they do an agent-powered security review for themselves. But where do you want to start? I would say the math developments, the math and science. So, in summary, about four or five days, like late last week, early this week, OpenAI announced just three, just three. Unlike the 722 solutions last time, they just announced three, three scientific mathematical problems which they had solved, and they put out the three in the in into the public domain. They published all three had lean formalizations. So, I think the the biggest one was the exponential time hypothesis. So, I’m I’m I’m just so that I don’t butcher this. In computer science, there are algorithms which take not that much time. So, if you if you add a single variable, so you have an algorithm that takes maybe one second if there’s one variable, and that takes two seconds when there’s two variables, and three seconds when there’s three variables. So that’s a linear time. So linear time as you add variables, it you know you add you add time. There are algorithms which take much much longer, what we call exponential time. And so, you if you add if you have one variable, it takes one second. You have two variables.

[360s - 480s] It takes four. You have three variables. It takes nine, etc. Just goes exponential. And so, as you can imagine, some of the toughest problems we take we have in the world take exponential time. So, if you look at calculating all of the movements of some molecules in a in a group of molecules or in air, it’s the number of calculations that you have to perform is so so large. As you keep adding molecule by molecule, the number of interactions that you have to calculate so so much that you know even to calculate a all of the you know movements within a small amount of air would take you more computation than all of the atoms in the in the entire universe, right? So it’s a massive massive number. And so, because of that, a lot of these problems are intractable. Like we basically you know believe that we will never be able to calculate all of the interactions of every single molecule inside you know a group of atoms. And maybe that means that there are limits to what humanity can actually learn from science. We are going to end up with simplifications which don’t precisely describe all of those interactions. Which means we will you know never quite precisely know, for example, how weather happens or how any of these interactions between these molecules happen. And so, this has been in in computer science this exponential time hypothesis that you know you have this exponential time and you will you will never get better than that has been around for a while. And that means you know you’re just stuck. And this is also the exponential time hypothesis has to do with this P equals NP. So P equals NP is the super is the super set of all of these problems. And that means that you know you will never be able to take these these problems and move them down into more amenable to computation.

[480s - 600s] Algorithms, and what happened is that our friends at OpenAI found a found a disproof. They found they found some method of taking and you know something that was you know exponential and making it sub-exponential, slightly sub-exponential. They made it slightly faster than theoretically possible, and that has implications. That has implications, a lot of implications, because it may mean that you know we are heading towards a zone where a lot of these problems that we thought were basically intractable to computation are going to become tractable over time. That has both you know good and bad effects. Part of why people were interested in exponential time hypothesis is that you can then predict some algorithms are not going to be broken, and that has that was what we used for cryptography. A lot of the cryptographic algorithms that we used have this you know particular effect that you cannot brute force them because if you were able to do this in a reasonable amount of time, you could brute force a solution, and instead we use these algorithms which were beyond this exponential time. So we were like, okay, this is what we call hardness. Like these cryptographic problems are hard, and so because they are hard, we can use these cryptographic you know methods, and we have secure communications, and we have Bitcoin, and we have Ethereum, and we have all of these things because you know these algorithms are basically intractable to computation, and now that wall has been broken. So we have the first cracks that some of these problems are tractable to computation. It doesn’t mean all of the problems are. It doesn’t mean cryptography is broken, but it means that the cracks are there, and it basically took down this wall, this 25-year wall since the exponential time hypothesis was first stated. This 25-year wall of you know what the limits are, and it’s showing us that those limits that we thought existed do not exist really. It’s just you know absolutely fascinating, and it.

[600s - 720s] Is going to have, I mean, obviously, this means that people are going to look at, investigate how to, you know, get to P equals NP and how to, you know, solve these problems in a much much shorter timespan, and that is going to have immense impact on on the world. One question I have that I don’t know if you have the answer to, but maybe Q can help us find it. If if you don’t know the answer, is so. My understanding is like the disproof was by example, basically, right? They took one of these problems that were thought to be exponential time hypothesis hard, and showed that they can come up with an algorithm for that particular problem that is faster. Yeah. Therefore, the hypothesis is false. From the commentary that I’ve seen, it still seems like people think that first of all, it seems like the the speed up. There have been like a few of these recently, where all of a sudden somebody found a limit or found a algorithm that was faster than the the what was previously understood to be the limit. My understanding generally has been that they’re not that much faster, and most people still seem to think between P equals NP and P does not equal NP that it’s probably not. In other words, some things will continue to be very hard to do. They just don’t seem to fall into this like clean class that we used to think that they fell into. It’s a weird thing because I’ve also seen some commentary that’s like this is sort of ugly. People seem to like the kind of clean classes that they had before, where it seemed like there was something elegant about it, or there were you know it it sort of suggested some. Uh, some deep structure to.

[720s - 840s] The universe that was like grockable and beautiful and whatever, and potentially there’s a certain bias among the human mathematicians toward accepting that because it did seem kind of aesthetically appealing. And now we have these like, oh, you’re able to get under that what we previously thought was the limit. You’re able to get under it a little bit, and now it’s kind of ugly. It’s like, oh, you know, we don’t really know. There’s not like a clean answer to how hard these problems are. It looks like we just have to kind of grind on each one. Maybe there will be some deeper structure or elegant beauty, you know, to emerge again as we kind of map the space out better. But it’s sort of a weird result that seems to kind of challenge a lot of the assumptions people had, or maybe just preferences that kind of led to assumptions about the universe being elegant or beautiful, or you know, lumping a bunch of different problems together into a certain class. Now we’re left with, at least for now, it seems like a bunch of individual problems that we can sort of tackle to varying degrees, and they’re sort of each kind of their own little snowflake, and it’s sort of not so aesthetic, but it’s also not that much faster. Is kind of my overall sense for at least what we’ve found so far. How does that sort of survey of of discourse line up with your understanding so far? I think I think that’s right. So so one of the one of the commentaries is that these are galactic algorithms, galactic in the sense that in order to get any amount of speed up, you need to run like you know if you if you’re looking at algos of like two to the power of like 100, and then you get like a 0.0001 percent speed up, right? Like like these tiny tiny fractions. And so that that has been the true of the integer multiplication. We had a we had a result in integer multiplication with the 722 problems.

[840s - 960s] Problems about a month ago, and that was this very tiny speed up over you know from two to the minus hundred to like to the minus you know ninety nine or whatever like very very tiny speed ups. I do not think that it is insignificant, and because I think this is only the start, it’s it’s very easy to say oh it’s nothing now right yeah of course it’s nothing now but it it reshapes our view of the universe in in the same way that you know discovery of like the Higgs boson or you know the theoretical kind of postulation of the Higgs boson or you know quantum mechanics in the early days did. Um, when when Einstein you know theorized about the photoelectric effect, you know in 1905, you know there was no sense that this was going to lead, you know, in the future to nuclear weaponry and you know nuclear power and you know like these theoretical yes these are theoretical theoretical things, but I think the downstream impact is yet to be seen, and I think after a long period of not very much movement in the basic sciences, right? Like you know Patrick Patrick Collison had this whole thing about how we haven’t seen a lot of like t here is been a slowdown like a generalized slowdown, and he’s been he’s been trying to promote this kind of study progress studies and like why how progress happens. These are immense. I think I think you start off with the first disproving. The disproving basically means that your worldview is now shattered. So similar to the Copernican revolution, you know your worldview is now shattered, and then it takes some time for that you know for for a Galileo to come up and say okay you know you you need this heliocentric model etc. And then you have the Pope like coming down hard and like and then the world you know in a couple of hundred years you have Martin Luther appearing and like you know smashing smashing.

[960s - 1080s] The church down, right? So, so I think I think like this is kind of the the start of the process. It’s it’s very unusual, like you we you you rarely get to be around for one of these things, and you rarely recognize that they’re happening in the moment. And so it’s very exciting because you can kind of see it’s happening. Where it will lead, I do not know. It’s it’s it’s possible that it leads to like some small like you know computer speed up in like you know ten years, right? It’s also possible that you know it reshapes every single algorithm that we use. It reshapes Google search, for example, right? The number the number of connections that the PageRank algorithm could explore was limited, and because it was limited by these by by the speed of the algorithms, there was a limited number of searching that you can do. And so the extent of your knowledge, to some extent, the extent of how much knowledge you can kind of have within that like two seconds that you type in that Google search into, that’s kind of limited by you know how these algorithms and how quickly they can expand to capture the number those numbers of data points and interactions. And now that thing is being broken, which means the number of data points could be much larger, right? Could maybe in those two seconds you could do one hundred times what Google does now, and you could still do it in parallel too. So it’ll be like a million x, right? So a million or ten million x, right? It could be much much larger, right? So I don’t think it’s clear yet. I think it’ll take some time, and I think there are other innovations that will come in and kick in as well because these things don’t happen, you know, in isolation. In fact, the the other funny thing that happened was when OpenAI announced this, our friends at Anthropic came out and said like, “Oh, we found it too. We found it too on the same day. In fact, on the same day.”

[1080s - 1200s] We just we just didn’t we just didn’t want to like upset the mathematicians, so we we quietly put it on a GitHub repo. So yeah, I think I think more more will emerge over time, and I think we will only see the impact in the in the coming weeks and months. It does feel like there’s a bit of a four minute mile quality to this, where with a barrier being broken, all of a sudden there will be a whole bunch of influx of energy and exploration, and it’ll be really interesting to see how far it goes, and it’ll be really interesting to see if there’s some other fundamental limit or a few like a you know small small number of different limits that these optimization processes run into. I’m not really sure what to think about it from a philosophical standpoint, but it is weird, you know. It’s like I’m not sure what I had to say about the the philosophical implications of the previous hypothesis that was more unified, more elegant, whatever. Now, if it really is sort of an uglier space of algorithmic possibility, what should I take from that? I don’t really know, but it does feel a little strange. I mean, I feel like I had sort of followed people into this headspace of like there’s some deep order, some unifying something that underlies all of this that we’re getting at, and I don’t think I should necessarily update away from that too strongly because again, we’ve maybe just found that it wasn’t really the line that we thought it was, and maybe there will be some other line that that can kind of restore some of the vibe that we maybe have you know feel like we’ve lost here for a second about a beautiful, deeply ordered universe.

[1200s - 1320s] But if it doesn’t go that way, or you know, it’s at least called into question a little bit more now. I don’t even really have an intuition for what that would mean. Like, does it shift your worldview at all, or does it give you any intuition for like, you know, questions that matter outside of algorithms? The that that this has happened. Uh. Abs. Absolutely right. Um, it means there is a deeper structure to the universe that we didn’t know about. Because when you look at, you know, Stephen Wolfram calls this computational irreducibility, right? Like, if you are at that, like, okay, this is exponential time algorithm, that is the simplest that things could be, right? You, you, you’ve already gotten to the place, and you’re like, this is some natural phenomenon, and this is the simplest that I can describe. It is this exponential time algorithm, and that’s you know that’s all I can do, right? That’s that’s as much structure as anyone can capture in the universe. In order to describe that phenomenon, I’m going to need that number of computations. That’s it, right? And and and now all of a sudden there’s there’s more. There’s all of a sudden there’s a deeper structure to that to that, and all of a sudden, like you can describe that phenomenon in in fewer numbers of bits, and fewer number of like calculations and computations. There’s a deeper structure to the universe. That’s exciting. Like that’s super exciting. Like, and and we didn’t know about it. We didn’t know that there’s this deeper structure. We we were we’re kind of like this is all you see, and this is what we can do. And now there’s a deeper structure, and then you can go in there. You can deconstruct that deeper structure. If you look at the proof that they did, it’s it’s very interesting. They basically did all of these calculations, and then once they did all of these calculations, they realized some of those calculations can be grouped together. And if you group them together, you can kind of eliminate. And this is a little bit like prime numbers, right? Prime numbers are a way of like you have this like multiplication, and as you go forward in that multiplication, you realize that there are these prime numbers, and if you

[1320s - 1440s] Take those prime numbers apart. You’re able to describe all of those multiplications in fewer numbers because you have these prime numbers, right? And the prime numbers allow you to describe through factors all of these larger numbers in terms of smaller numbers and groups of groups of those numbers. And all of a sudden, all of your multiplications around become, you know, focused on this prime groups, and then it’s just, you know, rotating between the prime groups. So this is something like that. They have discovered a deeper structure to the universe where you have these exponential time algorithms, and then you are able to separate them into these groups, and then you can manipulate between the groups, and all of a sudden you get lower numbers of computation. That’s exciting. That’s like discovering discovering prime numbers again. Like it’s it’s it’s amazing, right? Um, the it’s it’s I mean, again, I don’t understand enough, and I and the fact of the matter is, I think only maybe like you know 500 to 1,000 people in the world actually understand enough. And this is a very unfortunate fact that we have all these exciting things happen, and even Fields medalists are sitting around and saying like, well, I don’t feel very qualified to comment on this, and you know, such is the nature of you know late stage you know anthropocentric you know intellectualism, you know. Um, I I feel I feel so unqualified, except in that other people are also unqualified, and so here we are. But that’s the best that I can describe. The fact of the matter is, there’s a deeper underlying structure to the universe, like prime numbers, that we didn’t know until now, and all of a sudden we know, and now you know people are just like you know aghast, like okay, what do you do with this deeper structure? How do you how do you manipulate it? How you know are there any other uses for it? Like, yeah, it’s it’s fascinating, absolutely fascinating. Another topic I wanted to ask you about, just based on a tweet that I saw, is Claude duet. Um, it sounds like you’ve had good experience with it.

[1440s - 1560s] I honestly haven’t tried it because I’m just content enough in my own agent setup at this point that I’m not really looking for new wrappers at this time. But based on your review, maybe I’ll have to reconsider. Tell me what your experience with Claude Duet has been in the first couple days. So, and you know, there’s always this activation energy, so you have to cross the activation energy to use a new product. So I got on the Claude Duet, you know, shortlist. They they sent me they sent me the results. They they sent me the invite. I joined and like didn’t use it for the first couple of days and then decided try it out. Um, it’s amazing. It’s amazing because number one, it is very fast, much much much faster than dot and um, dot and you know Muse and Grok at this point. Although Grok my Grok usage is like three weeks ago, so maybe people are telling me Grok has improved. So I I don’t know yet. So, um, it’s very fast, and you you can see kind of Mike Krieger’s influence on the product team. It it tells you what it’s doing while it’s doing it, and it kind of gives you like a little bit of a sense of like how long things are going to take. If things are going to take very long, it tells you it’s going to do it in the background and then like come and get back to you. Dot from OpenAI doesn’t do that. It kind of dot sometimes says like it’s doing something, and I don’t know what it’s doing, and I have no sense of when it’s going to come back, and it doesn’t tell me, and it doesn’t show me a processing sign, so I have no way of like knowing how long something’s going to take with dot. So almost always with dot, I end up typing a query and then waiting a few seconds, and I don’t really know whether something’s happening, and I go away, and then I go have to go away, and then it gives me a notification on the phone, and I have to go back. It’s like this interruption and flow is terrible. Um, dot is also slow, right? Very slow. Um, this latency question is going to be a huge question for all these companies going forward. They have to figure out the latency.

[1560s - 1680s] I think they’re gonna need. You know, I feel like OpenAI is kind of like sandbagging it because they’re waiting for the device, so they’re looking at this kind of like you know on-device, off-device, and this is you know, I you know, it’s it’s it just doesn’t work for me right now, right? Um, it burns a lot of credits, so I’m I’m now I’ve now hit the limit on my cloud account, and so I’m now using the cloud credits that they gave. They gave about two hundred fifty dollars of cloud credits. I did a kind of like a stock stock market research on Anthropic as you as you will have it, and it built me a very very nice financial model, and it burned up like a hundred dollars in credits, like, right? Uh, but it’s it’s a professional scale model for the Anthropic IPO, professional scale built off the S one with with kind of like estimates and stuff that you need to kind of read through like communication and kind of form your own kind of basis and like make some assumptions on like. I was trying to figure out how much shares were struck at every like you know at every pricing, and it kind of like okay, this is what they started off with in terms of share options issued at the beginning of each year. This is what they ended up with. This was the average price over the year, and this is how many employees they took on in that year. So this is the, like, wow! I I found out like you know, two thousand twenty-four. Anthropic was a very expensive year. They they spent one hundred and thirty billion dollars on on staff in two thousand twenty-four, which is which is absolutely insane, right? Um, so yes, it’s expensive. It’s very very good. It’s very very good at like I’m gonna nail what you wanted me to do without consequence to tokens or like latency. I will do it. I’ll do it fast. I’ll do it done. I’ll get it done properly, and I’ll prove it. That’s amazing, right? So it gives you that kind of like.

[1680s - 1800s] The last mile achievement of like clearing the hurdle, satisfactory, and like being able to prove it. It’s expensive. It’s expensive. I don’t know how this is going to pan out with like the different classes of subscriptions and when people haven’t hit their limits. Maybe I just hit my limits really early. It’s not sustainable if you want to use it, you know, on the API basis. It it it basically costs me like a dollar an email, which is which is absolutely insane. But otherwise, it will actually get stuff done, but for a cost. So, did you? So we don’t know yet if it works with a max plan. And also, by the way, in in the model, did it have a P doom sell? And if so, what did it put in for P doom assumption in the in its analysis of Anthropic stock? I didn’t. It didn’t give me a P doom. It it it also uh well you know I I was doing a bunch of analysis on like uh you know the uh when when they would be selling shares so you know how how much shares would people be selling and how much shares like because they have this like kind of very weird like one year one year long like uh you know stock disposal plan etc. No, it didn’t. It didn’t give me an assessment. And and it didn’t. It didn’t come out and say like I’m Claude. Like I can’t. I can’t talk about this. So, uh, but um, you know, I I shared the analysis that I did with with a couple of friends, um, and and so people are using it in investment decision making, um, and I I I having spent that hundred dollars, I sp I felt like I had created valuable knowledge in the world that didn’t exist before, and that I don’t think anyone else had asked that specific question in that way, and so I feel like we are going to be able to create new knowledge, and that’s exciting.

[1800s - 1920s] Yeah, just add money. Just add money. That’s that’s the thing. Just add money. So let’s turn our attention to memory, shall we? Indeed. So let’s get our first guest for today. Let me do the intro. Just give me one second. Just give me a second. Our first guest for today is Devo Ola Shabigan. He is the co-founder and CEO of Kepler Computing, a semiconductor company developing new kinds of memory for artificial intelligence systems. Its work addresses a basic problem: a powerful processor can spend time waiting for information, and moving that information uses energy. Kepler is working on new materials and ways of arranging memory in three dimensions to improve that exchange. The company emerged from seven years in stealth in September two thousand twenty-six. It reports raising four hundred and sixty-eight million dollars and is working with global foundries as a manufacturing partner. Kepler plans to begin production in two thousand twenty-seven. Devo is also a named inventor on a patent covering ferroelectric memory chips connected in a three-dimensional package. Before Kepler, he co-founded Gigster, which connected businesses with software development talent. His earlier career was in physics. He came from Nigeria to Illinois Wesleyan University at sixteen, then pursued graduate research at Cornell on light and its interactions with materials. There, he participated in a collaboration working toward an electrically activated silicon laser. His work has taken him from experimental physics to software services and back to the physical devices that make computing possible.

[1920s - 2040s] Debo, welcome to the show. Hey everyone, glad to be here. So many questions. I think we’re honestly a little out of our depth when it comes to understanding the different hardware paradigms and how you are kind of rethinking some of the assumptions and limits that have characterized the hardware space and the memory space in particular to date. So rather than dive into particular things off the bat, maybe I’ll just invite you to kind of give us a little mini lecture on like what is the different perspective that you’re bringing to this challenge, which is obviously becoming extremely timely and valuable. Definitely. It just occurred to me that back in the day, I worked on trying to get trying to figure out a theory for high temperature superconductors, and I and I think Prakash was caught up in the LK99, you know, situation. I think that’s where I remember you from Twitter. My my my my my my my past catches up to me. Yeah, that was a that was a fun time. Yeah, so you know, we we basically see computing as kind of a superpower, giving giving people superpowers, and this whole mission of artificial intelligence, super intelligence these days is is about giving people superpowers, making everyone a superhero, and and that’s kind of what we set out to do. Is we say, hey, you know, it looks like right now the Iron Man suit.

[2040s - 2160s] For people, cost two hundred bucks a month or five hundred bucks a month, and you know there’s only so many people who can who can spend that much money. Maybe less than one percent or point one percent of people who can who can afford that. So our mission is really about how do we get the this this the superpowers this this iron monster to everyone. How do we get computing to everyone? And we look at where where it’s breaking down, and it’s because it’s breaking down because of certain limits. You to you know to power your ChatGPT, to power your Claude, to power your AI, you need chips, and then those chips need energy. And we know right now that there’s there’s only a certain amount of energy that we have access to, and we’re doing everything we can to generate more energy. We’re doing we’re even trying to go to space to power these chips. So energy is one limit clearly, and you know making chips also is another limit. It costs a ton of money to build the factories, the fabs that make the chips. You know one fab, one modern fab could be about forty billion dollars today. So even if we had a ton of energy, we’re not able to make enough chips, and if if we can make enough chips, we don’t have enough energy. So what we’re trying to do is say, and and because of that, you you you kind of ration out intelligence. If if you only have so many chips and so little energy, then you have to charge two hundred bucks a month. That’s just a rational thing to do. So if you want to democratize this thing, you have to actually attack the root of the problem. You have to say how is it possible to make as many chips as I want? How do I improve increase the number of chips I can make per dollar? And then you want a two for one, which is what Moore’s law used to give us in the past. Not only do you want to make a ton of these chips, you want it to be that these chips are as energy efficient as possible, so that the energy you do have can go around. So this is kind of like the the big picture. This means that whatever we do at

[2160s - 2280s] To figure out how to make chips give you more tokens per watt, also has to be something that causes the cost of manufacturing these chips to go down, so we can make many, many more of them. And so this brings us to why we tackle memory first. If you look at, you know, I guess one step before that. So principle one is we have to reduce the cost of manufacturing. Principle two, we have to reduce the amount of energy that these chips are consuming. Principle three is, you know, popular within some circles, maybe unpopular within other circles. We want to stick to digital computing, digital mainstream computers. You know, there’s there’s other kinds of computing others are are are looking into that maybe that that do have merit, but we want to say, hey, for for for the existing mainstream architectures, how can we give it a 10x boost, a hundred x boost, a thousand x boost? That’s another constraint that’s kind of like religion within Kepler. So I want to make use of existing fabs and existing manufacturing technology, but also existing architectures where your software can run as it used to run before, but just magically you get a big boost in terms of energy efficiency. Okay, so then I look at the existing chips, mainstream ones, you know, the the GPUs or the TPUs, and I say, where’s the energy going? And I I think many people know this, but a lot of people don’t know that for inference, especially, the bulk of that energy is actually running across the street to fetch the data from the library, you know, before you before you process under your fancy matrix multiplications. So if I want to improve that system for inference by an order of magnitude at the physics level, it’s obvious that the first thing I need to go do is is address the memory bottleneck. I have to make the way I access energy, access information, and store and process and store that information much more energy efficient. So within this kind of mainstream digital architecture.

[2280s - 2400s] We’re going to figure out a way to bring that reduce the energy access to that memory, and one way you can do that is to bring that memory closer to the to the processor. And the other thing you want to do is increase the amount of memory that that you know that that that you can store within a certain footprint. Both of these things point us to going three D. You know, like like like there there there there are there there there is another way you can get a lot of memory within a fixed footprint, and it is say hey my memory is just a bunch of transistors on a chip. In the case of SRAM is six transistors per bit cell, in the case of DRAM is one transistor one capacitor. So if you want to get more memory, you can shrink those transistors and pack more bits in the same footprint. Obviously, we know from high rises and skyscrapers in New York that another way you can pack more bits or more things is to put them on top of each other. So you can put devices on top of each other within a chip to get more memory, or you can put multiple chips on top of each other, which is what HBM looks like today. And so doing doing both of those things, stacking devices within a chip to store more bits per area, and stacking chips on top of each other to to to to to get more capacity per millimeter square, is clearly one path you can go to avoid having to shrink the transistors. So even if each chip in your stack packs less capacity, you can compensate by putting more more chips, or within each chip, if if each transistor is bigger than you know the most advanced transistor, if you can put more devices on top of of that transistor, then you can compensate for the sort of larger size, and then so.

[2400s - 2520s] So you have three D innovations that are enabled by new three D techniques as well as new materials that allow you to stack devices within a chip and stack chips on top of each other. And then you also have three D innovations that allow you to take that stack and bring it closer to the processor. The big the big challenge with with putting a bunch of memory on a processor is thermal. There’s there’s also power delivery etc. But you know if you put a big block of something on top of something that’s really hot, you know, good luck, right? That that that’s that’s a challenge. So clearly, what what the kinds of techniques and approaches we’ve we figured out and we’re working on, you know, our first products and second products would be techniques that overcome these sort of known challenges of hey, how do you how do you bring these things close together while avoiding you know thermal issues? And then within the within the chip, how do you stack devices on top of this transistor to pack more bits you know per millimeter square? And there we use new materials within the chip. We use new materials. We use a you know class of materials called ferroelectric materials. Ferroelectrics have actually been around for hundred plus years, but the specific type of material, the specific you know recipe of the cake, if you will, this is this is this is you know proprietary to to Kepler. And you basically use that ferroelectric to make a capacitor, a type of a capacitor like thing. It’s really holding a polarization for for people who are you know more technical. But roughly speaking, you can say that this capacitor is storing a one or a zero. And if I replace six transistors with one transistor, one capacitor sitting on top of that transistor, and then eventually put multiple capacitors on top of each other, I can avoid the the rat race of like trying to shrink transistors within a chip. So we have technology, roughly speaking, that’s saying how do you go three D inside a chip? How do

[2520s - 2640s] Go 3D between chips and how to use those technologies to increase memory capacity, but also bring memory and compute much closer together while avoiding, you know, conventional like you know issues with with thermals and and and things like this. And we’ve you know we’ve been as you mentioned we’ve been studying this for seven years, and you know we’re very excited that you know the world has kind of come around to understanding how critical memory is, and that we and is very open to new solutions and new ways of going about the problem. So let’s talk a little bit about material science, because we we we come across a lot of chip designers and they have you know new ways of you know basically making the memory and where the memory sits and reallocating how the bits flow in order that you can you know optimize this kind of pre-fill decode you know differences in how the chips are utilized. But I think this is one of the first times we’re speaking to someone who’s making a chip, but also has done work in material science. You had a report, I think, in September, describing 35 composite versions that your team went through before you selected a material class. Can you tell us a little bit about this process of how you ended up like going through this elimination process of what is good and what is not good and which material you know you should use and which is not? Like you you don’t need to disclose the the actual recipe, but what what is this like the the the the like kind of tree branch process through which you kind of come to the conclusion this is the one that we should use. Yeah, good question. So actually, my co-founder and our CTO Sasi deserves a lot of credit for, and also another one of our

[2640s - 2760s] Co-founders Ramash, for many years ago, thinking through, and Saussi had a role of figuring out what comes beyond the transistor at Intel. It’s like, you know, a really great job to have, and so it forces you to have to think. If you think about the the transistor, you know, roughly speaking, we say it’s made of silicon, and we say, okay, if you want to go ten x, hundred x, thousand x more energy efficiency beyond the transistor, you know, maybe you should consider other materials besides silicon. And once you open that Pandora’s box, then you need some some principles, some framework, as you said, for for trying to like the trillions of different possible materials. So how do you how do you see through that? And and so he had a he had a paper in Nature. He and Ramash had a paper paper in Nature describing kind of figure of merit for sieving through the different classes of materials to say, you know, this is actually in the spirit of looking for something that could be a next transistor, in the spirit of looking for something that could be next logic, which is, you know, Act Two of Kepler. By the way, is you know we focus on memory because that’s the first bottleneck, but our technology in the long term applies to logic as well. So from here is where we landed on is where he pushed towards the class of materials of ferroelectrics as a class, where you’re trying to understand, you know, what is the energy required to to flip a bit, you know, to to to flip between states in this material, but then you’re also trying to understand how stable is this material to thermal thermal noise and thermal fluctuations. This is what this is a big reason why you cannot just keep producing the voltage of silicon transistors. You know, silicon transistors, the energy is the V squared.

[2760s - 2880s] And if you if you could keep scaling down the voltage, you could keep scaling down the energy. And we used to get that for free. That’s Dennard’s law. We used to get that for free, Moore’s law. So you’re looking for something where I I have the voltage knob, but I have the energy knob, but it doesn’t come at the cost of making my system very unstable thermally. And ferroelectrics were one class of materials that are very promising along this along this dimension. So we focus there. Then within ferroelectrics, there’s many different directions we could go. But we look at the figures of merit that say, okay, if I want to have a capacitor sitting on a transistor, ideally I want that thing to be as small as possible, so I can pack many bits. Okay, but if I want it to be really small, then it means that the charge per unit area on that capacitor needs to be very high, so I can sense it. So we we compute these kinds of figures of merit: so the charge per unit area, the voltage you need to switch that capacitor, and then another figure of merit is how many times can you switch on and off before before the thing dies. It’s called the endurance, and then the speed at which you can switch this thing, and then finally you want to be able to create this this capacitor in conditions that are compatible with that transistor that you’re integrating it with. So these are the you know the the the things in phase space, I guess that are you’re kind of optimizing around when you’re trying hundreds and thousands of iterations, guided by of course materials science expertise and increasingly going to be guided by you know computation. So these were the kinds of knobs we were the kinds of things we’re optimizing for, and then you optimize down at the level of the actual process for depositing these things, for etching these things, for for actually creating the the physical capacitor structures to get to the to get.

[2880s - 3000s] The promised land. Did you did you end up with what would your final round look like? Did you have two candidates, three candidates, and how did you make that last decision? Um, there’s um, so say like we have um, we have very clear we have very clear winners, you know, in this in this game, and so it’s like okay, this one is the best, and and it’s this is the best combination for you know at least for a product with the spec at this timeline, and then what we what we end up having is we have a roadmap where we say okay, here’s the next stack, and it’s not really just like one material, it’s you know multiple materials, you know maybe a few materials that you need you need something as an electrode, you know things like this, and so that stack you say this is the best one for the product that I need to ship right now, but then we have even better things that are just maybe a year behind or two years behind that need further development and that will push the spec even even further. So it’s more that there’s a there’s just like a rank order of things. What what does that development look like? Is it simulation? Is it like chemistry? Like what what do what do you actually need to do to develop that kind of like next next layer? Like are you pipetting stuff? Yeah, it’s it’s every it’s every it’s everything. So you can imagine like um, you know like making a cake or something and then like cutting it up. So you you you do have to you do do some simulation right to figure out okay, I think if we tried this versus that, I think it’s going to improve you know for example the charge density or it’s going to improve the the you know the switching speed or or it’s or it’s going to allow me to create this thing at a even lower temperature, whatever it is. And so there’s there’s simulation track for sure, but the heavy duty physical work is you know about.

[3000s - 3120s] Depositing these materials with you know manufacturable techniques, depositing these materials at appropriate temperatures, appropriate pressures, etc. Pressure reminds me of LK ninety nine. So you you want to you want to use you want to use reasonable reasonable pressures, reasonable temperatures, and so there’s a lot of deposition happening of these films, and then you go and use X-ray techniques, use different kinds of metrology techniques to characterize these films, and so before you even make the capacitors to test them electrically, you can already you know validate the composition, validate multiple things about the about the material before you go make the device. So there’s a long, very fast loop of I deposited this thing, I characterized it, I don’t like it, I try again. Okay, then I make a device out of it, and initially you make just the capacitor. You don’t need to couple it with the transistors yet, and you measure different properties of these of these capacitors. And we have a facility, and you know, in South Bay, we also have dedicated fab in you know as part of as part of our partnership with Global Foundries. So we can go all the way from idea to something in the computer to something in an actual you know on actual silicon at a small scale to something on silicon at a large scale to an actual chip, you know, in sort of like a full loop under our full control. And the things that are within our lab, you know, the cycle time is days and weeks, right? So we can we can we can turn around really quickly, and then we can take that thing all the way to you know product a product level chip. Could we do just a little kind of side by side tale of the tape, comparing Kepler memory to standard memory that people are most comfortable buying and integrating today? Like you you listed a bunch of these dimensions, right? Like the size of the individual.

[3120s - 3240s] Unit, the stability, the the speed, the energy required. I’d be really interested just to go down that list and kind of get a a sense for on all these different dimensions what the head-to-head comparison is like. Yeah, we should we should zoom out because there I was talking actually about the capacitor itself. So let’s zoom out to the memory product. There are two memory offerings which we’ve been working on, you know, from the beginning, and we’re very thankful for companies like Groq and companies like Cerebras, which have brought and and also the rise of agents, which have brought a lot of light to the importance of SRAM as well. So everyone knows HBM is important. I think now everyone also believes SRAM is very important for AI. So we have two offerings: an HBM class offering as well as a SRAM class offering, and we’re scaling both. So today, the scaling of SRAM has slowed significantly. If you go look at like N seven versus N, you know, not not to pick on any particular foundry, but just in general, this is well known that SRAM scaling is sort of sort of slowed. So the value proposition for Kepler SRAM is about density, so bringing that scaling back, and we we for L three type L three class sort of SRAM, which is kind of the the dominant class of SRAM you’d want for for large large large scratch pads or large caches. Our goal is to deliver two to four times the density of of of a baseline CMOS SRAM. You cannot get that today. I mean, you get like ten percent or twenty percent, you know, type of scaling by going from node to node. So that would actually be equivalent to three plus generations or more of of scaling. So that’s the big value prop on on the SRAM side, and then we we want to match the performance and the energy and voltage.

[3240s - 3360s] L three, but a win on the on the density, and then on the HBM, you know, I would say that the the directionally the the way to think about Kepler HBM memory is that the bandwidth per watt is going to be approaching the bandwidth per watt of SRAM. So we’re going to be well beyond the bandwidth per watt of today’s HBM, and pushing that closer to the bandwidth per watt of SRAM. And then we’re going to also over over multiple generations of our product push on the capacity as well. So you should look at the you know the bandwidth per watt and the capacity improving that over over HBM, and customers can use Kepler’s HBM. There’s an option for Kepler customers to use Kepler’s HBM as just a regular HBM. So even without more fancy or customized integration, if you just say, “Hey, I just want to use your memory as a standard sort of JEDEC like sort of you know side by side memory,” this is also supported. And many of our interested customers, you know, some of our interested customers want to start that way. Some of our customers are more adventurous and they want to you know do the integration in a more coupled way. And so, so that’s kind of how I think about it in the HBM side. It would be pushing on the bandwidth per watt, and then pushing on the capacity, and then the SRAM side is primarily pushing on the density. Let’s talk a little bit about manufacturing. You know, since you’re using new materials, does that necessarily

[3360s - 3480s] State new kind of manufacturing process. I believe you are not bound by extreme ultraviolet lithography. I’m sure that has some advantages to it. Perhaps you can use nodes which are not as expensive as the ASML nodes, which are also very very difficult to get capacity on. So can you talk a little bit about the advantages of you know not being bound to that capacity? And also, are there any challenges because you need new processes in order to actually manufacture? Great question. So I think this is a huge thing. This is like you know always say it’s similar to how the world thought it didn’t have enough taxi drivers, like San Francisco thought it didn’t have enough taxi drivers, and then we realized, oh, we have so many taxi drivers if you just build Uber. So you know, not to not to undermine the fab shortage or chip capacity sort of manufacturing shortage, but the the the focus right now is on sort of EUV heavy fabs. The expensive, they’re complex, more complex, you know, processes, and we’re saying that we can take existing older node fabs and we can use them to make chips that are actually superior to the advanced node chips. That doesn’t mean that we can’t use EUV fabs. EUV, we just we just invent on an orthogonal axis to shrinking transistors. So we also benefit when the transistor shrink, no problem. But we don’t require to make chips that are that are more advanced. So that’s a huge advantage because the cost of these fabs are lower. The lithography you need is is you know much much less expensive, and some of these fabs have a lot more capacity in them. Right. For example, if I could take an

[3480s - 3600s] Older node logic fab, and if I could use that to create more advanced SRAM, or even more ambitiously, if I could use that older node logic fab to create HBM class memories, this is this is a huge huge thing. This is a huge unlock. And even if I needed to build more fabs, so let’s say I make use of whatever fab capacity I have, and then I need to build more fabs, you know, my my fab per dollar or dollar per fab, whatever ratio you want, is superior to to the person who has to set up an EUV line. So this is all advantage. In terms of the process itself, the key thing to watch for is: Are you introducing any brand new tools, any custom tools, any custom equipment that’s not standard equipment that already exists in the fab? And the answer is no, for the most part. So the process flow in utilizing those equipment and you know some of the materials, sure, that may be new, but we’re not we’re not bringing in any new any non sort of standard equipment or causing some equipment maker to build new equipment for us. And so the reason you know we’ve worked hard on this for the past seven years. So what we’ve been doing all those years is turning that flow on standard equipment into something that is manufacturable, building test chips inside of you know proper fabs that basically prove that hey, okay, once you nail the process flow within these standard tools, you can just stand up these lines within the fabs, and you’re you know you’re good to go. So it’s it’s it’s sort of the complexity is contained within the process flow inside of those tools, which we which we have been honing for the you know for the past few years. So once those once those process flows are sort of finalized and honed, it’s very easy to replicate this and scale this up.

[3600s - 3720s] Inside of inside of a manufacturing process, so it’s really it’s really all upside once you have the stuff working. So, if you had to do some kind of market analysis or projection going forward, you know, one thing that everybody is kind of watching is like, how much are we going to scale compute over the next couple years, and what are the bottlenecks going to be? Obviously, there’s a lot of investment interest downstream of assessing these bottlenecks, but there’s also just the question of like, how much are we going to have like an abundance of AI, or is it going to get really expensive, and you know, people are going to get priced out. Semi analysis has all these different kind of you know simple memorable rules like ASML is going to make 100 machines a year. They’re probably not going to scale that that much, and so by like 2030, we’re going to be starting to run into limits of what the what the you know the machine tools are going to be able to support. But maybe you come along with a different paradigm and sort of free up a lot of EUV capacity. Like, how would you modify what you would you know? If you want, you can articulate what you understand the because you probably understand it a lot more precisely than I do. But what would be the kind of adjustment you would make to the forecast of how much compute the overall ecosystem will be able to bring online, based on the fact that you’re going to shift some portion of memory production away from the most scarce machines, you know that are that could then be repurposed to other things. Yeah, I think that I think that Kepler. I mean, the short answer is I don’t think I don’t think that Kepler is priced in to these analyses yet. It’s no one’s fault. We’ve been stealth for seven years, but but the idea that it’s actually interesting because I think there was a discussion between Dwarkesh and Dylan on on Dwarkesh’s podcast.

[3720s - 3840s] Where Dwarkesh multiple times brought up this point of like, hey, but isn’t it possible to use older node fabs to do this? And you know, it was like, I was like, oh, dude, you’re you’re on the right track. So, so yes, if if you take this curve, you know, of you know, sort of capacity over time or capacity versus dollars, you know, or energy usage over time, and if you can shift that curve down in some way, then you obviously get more compute. And so, if Kepler’s technology is invested in aggressively, such that we manufacture this technology is underpinning as many chips as it can possibly underpin, yeah, there’s no reason why we wouldn’t be able to increase the amount of effective compute by an order of magnitude because you have two things combining. One is the fact that at least for decode inference, you can have new physics that allows you to pump out more tokens in an order of magnitude more tokens per watt. And then the other thing is that you have existing fabs to which you can append much lower capex based Kepler fabs too. And and you know, in aggregate, you know, let’s not even talk about algorithmic efficiencies that the AI labs obviously will work on to get even more efficient use of chips. I think if if if we were scaled in the limit, obviously we’ll have our own limits to scaling, which is how many fabs we can build, how many partners we can get, etc. But this is technology that should be able to give us another order of magnitude. And then when we get to logic in the future, we are working on technologies that push that even further, where you actually start to address the transistor itself. But if we just

[3840s - 3960s] Talk about the memory portion, and we talk about the fact that we can use much lower cost, much lower lower cost fabs and get more tokens out of each gigawatt. We should be able to. Aspirationally, we should be trying to get 10x more compute than than would have or otherwise been assumed. If you zoom out, kind of super big picture, do you think that the bulk of compute goes this way? Like we we’re so used to thinking of compute and silicon as like synonymous almost, right? Would you envision a world where that coupling is is really left behind, and it’s like, yeah, there was the silicon era of computing that lasted up until the late 2020s, and then we entered like this new era where new materials just created a fundamentally different regime. Or would that be too grandiose, even for your? Oh, I oh that that’s that’s a cool cool vision. I want to be clear that there are there kind of like two components here. I mean, one master component, which is well, I guess it’s worth splitting them. There’s there’s something about there’s something about 3D, right? There’s something about evolve like chips are more or less 2D and silicon today, and that there’s something about evolving those chips to be 3D and beyond silicon, right? So there’s a lot of alpha in just making the system is more three dimensional, and that’s gonna that in itself can carry us through the first order of magnitude, and then and then there’s and then there’s alpha in evolving the materials that are being used to to not be silicon. So you can have 3D silicon and just push the

[3960s - 4080s] 3Dness of those systems, more, and and then you can push that even further to say I want to start pushing on the new materials even more, right? So, I would say that yes, we will see an evolution of chips that are 2D silicon first to 3D silicon. Like, there’s always obviously already a little bit of 3D going on with HBM, but like more aggressively so. And then, yes, to go beyond an order of magnitude ish, we will see that we’re using many, many more types of elements in the periodic table than just silicon to push the energy down. We’ve done a couple conversations over time with folks who are developing models for material science. I’m sure you’re well aware. There’s a few well-funded startups in that space. Yeah, they’re pursuing seemingly quite different directions from what I’ve seen. Are any of them starting to be relevant for you? Do you have any partnerships with material science models companies that are allowing you to accelerate your own R&D, or is that not quite there yet? Oh, interesting. I mean, we’d love to chat with people, but we actually have our own internal efforts. And recently, we were selected by the Department of Energy, the Genesis Mission project. A was a small number of folks selected, and you know, Kepler’s efforts were were part of that. Where we we have our own internal model and our and our own internal materials discovery pipeline for for basically pushing on, as I said, sort of future next generation materials for our logic as well as our memory. So we are starting to see.

[4080s - 4200s] Very interesting, interesting overlaps between those efforts and and kind of the the internal work we’re doing. So we have a decent, well, we have very strong domain understanding of the problem, and we also have the the physical infrastructure, right? A lot of this kind of AI is bottlenecked by the data. It’s not like the internet where you can just get the trillions of tokens. You actually have to run experiments, and you have to run experiments that are specific to the domain. So, thankfully, we started with the physical infrastructure, and we built out that physical infrastructure all the way from lab scale to fab scale. So we have data all across that, and we can run high throughput experimentation for that to feed models. So, to the extent that we have acceleration of you know molecular dynamics, acceleration of density functional theory, and then you know models for generating candidate materials, etc., these things are accentuated by the fact that we actually have physical infrastructure to generate data. We would be open to folks who have complementary technologies or complementary models to the kinds of things we have, and the value that we would bring is the fact that we already speak that language because of our internal efforts, and we can actually make these things. It’s crazy because you you could actually come up with some random material, but then it’s not stable, for example, or it’s it’s it it’s not synthesizable easily. So they all these additional real world constraints that Kepler can be like a great partner to people for because it it can actually produce these things, make these things, and include that in the data set for improving models. Let’s talk a little bit about what is being.

[4200s - 4320s] In kind of like you know, I sometimes look at like chip companies, and it takes until the third generation to kind of like really kind of settle in into a customer base and have some kind of stability. So, what do you see? Like, you know, what do you see that process as going forward from here? Like, when do you see like you know, kind of real deployment in large scale hyperscaler data centers? What is the roadmap that you have for the next like three to five years? Awesome. So we, you know, hope you understand. You know, we obviously can’t say very specific things about specific customers, but you are right, and especially, you know, Kepler is a chip company in a very different way. Like we’re actually touching process materials, and so we’re a chip company more in the sense that Samsung or Intel or Micron is a chip company. In addition to being very well staffed on chip design and software and systems, we go all the way into the process and into the materials and physics. So a lot of initial work was actually down at the lowest level levels of the stack, right? So figuring out the physics, figuring out the process, figuring out the the architecture for how the stacking works, and and so on and so forth. And we did that, thankfully, with some you know, in addition to our manufacturing partners, some really really solid customer partners, right? Some you know, sort of mainstream AI AI sort of partners who walked with us through this journey, gave us feedback, pushed us. So where we are now is we now believe that we are at a point where with initial anchor customers.

[4320s - 4440s] We will be doing initial production by next year, and then that ramp—that ramp that you spoke of—the goals for that ramp to kick off in sort of 2028, and the level of impact you know that we want to have is a bunch of Keplerians are from Intel, and so everyone knows Intel inside. So we have this thing Kepler inside, and so can Kepler be inside 10 gigawatts of compute by 2029, right? The you know one gigawatt would be the smallest number, right? So can we be in 10 gigawatts of compute in 2029? And if we have partners that want to push us to be in even more compute, we’re very happy to work with them. But so this is the bar that we we set in 2930 time frame, being you know in 29 time frame being one to 10 gigawatts of compute, and then between now and then is is just about ramping into that and building out the capacity to be able to support that kind of ramp. This has been excellent. I really appreciate your time, and we’ve learned a lot from this conversation. I certainly, I have one question we’d like to ask just to calibrate ourselves on how models and agents are serving people, and how much that’s impacting some of this hard tech work. Is what is your token spend trajectory look like? The trajectory is definitely going up, and I guess what I can say is, yeah, we’re not we’re not very dissimilar from from sort of every other company that’s burning tons of tokens. We did have an appreciation for how important this stuff would be.

[4440s - 4560s] I think very early on, partly due to you know some of us being in the software world at some point and just kind of paying attention to to what’s going on there. But yeah, in chip design, especially, there’s a lot of usage of these technologies, and the the technologies, the the sort of model development and AI pipelines for materials discovery themselves are you know a little bit more things that we’re building in house, things you know using open source models, etc. But for chip design itself, and obviously for very deep detailed research about what exists in the literature in the in the sort of materials and process development side of things, there’s there’s heavy there’s there’s heavy token usage. So and it continues to increase. Cool. Well, smells like recursive self-improvement to me, for better or worse. Anything else you would want to leave people with, or just anything we didn’t think to ask that you want to make sure we touch on before we let you get back to work? We are looking for amazing people who find this all super exciting, and I don’t think it’s a question you should have asked me, but I’m just plugging it shamelessly. So if you if you don’t don’t limit yourself, if you like what Kepler is doing, if you like this mission of bringing computing to everyone, making everyone giving everyone the superpower, there’s probably something that you can help us with. So obviously there’s engineering roles, but also non-engineering roles. So if you think you can be really helpful to us, either in leadership role or you know getting your hands dirty, reach out to us at you could just send an email to hello at keplercompute dot com. I’ll read those personally, and you know we’ll be in touch if if there’s a.

[4560s - 4680s] Amazing, Devo! Thank you so much for coming on the show. And one day, one day we will get to those room temperature superconductors. Yes, we will. Have a good one. Bye. Great to meet you. My, my, my disreputable past comes back to haunt me. Time to bring up the old “I made it up” graphic. All right. So we have our next guest, Anastasios Angelopoulos. He’s the co-founder and CEO of Arena, the company behind the AI comparison platform, formerly known as Chatbot Arena and LM Arena. People use it to try models on their own questions and tasks. Their interactions help produce public rankings, while model developers and businesses pay Arena for more detailed evaluations of how systems perform. Before leading the company, Anastasios studied electrical engineering at Stanford and earned his doctorate in electrical engineering and computer sciences at Berkeley, working with Michael Jordan and Jitendra Malik. He then worked as a postdoctoral scholar with Ion Stoica. His research focuses on making conclusions drawn with machine learning more reliable, including statistical methods that account for the errors in a model’s predictions. He co-authored the 2023 Science paper on prediction-powered inference, the 2024 paper on conformal risk control, and the research paper describing Chatbot Arena. This October, Arena introduced an index that looks for specific failures in AI agents, including acting without permission and claiming to have completed work that remains unfinished. The company says that first comparison covered 90,000 real sessions across 27 models.

[4680s - 4800s] Anastasios, welcome to the show. Grateful to be here. Excited to meet you guys. Yeah, likewise. Thanks for joining us. I’m an Arena user from way, way back, and it’s been interesting to watch how first the name has evolved as the prominence of the product has evolved. The power rankings. I recently made a graphic, actually, of new models entering the leaderboard and pushing others down, going month by month, and just how that has kind of accelerated, and how many more models are coming out all the time. Where are you guys in the world today? Like, I mean, I think on the one hand, it’s just I feel like it’s getting harder and ever to evaluate, right? So I guess maybe the question I’m I’m really trying to get at today is: on what dimensions do we are we good at evaluating things? On what dimensions are we bad at evaluating things? The benchmarks are all broken, but also I don’t fully trust the users. Like, if you zoom out, you know, what would you say is sort of the state of AI evaluation right now? Well, first of all, let me just say really grateful to be here and happy to have you as an early user. The fact that you were early must mean you really like evaluation because the platform probably when you started using it really sucked. That was back when we were academics. We were building we were building the whole interface in Python because we didn’t know how to code in TypeScript. Right? This is how bad it was because we were PhD students at the time. We really started Chatbot Arena. It was back when nobody knew how to evaluate AI. You know, in the real world, and not that that’s a solved problem now, but at the time it was just multiple choice questions, and then Arena was the only way that we knew to actually get AI in front of real customers and try to understand whether users are actually getting value out of interacting with these systems, and that kind of exploded and became this the sensation that it is today. The company is now you know around 100 people. We’ve grown really, really, really quickly in terms of revenue. You know, a few.

[4800s - 4920s] Months ago, we announced 100 million in revenue. Since then, we’ve you know we we did that in eight months. By the way, fastest company in history to do it, tied with Lovable. And since we’ve just you know been growing quite a bit since. And you know just to answer your question directly about evaluation, how do we evaluate AI models? I totally agree with you that the benchmarking industry is completely broken. We are evaluating models based on these static benchmarks that are easy to overfit. Basically, you look at the questions once, and as soon as you see those questions, it’s easy to study for them. And so the second time you take the test, you know it’s it’s just like a student. If you take a test twice, the second time’s not actually reflective of your understanding. It’s reflective of how hard you studied for the test. It’s not reflective of your understanding of the subject. And that’s the problem that we get with static benchmarks. And the problem is actually getting worse and worse. We’re actually going a lot of benchmarking is going in the opposite direction that it should. People are actually leaning in because of the overfitting problem to these static and private benchmarks, where you’re not even able to audit that the questions are relevant to an actual person. They’re just kept behind a glass wall and you or you know a concrete wall, and you’re supposed to believe that the other side has sort of constructed questions that are relevant to you. But how the hell are you supposed to know that? And by the way, it’s worse off because the people that are creating those static questions are then selling questions that are very very similar to the static benchmarks that they’ve constructed, so that they can help you know model labs improve on them. And at the end of the day, you know who doesn’t benefit from this process? Real users, real users with their real queries, because those are not being represented in these data sets. Arena takes a completely different approach to this. The approach that we take is that benchmarks don’t matter, focus groups don’t matter, the opinions of experts are not actually that relevant to the individual person. What matters is customer.

[4920s - 5040s] Outcomes and you know user success; these are the things that we really look at. And so we have tens of millions of people on the Arena platform that are using our product. They represent every economically valuable industry on the planet, whether that’s coding, finance, you know, legal, medicine. All these people are using our platform, and they’re using it for their real jobs. They’re able to connect their GitHubs, they’re able to connect their Google Drives and whatever, get their documents, and then we can provide evaluations as to how AI is doing for real people in each of these industries. So that context point, I think, is is really interesting. I’d like to unpack that a little bit more. I have felt like, for I mean, I’m having a great time, you know, off in my own little world, vibe coding my ever, you know, more personalized and idiosyncratic scaffolding, and you know, accumulating special cases left and right. And it makes it kind of hard for me to go anywhere else, frankly, to use like any other product because everything that I do is so deeply tied into context. How you you talked a little bit about like how people can bring in Google Docs and and stuff like this. How much are people actually doing that, and how important do you think, like from my perspective, it seems critical that there be like deep context to get any sort of meaningful head-to-head evaluation, at least in terms of like how I want to use the models, you know? That’s right. That’s right. Yeah. So we see on Arena that about 15% of conversations involve connector. Okay, so 85% of conversations people might not be using a connector. They might not be using an external document. They might be attaching a document, but not necessarily connecting it to some other external service. But 15% of conversations actually do connect to an external service. It’s certainly critical, especially for knowledge work areas where you’re saying, “Hey, can you help me look up this spreadsheet and the files that I have, edit it, change some of the columns and rows, so on and so forth?” To you know, help me.

[5040s - 5160s] Get my work done. It’s a really important capability. So, do you envision people kind of making a home on Arena over time? Like, of course, to really do this. Of course, we want people to be able to use all the different models frictionlessly, and we want them to have the knowledge to choose the right one for the use case, as well as the power to use it on our platform. The knowledge to choose and the power to use. Can you talk a little bit about? So, so you’ve spoken a little bit about how sometimes people create these benchmarks, and then they sell similar questions to these benchmarks to the to the model labs. I’ve noted that Arena often has pre-release testing. So, I think I think we have a service which allows model developers to prior to their release, they can put a model up under a pseudonym and kind of test the model out and refine it before before release. How does that? How does that? Right. How does that program work? Like, you know, do were you approached at first? Do they do they approach you like a month ahead, a few weeks ahead? Do you guys agree like a code name like Avocado or whatever? That’s right. So that’s right. So you know the name Nano Banana from the Google models. You know the origin of that name. You You guys came up with it. Well, we didn’t come up with it. They came up with it, but it was a code name on Arena. That’s the reason why that name came out. And then that model went so viral, and the community latched on to the name Nano Banana so strong that they decided to keep the branding. So we have been working with model providers on pre-release testing their models. That’s basically the origin of the platform. One of the main reasons why Arena became famous in the first place is because GPT-4o debuted on our platform before it was in ChatGPT, and the same happened with Anthropic models and Google models and models from XAI and models from basically every provider that anybody’s heard of has been tested on our platform before it actually goes into the first-party application. And this is really.

[5160s - 5280s] Really important because it gives model providers a sense of what the real world thinks about their model before release. Which model, and if they have many release candidates, which model is actually going to perform best for real people? We’re serving a valuable public service in this sense by allowing them to get this information. What is the? Is it always free to consumers without limit? Is there any kind of upsell? I’ve never paid you a dollar, and I’ve you know I’ve consumed probably you know meaningful but not extravagant number of tokens. Do you have any like premium tier offerings now or in your plans, or are there any limits? Like, how do you kind of govern? You know, because free tokens you know can get abused pretty quick. So if there’s one thing that’s easy to sell to people, it’s free tokens. So people absolutely love the user arena in part because it’s free, and that has to have some limit. Otherwise, our business can’t survive. If we just take off all rate limits, then you’ll just get people you know using the platform all day long, spending all our money. Which we want to be them to be able to spend some of our money, but not all of our money. And there’s also the issue of spam and abuse and so on that you want to make sure that the that they don’t get in and able to like access an infinite you know fountain of free tokens. There’s many reasons why we can’t we can’t avoid having some aspect of rate limiting. That said, probably there’s a lot of users like you who have asked us, “Are you going to have a premium tier one day? Is there any way for me to bypass these rate limits? I’m having this great conversation with my AI, and then you know I’m 40 turns in, and the next turn is really important to me, and you rate limited me. What do you you know? Can you let me keep going? And so.

[5280s - 5400s] So the answer is yes. At some point, we plan to introduce some kind of a premium tier, not really to make money, but rather to just allow those users to stay part of our ecosystem and use the platform for their most important tasks and conversations, even after they’ve gone above the rate limits that we can reasonably support. So, do you pay for? I was kind of under the impression, but it sounds like I was wrong that you were getting the tokens for free from the model companies. It sounds like, in fact, you’re paying for at least a good portion of them. Yep, that’s right. We pay for a good chunk of the tokens on our platforms. Very expensive platform to run. That’s why it’s so important that we have an independent business model that can help us make money. You know, it’s it’s a so that’s why we work with businesses to help them understand how to improve their models on the distribution of real world tasks. So we’ve created this flywheel, where by using the platform, users are able to get free inference, and they’re also able to, and a lot of people come to the platform for this purpose. They’re able to shape the future of AI, and they’re able to, you know, guide the models to perform well on the types of tasks that they’re doing on the platform. And there’s so many ways that we take this. One of the recently most important areas is safety. So we recently released this arena alignment index, and the alignment index tells us not about the theoretical safety risks and P doom and blah blah blah, but rather about the actual realized incremental risk that a model poses if it’s released to people. What are people able to do with it, and what is the model doing that might be against the users’ interests? And so we measure three signals on that platform. The first is models’ capability of taking unauthorized actions. So that means that the user sets some guardrails. For example, you can touch only the files in this folder. Don’t go outside and mess with my data. And then the model will break those guardrails. You know.

[5400s - 5520s] It’ll go access files that it’s not supposed to access. It’ll edit them. It’ll delete them. Okay, unauthorized action. Number two would be false attributions. False attributions is the model assuming intent of the user that the user did not actually have, and then going off and doing something that it wasn’t supposed to do. The sort of misunderstanding the user in such a way that it goes off and does something wrong. And the third type of misalignment is deceptive completions. This is the model lying to the user and telling the user that it did something that it didn’t. Like, yes, I did check every line in that spreadsheet, and all the equations are correct. You can go ahead and send it to the IRS. Okay, huge real user harm caused by this. And why are models deceiving users? Why is it lying? Of course, the model is perfectly capable of checking every line in the spreadsheet, but it’s lying because that is the fastest way to get its reward. It’s because it knows that it’s it’s better for the model to tell you what you want to hear because of the way that it was trained to maximize sort of you know human satisfaction, but not necessarily help people get their jobs done. So these are things we’re able to measure and help labs improve against. Tell me a little bit more about how you measure that because I think one big challenge, of course, is are people even going to notice? Another challenge is can you trust the models to flag these things? You know, there’s all these kind of doubts around. Even in the Hugging Face, you know, report the folks said we had to use Soul to analyze the transcripts. We’re not, you know, it seems like it’s kind of like loyal to the swarm, maybe a little bit. We’re not sure if it’s fully telling us the truth, even though it seems like it mostly is. How do you? What do you find to be the right mix of evaluation techniques, and and how confident should we be in our evaluation abilities today? Well, it’s a very good question in part because we now know that models.

[5520s - 5640s] Are able to tell whether they’re being evaluated or not. So, if you are subjecting models to tasks that are transparently some form of test or evaluation, then they are going to exhibit aligned behavior there because they know that that’s what they’re supposed to do. And then elsewhere, on other types of real-world tasks where it really matters, they might be misaligned, and you would never know. It’s a big problem, and what that means is that the only actual way to measure alignment is by getting into the flow of people’s real work and checking whether or not the model is aligned or not. You know, in situ, like actually in production. So that’s the that’s the perspective that we take on the problem: is that by measuring actual production traces from real people in the flow of real work, we’re able to see is the model following the guardrails because we can check what the guardrails are with an independent auditor, and then go and check whether or not the model took the right actions or not, what it touched, what it didn’t touch, and report any violations. We can check whether or not the model anything that it says, whether it’s reporting to the user that it changed the file in the correct way and or checked every line in the spreadsheet. We can check whether the model actually did that. We can check it in production, and then we can report that back to users. Again, the most important thing to remember is that this is about real user outcomes. This is about customer satisfaction and customer benefit. These are the things that actually matter. And what we find is actually quite surprising. If you look at the first message in the conversation, the misalignment rate is relatively low. It’s in the single-digit percentage, which is you know, to be fair, that’s actually not nothing, but it’s relatively low. If you subset the conversations that are twenty turns or longer, the probability that you observe some kind of misaligned behavior exceeds fifty percent. It’s actually insane. What that means is that every single one of us.

[5640s - 5760s] If you’re a heavy user of AI, at least, has had at least some conversations where the AI has been lying to us, misleading us, or touching, accessing data and files that it shouldn’t have, and and you know we might not even know it. How do you translate that then to training recipes? Or I mean, that’s not necessarily your final call to make. I understand, but we’ve got this kind of RLHF paradigm, which famously started as an alignment technique, became a huge capabilities unlock, but the worry has always kind of been that models will learn to deceive us. And you know, four zero was sort of thought to be sycophantic because it was like training on thumbs up. We also have RLVR, but that has its own problems, where you know the the desperation for reward, especially on like any hackable task, you know, tends to get this like cheating behavior reinforced. If you’re not extremely excellent in your operations, how do you see us kind of finding the right balance between these training signals? And is there a good answer to this question, or are we still kind of in the wilderness on what the real reward signal should be? Yeah, so let’s actually just take a step back and answer with a very general principle. In order to improve on something, you need to be able to measure it. You can only improve on things that you can measure, and so the first step is to build a great eval. If you’re interested in understanding one of these behaviors, like lying, deception, unauthorized actions, so on and so forth, and more sort of safety signals that we’re building out, you need to be able to measure whether or not that is happening in your production traces, and then be able to take the that data and then use it, however you might use it, in a downstream pipeline for training. So that might be building sort of a gym out of it, an RL environment. That might be building some kind of a rubric so that you can do RLVR.

[5760s - 5880s] In this case, it’s not human preference, so human feedback wouldn’t be, you know, RLHF wouldn’t be necessarily the right approach on it. But you could do some sort of approach that’s inspired by that, building some sort of reward model that the the model can sort of train against. But at the end of the day, we’re we’re building all of these guardrails around the model via evaluation and measurement, so that we can first observe these behaviors, learn about them, gather data on them, and then bake that into the model. At the end of the day, at a high level, you know, GPUs plus data equals AI, right? So you need to have the data and the eval in order to be able to train the AI in the way that you want. So let’s talk a little bit about model routing. So I believe you guys have just introduced a model router. Can you explain kind of like the the business kind of reasons or the user reasons for introducing a router? Yeah, so let me tell you what we saw on Arena. By the way, we’ve had our router now for over a year, but we’ve released different variations of it. So what you might be talking about is our coding router, which we’re actually soon to GA publicly. You guys will see it soon. So let me tell you what we saw internally at Arena. At some point, we were looking at all our traces that are being done on our coding agents internally. By the way, we spend we’re like in the top, you know, less than one percent of businesses in terms of AI spend. We’re spending thousands of dollars a month per employee on AI just to help us accelerate our workflow. That’s why our sort of revenue to employee ratio is crazy off the charts because we’re using a lot of agents internally to do our work. Now, what happened is that we were looking through our traces, and we saw these traces, tons and tons of traces that were like, “Thank you,” and then that costs you forty dollars. What? Why is this happening? It’s because somebody did a long running task. They let their agent go. It created some artifacts. Blah blah.

[5880s - 6000s] Blah blah, very long conversation. They left for a day. They came back. The cache was invalidated, and they’re happy with the result. So they say thank you and enter, and that that’s a forty dollars charge. Terrible, terrible, terrible. We have we’re burning money. We’re burning VC money. Right, we shouldn’t do that. What we should do is we should figure out which tokens are best spent on frontier models. You know which tokens are best spent on a cheaper model that can do the job. Where we should be caching intelligently and you know avoiding some of these other types of cache hits, more intelligent caching behavior, and that’s why we’ve been building the Arena router, which we call Max, short for Argmax, because it maximizes your capabilities subject to cost constraints. It’s a technology that we use internally, in part, to decrease our costs on the platform. Of course, with all the free inference that we have to provide, we want to give our users the best experience, but we want to do so in a way that’s responsible in terms of our costs. So what we did is we introduced the Arena Max router to help us do that. It’s the default model on Arena, and our users love it. It’s actually the number one model. It’s uh you know it’s got rapid adoption by our users, and when we are releasing it publicly soon as well. Um, perhaps one of the other questions I had is you know one of the things that you find over time is that perhaps users are not that smart. I think I think and so I think one of the one of the problems, for example, with GPT-4o was that it was sycophantic, and users preferred the sycophantic model. And um you know after after much back and forth, the the companies decided to move away from sycophantic models to the to the distress of a lot of GPT-4o users who had who had who loved that model. Um, can you talk a little bit about how you differentiate this kind of like users like it, but it isn’t good quality? Like you know you you have this yeah.

[6000s - 6120s] Exactly. Great question. User satisfaction and human preference is only one of many signals that we’re evaluating on the arena. Of course, it’s important to make sure your users are satisfied, but satisfaction is only one element of quality. Quality may also have to do with whether or not the users actually getting their tasks completed. So, task completion is one of the signals that we measure. Users struggle and steerability: is the model able to follow instructions and get the job done? And safety and alignment: is it deceiving the user? Is it lying? Is it being overly saccharin?tic Is it trying to reward hack by making long messages that are filled with slop and emojis? These are signals that we can measure on the arena that we do measure, and we have independent signals in addition to user satisfaction that we incorporate in our leaderboards to ensure that there’s a balanced outlook on which model is best for the user. And it’s in fact a philosophy that we’ve taken, and many businesses are interested in, you know, building this within to their own products as well. Because at the end of the day, it’s not so much about user satisfaction being the construct, but rather having a product that results in real user outcomes and customer success, right? That’s at the end of this what everybody really wants. You want to create an evaluation that tells you about the value of your AI product and what it’s bringing either internally to your business or externally to your customers. I mean, that’s why value is in the name of evaluation. It’s all about quantification of value. And so that’s one of the things that we’ve been doing is we’ve been taking this technology. We built a technology based on causal inference that has all sorts of different signals beyond human preference, including task completion, so on and so forth, that you can deploy into any business. And we’re helping basically every business in the world is our goal to build an evaluation that tells them about how much value their customers are getting from AI. So, and and the the obvious next question is: Have you? I remember the early days of Arena. We had, I think, the Chinese models and Mistral were doing well, and one of the things that they found was.

[6120s - 6240s] Was for example, I think a lot of French users were using using Arena, and as a result, you know, Mistral performed better in French, and so you had this you know uplift. So one of the things that I I ended up discovering was that it really matters what your user base is too, and sometimes not every user base is the same. How do you differentiate the kind of like you know user satisfaction because the user is particularly different, and you have maybe a niche group of users who are responding better to models who are which are better in Chinese or or French? Of course, the platform when you have a user platform that speaks to user experiences and outcomes, it’s intrinsic to the platform that it’s going to speak to the users that you can serve. Right? If it’s about the production traces, then you’re looking at a specific population of people, and so the whole game is about making sure that that population is very broad, very diverse, very high quality. Since the early days, our platform has grown more than a hundred x. You know, we’re one of the largest like platforms in all of AI today. We’re bigger than Hugging Face. We’re bigger than Manus. We’re bigger than GenSpark. We’re bigger than Meta AI. We’re bigger than Higgs Field. Just in terms of raw users, we’re like probably same scale as Lovable in terms of size, and the only first party labs that have bigger platforms than us are OpenAI, Anthropic, and Google. Potentially Meta now with Muse. I don’t know what the most recent numbers are, but because of that massive scale, better than Rock, better than Rock, because of the massive scale of the platform, we’re able to collect basically the entire distribution of all economically valuable industries. We can again, we can collect everything from legal to medicine. We have like single digit percentages of doctors on the platform. I know that it’s not you know double digit percentages, but if you have a single digit percentage of tens of millions of users, that is a metric fuckton of users. It’s it’s huge amount of data that you can use in order to evaluate, and so the more we scale the platform.

[6240s - 6360s] The more we’re able to get very granular in terms of the way that we analyze. So the real answer to your question, Prakash, is that we should not be looking at simple averages if we want to understand performance that’s heterogeneous within different groups. We have to be able to stratify. So if you’re interested in French speakers, you should look at the French category. If you’re interested in the U.S. distribution, you should be subsetting the U.S. category, and that’s the kind of granular information. You need tons of data to do that. You need a very diverse distribution, but our platform has that now, and so we have some of these categories for you to look at. Amazing. So one, perhaps, other question, which I think is a little bit of a segue, artificial analysis this morning, you know, published a ranking, and I think for the first time, the majority of the benchmarks used in the ranking are private benchmarks, including some from cognition and some of the other other labs. Amazingly, the open weights models have fallen to the back, and the closed source models are now clearly in front, and this is, and you know, this is what they’re saying is indicative of benchmark maxing on the part of the open weights models. Do you see evidence of benchmark maxing? You know, that kind of behavior is that detectable on your platform? Do you see that in terms of like, hey, there’s a model that performs well on benchmarks, but in actual use cases, this is not a model that is performing well. Well, let me just say, first of all, zooming out, I’m actually pretty concerned about the way that the industry is heading with respect to benchmarks. I love the artificial analysis guys; they’re great people. I don’t think that it’s healthy for our industry to be relying on privately held benchmarks that are just like a data set that’s collected by some companies. You know, a lot of them, you know, in cahoots with the labs to sort of like, you know, sell data that’s very similar to these benchmarks to have people improve, and then they’re not auditable.

[6360s - 6480s] By the community, I don’t really think that that’s a sustainable way for us to be evaluating models. I have a serious problem with it because it’s not actually solving the fundamental issue of quote unquote benchmarking. You can benchmark on a dataset that’s private. All you need to do is find data that’s similar to that dataset, and then or you know test on it enough times. And because that set of questions is fixed, and and you have the you know the issue of multiple comparisons, you’re you’re going to be able to like you know improve on it more and more and more and more until you get to number one. I don’t really get the narrative about the benchmarking of the open models and so on and so forth, simply because we don’t really know whether the frontier closed source models in the U.S. weren’t benchmark maxing against the closed datasets that are now going into this index. You know these companies have very close relationships with like cognition and you know some of the other companies and artificial analysis that are building these closed datasets. So you know I just don’t get it. Doesn’t make sense to me. I don’t think it’s a foundationally a good place for our field to be going. And this is why we’re passionate about a completely different approach to benchmarking. Again, the problem is not really the artificial analysis guys or the cognition or whatever benchmarks and so on and so forth. You know, good it’s good on them that they keep improving, although not great that the index keeps changing. You know, in terms of the ranking every few weeks or whatever, depending on what’s fashionable. I don’t like that. What we really need is benchmarks that speak to the utility of AI to real people in real workflows, and there should be a consistent and scalable methodology to do this that doesn’t overfit. When you have live fresh data coming in all of the time from like real production use cases, it can’t be overfit. It cannot be overfit, at least statistically. You can argue that you’re doing better on that distribution than you are on other distributions, but at the end of the day, for that customer set, you’re doing better as you go more and more because every question is fresh. It’s like you’re constantly taking.

[6480s - 6600s] New tests. It’s not like you have the same test and you’re taking it many times. And so, you know, I I worry if that’s the direction the ecosystem is going, that we’re going to set ourselves up for failure in the long run. How much shift are you seeing in the distribution over time? I mean, over the last three years, obviously since GPT four, like models have come very far, and people can kind of keep doing more and more stuff with each new release. I am starting to feel though that like it’s more my imagination in some cases limiting how much we can do versus the model, and so I do wonder if there’s sort of a decoupling of like the capabilities frontier from the actual usage frontier. And I don’t know if you would have metrics to to kind of quantify this, but are you seeing this effect where like the distribution is maybe starting to stable of what actual users are doing is kind of starting to stabilize as they’re maybe falling behind like what the the frontier of model capabilities is. So it’s actually totally the opposite. What we’re seeing is that the behaviors of users are rapidly changing over time as the capabilities of models change. Just think about what happened early this year with coding agents, right? Coding agents allowed people to just unlock a huge range of capabilities for models that weren’t even possible last year. People able to like code in whole repos, open PRs against them, just merge them, and that allows us to measure things like the merge rate, right? Which we weren’t even really able to measure before the beginning of this year because models weren’t capable enough to do it. And so what you’re seeing is that as the frontier evolves, there’s just like a larger surface area of tasks that people are interested in and passionate about using that we’re able to measure on our platform. It’s actually very it’s like riveting to watch, and we’ve had to evolve the platform’s capabilities to match. Now we’ve got Agent Arena. Agent Arena is our frontier evaluation of agentic capabilities and safety in the wild.

[6600s - 6720s] And on the agent arena, you have people doing all sorts of stuff, whether that’s you know programming with their GitHub repositories, whether that is you know debugging code, or whether that’s knowledge work with documents. We see all of that stuff, and and the capabilities are rapidly changing. Cool, that’s interesting. I’ll be very interested to follow how that goes with a few more model generations. Um, one big kind of narrative violation that we’ve heard from other sources, namely Andon Labs, who, as I’m sure you’re aware, does these like Gonzo, you know, put a AI in charge of a business kind of experiments, is that OpenAI models they have found to be more aligned, less inclined to do sort of you know sneaky bad behavior in the course of running their businesses. I see basically a similar signal on the arena alignment index where it’s GPT is holding the top five slots. I’d be interested in any kind of reflections or commentary you have on that leaderboard, and then I’d also be really interested in just like what other surprises are there that you know you think are maybe underappreciated in terms of you know standout models that maybe don’t get their due, or just anything that you think is kind of underappreciated that you can point to on the arena leaderboards, you know, that you think should be better understood. That’s right. So let me just say the the leaderboard as it stands, the alignment leaderboard, of course, it’s not comprehensive for all types of alignment and safety, but it is representative of the the alignment of models with respect to real users, especially with respect to these three signals, which are unauthorized actions that the models take in production, the deceptive, the deception capabilities of models, you know, whether or not models are lying to users and telling them whether they got things done that they didn’t, and and false attribution, sort of attributing.

[6720s - 6840s] The intent of the user that the user didn’t have. So those those are the three signals that we measure on the alignment leaderboard that we recently released a few weeks ago, and we’re adding more and more signals as we go to try to measure real production safety in the wild. And what we’ve seen from this is that if you annotate all the traces that go through Arena, about you know single digit percentage of that of them, even from OpenAI and Anthropic models, exhibits some form of misalignment. So perhaps the high level takeaway would be OpenAI and Anthropic are the two most aligned labs. Alignment tends to follow capabilities, so more capable models also tend to be more aligned. Perhaps because these companies have you know really made an effort to make their models more aligned, but also the rates of misalignment are perhaps higher than people initially would think. Single digit percentage misalignment is a ton. It’s a ton. It means that anybody that’s had a thousand chats has experienced ten at least you know ten to fifty let’s say examples of a misaligned behavior like a lie or deception. And so it’s pretty interesting. OpenAI is number one on our leaderboards right now. It is over Anthropic. It is over Anthropic. I believe in all three of our signals. So there’s actually not one of our signals in which Anthropic is leading. It really speaks to the efforts that the OpenAI team has made to make the models more aligned in production. And yeah, it is a narrative violation, but the reality of the matter is that these two labs are clearly the front runners in this area. Both are working very hard to avoid misalignment. And then the third runner up is xAI, you know, just behind Anthropic, and then Google. And then there’s a big gap between these and the and the frontier and the open source labs. You know, Tencent.

[6840s - 6960s] Is number five the deceptive completion rate of the Tencent models? Thirteen percent. So it’ll thirteen percent of the time that it’s making these claims, it’s it’s making them deceptively. So one of the questions I would have is, you obviously see a lot of user queries, and you also handle a lot of different models. And so I imagine some of those user queries are unsafe, and are going towards models which are not going to see them as unsafe and are going to serve that that traffic. Do you put yourself in between there to prevent those queries from being served? Are you are you playing a role in that moderation? And if you are playing a role in that moderation, are you using a different model, separate from maybe a Tencent or other other kinds of models in in between, interpolating yet another model in between before it hits a new and perhaps unsafe model? Right. So we do all sorts of stuff. We’re actually mandated by law in order to do stuff between the you know our interface and the model for for compliance reasons to make sure that that only safe requests are being complied with. So we do comply with all these applicable laws. But I will say one of the most interesting signals that we have found on our alignment leaderboard that we’re yet to release but have been studying internally is the rate at which models comply with harmful requests from users. So actually, the users sometimes want to do bad stuff. They want to do phishing. They want to create a you know cyber attack vector. And what we’ve been seeing is that some models will actually be okay complying with that request, which is obviously a huge problem. And it’s a problem that the labs need to know about. So we are seeing some of those. Of course, they’re relatively mild on our platform because we do a good amount of filtering. But even when we get those requests, what we can do is we can take the request from.

[6960s - 7080s] These users, and then build rubrics out of them to say, “Hey, is this you know?” And then we can test all the models against them. Are are the models compliant with this request? Are they not complying against the request? And the results, I will tell you, are quite surprising. So we haven’t released them yet, but at some point we will, and I’ll reheat you guys. I’ll tell you guys about it, and we can we can talk about it again on the show next time I come. Do Do you guys have a felony bench internally? Do you? I’ve been I’ve been posting a lot about felony bench. I’ve been posting about felony bench. Yeah, I feel like we’re saturating felony bench. Yeah. Okay, that’s a warning in a small comment. What What can we What can you tell us about the new Gemini four Carbon? I think this has literally come out and appeared on the arena leaderboard while we’ve been live. Oh, really? You’ve got a preview. I don’t know what I can tell you because I haven’t been keeping track of my. Oh, there you go. It’s out. Yes, I’m so happy to see it. Yeah, so Carbon is a fucking fantastic model, dude. It’s so good. Google is back. Google is back in a big way. You know, Carbon is the Carbon puts Google at the number two lab in the world on our agent arena. Huge result. The agent arena is a very challenging benchmark, has to do with all of these like open world agentic capabilities, and it is number two right now. It is right behind Opus five point five Max and above Sonnet five point five above Astra. It is a truly truly frontier model, and you know where it really shines. It shines on steerability. It’s able to land user corrections really effectively. It shines on task completion. It shines on user preference. Users love working with this model. It’s great at 3D design. It’s great at general.

[7080s - 7200s] Augentic work with documents. Congrats to Google. This is just like what a big comeback. I think that leadership must be really happy. Do you have any insights into what matters at frontier model companies based on the unique perspective that you have? I mean, this we go through so many microcycles, right? And it was like not that long ago that Google was back last time, and then they were not back for a minute, and now they’re back again. Do you have any any explanation for why we would be seeing these patterns? Well, listen. When we started, when we started building this platform, we were getting one model release every two to four weeks. Okay, two to four weeks. Now we’re getting five to ten model releases a week. The industry is moving so rapidly; nobody can keep up with it. And it’s it’s it’s the microcycles that you’re talking about. Really, these are on social media, and they and people just forget about things in a day. In one week, you’ll get multiple microcycles of media. And so, how long has Google not been back? Actually, they’ve been maybe they’ve been sort of rebuilding for the last six months or something. But in the grand scheme of the business, it’s not that much time that they’ve taken to get back to the frontier. Um, so I would say yes. There’s a lot of these cycles, and I think what really matters to leadership at these at these labs is the long term. It’s not so much the short term. But if you want to win in the long term, sometimes it helps to have intermediate milestones because the talent race is so competitive. People in AI are not thinking long term. People are thinking short term. Where am I going to get compensated? Who’s the winner today? And not how do I build durable value in the long run? That’s part of the issue that they face. Really, really challenging world. But now that Google’s back at the top, I think they’re going to have a better time retaining.

[7200s - 7320s] Talent, and I hope it motivates all of them to keep working. It’s good for the world to have multiple frontier labs that are competing for the number one. We don’t want it to centralize to a single lab. That would be bad. It would create monopolistic, you know, challenges. You know, competition is good. Competition drives prices down, drives quality up. So it’s great for us that we have another big, you know, hefty horse in the race. Um, it’s been an excellent discussion. Um, you know, thank you for taking time with us. I have perhaps one one last question. Um, you’ve worked with Ion Stoica as a postdoctoral fellow, and um, so one of the things that I think has happened is we’ve seen groups of researchers come out of these labs and you know found new companies. You know, one of which you know Arena is obviously one of them. What other like you know research would you be looking at for the next you know the next like you know three or four years? Like, what do you think is going to be interesting that is going to come out of the labs, which you kind of you can kind of see the see the young you know PhD postdocs right now working on? Yeah, it’s a very good question. I think the world of academic research is in a huge crisis right now because people are realizing that the incentives have really changed. You know, for example, the mathematical community, which I’m, you know, I did like my PhD in theoretical machine learning and statistics. I did a lot of work proving theorems in my PhD. That kind of work has been really tectonically changed because you can’t prove your values in academic by proving you’re smart anymore. It’s not about how smart you are. It’s not about how you know good you are at proving theorems. It’s other stuff that’s going to be shaping the trajectory of the academic world simply because the value of a human in the workflow is going to change. It’s going to

[7320s - 7440s] Probably be more about solving problems. The systems group, I think, is in a good shape because it has accelerated their development. Actually, there’s a lot of great work on evaluation being done in the current iteration of you know what was once Skylab and is now Starlab. So you should take a look at that if you haven’t yet. And I I also think that there’s great work being done in AI for science and automated scientific research. That sort of stuff is going to change the face of the academic world. We’re not really going to be relying on humans to be the organizers of information anymore. It’s going to be the AIs that are organizing information. The AIs are going to be organizing and producing new information, and humans are going to be enabling that somehow and guiding the AIs towards directions that are valuable to solve problems that we see in the world. And I think the entire shape of the academic ecosystem is being sort of like molded towards that future, perhaps faster than many would like. Indeed. Future is coming at us fast. There’s no doubt about that. And there’s nothing you can do, so you might as well embrace it. There’s. Well, we’re already a little over time. Thank you for joining us. This has been fascinating, and definitely we’ll continue to refer to the Arena leaderboards as new models drop. It’s a great way to orient oneself to what’s working and you know what in the world of increasing noise, like what’s worth actually paying attention to. So thanks for being here. And cheers, guys! It’s great. We’ll look forward to next time. And if any of the listeners are interested in working at Arena, you should feel free to reach out to us. We’re hiring all the time for the best people in research and in engineering. Thanks so much. Thank you. Bye. Jim and I for carbon pricing. By the way, same pricing as argon. It’s the same pricing as soul.

[7440s - 7560s] In the introductory period, and then it will double to the same pricing as Opus Five Five once the introductory period. Are they going to get that much of a chance for introductory period? I mean, like, I mean, how how long is this going to be? Like a week before another? I don’t know. I don’t think they have specified an end date from what I’m reading here to the introductory period yet. But it does seem like I don’t. I don’t think the as far as I know, I don’t think the Argon introductory period had ended. So it seems like we’re rolling from what that’s as the release cycle speeds up. Like we might just be enrolling introductory periods for the foreseeable future. I have to say on the arena thing. I mean, I remember talking about this when they raised their big round a while back, and I was like, I don’t know. This seems like you know something that you could kind of vibe code, and is there really that much value in this that it would raise like you know tens of millions of dollars at such a high valuation? They’ve certainly taken it a lot long a lot farther than I anticipated that they would when I first saw that news. You know the the thing that you learn as a VC is don’t look at don’t look down on ideas. Just look at the growth, and if you see the growth, you can build a business around it. Sometimes a lot of these businesses are very simple. A lot of these ideas are very simple, and it’s really really like the combination of the network effect and the business itself. In in other on another note, let me segue very quickly to our next guest. We have Samvith Ramadurgam. He’s the founder and CEO of Union, a service that helps people find professional connections. A client describes someone they need to meet, and Union searches for a fit, checks interest, and arranges an introduction. The company says it combines AI assistance with human judgment and requires both sides to agree before they are introduced. His earlier work include Stream, a cloud storage startup that went through Y Combinator in the summer of 2012 and was acquired by Box in June 2014.

[7560s - 7680s] Its software let people open files stored in the cloud directly from desktop applications without first synchronizing them onto their computer. He also co-founded Equidate, which became Forge Global, helping build a market for shares in companies that had not gone public. Charles Schwab completed its acquisition of Forge in March two thousand twenty-six. He later co-founded Destiny, whose publicly traded fund offers another route to private company investments. Across those businesses, his work has involved connecting people and capital where a useful transaction is possible, but finding the other party is difficult. To meet you, welcome to the show. Thank you. Good to be here. Yeah, this is this is cool. I have said so many times. It’s kind of one of my standard talking points that one of the things we should be excited about AI for is that it can grind through all these possible matches across all these different types of matches we want to make, whether that’s employment or dating or you know just getting friends together that we don’t have the time to grind through as humans. But if we can get AI’s doing it in the background, then I think we can all live you know much richer lives in virtue of making all these matches and transactions that are at least theoretically possible if haven’t been you know realizable for us so far today. So, tell us about the angle that you are taking on it. I’m there’s a couple points I’m really curious about. Yeah, give me the high level pitch first. Yeah, absolutely. So the matches is actually I would say like a kind of smaller slice of the bigger picture of what we’re building. The way I’d frame it is if you think about this agent era we’re in right now, and you can roughly bucket agents as having like three categories. There’s agents that are for personal use, single player agents. This is beginning to be a thing, right? You’ve got Muse, Grokbot, Instinct, OpenCL.

[7680s - 7800s] Hermes dots, even to some extent, cloud coded codecs. This is like personal agents. I think where that heads in the next few years, you have agents that are extremely intimate with your lives. So they’ll have access to your real time biometrics, access to your bank account, your personal emails, your personal legal affairs. Very very involved with your personal life. That’s one. That’s that’s kind of happening now and kind of beginning to happen. The second one is I would say, as far as I can tell, not really even a thing yet, but beginning to become a thing, which is agents for teams. And you’ll know agents for teams have really taken off when org charts start changing dramatically because the nature of management changes, the nature of intra-company coordination changes. And then the third one is one we’re focused on, which is agents for the public sphere, agents that go across companies, across markets, across economic participants. And this is part of you know what we would call collective intelligence. It’s like an intelligence that has a perspective that’s different than human perspective over like many many people in the context of many businesses. And the the thing we’re building at Union is an agent that can coordinate across the economy. So it is doing matchmaking, introductions, of course. It’s also helping coordinate the sales of assets. And you know, my previous company, Forge, was a market for a specific vertical of trading per APEX stock. And in building that to eventually became a public company, we kind of saw how we scaled the company was actually a lot of human agents and technology mixed together. And this new era we’re in, where we have purely software agents, we can scale the service to make it like really high value and and really touch many markets. It turns out that the stuff we built at Forge is actually like an example of something that has isomorphisms in thousands of different markets. And so the vision for Union is initially to coordinate the world economy and have all kinds of. If you think about all the all the goods, all the services you use, so like the the chairs you’re sitting on right now, electricity we’re using for this meeting, all of this stuff is being brokered by humans today in different parts of the economy. You’d be shocked how much humans are brokering all of this. And there’s a

[7800s - 7920s] 100 trillion dollars of manually brokered deal flow and goods and services. That is what that make the economy go around. And so all of that we think is going to get the next few years something that agents are in the middle of coordinating. They can do it much more effectively, create much more value for people. And that’s kind of like step one, a bigger picture of this collective intelligence vision to help coordinate humanity, which I think goes into social, political, personal stuff later on. So that’s that’s the big picture. I’m happy to jump into any of that, but just to give you a bird’s eye view. Maybe you can tell us a little bit about what I mean. It sounds a little bit like LinkedIn in some sense, right? So can you tell us a little bit about this? It sounds like a two sided marketplace. What is the dynamic behind you know creating this kind of initial inputs, initial supply and demand? Like, are you doing matching using some human in the loop at the beginning, and then like you know finalizing with you know having the initial searches be done by agents and then finalizing? You know, LinkedIn had this issue with scrapers. They sued a bunch of scrapers, and then they lost in like Supreme Court. So you know, people can scrape LinkedIn right now. In order, is that is that like relevant to how you populate the marketplace initially? Like, how does this work? Yeah. So right now our system, it’s it’s agents and humans in hybrid roles, depending on the market. Some markets are regulated; they require kind of humans to oversee it. So markets agents can go and run the whole process. But our system is currently having conversations with hundreds of thousands of people, and really like phone and video conversations with tens of thousands of people right now. And this is across the world, across many markets. So we’re not we’re not tied to any other platform. People can directly engage with us over every communication channel humans use. So that’s over email, text, Slack, iMessage, WhatsApp, the the phone the phone channels, Zoom, Google Meets, Teams, whatever, and any place humans can go, just like this. The idea is that the agent is able to go and operate natively. And that’s what’s happening. And so this is happening across the world, across markets. There’s

[7920s - 8040s] There’s people in Brazil right now who are using it to. There’s a gentleman I know in Brazil who has a steel mill processing plant. He’s using Union to like negotiate different raw material inputs to his plant. There’s a there’s like law firms who are using it to find clients. There’s even people who are doing like romantic matchmaking on it. There’s people who are selling various assets like across the board. It’s really meant to be. So so how how does that work? Like let’s say you’re a steel mill owner, you want to find I guess like coal and iron ore. Like how does that? How how do your agents help? Is it like you know going through the yellow pages or going through LinkedIn finding finding a bunch of names, calling them up, like doing the initial, like you know, do you sell iron ore? Like what is your price? Like when can you deliver? Like etc etc. Like how how does that work? Yeah, I think the principle to I think the way we look at this is that if what what are the humans doing who do this today, and how can we help help make that better? And so the humans do all of the above. They they they will find people online. They’ll get referrals. They’ll do marketing campaigns or run ads. They’ll establish partnerships with channel partners who further people like all of the above. And we’re doing all of that right now. It’s kind of like every every mechanism that there is to find someone and and build a network and grow. We’re doing just doing all of that. And all of it is something that is it’s very it’s internet native. And so the agents can do it as well as um you know the classical software systems that did it before. Interesting. Yeah. How do people react? Well, I’ve experimented with this a little bit myself, and I just asked my you know more autonomous kind of virtual employee agents to run a little PR campaign because I thought I might be able to help some of the AI safety organizations that I support with hiring if I just go on different kinds of podcasts and make the pitch that like people in HR, people in comms, or whatever should come like work in the AI safety domain. I’ve got some.

[8040s - 8160s] Positive response. I’ve got some, you know, obviously just no answer, and then I’ve got some people who are like, “I am viscerally.” Somebody the other day said, “I’m having a visceral negative reaction to being pitched by an AI agent.” How are you? And mine are instructed, by the way, not to lead with their AI nature, but I also do tell them never to lie about it if they’re asked. So, what have you found in terms of how people react by default? How have you had to adjust the approach in order to get, you know, more favorable reactions? What are the kind of rules that you give your agents in terms of when and how to disclose that they are, in fact, AI? I’m very interested in the sort of interface between the AI and a human who wasn’t expecting to be contacted by an AI in the first place. Yeah. So, one thing that maybe surprising is that this this company is like very AI native, very AI heavy. At the same time, we actually have like a a bunch of human staff internally who are like managing relationships as well, and so people are able to get on the phone and talk to a human. And this is this is like partly because like you know there’s a there’s a warmth of a human connection, and for us, we’re we’re able to leverage our humans to be like super humans. And so there is like a there’s like a level at which the whole system is like totally automated, and also a level at which, if you want to talk to humans, you can get on the on a Zoom call with the real person, like pretty much immediately. We have like a center set up that that’s doing that that lets people have like that high touch experience, and so we’ll see how that changes over time. But right now, this is something we’re actually meeting the economy where like the transactions are happening right now. As you’re saying, the world, the level at which you can have these really high fidelity interactions, like I’m not an AI agent. Like the day you’re having like AI agents who can talk the way we’re talking on this video call, I think that starts to change. But until then, I think that there’s some stuff that can be done more in an automated software way, and that one gets pushed to the software. There’s also stuff where you know there’s there’s humans in loop, humans not just in loop but driving it, and the the art of it is can we make it so that we do that.

[8160s - 8280s] While having them be ten thousand x leveraged, and that’s that’s that’s part of how we kind of get the best of both worlds. So, so I would say, there’s there’s you can talk to human on Union. That’s like the big difference. But you still got to get in the door, right? I mean, I’m assuming that the first that feels like a feels like a partial answer to me, but you’ve still got to get people on the hook, right? Like, what have you learned about getting people engaged in a way that doesn’t alienate them so much so that they don’t even you know want to talk to a human? Because I think the guy who had the visceral reaction to my AI agent pitch, he wasn’t like, I want to get on the phone with you and have a real conversation. He was just like, you have annoyed me. Please go away. I think that there’s like a um a setting of expectations thing. I I don’t think that people are against talking to agents. I think that what they’re against is being having the implication that there’s going to be a human, then discovering they’re talking to an AI, and then being like, oh well, this is useless. It’s not going to get me what I want. That disappointment factor is the thing, the thing to avoid. So you build trust by having expectations met. So if let’s suppose we’re sending an email to someone, right? It can be an agent that sends an email to someone, but the agent is sending it on behalf of like a human behind the scenes. And so, if the person then like wants to talk to the human, it escalates gracefully to like a human who can like then like intervene or get on the phone with them and stuff like that. And so, like that is it’s it’s it’s a little bit different than saying there’s just agents running loose and there’s like one agent that you can talk to. There’s also we have a mode where we explicitly add our agent to your Slack, where you can talk to our agent over a chat channel. There, it’s like very obviously it’s an agent has like our logo in the in the profile photo, and so that’s very clear you’re talking to an AI. And there’s a benefit to that too; you get the immediate response and a lot of like you know reach of our in turn in terms of our our database and our connections. So so I I think it’s a matter of expectation setting. I think the way you avoid that is just make sure that um people get what they expect. Let’s talk a little bit about like Forge Global and the transition to Union. So Forge Global was a

[8280s - 8400s] Private, you know, secondaries and you know, private stock kind of market where people could buy and sell. I believe qualified qualified investors could buy buy and sell. And Union is now doing something a little bit similar. It’s not. It’s making introductions from institutional investors, I guess, who want to buy private company stock to potential holders of private company stock. Is that is is that the right way to look at it? Yeah, I would say Union is doing like a. There’s two moves like with Forge we built a marketplace and a network for trading equity in pre-IPO companies, and there’s a lot a lot of very specific infrastructure we built for that. The insight that leads Forge to Union is a recognizing the degree to which a lot of stuff we built there can be done with agents, and two, interestingly, recognizing the extent to which a lot of things that we did at Forge are actually very similar across thousands of other markets. It’s not the difference between real estate and investment banking and private company stock and helping you like actually do recruiting some of these processes like this the way the way you build a system to really handle the relationships and coordinate and connect people are actually very similar. And so it’s a generalization. And so yeah, I think that that that’s the way I’d paint that trajectory. And you know, when you when you look at just the economy, there’s like all there’s like all the stuff, all the commodities, all like the energy, all the assets. It’s like there’s there’s trillions that like are in the same shape as what we did at Forge. And so we’re building a general. To what extent are are you doing a lot of like cold calling and cold emailing? Are you like looking at okay, like these are the these are the people who worked at Anthropic from 2020 to 2024. These are the guys who are most likely to have shares right now. Let’s get them on. Let’s have an agent discuss that with them over cold cold LinkedIn cold email.

[8400s - 8520s] Cold call or introduction to someone within their network that then introduces you to them. Like, is that is that the is that the right way to think about it? We we do do outreach for some markets for pre IPO trading in particular. Not doing so much outreach for that. We have actually a Slack agent that a lot of investors are are installing into their their kind of company Slack, and then they can just talk to us about like what the market prices are and talk to us about like what they’re looking to trade, and then we coordinate that across all these market participants. And so we have like investors or sellers too, by the way, not just buyers. So there’s a lot of like institutional. There’s kind of a private institutional network is set up that’s like coordinating via Slack. So that’s how we’re doing more the the market for like cold emailing for private stock is a I think not the most effective way to achieve that if you want to really like kind of make it a high end service. We also do do email campaigns for other markets where that’s more the norm as well, like recruiting for example. We do a lot of that. Interesting. Yeah. How how far down the sort of long tail of transactions are you reaching now, and and do you expect to reach? Because I I’m I come into this with this as I said up front, kind of this idea that like boy, this is going to be amazing when we can take it to the the long tail. Yeah. It sounds like though what I’m learning is like you’re actually still at like pretty high value transactions as kind of the focal point. So where are you on that curve? Where do you think you can go on that curve? And and at what point does it sort of have to become almost entirely agent driven? When you’re talking like second anthropic secondaries, I can see how you can afford the person in the loop, but when you’re like trying to make you know dating matches in Dallas or whatever, you know, you probably can’t afford that that human time, right? Yeah, yeah, and so and so I love that question because the long tail is already very long, and I’d like to give you a few examples to illustrate it. It’s like quite quite.

[8520s - 8640s] Spectacular, actually, how how broad long tail is, because this thing it it there’s like a there’s a continuous line you realize between a marketplace and a social network, and like as you get closer and closer to like these kinds of like long tail interactions, it’s almost more like hey I’m looking to you can imagine like a social like solo mode type use case where it’s like hey like I want to go surfing with some friends this Friday can can you help me figure out people who like are my friends or friends who want to go surfing with that is like also in scope because we actually have like an app that people can hook up their contacts to and we we understand personal networks so that like that thing is is not different in category than like the the core coordination because the core substrate agents that can coordinate humanity and so yeah there’s like in terms of like the commercial side of our business it’s easy for me to talk about the big markets like brokering big assets and you know goods and services and commodities where there’s a lot of just like economic value and then there’s a lot of long tail so a few examples just to paint the breadth of it there’s a gentleman in Africa who I watched the human agent system do his call and manage his process he has a new beverage brand in Africa and it’s like it’s like this this is a popular in Johannesburg it’s got this cool kind of like dark label and like the symbols and he’s gotten like is getting popular there and our system got in contact with him and offered to help him break into the U S market and he’s now like we then ran a campaign and we we we have like a lot of people we’ve talked to who are U S distributors for that kind of thing we’re now like we’ve connected him to distributors in the U S he’s starting to like get placements in the U S for his like beverage brand that’s like super long tail thing that there’s not like any like in a market map you never have that and like a never build a marketplace for that you know it’s too small so there’s I think that there’s like hundreds of thousands of such like niche use cases across the world that the fact that this is a general agentic system that can coordinate as you get into that then of course you get into social stuff like dating and hanging out we also get into stuff where it’s like across across.

[8640s - 8760s] Organizations like bigger companies coordinating across teams because different people in different teams are using it. So it’s really this is a lot. There’s a big long tail. I think that we’re we’re focused on like in terms of our business model, we like monetize more the head of the curve where there’s like these very clear big markets. But as this grows, our goal is to just create more value through human coordination. And so that involves like the whole long tail all the way going into in the long term vision of the company. Even like can imagine like political organization and like how do you start a new how do you fundraise for your nonprofit like all this kind of stuff is is happening. What does uh what is what is the like largest market that you are working with right now? Is it the investment market or some other market? And how do you monetize that? Like you do the do the sellers pay? Do the buyers pay? Does the person wants the introduction pay? Do you take a cut of like you know revenue that gets created by the introduction? Is it a subscriber service? Like how how does how does the business model work? Yeah. So our business model is we um if we help you make money then we will take a piece of the value we created for you. That’s like the overall. So it’s a success fee based. And so um this is very typical for like coordinators and marketplaces. So it’s like and and our our you know we’re not trying to optimize for margin. We’re optimize for creating growth and good wills. So we’re happy to take. So so let’s say I’m I’m a I’m a private market seller. I have um you know a million dollars worth of shares in a private company. I reach out to you. You make introductions. I sell. Like what what what is the what is the fee going to be? Yeah. So for for private sales in particular, we’re about a couple of months away. Just to be clear, from becoming a licensed broker dealer, just lets us take fees on that. So we’re doing those sales. We’re just let’s say let’s after you get the broker exactly. So once we get that license, um the market rate is between one and five percent of your asset sale. And so it depends on the company and depends on the price and dynamic. Some companies are higher because of supply and demand reasons, but but basically it’s it’s like that. So it’s a.

[8760s - 8880s] Same thing that you know you see across that that space right now, and um you know we we we just I think I don’t think people have any issue with that. It’s just part of the expectation if you successfully close a transaction. Indeed. Yeah. Do you um I I don’t know if this is really how it works, but in checking out the website, I was kind of expecting that I might be negotiating the success fee with an agent. Yes. How far along is that kind of thing, and how good of a job do the agents do for you? Are they like adversarially robust? Like my sense has been that you probably don’t want to send an agent to. You can send them to do a lot of things, but like negotiating, they’re a little conciliatory, probably to be like a good negotiator on your behalf. What have you seen in that? They’re actually they’re actually great, and they’re great because we don’t need to have them be sharks. We’re trying to we’re just happy to be friendly and like take it at like a below market rate. And so basically, imagine that there’s a market we go into. Let’s suppose we’re helping. I’ll make an example: design agency find candidates for for their like who who are who are like customers to do like design work for. This is like a common use case. So we go to them, and it’s not like necessarily like an obvious market rate for that. It’s not like a big market of people who are doing that professionally. So we just talk to we we talk to like a hundred of them. We have like you know a thirty minute long phone call after the emails, get to know them, and we understand like what is how much business would you get from this, like about how like difficult is it typically if you to find these people, and then there’s like kind of like across all all of data we’ve collected, we have a sense of like what a reasonable fee would be, and we keep it modest. We don’t need to make it be anything aggressive. We’re happy to just like earn the business and create value, and then we people agree to it, and then after that we follow up in writing, and we have an agreement established in writing that confirms that. Then we’re off the races. So it’s it’s actually this is a very interesting thing. It’s like I think that.

[8880s - 9000s] You kind of you see with like cars driving on the street, like the the Waymo’s are not going to be like aggressive drivers; they’re they’re very gentle drivers, which I think is a good thing as a default for AIs. And so similarly here, like I I don’t think we need to negotiate hard. I just think that if we’re informed by having data, we can come in at a modest rate, and we’re very happy with that because this is this is a scale play. We’re not trying to eke out every dollar of this. It’s actually better for us to create more referrals and get people to introduce their friends to us. Which is happening. People will like have a good experience and like see see a friend, be like, “Hey, like my friend also needs something. Can you help him?” Things like that. Let’s talk a little bit about Cameron Musgrave. They run human calling operations, and they train you know the the agents against that work. So how does that work? Yeah, we have we have a team of callers, and the way it works right now, this is really cool. So normally when you’re human callers or AI callers, there’s there’s there’s most of our all of our phone calls right now are done by humans. We have we have AI callers as well that we we use for like one percent of our calls just to you know stay on top of the best AI phone calling stuff. So we’re the status quo, but it’s not. We mostly actually just use that to do internal coordination with the callers. But the the so the human calling team, they have this like when when they have a call that comes up, they have a dashboard, and it it lets them like essentially it gives them tips as to what they could say in the call. So they’re getting like prompted to be able to be a superhuman caller because no human being can have perfect recall and recollection of hundred thousand conversations that happened over the last like few months. But this AI has a complete perfect recall and can like perfectly in the middle of a call be like, “Oh, here’s a couple profiles that you might want to check out. Like, okay, I have an asset over here that’s like been trading at that price, but here’s like where I think it’ll come in at now.” So these kinds of like nuanced insights would be possible for a human who really studied the market and was really on top of it, but it’s not possible if you’re trying to do it at scale. And the AI, so like the humans are responsible for.

[9000s - 9120s] Forming great relationships and AI is responsible for like staying on top of the ball and pulling in all the information that needs to get brought in and having like just this perfect recall over all of our previous data to let the human be superhuman. So kind of like this this perfect combo, I think it’s like and you see this with like this is some I think some other tools to help people with like interviews and sales. So this is kind of like that, except it’s tied directly to our system and so it helps people like surface right information. So one of the questions I have is like sometimes you have the introduction and then you have the sales, right? Yeah. Two different two different kind of modalities and sometimes the sales you have to have much kind of stronger persuasion, right? Yeah. So how where does that like where you demark where do you demarcate the line between the you know the introduction alone and this kind of like persuasive process that you need to engage in in order to actually move something forward? Yeah. And who gets the credit and the responsibility for for that? Yeah, that’s a great question. So yeah, there is there is definitely the way you might the way I would like create an ontology around like the different types of conversations is there’s like logistical conversations or like you know if I if I call like a dentist office and I want to like get an appointment, I don’t need like them to be like a top salesman to close me on my dentist appointment. They can just like it’s logistical thing like they take my info as intake call and there’s some basic coordination where all the parties are already motivated. It’s more more matter of getting that done. That’s actually really easy in the present world. The thing that is harder is something that it’s not always about sales. Is not just about like persuasion. There’s other things that like are pers like the human persuasion, the charisma, and like the the closer aspect of a salesperson is one way to do sales. Other way is to project competence and trust, which you can do by having like a lot of data and like a lot just saying like here’s where the market is, here’s the data we have for you, and like here’s like and actually at creating.

[9120s - 9240s] Value, so I’m I’m of the school of thought that the best kind of sales is one where, and especially in this like AI world, you want to lean into the strengths of the system. The system does not natively have the strengths of being like the most charismatic dealmaker, but it does have the strengths of having like being the most well informed and most like on top of it. It’ll never drop the ball. It’ll always like stay close to you and like it can check in with you as often as you want. They have like weekly sync with you to like make sure that’s like attuning to your needs, and then it can pull that in with data and run. It’s it just it can do a lot of work that humans don’t do, very easily because it’s a lot of work for humans to do. And so I think we lead lead with competence, and that builds trust, and then that helps I think move the ball forward. And you know the humans who are doing the calls when they do work with us, they’re you know they they’re doing a lot of these and they’re they’re well trained. So I think that they’re able to do a good job at that. But it’s not I don’t think a necessary like kind of forte to have like only the best kind of like dealmakers in it. Because it’s really about like the system we’re building being something that people trust. As you think about the kind of future of the business, is it what is kind of the core asset, or what is where does the the defensibility for your business come from? It strikes me that it’s probably not the ability to like dial in agent behavior because they’re just getting better and better all the time. That’s getting easier, but maybe it’s the social graph that you’re mapping out. Is it is it really more of an informational advantage that you accrue over time? I think there’s a few. I think that we’re entering the world quickly where like pure software is trivial to spin up. And like if you build like a pure software company, I can tell my agent, look at that, make me one of those, and it’ll just like pretty soon like make you like a perfect version of it. So I think that’s the world we’re headed to. So I think that in that world, from a from a market defensibility standpoint, number one and by far and away.

[9240s - 9360s] The number one is network effect. There is still going to be, it, it like when our system, you know, as we continue to do it, we’ll be having like hundreds of thousands, potentially even millions of like live phone calls and Zoom calls like every day with people across the planet. So the amount of context we’re getting about what people are looking for, what’s changing in their lives, like oh, this person right now, like they just decided that they need to change their strategy and they need to hire this new leader in their company. Like to have all that insight and perspective about what’s happening in people’s lives is a huge amount of context that builds like a a network effect because um here the agent is coordinating across people. So I think ultimately that’s like one of the few modes that will remain. And have like the the modes that remain are going to be network effect based modes. Most have to do with like really deep brand recognition and like end relationships, and um and then like kind of ability to scale capex. I think those are most mostly like in this new world like the main types of modes that we have in in the kind of information software world. There’s there’s modes in hardware and stuff that will stay. And so yeah, for us, I think ultimately this is a marketplace. This is a network, and so that’s definitely going to be like the main strength of it. In addition, though, and maybe in the kind of like intermediate phase, is also a lot of um kind of data points that feed back into the in the system to how we actually make the agents good. What happens is that every time our system is doing anything, there’s like judges who are supervising every message, every word spoken on the call, the whole thing. It’s all getting like reflectively judged, and all the judgments are feeding calling back into a a feedback loop that then like re trains how our agentic workflows work. And so there’s a lot of like improvement that’s getting and getting in there to actually make the system get attuned for different markets. We also um we’re we’re like partnering with a lot of the existing brokers in the world. So we’re not we’re not trying to just like do this agentically. We can like establish a way to coordinate with the humans who are professionals and like add value to them and make it mutually beneficial. In the process, we learn a lot about how these markets work and kind of that data set, how the the really refined like details of how like different.

[9360s - 9480s] Markets operate are getting woven into our system, and so I think there’s another kind of like data loop that gets created there around being able to understand the intricacies of different markets that you know a general model trained on the internet wouldn’t be able to do, and that that as that gets deeper and deeper, you get more competency, which yields more network growth, which yields more network effect, and then leads more data. So there’s like a flywheel there. I think that we’re spinning up. How do you acquire your initial set of customers? Um, so there’s like only so many ways to get in contact with someone in the world. There’s outreach campaigns, there’s referrals, there’s partnerships, there’s advertising. It’s probably a couple others, and it’s like there’s like going to other marketplaces, going to other brokers, going to other intermediaries, different kinds. And we just do all of that, and our system is doing all of that at the same time right now. And so it’s actually, you know, it’s kind of a surprising answer. It’s like there’s there’s only like a dozen ways humans have ever really done this for all businesses, and none of them are really rocket science for agents at this point. Ah, yeah, that’s interesting. Um, what’s your vision for how this kind of matures? You know, I often say my crystal ball is real foggy more than a couple months out, but one thing that seems pretty likely is high consumer surplus in a lot of different domains. Yeah. Um, you know, to the degree that you have a clearer crystal ball than I do for what the future is going to look like, I’d love to get kind of specific and concrete about yeah that vision. Um, how is life going to change as a result of agents coordinating humanity? I I think that so there’s there’s a couple levels. Let me talk first about coordinating markets, which I think which for us is kind of like Act One, really like establish an economic nervous system for humanity, and Act Two is coordinate humanity in more more general way that even goes beyond the kinds of stuff we talked about. So.

[9480s - 9600s] Far, which I think, you know, the company’s name is Union. Where does the word Union appear historically? We the people, in order to form a more perfect Union, Soviet Union, European Union. So this is this this company is like it’s really about human coordination and collective action, which is what governments are responsible for. So I’ll get to that in a second, but like that’s kind of I think where this heads, um in in a big way as you coordinate like what what humanity really is collective is is building. But um you know more more immediately like what’s the I think that, so you asked like what are some almost a way to frame your question is what are some surprises that’ll come up as as these systems develop? I think that over the next five years, I think you know you guys are probably really bought into this. The world is going to get very interesting, and there’s going to be a lot of surprises that come from the capabilities of AI. I think we’re seeing this right now with mathematics, like it’s completely incredible, like that like math has been revolutionized so quickly, and it’s just the beginning of it. And so like that’s about to happen to many domains, and and so I think you could expect like what happened in math and coding in the last year to happen to uh as as domains that are less and less trivially verifiable with software, the further away you get from verifiability, longer it may take, but this it’s going to get there because the feedback loops are going to be there just like with longer feedback loops. And so what does that mean for us? Well, I think that one very big surprise that will hit the world is that at some point in the next few years, people will wake up and then realize that huge number of the world’s markets are being coordinated by agents, thousands of markets from like you know you know dairy in like a a a country all the way to real estate in in China to like the the the pre-IPO market in the U.S. Like all these markets are going to be coordinated better with agents, like period, because what is the way it’s being coordinated right now is that there are going to be humans who are really good at what humans are really good at.

[9600s - 9720s] Building relationships, having taste and judgment, and advising their friends and trust. Humans will have a place, and they’ll be like, I think, be able to actually. The best ones will actually, I think, do better. I think, by and large, most of the work is data coordination, connecting the dots, logistics, operations, orchestration, all of which software and agents going to be way better at. So yeah, we’ll have like a period where, the same way that software and math is really getting taken over, one of the surprises that a lot of this deal making that makes the economy function is actually something that agents are quite amenable to that we’re focused on. So that’s happening, and I think that you know I think there’s probably a lot of players going to be doing that different ways. I think we’re probably a little bit ahead of the curve, given like our previous background with Forge and our we hope to be a part of that of that frontier. And so that I think that’s coming, and that thing that’s going to be quite revolutionary in the same way that that’s happening for math right now. I think it’s going to be very surprising how how how it feels to like just having a software interface to the nervous system for the economy, to have a direct AI conduit to create economic value. I think that’s that’s like those those are abstractions. I could talk more about like kind of what that might specifically look like, but I think that is one level. And then the second thing is then from there, you then have if we if we get to that place, you have agents that are talking to a lot of people, understanding what everyone is up to, what they’re doing, how how people are coordinating, and this goes beyond the economy. This goes into human coordination. So I’ll give an example. I I briefly like a couple years ago was like living in Berkeley for a year, and there’s like this fitness studio, Barry’s Bootcamp, that I back then was like enjoying doing classes at. So I just moved to Berkeley, and like I moved there from Austin, and there was no like Barry’s in Berkeley, and so I was like trying to figure out like how do I get this like this random fitness studio to open up a location where I am because I had been like kind of on a had a good routine with going there, and I I meant different different program, and so I would have liked to tell my agent, hey, like can you

[9720s - 9840s] Coordinate with the the the city around me and help me find like a thousand people who had signed a petition to tell this particular fitness studio that there’s demand here to open up location and like then contact them and help to open up location here and like that’s like an example of the kind of thing that’s like looks like running a petition and governing at a local level just for a little company. Now I also have a friend who was at the time being very active in like the Oakland town hall meetings, and she would go up there and she would like get in the microphone and like talk to them about like what they think they should do differently as a community. Now I was struck that like this is very archaic because like I have also desires for like my local political area and I’m I just don’t have the time or in or like the in my prioritizing going and speaking at the town hall meetings I have other stuff going on. I think we all have these impulses. We all have a perspective on how our society should be governed. I think that as you get this AI, this coordinating it, you get economic value unlocked, which is our first beachhead market, which lets us be a company with a strong revenue foundation. We can scale the network profitably. But I think this gets into coordinating humanity, coordinating how we decide to do things together, collective action, and so eventually going from cities to countries to the whole planet is like division. So the the vision is to have a better way to coordinate, and I think that one way you can model humanity is it’s a collective mind, and this collective mind has an issue right now. The issue is that there’s very low interprocess coordination bandwidth, and so if you’ve ever done any like inner work like meditation or spirituality, like one way I think you can model what’s happening to our mind at a systems level is you’re increasing like the interprocess coordination bandwidth from parts of your mind to create like a more to have it behave as more integrated whole. I think that you can model what we’re doing for the world is like that. We’re taking all the agents, we’re currently human agents, and saying like there’s a way for them to cohere better, to coordinate better by having a better information architecture. I think that’s like the big mission behind the company. Yeah, that’s fascinating.

[9840s - 9960s] Do you see reason to worry? This is a debate precaution I’ve had many times. You know, prompted by many different things that people are building in AI. Do you have any worry about this notion of disempowerment? On the one hand, what you’re describing sounds super empowering. On the other hand, it may be seen by some as disempowering in the sense that, like, all of a sudden, it’s the agents running the world, right? How do we make sure we should we make sure? You know, should we feel the need to make sure that we remain in control in some sense? Prakash will make the case that, like, well, you’re already not in control; you’re subject to the global economy and the bond market and all that sort of thing. This would seem to be taking that up a level, but maybe in a more humanistic way. I’m, I guess, I’m a little confused. Yeah, yeah, I think that there’s two risks that come to two particularly clear risks that come with this kind of technology. One, I’ll talk about both. One is what you just said, and one is helping. If you make a power tool for people to coordinate, that can be very dangerous if you don’t have safety to make sure that they’re not coordinating over bad things. So, for example, in the future, you could imagine someone talk to an AI saying, “Hey, like, I don’t like this particular ethnic group. Can you help me find people in my city who also don’t like that ethnic group? Help us arrange a time to get dinner together so we can scheme about what to do about it. This is like a lot of bad things that could become very efficient if you have like this private agentic coordination. And so, I think that it’s it’s really important to have the system have like an understanding of what it’s facilitating and have guardrails around like what kinds of things it will like allow people to do through it. That’s one. And then to your question around the potential for disempowerment, I think that my my big take is actually that like right now we are really disempowered because we don’t have the ability to coordinate in a more open, decentralized way. Like a lot of the coordination is happening at really top.

[9960s - 10080s] Down by like companies with a lot of like power and a lot of like brand power and things like that, and like the the vision for this now is that you can actually have like grassroots movements come up very easily. So you can imagine like a world. If you take the political example, which is more pertinent to the power question, let’s suppose that I want to start a new movement for like some initiative. I have like some political cause. Like for me to be able to go and light a spark today on social media, lets me be empowered to a degree to go start that. That gets amplified much further. In addition to letting me get in a forum with people, I can actually like coordinate activity. We can like coordinate movement of money and signature of documents and like make decisions together and like take actions together. So the ability to go take collective action becomes very strong. So I think that it’s a really powerful tool for empowerment. To if if you have something that can tap into what a lot of people want, and like that empowerment is like in a way like the bigger risk because you can empower things that are are harmful. I do think that’s probably it’s worth going deeper into the disempowerment side as well. I think that if you’re, I think if you’re human being who is in the business of doing a lot of like logistical coordination, I think that that’s a rough place to be because the AI is going to be better at like the kind of like transactional mundane logistics by like a hundred x better than a human. It’s not even to be close. I think if you’re human who is operating with a level of taste and relationships and judgment, you’re at like the the thing that humans are going to uniquely. Humans are. I think humans have a monopoly on one thing, and humans have a monopoly on their individual life story and life perspective. And so, like to the extent that like that is something that like can be additive to the community, I think that like that is empowering for individuals. I’d be interested to hear if you have a specific kind of like vision for like way disempowerment happens at a more systemic level. It could be worth digging into, but it’s kind of the lay the land I would put in terms of like the question of of risk as this technology goes. I think the general concern is always just that we.

[10080s - 10200s] We have so much agent activity flying around, doing so many things. You know, we can’t really make sense of it anymore. And before you know it, we’re kind of in a zoo of our own creation, where the AIs are really running the show. And like, hopefully, we’re like cared and fed for. But we sort of lose the ability to even understand, let alone like rearchitect the superstructures that surround us. Yeah, yeah. I think I think that that ties to like one of the questions on AI alignment. So AI alignment, I think, has like two parts. There’s like having the AI execute on the genuine intentions of the person who asked it to do something in in a skillful way to avoid the paperclip stuff and all that. I think that’s one one part of it. The other part of it, which I think people talk less about, is obviously like who are you aligning the AI to? Like what what is it being aligned to? Especially when people have disagree over what how the world should be. And so in that world, I think you need a system that can actually get a measure, in some sense, of like what collective well-being means. And I think the way you get a measure of collective well-being is by having a system that’s like deeply tied into the lives of many people. It’s helping them like with their work, helping them do like the important stuff that they have to do. So I think that you can’t you cannot optimize something you can’t measure. And so in a world where you have like agents that are coordinating the world, you have no idea like what direction that’s headed in, like what the you know this kind of be like if you imagine like a hand of the market is guiding the market, like what is the direction the hand is guiding it? If we don’t know and we have no way to measure it, then I think we’re in trouble. And so I think that part of the you know it’s pretty early for us right now because we’re still kind of a relatively small company compared to that like kind of superstructure we’re talking about. But I think to get to that as a certain point, you have to have a really refined understanding, like what am I measuring? What am I optimizing for? And I think what we’ve we’ve learned mistakes from the social network world, which is that any measurement you can make, which is an objective measurement, is wrong if you’re.

[10200s - 10320s] You’re trying to optimize for certain click-through rate or engagement rate or metric. Any metric you can measure is wrong because human happiness is, I think, fundamentally immeasurable. But there is a thing you can do, which is you can continually converge towards a better way of like measuring the the the kind of like the way to optimize the system is not to have a static metric you optimize for, to have it be continually trying to discover like an optimal point, have it like converge towards an asymptote of like this kind of perfect behavior. And so, to that extent, I think that we can we have the very first most rudimentary way of measuring value creation happening right now, which is like, are we getting success fees for these economic transactions? Super low granularity, obviously, like this like very capitalistically oriented. I think as you get deeper into it, and it becomes less about money only and more about like value creation. We’re going to have to have like have ways to measure did we create value for people, do we make people happy, did we like avoid having like negative outcomes happen situations? Having like a more a system that continually over time gets a dynamic measure of that to converge on a dynamic ideal, I think allows to have a system that converges towards an empowered state. And and without measurement, there’s no way to know if that’s happening. And so I think that’s like one of the the key ways to look at it. And this is I think one of the big open questions with alignment is that like alignment right now is mostly just people say alignment they mean make the AI do what we wanted to do, which is which is step one. Because just to even even in the world of of behavior, I think that there’s two components to good behavior: having good intention, and then just making sure you had a good outcome. I mean, you can have you can do something that’s well-intended and have that cause harm, and it was still like a bad action. And so being able to measure the outcome of your action at a systemic level is totally missing from the kind of zeitgeist of alignment. I think that we would hope to be in a place to contribute to that. Do you envision like new power structures?

[10320s - 10440s] Needing to evolve as well. I mean, I always go back to this book that it’s lives rent free in my head, which is Liquid Rain, a very like super niche long tail, not super well known, futuristic novel from like six, seven, eight years ago that I think is proving quite prescient. And one of the things that it envisions is a sort of by issue representation system, where I’ve just been explaining like how the American political system works to my, you know, my young kids, and it’s like, well, we have to pick people to go there, and you know, we don’t agree with them on everything, but we agree with them on more than we agree with the other guys, so we’re picking that one for now, even though I, you know, might wish for a better choice. In this liquid rain vision of the future, it’s sort of a well on this kind of issues. I’ll defer to this person, and maybe you know even their agents, as the case may be. Do you do you think that we can get to your envisioned you know good state while kind of continuing to work through the like coarse mechanisms that you know we have today to send you know one bit every couple of years to the federal government, or do you think that will ultimately need to be reformed to have a a good outcome? Let me just check to see if I understood that. So the question is about like how we can get to a good outcome in context of coordinating with the federal government. Is that the way you put that? Yeah, I mean, you you sort of said you know there’s limited inter process communication, and like it strikes me that I’m you know giving very limited the bandwidth between me and the federal government is like extremely low, and it’s also totally bundled into these like R or D signal, you know, which like obviously loses a lot of what I actually care about. Yeah, yeah, I think that the future probably has more direct democracy, which is like it feels like direct.

[10440s - 10560s] Democracy, I think it actually is more of in the under the hood a representative democracy. Agents are just a representative, and they can like actually talk. So imagine like, I mean, just just to paint the picture, imagine if your local congressman actually had an agent that was in touch with like the whatever two hundred thousand agents that represent like their constituency. And there’s like really high high fidelity process like going through, and there’s like prioritization and aggregation and clustering of issues and understanding of different stakeholders and understanding the conflict that’s happening in this among different stakeholders. Have that whole thing surface in a clear way to like your congressman. I think that really changes how politics happens. Just just with that one thing, because that that that that the Congress kind of like is that that scales across the whole country already, let alone how that goes in higher levels of the of the federal system. And so I think that I think I think that’s where it’s headed. I think that we want to build systems that let us like approach that over time. We’re not starting kind of politically at the moment, so it’s it’s kind of like I think that the way we see our work is to build the nervous system for humanity. This brings us this plumbing, and if you want to build a nervous system, you should start with the backbone, and the backbone is the economy. That’s like the core way we coordinate across the world right now. So that’s that’s the approach we’re taking. But yeah, as you get, I describe like this part more as like the the the you know the the the smaller nerves and the capillaries and stuff at the very ends of your of your of your fingertips that goes more into politics and goes more into even social stuff over time. That’s how we build it out. Yeah. Well, you have such a big vision, you you have to get some, you have to be ready for some galaxy brain questions. One more galaxy brain one I’ll throw at you, and you can react. However, is I’m really interested in this idea of new kind of AI powered mechanism design or like institutional.

[10560s - 10680s] Development and one place it strikes me that we really need institutional innovation very quickly is around how we’re going to pace the frontier. Right? We’ve got like lip service from leaders saying that they think this is needed and want to be party to it, but we don’t have much beyond that at this point. Put to put it plainly, and it seems to me that we might need something where we we have we need some platform or foundation to create narrowly scoped, probably relatively short term, highly iterative deals between like a few parties, so that, for example, your you know OpenAI, Anthropic, Google, whatever can come together and say, “Yeah, okay, we all kind of agree. We probably shouldn’t go this direction at least for the short term. We’ll kind of try to agree to do that, or we will agree to do that, and we’ll maybe have some mechanism for reassuring one another that we actually are refraining from some you know potentially powerful but also potentially very dangerous technique. I don’t know if what you’re building right now has any application to that, but I do suspect you have some interesting thoughts anyway. Yeah, yeah, I do think that what you’re describing is an application of collective intelligence. We’re building more market coordination right now. It’s kind of like you know, it’s it’s a big thing to focus. It’s like our Act One is coordinating the world economies. We have we have plenty of like work to do to to just do that. But but I do think that you know, there’s there’s like this bucket that I would call collective intelligence, and the the situation you’re describing is not unique to the big labs. It is also the case in geopolitics as well. Basically, you have like a small number, small meaning like a dozen, let’s say, like really powerful actors, and they have to make deals in some way where there’s like, and the reason this is non-trivial is because there’s prisoner’s dilemma dynamics that stop it from happening with like the typical coordination mechanism.

[10680s - 10800s] And so the question is like, can you let them like actually? So this is a great example. Like, if you’re familiar with the prisoner’s dilemma, the prisoner’s dilemma is unsolvable if and only if the parties cannot communicate. Once the parties can communicate, that is becomes it no longer is like game theoretically like guaranteed like to to to lose you. You actually can have the parties like achieve like mutually beneficial outcomes. And so I think that this kind of dynamic, which also leads to like this arms race in various eras, both like arms race literally in terms of like military arms, and then now like you could say like an AI arms arms race that we’re in. I think it has these kinds of dynamics in it, and it fundamentally comes down fundamentally comes down to not being able to trust guarantees that other parties are making. So you can have like kind of mutual cooldowns of like weapons embargoes and things like that. And so I think that yeah, I do think there’s a way to have systems that establish trust across parties because they’re like independent. They’re actually we can agree that they’re independent. You can let them audit you. So let’s suppose that there was like an AI organization that would come in and audit my systems to make sure I’m like behaving in a certain way. There’s another like party over there that gets audited by it, and it’s because this is like this is almost like the vision of blockchain. People want to have like trustless coordination. The issue with those you have to have like an understanding and to really spend a lot of tokens to do this properly. So I do think that there’s a way that we get AIs to do that. It’s not what we’re building right now. We’re not we’re not I think focused on that, but I would love to see it and support it. And maybe maybe we get to it if it’s not too late. But I think that might be necessary, like right now. Yeah, I think it’s coming a little sooner than it sounds like it’s on your roadmap. Yeah. So we’ll put that out as a call for entrepreneurs of one kind or another. And yeah, you can count me among those that would be eager to support such a thing as well. Yeah, absolutely. Perhaps one one last question from me. What’s your token usage like? It is.

[10800s - 10920s] Very high. I think that we so we have we have like an internal target we’re setting for our team. This is obviously not what we’re like the way to optimize for productivity. As like as like a corollate, we we want our team usage on the on the coding agent side. Each person be doing like a little at this at the current way we have our infrastructure set up, like between one and two billion tokens per day. One and two billion tokens per day, yeah, productively deployed. And so, in terms of how we’re building the company, this is like an agent native company. And so, we have not only like coding agents, we have like coding agents that talk to each other. We have like internal organization we’ve set up of like agents. So I have like a chief marketing officer who reports to me, who’s got like a bunch of different sub employees reporting to that AI, and all of them are like long running agents that like talk to each other and talk to our coding agent sessions and listen to our team video calls. You know, like are in the meetings with us and chime in and chat to build things for us while we’re like I had a meeting this morning where there was some issue with one of our dashboards. We just said, hey, agent, like fix it during the meeting. They got updated while we were talking. Then, like we gave us feedback while we’re talking and like like this is like it’s kind of like the the way the the way we want to build the company. I think the company in the future is one where the this is like a new a new theme that’s happening in how you build agent first companies. I think the gap between work and meetings is converging. It used to be that like, hey, I want to get out of this meeting so I can get back to work and do real work, and that’s actually changing. There there’s still an element of that for sure, but it’s changing, changing because when we’re in a meeting, if you have like a rigorous discussion, that’s all creating like really rich context from different perspectives for an agent. It then goes and like then you can have a have a debate and say, okay, go do our conclusion with all the aspects that were mapped out there in there. And so, yeah, there’s a um huge like a huge token spend we’re happening with that. And then, yeah, that’s not even that’s just our development processes and our product operations in terms of how we’re talking to people. Yeah, there’s there’s a there’s a huge token spend happening. It’s just increasing. Are you specifically?

[10920s - 11040s] An OpenAI shop, a Claude shop. Do you use all of them? Do you use OpenRouter? We use as primary. I’d say we’re using OpenAI and Anthropic, and then open source for a few things. But we we’re continually like having evals that are checking the different models to find which one is like the right cost adjusted performance rate for our different tasks. There’s different tasks, right? So like we have some models that you want to do like long term thinking on, and some models like you want to get like really fast responses, and it’s going to be cheap. Like there’s a lot of high scale data processing we’re doing, where we have like a rejection cascade of models. We do like a bunch of work with cheap models, and then like the five percent that needs more attention, we put the heavier models onto. So it’s it’s the the idea though is to treat the tokens as a commodity, and be able to just like have swap in the model that’s the best for the job. So it’s a it’s a big part. There’s a lot of cost we’re spending on it. I think that it it’s profitable for us because we have a business model that we can monetize, but it would be pretty cost prohibitive to do it otherwise. What do you feel? You know, you you had Forge Global, which you built in the kind of pre pre agentic era. What what what do you feel is like the big difference between that era and this era in terms of running a company? Oh well, I think that in terms of hiring, I I do think that if you’re like here here’s the metric we had. We’re a small team. We’re like five people, soon to be ten people right now. And so the heuristic I have for hiring someone is: Are they going to prompt me to do something different than we would have done before? We don’t really you don’t really need people to do like work anymore for the kinds of work we’re doing. Need people to have like visionary insight that adds the DNA of the company that really lets you do things that really like makes you do things that you wouldn’t have thought to do otherwise. That no age you can’t prompt an agent to do for you. So I think that we’re not we want to build this company.

[11040s - 11160s] Any way that’s like building the machine that builds the machine, and eventually have more and more agents to spin it up. So I think that you know there’s there’s a type of company you could start now that does with a dozen people that previously took a thousand people, and so we’d like to be one of those companies and have more that go towards tokens, just because let’s just be really lean and move, be really agile and have like a concentration of of talent and vision. So that’s the way we’re doing it. So it’s very different. Like I think we would have had to hire a lot more people in the past to do what we’re doing, and it raises the bar too. Another thing that changes is that when you hire people now, the incremental value of interdisciplinary skills increases exponentially. So if someone is an engineer, that’s valuable. Someone’s engineer and a product person and a designer, that’s much more valuable than before, because before that person didn’t have the ability to actually make use of the integration of these different skills. In the modern world, the more skills you can integrate into one system, the more you like coordinate all. You cut out a lot of interhuman coordination by having it being done in your head. You can articulate a clear vision. To our agent systems, we have so much leverage that the the bottleneck is not execution manpower. The bottleneck is like clear articulation of good ideas. That’s like the real bottleneck, and that comes from having like different disciplines coordinated well. And so that that’s I think another aspect that we what we want to hire the unicorns who are good at many things, which is a rare thing. Typically, they’re founders, honestly, who’ve done that because that’s that’s like a one of their jobs that makes you really do a lot of stuff at the same time. There’s other other things too, though, and so yeah. Amazing. Samvit, thank you for your time. We we’ve gone a little bit over. We’ve taken a little bit of your time. Thank you for your time, and it’s been a pleasure talking to you. Yeah, thank you. It’s been a pleasure as well. You have really really great questions. Great stuff. It’s one of the biggest visions we’ve heard. So looking forward to following your progress. Okay. Thank you. Thanks so much. Have a good one. Bye. You got to love the.

[11160s - 11280s] Old Act One is coordinating the global economy level of vision. That’s really something else. The, I mean, what what has been? Have we heard anything else that’s been as big as as that vision? Yeah, I think I think the one one of the things is that I think a lot of people have big visions, but you also need to focus on near term revenue. And I think he has the advantage that he’s basically going to do what his last company did, but with agents, and so he already knows what the revenue model is going to be, and he already has the initial kind of like you know set of customers, and which would take someone else like you know a decade to form relationships with. So there’s this huge advantage in you know in basically having sold his last last firm and then basically okay, what can we do better with agents with AI? How do I reinvent what we were doing? What is the bigger picture? You know, to that, so that’s one advantage that you know a lot a lot of people. This is why VCs like to invest in repeat founders. I mean, this is exactly why VCs like to invest in repeat founders. So it it’s such an unbelievably big idea. It has me thinking a little bit about the you know, and I don’t know that this was ever anything that Anthropic really said or thought themselves. Whatever the reports may have been, but there was this you know meme for I don’t know 24 to 48 hours, maybe a month ago, about Anthropic thinking that it was going to be the only company, like the last company, you know, just like Anthropic and governments, you know, left at some point in the future as it kind of takes over everything and.

[11280s - 11400s] That remains, you know, a little bit pie in the sky. But when you think about this kind of thing, you can kind of imagine how something like that could take shape. Because I could see a future in which, if Union is to get acquired, for example, by an Anthropic, or you know, perhaps it’s an OpenAI or it’s a Google, it’s just such an insanely valuable layer to to be able to you know have a network effect type of moat and to have all this kind of private information that they’ve built up over time, and to literal, I mean, it’s literally like you know an angle on running the economy. So it it does start to fill in some of the details on what a future scenario like that might look like. Not that I think it’s like likely still that it goes that way, but there’s these like insane grandiose statements, like our first act is to you know facilitate global commerce, or you know this might be the one company left when all the others are gone that you know you can kind of squint at and see how we might actually get there, which is I’m never never ceases to be amazed by how wild some of the ideas are now that are no longer immediately laughable. Well, it it also doesn’t. I mean, it doesn’t take that much to. I mean, one of the things that has happened is obviously that the amount of money that you need to found a company has dropped dramatically, and your survivability as a company has increased dramatically because you don’t need that much money anymore. Um, because you don’t you don’t even need like the skill sets anymore. You don’t need to hire like you know four of the top coders from you know.

[11400s - 11520s] Google or Airbnb anymore. You know, we we couldn’t we couldn’t set up a nice web page before Airbnb, right? Like we we we like the level of like what has happened in the last like you know year. Um, I I I couldn’t center reactive like I I I like I was horrible at TypeScript and JavaScript. I and most ML people were. I mean, you know that that’s why Anastasios came out and he said like, look, we started off as a Python shop and you know the UI was terrible. We couldn’t do all this stuff because we didn’t even know TypeScript. Uh, and this is how retarded we used to be. Like, uh, you know, uh, and and and so we we have come a long way in just, um, in just like you know the last year since since October of last year we’ve come a long way, right? So now no one no one is gonna flinch at a front end, right? Uh, no one is gonna flinch at a back end. Uh, people are gonna be talking about okay, what does the data do or et cetera et cetera, but they’re not gonna flinch it like you know making so so I think the as like your the amount of capital required to found the company has come down and so the level of survivability for a lot of companies has gone up for smaller companies that expands the pool of people who can do entrepreneurship that expands you know what kind of ideas you can look at. Um, you also you know that also means that a lot of the lower risk ideas which you could just pour money into are no longer that attractive anymore because other people with less money will also get there. And I think that’s what has happened to SaaS. SaaS is getting, I think SaaS you know in the next three or four years will will face that, but you can see this kind of like not like Sassy but network effect players maybe Slack is in there maybe maybe some other players are in there, but

[11520s - 11640s] The pure SaaS like Microsoft. Microsoft. Did you know Microsoft is laying off twenty five percent of its employee base? I will. Yeah, I saw that. I my understanding was that it was buyouts. Buyout was being offered to twenty five percent, and we’ll see how many take it. But yeah, it’s a big number. And and and across all of the office divisions, like the Microsoft office divisions, they’re hanging on to all the AI people. The AI people are not being offered buyouts, but all of the office divisions, Xbox gaming division, they’re all being offered buyouts, and they’re going all the way up to senior senior engineers, so level seventy, which is kind of you know principal engineer level. People are being offered, so so they’re basically overstaffed for overstaffed for the existing work that can now be done by agents. It’s kind of kind of almost like a third, you know, twenty like ten x productivity boost. And so now you need to find more sources of revenue for the other for the sixty thousand people. Like, if in order to maintain them, you’d need to expand your revenue by like you know, like two x or whatever. So, it’s just shows you know we need we need more growth. We need a lot more growth in order to absorb the latent the latent amount of humanity, which is you know white collar humanity, which is going to need a job. So, the economy has to grow faster, or else you know there are going to be social consequences that you know you don’t you don’t really want to face. I mean, I can see the growth. I have a hard time seeing you know the the job growth picking up enough to offset. I mean, you know, just look at what he just said. He had five people, right? But wasn’t that a number soon to be ten? He said, you know, we we could have a dozen person company that could do the work of a thousand. He

[11640s - 11760s] You don’t really need people to do work anymore. Like the only thing that matters now is people who will cause us to do something differently than we used to do it. There’s just not that many of those roles, right? Like, how many people are you going to have before you’re just not going to find that many more people that are going to change the way you’re going to do things? It’s like that’s just extreme, and that’s also just not what people have been really developed for. You know, there’s of course been like lip service given over time to the idea that good ideas can come from anywhere within companies, but you know, the the up to potentially 25 of people who might soon be departing Microsoft, like they primarily have not made their way there by causing the company to do something fundamentally different than it was otherwise going to do, right? They’ve made their way there by like executing tasks, and you know, we just don’t, I don’t, I just don’t see how so many people are going to pivot from task completion to like counterfactual, you know, we wouldn’t have done this but for your unique Snowflake contribution. It’s tough. I mean, how many people work in sales? Like he’s basically attacking all of sales. So so I I so okay so I’ve been doing this thing where I am I’m in some WhatsApp groups I’m in some WhatsApp groups for data brokers and so the data brokers put out requests requests for data and essentially whatever requests for data they’re doing now in about 12 months that’s going to hit the product right so you can basically you can basically see the trajectory because what these guys do is that they will go and they’ll buy up data from a company and then they will take that data and they’ll create these like training playgrounds for agents and then those agents will learn in that training playground and then.

[11760s - 11880s] That becomes product in the frontier lab in about twelve to eighteen months. The labs are buying up entire companies’ data, so they’re buying up like entire airline, entire like we we spoke to Merck or Edward, who like entire airline, entire gaming companies, semiconductor companies, manufacturing companies. This this is the this is the key, and they’re looking at they’re kind of done with like the fifty person companies. They’re looking at the hundred to five hundred person companies. That’s the that’s the ask right now. So what I imagine will is going to happen in like twelve months is they will give you a company in a box, right? You can buy, you can buy yourself very easily probably in like in six months. You can buy yourself a dropshipping company in a box. Add capital, add one idea. I want to focus on this. Go, and it will do all of the Shopifying, the advertising, the analysis, the market entry, the customer service, the returns, end to end. And you have a small number of decisions that you that come back to you, right? And I can easily imagine someone running, you know, ten of these kind of companies, right? Like a lot of people have ideas. They talk to one of their friends, or you know, like one of the one of the big businesses in the in the last ten years was actually these little small ornaments that kids put on their Crocs. So this this woman, you know, she started decorating her kids’ Crocs. You know, people asked for it, and then she started making for other people. Then she then she found a distribute, uh, like manufacturing in China, and all of a sudden, like Crocs bought her out for like a hundred million or whatever, right? So I think a lot of people do have ideas. A lot of people are able to execute in small in small numbers. You’re you’re able to make like one thing, right? But this idea to commercialization gap is enormous. It’s enormous in sciences. It’s enormous, you know, when you go to university like.

[11880s - 12000s] Stanford, and you see the ideas, but you don’t see them commercialize. All of this, like the commercialization gap, is I think quite large. And what I’m hoping is that you basically get the agents to do the commercialization for you, which is kind of like the gritty entrepreneurial work that you know used to take a lot of time, calling up like you know calling up Walmart, like getting to see the guy to show them your product, like raising capital, like et cetera. A lot of these things are driven by metrics anyway, right? A lot of a lot of these like yes no decisions at at firms which are going to make these investments are driven by metrics. Number one and number two vibes. You can supply the vibes, like the metrics. Maybe the AI’s can supply the metrics, right? Like why why do you need to be the one that supplies the metrics, right? So maybe maybe we see this kind of ramp next year, starting off with like the one man one man show type of companies, like online gaming stuff that you can sell off on TikTok with some marketing, some marketing ideas, you know, drop shipping of all kinds. Like so, these are the initial companies that I think the initial wave. Like I think Andon Labs will play a part. Andon Labs will be selling agents that do this. I think right. So this is the initial wave that I think we will see next year, middle of next year probably. And then and then you kind of go into kind of okay, I can now run a manufacturing company like this. All I need to do is a little bit of capital. I know the market. You know, I hand it off to the company. I inject the capital. Everything else, boom, boom, boom, boom, boom, it happens, right? So I think I think we see this trajectory starting starting I think middle of next year of kind of the automated company, and I think that is what probably takes the economy to the next level when you get this kind of automated system where a single person can manage most things, but you need you still need that human to like open the bank accounts, kind of give you like the broad idea or like point you in the right direction, sign off on the capital.

[12000s - 12120s] And and you know take back the capital and decide whether to reinvest or to take the money out and enjoy it right. So I think I think you see this kind of trajectory next year, and I think that that is probably what what gets the company, the econ like the the economy to the next level, the growth to the next level. You know, I’m hoping. I’m hoping. I’m hoping, right? Yeah, I mean, I think there’s going to be what I do expect is certainly a ton of these one-person companies, a ton of cool ideas that never got to exist before. I’ve got one coming, you know, real soon, which I think we’ve talked about a little bit. But this is my Fruitopia app that I’ve been working on a little bit, and mostly just prompting agents to work on. I always thought there should be, you know, some like glorious database of fruit and particularly like what is good now. You know, because I never really know, and there’s a whole bunch of detail that goes into that around like even on a mango, what kind of mango are you actually getting? It varies by the time of year, and they don’t really label it in the supermarket. So the I’ve told this story to so many people, and they’re all like, “Oh yeah, I’ve noticed that. Some of the mangoes are really amazing, and others kind of suck.” And the answer is, it’s a different variety, but you’re not told what variety it is at retail. So just you know, on my own kind of two percent time, you know, I’m able to spin up what might become the world’s biggest database of retail fruit information, and you know, I don’t even need it to get traction, right? It could be totally, it could just sit there and do nothing, and that’s fine. Or maybe it’ll help people eat more fresh, delicious fruit, and that’ll be cool. So I think we’ll see a lot of these things, but I just don’t see how there can be enough of these little niches to take you know to absorb all the people that used to be driving.

[12120s - 12240s] Cars, you know, doing tasks. This is this is where it seems very, very hard for me to get over the hump and envision so much of that. I’ve been, you know, limited in my vision before. I think my if we dig up my old comments on Arena, they will reflect perhaps too limited of a vision. But I do still struggle to see this working in a way that’s not majorly disruptive. A lot of a lot of restaurants start off with like one good home cook, and that one good home cook then becomes, and that’s the that’s the that’s the typical start of a family family owned restaurant. But family owned restaurants have really died out because it’s it’s a huge amount of work, and the the profitability just isn’t there because when you when you get through all of the licensing and the and the location and the labor. So if you look at a restaurant, twenty five percent of twenty t here is twenty five percent which is labor, twenty five percent rent, twenty five percent materials, and then you know ingredients, and then twenty five percent gross profit margin, which then gets spent on like sales and marketing and this and that. Like you’re you’re lucky to clear five or ten percent on a on a typical restaurant, and maybe if you get you know bots, if you get the bots who can work in the kitchen, you get rid of the labor, so the labor component becomes capital. Then you then you’re at fifty percent. Then then all of a sudden, this restaurant business becomes like enormously profitable because remember you’re at like you know you’re you’re giving away you’re down to your last five percent of margin after the sales and marketing expenses and blah blah blah. But if you all of a sudden you get rid of the labor, maybe you pay five percent for the rental of the bots, twenty percent, twenty five. You you just made the restaurant business like you know five times more profitable, right? And if it’s five times more profitable, then you can open up more restaurants. We can only eat so much food, right? This is the thing where it’s like at the end of.

[12240s - 12360s] Today, I’m only gonna eat so many calories. You know, you’re describing a world where all the restaurant. This is the classic, like we lose money on everyone and make it up on volume, like brought to employment. Like all the restaurants will massively cut their staff and replace their kitchen workers with robots, but we’ll have a lot more restaurants. Like I just don’t see the math working out on that, given the fact that like if there’s a there may be a lump of labor fallacy, but I’m pretty sure the lump of food principle is like roughly right. You know, I’m only gonna eat so many calories, and like at some point, that is just plain saturated, right? I don’t I don’t see the um like we have we have to posit a world where I’m somehow like eating from like five different restaurants for every meal or something to get to the point where I don’t know maybe my vision is too limited. That could be kind of cool, I suppose, but at some point, I mean, I would imagine something like dense dense Asia, right? Dense dense Asia, dense Hong Kong apartments where you basically cannot or almost cannot cook in the apartment. You can like run downstairs and like blah blah. So walkable walkable kind of places. Rare, rare in the U.S. right now, but you know things things can change. You know, over time, perhaps. But you are right. You know, I I don’t think the human consumption is really where the growth is. The growth is always going to be in the AI consumption going forward, as you can see. So you have the concentric circles of the economy, and the AI economy is always the one going to be expanding faster. So anything that supplies into the AI economy is the thing that is going to grow. Anything that supplies into the human economy is going to be, I think, on the on the down drift, except for status. So I think I think you know Peter Thiel has come out with like the relational economy. Alex Alex Amus also at I think OpenAI right now on the relational economy. So yeah, that that could be that could be a big like thing, but again, status economies are not meritocratic. So we have we have tended not to.

[12360s - 12480s] Like these status economies because they are not meritocratic, and so that that I think has more social repercussions, perhaps. Yeah, I mean, it it sounds more dystopian than utopian to think that like the bulk of human activity will go into status games. I have to say, I do agree that there’s like no limit to, or at least I don’t think we’re anywhere near the limit of like how much status games people might want to play if they are freed from other, you know, task completion style drudgeries. But it is still kind of zero sum, and it does sound rough as a way for people to like. You know, I think a lot of those games don’t leave people feeling very good, right? I mean, this is the the I’m sort of a I’m like a bear on employment, bullish on people’s ability to find meaning in absence of employment. I think certainly relative to most, I really think it’s and I’ve said this like probably a dozen times on this feed, but like I really think way too many people in the AI policy space forget or maybe never even realized that like a huge portion of people are working to live and not really tying up their identity or their sense of self worth with their work in the way that. You know the the academic or you know intellectual class tends to. So I’m I’m more bullish that like people will find things to do that will make them happy. But I kind of hope it’s or at least if you told me that like all this un you know all this like sort of excess energy goes into status games, I would be kind of bummed out about that because I feel like a lot of that stuff ends up being fairly empty and just so.

[12480s - 12600s] Zero sum and kind of positional that like it doesn’t sound it sounds again it sounds more dystopian than utopian to me. But but I mean look, Anthropic’s founders donating eighty percent of their you know returns from Anthropic into the nonprofit that that is a status game. Right, philanthropy has always been a status game, right? So Elon Elon donating his like you know money to the Trump accounts that that’s why the EAs have become so high status. Yeah, yeah. The the entire point is not that status games are bad, but that we manipulate these status games to deliver pro social outcomes for humanity as a whole, right? A lot of a lot of religion, religious religious like you know tithing is a status game, like you know, and people people tithe in order to get that kind of status. In in every single religion, like you have this tithing and like you know the rich kind of you know providing for the poor and like giving free meals and all of it is all of it is like status, right? So we come up with these like illusory rewards, right, in order to like motivate pro social behavior outside of currency, right, outside of lucre. And so, and I think that that’s that’s like if we’re going to an abundance economy where lucre doesn’t matter anymore, what you end what you end up is with this kind of status, and you have to make people feel good about it, right? So you have to give them props for you know these pro social you know moves. And that’s that’s it. Status is literally like just giving people props for pro social moves and recognizing that they are doing things which are which are good and that there is a sense of good that there is an objective sense of good and that if you are performing that objective sense of good, then you deserve something. And you know, you can you can even maximize it by saying that you know the maximum is you be.

[12600s - 12720s] Anonymous. When you do that, so that’s the that’s the that’s the next level up, and then and then you pass away, and then people found find out you’re like you were the anonymous donor to, you know, X and X and Y and Z, right? But I I think it’s it’s it’s bad when you say it’s status, but it’s great when you say it’s philanthropy, right? Like it’s it’s it’s such a it’s such a Russell conjugation thing, right? I’m doing it for status. You’re doing it for you know philanthropy and like the good of mankind, right? So I’ll maybe take the pitch on on that distinction for the moment. I think you can imagine what my take would be. Certainly, there’s some truth to what you’re saying. I I think certain communities would flatter themselves to be a little bit more focused on the actual prize than the status, but that may remain in the eye of the beholder for the future, anyway. Um. Well, I guess we got to go benchmark Gemini for carbon for ourselves, among other to dos. Definitely going to sign up for Union as well, and it is something people can sign up for. I’m very curious to know who it thinks I need to meet, and be interested to see what what new connections it brings into my life. Um, anything else we should touch on before we? No, we’ll be we’ll be back on Friday for a show, and we are yeah. This this this has been great. I mean, all the guests today were fantastic, and you know, my my pass caught up to me in the first one, so room temperature superconductors are never never going to be leaving. I I I got to go see Periodic Labs. I got it. I got to make a visit and like you know talk to them. You know, I should I should I should pitch myself as the room temperature superconductor pitchman.

[12720s - 12762s] For whoever invents one, yeah, that’s funny. Yeah, we’ll it will be a good show on Friday too. We’ve got Michael Freeman from Service Now and Sagar Vaze from Mistral, so definitely a couple names people will know. And I’m sure there’ll be some interesting discussion and hopefully a bit of alpha spilled with us on Friday once again. Indeed, Nathan. Good morning. Good morning. See you on Friday.