Toyota, Tech Chiefs Pivot to Local AI Harnesses as Token Bills Mount

Token成本高企，丰田与科技高管转向本地AI框架 暴力扩展基座模型导致失效率叠加与成本失控，迫使企业转向特定任务小模型与确定性评估框架。 圆桌讨论探讨了从暴力堆叠基座模型规模转向模块化智能体架构、领域专用小语言模型以及结构化框架工程的趋势。与会嘉宾分享了实际工业落地场景，包括多模态儿科诊断、丰田加速汽车材料测试，以及机器人领域优先考虑高熵数据清洗而非单纯数据量的主动学习策略。在肯定推理能力和本地计算集群进展的同时，发言者对生产环境中缺乏监控的多智能体链路表达了深度怀疑，强调了误差传播以及人机协同治理带来的高昂验证成本。会议最后强调，必须坚持使用干净的基准真值数据集、为领域专家提供多层级视觉可解释性，并遵循严谨的软件需求规范，而非盲目进行毫无节制的感觉式编程（vibe coding）。 AI研究生态的核心特征体现为传统科学严谨性与企业应用落地所要求的极速部署节奏之间的内在张力。 (0:04) 人机混合团队在业务压力下本质上具有不可预测性，因此在实际部署前，必须通过合成数据模拟和严格受控的执行框架来化解协作风险。 (5:27) 严谨的AI工程要求在引入深度学习或大语言模型之前，先建立表格机器学习基准模型，在单纯的性能提升与可解释性及算力成本之间取得平衡。 (7:07) 通过编排、动态路由和自我反思循环，成本更低且高度专业化的模型协同组合能够超越单体前沿基座模型。 (11:20) 针对高熵决策边界的主动学习，相比暴力数据收集或大规模遥操作数据采集，能带来显著更高的准确率提升。 (16:08)

Industry researchers argue that enterprise artificial intelligence has reached a point where raw model scaling yields diminishing returns, shifting competitive advantage toward domain-grounded orchestration, small language models, and rigorous evaluation harnesses. Diminishing Returns on Scale and Brute-Force Data (13:00) While foundational AI research initially advanced through massive compute clusters and sprawling data sets, panelists emphasize that physical and financial limits are challenging the brute-force paradigm. Pedro Alves highlights this bottleneck in robotics and computer vision, noting that while the first million images reached 90% accuracy and the second million achieved 95%, climbing from 95% to 96% required jumping to 10 million images. By adopting active learning to target high-entropy decision boundaries, just 50,000 curated samples outperformed the lift of the prior 8 million random images. Alves warns that robotics startups burning tens of millions of dollars on 50,000 monthly hours of teleoperation at $100 per hour risk repeating early machine learning mistakes. Simultaneously, enterprise infrastructure costs have forced a shift away from sole reliance on hyperscaler APIs. Michael Caiola points out that organizations often face severe capital expenditure trade-offs, from GPU pods costing over $500,000 with annual depreciation to cloud-based setups burning through half a billion tokens in a single day. To mitigate runaway API expenses, teams like Ravi Ummadisetti's at Toyota North America run local NVIDIA DGX clusters to evaluate internal coding agents and implement soft usage caps on developer workloads. Model Orchestration and In-House Small Language Models (11:20) Rather than training foundational models from scratch—a race Michael Caiola argues is already dominated by Google, Meta, and Anthropic—enterprises are prioritizing harness engineering and model orchestration. Pedro Alves compares current large language models to a Ferrari trapped in first gear at 30 miles per hour, arguing that proper routing, query decomposition, and ensembling act as the transmission that unlocks higher speeds. In practice, coordinating multiple cheaper models often outperforms expensive flagship frontier models on complex tasks. For regulated sectors and enterprise operations, deployment strategies are leaning toward smaller, domain-specific models (SLMs). James OKane observes that Fortune 100 enterprise clients primarily deploy specialized in-house SLMs, reserving large frontier model calls only for rare edge cases. Ravi Ummadisetti explains that Toyota addresses specialized materials challenges—such as compressing a five-year automotive paint formulation and environmental qualification cycle—not by relying on generic interfaces, but by integrating foundation models with historical formulation data, factory robotics telemetry, and domain ontologies. Production Bottlenecks in Multi-Agent Reliability and Verification (29:31) Despite widespread industry enthusiasm for agentic frameworks, panelists express caution regarding autonomous multi-agent deployments in mission-critical environments. Michael Caiola warns that multi-agent systems risk compounding black-box opacity, as errors propagate multiplicatively across autonomous chains rather than resolving linearly. While developers may find multi-agent harnesses useful in code sandboxes, James OKane notes that verifying model outputs often costs more than generating them, causing human-in-the-loop oversight to degrade into passive rubber-stamping. To balance operational automation with diagnostic reliability, Toyota deployed its GearPal agent across manufacturing assembly lines to resolve equipment malfunctions using terabytes of machine data. Rather than trusting raw generation, Toyota monitors agent performance by analyzing execution traces and user feedback, using graph-based skill representations to continuously evaluate and improve retrieval strategies. Alves also cautions against naive self-reflection prompts like asking a model 'Are you sure?', which frequently induces a Monty Hall-style trap where models second-guess correct answers and consume unnecessary inference compute. Pragmatic Tooling, Multi-Tiered Explainability, and Rigor (45:53) Gaining adoption from skeptical domain experts requires replacing vague generative text with layered, empirical explainability. In clinical projects predicting pediatric congenital heart conditions, Caiola's team pairs structured health records with raw echocardiograms to produce rank-ordered diagnostic shortlists for genetic testing. To convince clinicians who distrust predictive models, the team employs a three-tiered validation strategy: establishing accuracy on benchmark datasets, projecting embedding spaces into visual phenotype clusters of historical cases, and mapping vision transformer attention patches directly onto anatomical features. For smaller organizations adopting development techniques like 'vibe coding,' Alves and Ummadisetti argue that AI should be treated as an 'outskilling' amplifier for skilled engineers rather than an upskilling shortcut for novices. Ummadisetti advises teams to invest two to three weeks drafting thorough specifications in Markdown files before code generation, running automated verification loops across pull requests. The panel warns against treating models as inexplicable 'magic,' asserting that deploying unchecked outputs without disciplined domain validation reduces critical engineering to an operational gamble. The panel concludes that future enterprise gains will rely on organizational discipline, ground-truth data formalization, and rigorous evaluation frameworks rather than simply waiting for larger foundation models. However, establishing low-latency, real-time verification and maintaining meaningful human oversight in regulated production pipelines remain unresolved industry challenges.

Ai4研讨会上的多位产业与学术专家指出，当前人工智能的前沿突破正从盲目扩大模型参数转向注重工程套件、主动学习以及任务编排的务实落地；企业若缺乏严格的评测基准与领域专业验证，单纯依赖黑盒大模型等同于承担高风险的技术投机。 告别蛮力扩展与数据收益递减 (16:08) Pedro Alves指出，盲目依赖海量计算与数据的扩展策略正面临边际收益递减的硬约束。在计算机视觉早期实践中，前100万张标注图像可带来90%的准确率，追加100万张提升至95%，但从95%增至96%却需要800万张新增数据。通过采用主动学习机制，针对模型决策边界的高熵不确定区域定向采集5万张样本，其效果便超越了800万张随机样本。Alves批评当前机器人领域部分初创企业每月耗费数千万美元盲目采购5万小时单价超100美元的遥操作遥控数据，重复了粗放采集的误区。在数据极度稀缺的医疗高危场景中，研究人员亦通过物理信息神经网络（PINN）与语义损失函数将先验物理定律硬编码至模型中，使得仅需1000个样本即可实现传统模型需要10亿个数据点才能达成的泛化能力。 与会专家进一步质疑建设兆瓦级巨型数据中心的合理性，指出企业用户很少需要无所不知的通用全能引擎，更不应为了通用基准上1%的提升承担极高的延迟与成本。相比于让企业陷入全量微调绑定的静态权重黑盒，模块化的解耦架构与轻量化小语言模型（SLM）能够让企业在新一代前沿模型发布时无缝替换局部组件，从而以更低算力实现高可靠性的业务闭环。 模型编排与企业套件工程 (11:20) 专家们强调，基础模型层的技术角逐已被极少数超大规模厂商主导，非头部企业自研底座并不具备经济可行性。Ravi Ummadisetti指出，企业落地的核心差异化优势在于“套件工程”（Harness Engineering），即将企业专有数据封装为确定性的领域工具与API，交由具备高级规划与推理能力的外部模型调用。在丰田汽车北美公司，涂料研发传统上需耗费5年时间（1年配方研发加4年户外耐候与腐蚀测试），工程团队通过构建融合历史配方、实验测试日志与产线喷涂机器人遥测数据的逆向学习仿真套件，大幅缩短了材料验证周期。此外，丰田还通过本地互联的8节点NVIDIA DGX集群处理代码智能体推理，并对云端API调用设置软性Token限额，避免产生失控的开发账单。 Pedro Alves用汽车传动系统作比喻，认为业界造出了顶级的“法拉利”大模型，却因缺乏调优的变速箱而被困在一档以低速行驶。模型编排层的作用便在于根据任务复杂度动态分发查询：基准测试显示，由两个廉价小模型组成的集成架构完全可以击败昂贵的顶尖模型；简单Prompt技巧如让模型在输出前自我反思“你确定吗”，能补足遗漏的推理步骤。然而，盲目的多轮询问也可能诱发“蒙提霍尔效应”（Monty Hall Trap），导致模型因误以为自己出错而改写原本正确的答案并空耗算力，凸显了结构化编排规则的必要性。 多智能体生产陷阱与验证瓶颈 (29:31) 针对2026年初业界对“Agentic Harness”的热炒，Michael Caiola对多智能体在国防与医疗等关键任务生产环境中的部署表达了强烈质疑。他警告，在链式自主运行的多智能体系统中，单点错误率会呈乘数级级联放大，原本单一模型的黑盒不透明性在数十甚至数百个交互智能体间演变为不可调试的灾难，且当前行业严重缺乏隔离单组件与评估双智能体交互的严格测试与评估（T&E）标准。James OKane亦证实，其实验室虽搭建了由35个智能体协同的沙箱测试框架（其中超过一半专门担任验证器、单元测试员与红队对抗者），但此类多智能体系统在高度受规管的行业中仍局限于研发原型阶段，不具备无人值守上线的成熟度。 在落地实践中，丰田通过在制造工厂部署辅助智能体GearPal展示了受控落地的可能性。该系统在协助车间工程师排查设备故障时，并非放任黑盒运行，而是将无结构维修文档映射为图谱技能表示，并通过运行时执行链路分析（Trace Analysis）与现场工人点赞/点踩反馈相结合，由评估引擎自动定位是检索策略失效、工具调用故障还是Prompt合成缺陷。OKane指出，在高度受规管的财富100强企业中，人工介入评估的成本往往数倍于AI生成本身的开销，若缺乏体系化的人机协同工具，严密的人类监督极易退化为被动无意义的“橡皮图章”。 拒绝黑盒崇拜与分层可解释性 (45:53) Michael Caiola针对医疗与生命科学领域提出了严苛的基线建模法则：在引入昂贵的深度学习或生成式LLM之前，必须先构建最基础的传统机器学习模型作为基线；若不能证明深度模型带来的性能增益足以抵消算力开销与黑盒不透明性惩罚，便不应盲目推进。在儿科先天性心脏病的基因诊断项目中，Caiola团队将结构化电子健康档案（EHR）与原始超声心动图相融合，直接预测遗传分类概率并输出排序候选清单，帮助临床医生优先订购针对性基因面板，避免冗长且昂贵的泛化筛查。为了化解临床医生根深蒂固的怀疑心理，团队设计了三层解释流水线：首先通过测试集准确率设立门槛，随后将高维嵌入空间投射为可视化表型聚类以证明病例关联，最后将Vision Transformer注意力补丁映射至超声心动图的具体解剖结构上，使算法决策依据与医生的生理直觉对齐。 研讨会最后探讨了初创与中小团队在“氛围写代码”（Vibe Coding）风潮中的认知误区。Alves认为AI编程工具的本质是优秀架构师的“技能放大器”（Outskilling），而无法填补平庸开发者在系统架构直觉上的先天欠缺。Ummadisetti建议开发者将80%的精力前置到编写详尽的Markdown需求规范与自动化校验闭环中，而非盲目将未经消化的错误提示词甩给模型。专家严厉警告，生成式AI输出的流畅性极易让缺乏专业底蕴的人“角色扮演”为伪专家，若工程师无法向外行拆解系统底层为何运作及何时失效，将无法解释的黑盒输出奉为“魔法”，本质上就是在关键业务决策中进行毫无工程严谨性的赌博。 总体而言，专家们认为当前AI产业的重心已从算法层面的前沿参数扩张，转向企业内部的数据就绪度、编排套件工程与评测体系建设。然而，跨智能体复合错误的数学级联、高昂的人工核验成本，以及面对严苛业务场景时可解释性工具的匮乏，仍是制约高度自主化系统在实体工业与受规管生产环境中规模化部署的核心阻碍。

## Background 本场圆桌由 Ai4 频道于 2026 年 8 月 22 日发布在 YouTube 平台。视频汇聚了来自丰田、波音、Credence、Arcova 及 Thoth AI 等顶尖企业与科研机构的 AI 科学家及负责人，围绕大语言模型、智能体架构、可解释性与扩展效率等前沿研究突破展开深入探讨，旨在剖析科研前沿如何指引企业应用落地与未来技术演进。 ## TL;DR 本期圆桌聚焦 AI 前沿科研与产业落地的交汇点，深入探讨了模型编排路由、智能体系统工程（Harness Engineering）、端侧小模型与主动学习等核心技术趋势。专家们警示了盲目堆砌算力、消耗海量 Token 以及多智能体串联带来的黑盒与错误复合放大风险，并结合丰田制造车漆加速研发、医疗心超多模态诊断等实际案例，强调回归基础基线模型与严格评估体系的重要性。针对氛围编码（Vibe Coding）与人机协同，嘉宾们指出企业应重塑数据底座并强化严苛的需求工程，警惕脱离底层理解而将 AI 当作“黑盒魔法”的伪专家虚假繁荣。 ## Atmosphere & style 整场讨论兼具前沿学术严谨性与一线工业实战的务实冷静，主持节奏流畅，嘉宾之间既有深厚认同亦有针锋相对的真知灼见。面对当前过热的 Agent 炒作与算力军备竞赛，科学家们频频抛出“法拉利挂一档”、“造火箭还是开汽车”以及“赌场掷骰子”等生动犀利的比喻，在幽默风趣的交锋中流露出对底层工程机理、可解释性与科学严谨性的深刻敬畏，给观众带来一场去伪存真、极其清醒且信息密度极高的思想盛宴。 ## Key takeaways - [21:59](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1319s) 企业级 AI 落地的核心竞争力在于将私有数据封装为标准化工具的“治具工程”（Harness Engineering），而非盲目追逐参数量更大的通用基础模型。 - [11:20](https://www.youtube.com/watch?v=n3OYFFujUgg&t=680s) 大模型在很多场景下的瓶颈在于缺乏高效的查询分解、路由与自我反思机制；经过良好编排的小模型组合完全能够匹敌甚至超越单体旗舰模型。 - [29:33](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1773s) 严肃工业生产环境对多智能体系统（Multi-Agent）需保持审慎，因其存在错误率层层复合放大、系统黑盒化导致难以 Debug 等严重隐患。 - [16:09](https://www.youtube.com/watch?v=n3OYFFujUgg&t=969s) 具身智能与计算机视觉领域正面临暴力堆砌数据的边际效益递减，采用主动学习（Active Learning）定向采集决策边界样本能大幅降低数据采购成本并提升模型性能。 - [07:07](https://www.youtube.com/watch?v=n3OYFFujUgg&t=427s) 在医疗等高敏感专业领域，AI 落地必须建立在传统基线模型与严格可解释性之上，通过潜空间聚类及 Patch 级图像特征高亮打破临床专家的认知偏见。 - [17:34](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1054s) 企业自建微调模型存在高额算力与迭代沉没成本，采用解耦的模块化系统和后处理技术更有利于未来基座模型的无缝热替换与敏捷评估。 - [26:27](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1587s) “人类在环”（Human-in-the-loop）面临高昂的专家校验成本，若缺乏实时可验证机制，极易导致人工审查沦为形式主义的“橡皮图章”。 - [19:32](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1172s) 端侧智能（Edge AI）结合符号编码器（Semiotic Encoders）与物理约束等先验规则，可实现在极低资源与千级别小样本下的精准建模与连续学习。 - [47:55](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2875s) “氛围编码”（Vibe Coding）的本质在于让顶尖架构师产能倍增（Outskilling），企业应将 80% 的精力投入前期需求工程，通过详尽的规格文档构建自动化审查闭环。 - [51:37](https://www.youtube.com/watch?v=n3OYFFujUgg&t=3097s) 需高度警惕借助 AI 光鲜回答“假装领域专家”的虚假繁荣；缺乏对底层机理与失效边界认知的黑盒依赖，本质上等同于碰运气赌博。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=n3OYFFujUgg&t=4s) 圆桌开场：连接前沿 AI 科研与产业落地的桥梁 主持人 Jason Hiner（The Deep View 主编）宣布圆桌讨论正式开始。他指出，尽管科学界的研究往往遵循着严谨而稳健的步伐，但当前科技产业与社会所经历的惊人变革，其底层驱动力无一例外全部来自前沿的科研创新。从大语言模型（LLM）、世界模型到多模态系统，技术在高速演进的同时，也带来了模型可观测性、可解释性以及鲁棒性等一系列现实难题。本场圆桌汇集了来自顶尖工业界和科研机构的幕后科学家，旨在剖析最新研究突破，帮助大家理解 AI 技术的底层机理以及未来科学与产业的发展走向。 ### [01:58](https://www.youtube.com/watch?v=n3OYFFujUgg&t=118s) 专家阵容介绍：从基因组学、具身智能到人机协同与医疗 AI 与会专家依次进行了自我介绍，展示了横跨多个学科与工业领域的深厚背景： - **Pedro Alves（Thoth AI 首席技术官）**：拥有约 25 年的 AI 与机器学习经验。早期从基因组学研究起步，随后历经计算机视觉、社交网络以及华尔街量化交易等多个领域。目前专注于大语言模型编排（LLM Orchestration）与智能路由（LLM Routing），并在机器人领域探索如何利用主动学习（Active Learning）将机器人的执行错误反馈回数据集中，为具身智能（Embodied AI）构建更高质量的训练数据。 - **Ravi Ummadisetti（丰田北美智能体 AI 与产品研究负责人）**：软件工程师出身，深耕工业界 15 年，专注于利用 AI 解决实际商业业务问题。目前全面统筹丰田在制造、供应链、研发等各业务部门的 AI 智能体研发与工程化落地，负责从前沿 Agent 理论研究到生产级系统部署的全生命周期。 - **Michael Caiola（Credence 高级生物医学 AI 研究科学家）**：典型的硬核科学家（"hard scientist"）背景，拥有数学博士学位。曾在非人灵长类动物湿实验室从事神经科学研究，后在 FDA 从事人脑临床转化研究，并拥有深厚的联邦政府科研经验。他致力于在机器学习和所谓“魔法般 AI”爆发之后，重新为 AI 领域建立严谨的科学实验方法，专注于为客户构建风险感知（risk-aware）能力。 - **James OKane（Arcova 人机 AI 服务总监）**：学术背景主要为社会心理学，专注于犯罪预防与人类信息处理机制；在工业界拥有超过 25 年的技术风险管理与网络安全经验。目前在 Arcova 负责面向企业员工的下一代劳动力架构设计与工程（nextgen workforce design and engineering），同时也是一家专注于人机通信的 Pre-Seed 轮初创公司联合创始人。 ```mermaid flowchart TD accTitle: Panelist Backgrounds and Focus Areas subgraph Panelists["圆桌嘉宾与核心研究方向"] P1["Pedro Alves<br/>(Thoth AI)"] --> F1["LLM 编排路由 & 机器人主动学习数据集"] P2["Ravi Ummadisetti<br/>(Toyota)"] --> F2["全产业链 Agent 研发与工业工程化落地"] P3["Michael Caiola<br/>(Credence)"] --> F3["生物医学 AI 实验规范与多模态临床诊断"] P4["James OKane<br/>(Arcova)"] --> F4["人机协同模拟、任务分解与安全治理"] end ``` ### [05:27](https://www.youtube.com/watch?v=n3OYFFujUgg&t=327s) James OKane：人机协同系统的复杂性模拟与基于 LLM 的任务规划 在谈及目前正在开展的核心研究时，James OKane 率先分享了 Arcova 在“人机系统（Human-AI Systems）”稳定性与可预测性方面的探索。由于人类与 AI 混合系统的动态交互极难预测，他们正在利用合成数据构建仿真模型，重点探索将大语言模型作为规划器（Planner）。 在此架构中，LLM 负责将复杂业务目标进行任务分解（Task Decomposition），并根据能力画像将子任务动态指派给人类或机器，进而组建“人-人”、“人-机”、“人-AI”或“机-机”等多元协作团队。为了确保系统的安全性与严谨性，团队在仿真环境中构建了多层安全控制、领域本体知识库（Domain Ontologies）护栏以及决策逻辑层，通过数学与计算建模来模拟未曾预料到的协同模式，为后续的真实场景测试提供理论验证。 ```mermaid flowchart LR accTitle: Human-AI Team Task Orchestration Simulation subgraph Input["任务输入与规划"] T["复杂业务任务"] --> P["LLM 规划器"] P --> D["任务分解<br/>(Decomposition)"] end subgraph Assignment["动态团队分配"] D --> A1["人类 ↔ 人类"] D --> A2["人类 ↔ AI/机器"] D --> A3["机器 ↔ 机器"] end subgraph Guardrails["受控仿真环境"] A1 & A2 & A3 --> H["领域本体护栏<br/>(Ontologies)"] H --> S["安全控制与决策逻辑层"] S --> O["数学/计算模型验证输出"] end ``` ### [07:07](https://www.youtube.com/watch?v=n3OYFFujUgg&t=427s) Michael Caiola：从传统基线模型出发的儿童先心病多模态医疗 AI Michael Caiola 深入剖析了医疗 AI 落地面临的信任挑战。他指出，当前许多人盲目跟风，误以为 AI 就等于 LLM，但在严肃的医疗领域，临床医生绝不会接受一个缺乏临床依据和解释性的黑盒工具。 Credence 正在研发用于小儿认知健康及先天性心脏病的多模态深度学习系统。患儿往往需要根据超声心动图（Echocardiogram）与基因诊断结果来制定差异化治疗方案，但传统的基因检测流程既耗时又极其昂贵。团队通过将患儿的电子健康档案（EHR）与超声心动图影像数据进行特征融合，不仅能实现基因诊断分类，还能给出按置信度排序的潜在疾病候选清单，指导医生优先进行最关键的筛查项目，直接排除了前 10 到 30 个不必要的检测项，大幅削减了医疗成本并改善了患儿的治疗预后。 谈及科研方法论时，Michael 幽默且直率地强调了回归基准的重要性： > "我们总是主张先构建最基础的基线模型。如果我们连一个哪怕做得很烂的传统机器学习模型（a shitty machine learning model）都打不过，那为什么还要强行套上复杂的 AI？" 他告诫行业不要为了追求炫技而直接把 LLM 扔进场景，必须先从传统机器学习和深度学习一步步搭建，严谨评估算力消耗、性能增益与黑盒风险之间的平衡。 ### [08:58](https://www.youtube.com/watch?v=n3OYFFujUgg&t=538s) Ravi Ummadisetti：丰田利用领域知识护栏将 5 年车漆研发周期全面提速 Ravi Ummadisetti 介绍了丰田在工业制造领域利用 AI 重塑材料研发的硬核案例——汽车车漆工程（Paint Engineering）。丰田车身的所有外观色彩均为企业自主研发并拥有专利。在过去，市场部门提出一种全新色彩需求后，科研团队需要耗费大约 5 年的研发测试周期： - **第一年**：科研团队进行大量化学配方调制实验，确定基础成分配比。 - **后续四年**：进行严苛的环境耐久性测试，例如将涂层长期放置在极寒降雪区域与强紫外线暴晒环境中，测试车漆是否会出现剥落、化学变质或腐蚀等问题。 丰田企业 AI 团队介入后，将历史配方数据、真实喷漆机器人的喷涂作业参数以及色彩腐蚀机理全面整合，构建了车漆工程专属的领域知识技术护栏（Domain Knowledge Harness）。借助这一工程护栏赋能给前沿基础模型，并为其配备专用的分析工具与技能（Skills），使模型能够深度挖掘车漆性能趋势、模拟环境腐蚀反应并加速配方优化，成功将长达数年的实体试验大幅压缩为数据驱动的仿真研发。 ### [11:20](https://www.youtube.com/watch?v=n3OYFFujUgg&t=680s) 大模型的“变速箱”隐喻：编排、路由与自我反思 理解当前大语言模型（LLM）究竟能发挥出多大潜力，是当下 AI 研究中极具吸引力的前沿方向。在诸多行业基准测试中，经常会出现令人玩味的现象：某些成本极低的廉价模型在准确率上竟然直接超越了昂贵的旗舰模型；又或者仅需在模型给出答案后多追问一句“你确定吗？”，模型的正确率就会显著提升，这说明模型本身缺失了某些关键的自我验证机制。更有研究表明，在某款代号类似“Table 5”的新一代顶级大模型发布仅两周后，研究人员通过将两个廉价模型进行集成（Ensemble），就在性能上击败了该旗舰大模型。这些现象背后，直指查询分解（Query Decomposition）、模型路由（LLM Routing）以及大模型自我反思（Self-reflection）等核心技术环节的不足。 ```mermaid flowchart TD accTitle: LLM 编排与路由机制流程 subgraph InputStage["用户请求与分发"] Q["复杂用户查询"] --> R["LLM 编排器 / 路由器"] end subgraph ExecutionStage["专业化执行与工具调用"] R --> M1["廉价专用小模型 A"] R --> M2["廉价专用小模型 B"] R --> Tools["外部工具集"] end subgraph VerificationStage["评估与输出"] M1 & M2 & Tools --> V["LLM 验证器 / 自我反思"] V --> Out["高准确度最终输出"] end ``` 这种现状引发了业界对“我们是否真正正确使用了 LLM”的深思：能否专门训练一个善于调度工具、懂得协同分发与结果校验的 LLM 编排器（Orchestrator）、路由器（Router）或验证器（Verifier）？这就像人类制造出了一辆性能强劲的法拉利跑车，却没人发现它其实配有变速箱，所有人一直挂在一档以每小时 30 英里的速度缓慢爬行，还误以为这就是极限；直到有人摸索出换挡的门道，才惊觉这辆车原本就能飙到 100 英里以上。针对模型编排与协同的研究，正是那套能让现有模型跑快五倍的关键“变速箱”。 - **模型表现倒挂**：在特定场景下，经过良好编排的低成本廉价模型组合，在准确率上足以匹敌或超越单体超大规模旗舰模型。 - **自我反思增益**：简单的提示词反思机制（如“Are you sure?”）即可触发模型的重新检验，大幅修正推理盲区。 - **法拉利变速箱比喻**：当前大模型的能力瓶颈很多时候不在于基础参数规模，而在于缺乏高效的工具调用、查询分解与编排架构。 --- ### [13:07](https://www.youtube.com/watch?v=n3OYFFujUgg&t=787s) 算力与成本瓶颈：企业级落地中的控本与本地化集群 当前 AI 浪潮很大程度上建立在“只要堆叠更多算力，模型表现就会更好”的扩展定律（Scaling Laws）之上，但行业正迅速触碰数据和电力的双重物理天花板。从企业级开发者的实际应用来看，许多开发者习惯于直接向云端大模型抛出 Prompt 试图一键生成应用，这导致了惊人的 Token 消耗与飞涨的账单。为此，丰田北美（Toyota North America）等企业开始在工程治理上实行“软性配额限制”，追踪每位开发者执行具体任务时产生的实际成本，并在内部软件生态中构建了企业专属的“Toyota GPT”。 为了在控制云端 API 巨额开销的同时支持前沿研究，企业团队一方面与 OpenAI、Anthropic 等头部机构的 Alpha 产品线紧密合作，探索前沿功能；另一方面积极构建本地算力基础设施。例如，通过将多台 DGX 工作站（如 8 台 DGX Spark 集群）进行整合，搭建本地离线推理集群，专门用于代码智能体（Coding Agents）的 Harness 工程构建与推理测试。这种本地化部署让每位核心研发人员都拥有专属的算力单元，既能摆脱对昂贵前沿模型（如类似 Table 5 或 Claude 系列模型）的高昂依赖，又大幅加快了研发迭代周期。 - **Token 消耗管控**：企业内部对开发者设定软性成本红线，量化每个开发任务的调用开销。 - **私有化模型基座**：构建内部定制的 Toyota GPT，结合自身业务软件生态开发专属功能。 - **本地硬件集群替代方案**：使用 8 节点 DGX 集群承载编码智能体的推理调用，与顶级商业 API 进行效果和成本的横向对比评估。 --- ### [16:09](https://www.youtube.com/watch?v=n3OYFFujUgg&t=969s) 机器人与主动学习：告别暴力堆砌数据的边际递减 在机器人灵巧手操作（Hand Manipulation）等具身智能前沿，大量资金与初创团队正蜂拥而入，试图通过“大力出奇迹”的方式采集海量动作数据以突破技术瓶颈。然而，这种单纯依赖数据规模的暴力堆砌正在重演计算机视觉早期的边际效益递减困境：在图像识别模型研发中，标注前 100 万张图片能让准确率迅速提升至 90%，再增加 100 万张可升至 95%，但要想从 95% 艰难提升到 96%，却往往需要从第 200 万张一路标注到第 1000 万张。 ```mermaid flowchart LR accTitle: 主动学习与海量盲目数据采集对比 subgraph Passive["盲目批量采集（边际效益递减）"] direction TB P1["耗资数千万美元<br/>采购 50,000 小时/月"] --> P2["新增 800 万张冗余样本"] P2 --> P3["准确率仅提升 1%（95% -> 96%）"] end subgraph Active["主动学习（精准切中决策边界）"] direction TB A1["分析模型决策边界<br/>定位最困惑区域"] --> A2["定向采集 50,000 张高价值样本"] A2 --> A3["性能增益超越 800 万张常规数据"] end ``` 突破这一困境的关键在于引入主动学习（Active Learning）。与其盲目抓取海量数据，不如深入分析模型的决策边界，精准找出模型在哪些分布区域最为困惑，并定向采集这些高价值样本。实践表明，针对决策边界采集的 5 万张精选图像，对模型性能的提升效果甚至远超此前盲目堆叠的 800 万张常规数据。当前许多机器人公司每月花费数千万美元，以每小时高达 100 美元的成本采购 5 万小时的操作数据，这种由于缺乏“智能数据筛选”而陷入的停滞，凸显了主动学习在具身智能领域的巨大科研与商业价值。 - **计算机视觉边际递减案例**：100 万样本达 90%，200 万样本达 95%，而 200 万至 1000 万样本仅带来 1% 的微弱提升。 - **主动学习价值倍增**：聚焦决策边界模糊区域的 5 万个高质量样本，效能超越 800 万个盲目采集的常规样本。 - **机器人数据采购痛点**：部分企业每月耗费数千万美元盲目堆砌 5 万小时数据，因缺乏精准筛选机制而陷入进展停滞。 --- ### [17:34](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1054s) 盲目扩大数据中心的质疑：模块化架构与专用小模型 针对“不断建造更大规模的数据中心是否一定能带来更优 LLM”的议题，学术界与工业界正在重新审视其必要性。虽然模型能力在近期并未彻底停滞，但盲目通过消耗海量电力、扩张数据中心换取的基础大模型，本质上只是提升了“通用组件”的泛化能力。然而对于绝大多数终端用户和垂直业务场景而言，实际需求往往具有明确的风险偏好、特定的决策支持目标或严格的自动化生产流程，并非纯粹依赖通用智能。 ```mermaid flowchart TD accTitle: 紧耦合微调与解耦模块化架构对比 subgraph TightCoupling["静态微调模式（维护成本高）"] Base1["特定基础模型 v1"] --> FT["深度全量微调"] FT --> Lock["绑定在静态模型<br/>新模型发布时整体推倒重构"] end subgraph ModularArch["解耦模块化模式（灵活敏捷）"] SLM["任务专用小模型"] & Prompt["后处理与提示链"] --> Router["插拔式评估管线"] Router --> Eval["无缝热替换最新基座（v1 -> v2）"] end ``` 在垂直落地中，哪怕通用底座只提升 1% 的精度，往往也伴随着不成比例的高昂算力成本，此时针对特定场景优化的小语言模型（SLM）具有更高的性价比。此外，深度的全量微调（Fine-tuning）极易导致技术被困在静态模型中，一旦底层模型迭代（例如类似 Fable 5 升级至 Fable 6），整个研发管线就面临推倒重来的窘境。相比之下，采用后处理（Post-training）技术与高内聚低耦合的模块化系统（Modular Systems），能够让企业在底层技术突飞猛进时，实现核心模块的无缝热替换与敏捷评估，更加贴合下游用户的真实业务诉求。 - **通用泛化与垂直业务解耦**：终端用户关注的是特定任务的准确性与低成本，而非通用大模型无差别的大规模计算冗余。 - **微调瓶颈与沉没成本**：过度依赖特定大模型的深度定制微调会导致系统架构僵化，在新模型发布时产生高额替换成本。 - **模块化设计优势**：将系统拆解为小组件，保持模型可插拔性，以便在新型基座模型发布时快速迁移并进行 A/B 测试评估。 --- ### [19:32](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1172s) 端侧智能与符号编码：Mindstone 项目与小模型的探索 在人机交互与端侧智能（Edge AI）领域，研究重心正在从中心化巨型模型向轻量化、自进化的端侧微型架构迁移。开源项目 Mindstone 展现了这种全新的技术路径：它聚焦于构建能够在端侧长期运行、通过与现实环境互动经验不断自我学习的大模型。在该范式下，端侧模型依托“教师-学生模型”知识蒸馏（Teacher-Student Knowledge Distillation）机制充当学生角色，既能在设备本地完成推理与演进，又能与其他智能体通过联邦学习（Federated Learning）进行知识同步，从而在极低资源开销下实现连续学习。 ```mermaid flowchart LR accTitle: Mindstone 端侧智能与符号编码架构 subgraph EdgeDevice["端侧本地学习（Mindstone）"] GNN["图神经网络 (GNN)<br/>三元数据结构"] --> SLM_Agent["端侧小模型 (学生)"] SLM_Agent --> Exp["持续经验学习"] end subgraph Communication["符号与知识同步"] SLM_Agent <--> Semiotic["符号编码器 (Semiotic Encoders)<br/>共享概念模型"] Semiotic <--> Fed["多智能体联邦学习 / 教师蒸馏"] end end ``` 在人机通信与底层表示层面，结合图神经网络（GNN）与前馈循环（Feed-Forward Loop）等架构原语，研究人员正在构建轻量化的“符号编码器”（Semiotic Encoders）。符号学（Semiotics）通过“符号”这一媒介理解人类或机器解释世界的方式，借助极简的三元数据结构，让两个独立智能体之间建立共通的理解模型。这种结合先验规则与物理约束的方法（如物理启发的神经网络与语义损失项），在医疗等高风险、缺乏公开标注数据的专业领域展现出显著优势——无需依赖数十亿（Billions）级别的巨量样本，仅凭数千个（Thousands）带有结构化规则的数据样本即可完成精准建模，开辟了算力受限环境下的全新研究方向。 - **Mindstone 开源项目**：探索端侧长期运行的自适应小模型，结合教师-学生知识蒸馏与联邦学习机制。 - **符号编码器（Semiotic Encoders）**：利用图神经网络与三元数据结构，在不同智能体之间建立极简的共享语义世界模型。 - **先验规则缩减数据规模**：在医疗等高门槛垂直领域，引入物理启发网络（Physics-inspired Neural Networks）或语义损失项（Semantic Loss Terms），使得模型在千级别样本下即可达成以往十亿级数据才能实现的关联效果。 ### [21:59](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1319s) 模型竞赛 vs 治具工程：企业应聚焦私有数据与工具调用 主持人 Jason Hiner 抛出了困扰企业的核心痛点：运行速度、计算成本、投资回报率（ROI）、数据主权（Data Sovereignty）、隐私泄露风险以及模型幻觉。他向在场专家发问：行业的解法是否正在发生范式转移——不再是一味盲目追求参数量更大、算力消耗更高的巨型模型，而是转向结合小语言模型（SLM）、领域专用模型（Domain-Specific Models）以及本地/端侧（On-Device/On-Prem）计算来保障数据安全并提升响应速度？ 丰田北美的智能体 AI 与产品研究负责人 Ravi Ummadisetti 明确回应，从企业落地视角来看，核心战略应当是“聚焦自身拥有的存量数据，而不是紧盯模型本身”。底层基础模型每天都在快速迭代，尤其是工具调用（Tool Calling）和推理规划能力——例如使用 Claude Code 时，若将推理等级调高，模型会在执行动作前进行极其详尽的前置调研并制定出严密的执行计划。企业要做的是将私有数据封装并暴露为标准化的“工具（Tools）”喂给模型，从而极大提高代码生成和业务执行时的置信度与准确率。Ravi 强调，单纯比拼模型推理本身已经不是核心竞争力，围绕模型搭建的支撑体系至关重要，下一波技术浪潮将是基于垂直领域数据的“治具工程”（Harness Engineering）。 ```mermaid flowchart TD accTitle: Enterprise Harness Engineering Flow subgraph DataTools["企业私有数据与工具层"] D1["企业存量私有数据"] --> D2["封装为标准化 Tools / API"] end subgraph AgentHarness["智能体治具与规划层 (Harness)"] H1["前置深度调研与规划 (Reasoning)"] --> H2["任务调度与工具编排 (Orchestration)"] end subgraph FoundationModels["基础通用模型层"] M1["商业大模型 (Claude / GPT 等)"] M2["本地专用小模型 (SLM)"] end D2 --> H2 H1 --> M1 H1 --> M2 H2 --> OUT["高准确率/高置信度业务执行"] ``` --- ### [24:43](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1483s) 火箭还是汽车？大模型编排与垂直领域小模型落地 Thoth AI 首席技术官 Pedro Alves 用了一个极其生动的比喻来剖析大模型厂商与企业客户之间的错位：“造火箭的大厂绝不会停下脚步，他们只会造出越来越快的星际飞船；但他们不会告诉你究竟是该开一辆车去杂货店，还是造一艘巨型火箭去商场。”基础模型巨头没有责任替具体业务定义需求，企业必须自行算清 ROI 账本： 1. **匹配合适动力引擎**：究竟需要千亿参数大模型（LLM）还是轻量化小模型（SLM）？ 2. **明确业务落地方向**：到底是要“开车去超市买菜”，还是要“飞去月球开采难得素（Unobtainium）”？ Pedro 指出，企业需要弄清楚自身拥有哪些数据、希望从中产出什么价值，并基于工具链构建真正能为业务“带来实质性突破（Move the needle）”的应用。针对不同任务与工具之间的调度难题，Pedro 认为大模型编排（LLM Orchestration）领域的研究正是为了解决这一痛点——让系统根据任务复杂度自动判断并路由，无需使用大模型时就降级调用小模型。企业用户无需从零自研这套复杂的编排基础设施，这正是 Thoth AI 这类软件平台所承担的角色，终端用户只需把精力放在具体业务场景的价值挖掘上。 - **模型选用与场景对齐**：根据计算成本和业务精度在 LLM 与 SLM 之间进行动态权衡，避免大炮打蚊子。 - **编排层软件自动化**：借助编排框架自动实现工具连接与模型路由，降低企业底层工程负担。 --- ### [26:27](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1587s) 验证成本困境、算力军备竞赛与小公司的生存抉择 Arcova 的 James OKane 分享了服务 Fortune 100 等强监管行业客户的一线观察。这些行业正在加速将垂直领域的小语言模型（SLM）引入企业内部私有化部署，极少需要向外调用庞大的通用大模型。然而，在引入人机协同（Human-in-the-loop）时，整个行业尚处于早期婴儿阶段。最大的瓶颈在于**人工验证成本过高**——AI 生成内容的成本极其廉价，但业务专家校验真伪的成本却异常昂贵，导致实际操作中人类专家往往沦为“盖橡皮图章（Rubber Stamp）”的形式主义审查。 Credence 资深 AI 研究科学家 Michael Caiola 紧接着指出，推理阶段的实时可验证性（Inference-time Verification）是当前行业面临的重大缺口。若一个模型需要耗费大量人力在离线状态下逐条审计才能判定对错，那就无法在实际业务中建立信任。谈及模型生态格局，Michael 直言不讳地断言：“如果你现在不是 Google、Meta 或 Anthropic，你已经输掉了基础通用模型的战争。” 对于中小企业而言，算力生存现状十分残酷： - **自建算力硬件门槛**：购买单台/单套先进 GPU 的门槛动辄数十万甚至数百万美元，且伴随高昂的年均维护成本，同时还要面对硬件厂商每年推出的万亿级新芯片换代周期。 - **公有云 API 依赖隐患**：目前很多处于“野蛮生长时代（Wild West）”的企业管理层不惜代价推进 AI，甚至出现“单日消耗数亿美元 Token”的极端账单；一旦业务完全绑定在巨头生态中，上游稍有提价就会给企业利润带来重创。 - **微调与私有化运营负担**：选择 SLM 微调路线的企业，必须面临模型私有化托管、持续更新、算法团队维护等一系列沉重负担，导致大量企业目前只能“一瘸一拐地艰难前行”，暂时靠购买订阅账号应付了事。 --- ### [29:33](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1773s) Agentic Harness 狂潮与反思：智能体串联带来的生产环境风险 主持人 Jason Hiner 提到“治具（Harness）”与“智能体治具（Agentic Harness）”已经成为当前的绝对热词，并询问全行业在过去六个月内将研发重心大规模转向智能体系统背后的深层原因与实际变化。 面对全场的智能体狂热，Michael Caiola 给出了极其冷静的反调：“我目前还没有为 Agent 的过度炒作买单。”他坦言在日常软件开发或决策辅助工具中，智能体确实能够提供很大便利，但在真正严肃的**生产环境（Production）**中，现有形态的智能体远未准备就绪： ```mermaid flowchart LR accTitle: Multi-Agent Error Compounding Risk subgraph BlackBoxSystem["多智能体串联黑盒系统"] A1["Agent A (错误率 e1)"] --> A2["Agent B (错误率 e2)"] A2 --> A3["Agent N (错误率 en)"] end BlackBoxSystem --> ERR["复合错误率指数级放大\n故障链路极其难以排查与 Debug"] ``` Michael 指出多智能体系统在工业生产落地中的致命缺陷： - **错误率复合放大**：当系统中串联起成百上千个智能体时，每个智能体自身的固有错误率会层层累积、乘积放大； - **黑盒化与可解释性归零**：在国防军工、医疗诊断等高风险关键领域，用户原本就对单体黑盒模型存有疑虑，若将上千个黑盒相互连接，一旦发生崩溃或逻辑故障，内部 Debug 追踪将变成无法理清的灾难； - **缺乏测试评估治具**：在针对单个组件的边界表现、以及多组件协作的系统级测试评估工具链（Test and Evaluation Harnesses）真正成熟之前，盲目将多 Agent 系统推向关键业务生产线是非常不可靠且极具风险的。 ### [31:59](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1919s) 人机协同的演进与丰田智能体评估体系（Evals） 讨论从人机交互模式的变迁切入。过去一两年中，“人类在环”（Human-in-the-loop）是行业内最热门的话题，而如今人们开始频繁讨论“人类近环”（Human near-the-loop）乃至“人类离环”（Human off-the-loop）。针对 Michael Caiola 认为当前第一代（V1）智能体尚未成熟、难以成为现阶段可靠方案的观点，丰田北美（Toyota North America）的 Ravi Ummadisetti 展开了友好的辩论，表示自己“一半同意，一半反对”。Ravi 强调，任何技术如果无法被最终用户实际使用，其研发就毫无意义，因此用户的满意度至关重要。 为了解决智能体落地中的不可靠问题，丰田正在构建自研的智能体评估体系（Agent Evals）。以丰田制造车间的核心用例“Gearpal”智能体为例——在汽车制造场景中，一旦产线停工，整个工厂就无法造车。现场工程师遇到故障时会向 Gearpal 提问，智能体随即检索向量数据库中数以 TB 计的数据，调用各类工具生成解决方案。为了验证回答质量，团队直接深入产线一线，收集现场工程师的点赞或点踩反馈。与此同时，后台基于 Lang 引擎自动分析运行轨迹（Traces），通过图谱（Graph-based）方法将非结构化数据转化为“技能”（Skills），并反向向开发团队输出诊断建议（例如提示当前向量检索策略存在缺陷需要调整），从而实现根据用户真实痛点驱动的自动化闭环迭代。 ```mermaid flowchart TD accTitle: 丰田制造车间智能体工作流与自动化评估闭环 subgraph S1["产线交互层"] A["现场工程师提问"] --> B["Gearpal 智能体"] B --> C["调用工具与向量检索 (TB级数据)"] C --> D["输出故障排查方案"] D --> E{"工程师反馈<br/>(点赞 / 点踩)"} end subgraph S2["引擎评估与优化层"] B -. "运行轨迹 (Traces)".-> F["Lang 运行时评估引擎"] F --> G["图谱方法:<br/>非结构化数据转技能"] G --> H["自适应优化建议<br/>(如调整检索策略)"] H -. "闭环修正".-> B end ``` - **现场停线风险**：汽车制造车间停机成本极高，智能体必须直接服务于保障产线运转。 - **数据与评估机制**：Gearpal 需要实时检索 TB 级别的向量存储，并通过 Lang 运行时引擎对底层 Traces 进行根因分析。 - **图谱驱动演进**：利用基于图谱的方法将非结构化数据沉淀为结构化技能，并将其作为评估智能体的数据基准。 --- ### [35:18](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2118s) 多智能体测试套件与社会学概念的工程化落地 从智能体的实践落地出发，专家们提出了两种不同的观察视角：一种是战术性的，另一种则是战略与概念性的。在战术层面，团队构建了一个由 35 个智能体组成的测试套件（Harness），其中超过半数的智能体专门充当验证器（Verifiers）、测试器（Testers）和对抗节点（Adversaries）。尽管这类系统目前主要用于前期的原型开发和研发提速，尚未直接部署进生产环境的软件流水线中，但已经展现出显著的辅助价值。 在战略与跨学科层面，智能体系统为社会科学家提供了将抽象社会学理论工程化为多智能体系统的契机。例如犯罪学中的“集体效能感”（Collective Efficacy）概念——指社区居民自我组织以抵御犯罪、维护社会规范与秩序的能力。这一理论正被迁移到多智能体系统的安全架构中：探索多智能体系统能否通过内部自治通信实现自我防卫，形成数字层面的“集体效能”。不过专家们也一致认同，在强监管行业中，智能体的全面落地仍需保持极高的审慎态度。 - **多智能体线架构**：开发阶段运行包含 35 个智能体的测试平台，半数以上承担对抗与校验职能。 - **跨学科理论迁移**：将犯罪学中的“集体效能感”模型引入多智能体通信与安全防御体系。 --- ### [36:56](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2216s) 幻觉检测、智能体通信与集成模型（Ensemble）的权衡 针对多智能体协同与模型可靠性，Pedro Alves 指出了当前研究的三大核心方向，首当其冲的是幻觉检测。当大语言模型（LLM）产生幻觉时，其神经网络内部神经元的激活放电模式（Firing patterns）会出现异常变化，这与人类产生幻觉时的神经生理现象高度相似；同时，模型输出的特定用词和语言组织模式也会呈现出统计学异动。 ```mermaid flowchart LR accTitle: 多智能体协同通信与集成智慧机制 subgraph Agents["基础模型集群 (单体性能普通)"] A1["LLM 智能体 A"] A2["LLM 智能体 B"] A3["LLM 智能体 C"] end subgraph Protocol["结构化规范通信层"] P["多向协同与交叉校验"] end subgraph Output["最终输出效果"] O["超越单一前沿 SOTA 模型的综合表现"] end A1 <--> P A2 <--> P A3 <--> P P --> O ``` 在智能体通信层面，机器学习中经典的“集成学习/群体智慧”（Ensemble / Wisdom of crowds）理论同样适用：只要为智能体之间设计合理的通信协议，由数个性能相对平庸的模型组成的智能体网络，其实际表现完全可以超越单体最顶尖的 SOTA 大模型。但与会者警惕当前市场上某些低效设计——中央调度器将单一任务分发给 500 个智能体，在白白消耗了 300 万 Token 后产出的却是完全错误的结果，这种架构本质上只是在迎合模型供应商“兜售 Token”的商业利益。 Michael Caiola 进一步指出，除了单次推理的准确率，模型的“输出波动性”（Variability）在工业级应用中同样致命。企业用户在实际业务中不可能针对单次推理重复测试 100 次，因此必须向终端用户透明呈现该模型的置信度与失效率（如“每 10 次或 100 次可能出错 1 次”），从而建立合理的人机协作信任心智。针对学术界提出的“自洽性重问（Self-consistency）与多次重采样取均值”方案，专家们调侃这往往伴随着更高的算力成本；若直接质问模型“你确定你的回答正确吗？”，模型极易陷入类似“三门问题”（Monty Hall problem）的过度反思陷阱，因为怀疑自身而重新计算，最终反而输出错误答案。 --- ### [40:06](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2406s) 未来技术演进：数据底座重构与个人专属智能体 主持人 Jason Hiner 向在场研究专家抛出核心问题：站在 AI 研究的最前沿，未来几年普通从业者和企业应当关注哪些即将到来的能力突破与底层变化？ Ravi Ummadisetti 从企业与个人两个维度给出了前瞻性建议： 1. **企业维度（构建 Harness 与改造数据底座）**：企业需要覆盖供应链、研发和制造全链路，积极构建由技能（Skills）、工具（Tools）及 Markdown（MD）文档体系深度融合的智能体底座（Harness）。当前许多企业的数据架构根本无法直接被智能体消费——例如直接在超大型数据库上运行 Text-to-SQL，往往需要等待 5 分钟才能返回结果，延迟极高。企业必须提前重构底层数据，打造适合智能体快速检索与调用的数据基础设施。 2. **个人维度（构建个人专属自动化知识智能体）**：无论从业者处于风险投资（VC）、汽车制造还是医疗健康领域，每个人都应当建立专属的个人智能体。通过智能体自动化管线，每天实时跟踪、消化行业前沿白皮书、学术论文与专利技术，将最新研究成果持续沉淀并反哺给个人工作流。 ### [42:10](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2530s) 企业落地困境：从盲目消耗 Token 到组织内部协同 面对模型推理与应用架构的选择，嘉宾们强调必须在“高算力开销”与“实际工程需求”之间取得平衡。如果无论遇到多么简单的问题，都一律在云端拉起高强度的复杂推理模型，不仅极其浪费 Token，在实际商业落地中也根本行不通。因此，人类工程师在架构设计与分流控制中的把关作用依然至关重要。 Pedro Alves 从 CTO 的视角剖析了自身“研究思维”与“务实思维”之间的撕裂。在研究层面，各种层出不穷的新模型确实令人兴奋；但如果放眼企业“下一步该怎么走”，答案往往缺乏学术研究的光环，而是高度务实的组织协同。当前许多企业对 AI 的态度依然停留在盲目砸钱阶段，甚至允许不同部门每个月烧掉数百万美元的 Token，结果导致内部十几个团队各自为政，重复用 Token 去解决完全相同的业务痛点。Pedro 直言，企业真正需要的不是堆砌新技术，而是成立专门负责落地协调的跨职能团队，打通技术提供方与业务使用方之间的壁垒，把精力放在如何实打实地跑通投资回报率（ROI）上。 ```mermaid flowchart TD accTitle: 企业 AI 落地模式对比 subgraph Inefficient["低效模式：缺乏协同与资源浪费"] A["高层盲目预算投入<br/>'月耗百万 Token'"] --> B1["部门 A 独立调用大模型"] A --> B2["部门 B 独立调用大模型"] A --> B3["部门 C 独立调用大模型"] B1 & B2 & B3 -.-> F["重复造轮子 / 缺乏 ROI"] end subgraph Pragmatic["务实模式：组织协同与统一工具链"] C["统一 AI 协同团队"] --> D["梳理企业共同业务需求"] D --> E["共建针对性 AI 基础工具"] E --> G["可验证的商业回报 (ROI)"] end ``` --- ### [43:40](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2620s) 模型前沿与评估危机：真值数据、扩散架构与合成数据 Michael Caiola 随后梳理了技术演进与行业落地的核心动向。尽管大语言模型（LLM）凭借先发的数据科学沉淀占据了绝对主流，但在文本生成等细分领域，扩散模型（Diffusion Models）等新架构正展现出更强的竞争力，值得持续关注。然而对于终端企业而言，来年最紧迫的挑战并非追逐基础架构更迭，而是建立能力评估机制。企业开始意识到，绝不能把类似 Claude Code 这样的前沿编码工具随手扔给单一开发者后就任其“野蛮生长”，而是需要系统性地组建能力评估团队——这在某种程度上重现了十年前数据科学团队的建设周期。 在国防与医疗等高敏感垂直领域，另一个核心痛点在于客户往往缺乏“真值”（Ground Truth）与验证数据。如果客户自身无法提供基准数据集，也不清楚该如何判定工具的正确性，算法提供商根本无法凭空证明模型性能。因此，行业正在形成新的共识：企业客户必须先理清自身数据资产并构建评测集。此外，针对业界关于“LLM 撞上性能天花板”以及“人类数据即将耗尽”的悲观论调，Michael 明确表达了乐观态度，认为高质量的合成数据（Synthetic Data）与精细化清洗将填补缺口；人类应当充分发挥机器挖掘未知特征模式的长处，再配合专家复核，这种人机协同增强在未来很长一段时间内依然是技术发展的主线。 --- ### [45:59](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2759s) 医疗 AI 的可解释性破局：从特征嵌入到图像 Patch 的可视验证 现场问答环节开启后，一位观众提出了医疗场景落地的经典难题：许多临床医生甚至连线性回归或决策树都难以接受，面对黑盒属性更强的 LLM 或超大规模深度学习模型，究竟该如何向医生解释其输出结果？ Michael Caiola 结合一线临床经验指出，医生群体普遍具有强烈的经验偏见与认知防线。他分享了一个生动的真实经历：他曾与一位临床医生共同研读一篇关于某罕见病新临床试验的顶刊论文，该医生当场嗤之以鼻表示“这种病我亲自治过 5 例，这篇论文完全是胡扯”，哪怕该研究基于多达 10,000 名患者的大样本临床数据。为了打破这种固有偏见，算法团队在向医生交付系统（如心超超声心动图分析项目）时，必须采用层层递进的“逆向证明”可视化策略： - **基准测试匹配**：首先在权威评测集上展示模型与临床诊断标准的准确率对齐，建立最基础的准入门槛； - **潜空间聚类展示**：在特征嵌入空间（Embedding Space）中将具有相同基因诊断的相似病患聚类可视化，直观向医生证明待测病患与已知同类病例的空间关联； - **图像 Patch 级归因**：借助 Vision Transformer 等视觉模型的注意力机制，精准高亮标出是心超图像中的哪一块局部图像切片（Patch）触发了该项诊断判定，让模型的决策依据与医生的病理直觉形成共振。 --- ### [47:55](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2875s) 氛围编码（Vibe Coding）的现实：从放大核心人才到严苛的需求工程 现场中小企业代表随后发问：在没有预算组建庞大算法研发团队的情况下，中小企业该如何运用 LLM 或“氛围编码”（Vibe Coding）去构建确定性系统？ Pedro Alves 率先指出，业界对于 AI 编程存在认知误区。Vibe Coding 的核心价值不在于“技能托底”（Upskilling，试图把平庸的工程师变成卓越专家），而在于“产能倍增”（Outskilling，让一名卓越的工程师借助 AI 拥有 10 倍的工程输出）。LLM 本身缺乏真正的创造力火花，当开发者试图将一篇前沿论文转化为落地系统时，模型往往无法自主推导逻辑下一步；但只要人类架构师给出清晰的顶层方案，LLM 就能极其高效地完成具体实现。 ```mermaid flowchart LR accTitle: Vibe Coding 研发与需求闭环 subgraph SpecStage["前期准备：需求工程 (2-3周)"] A1["撰写高精度.md 规格需求"] --> A2["定义不可逾越的业务逻辑边界"] end subgraph DevStage["自动化开发与双模型审查"] B1["代码生成模型 (如 Codex)<br/>执行编码与提 PR"] --> B2["评审模型 (如 Claude)<br/>对照.md 需求自动化审查"] end SpecStage --> DevStage ``` Ravi Ummadisetti 补充指出，中小企业应当把 80% 的人类精力投入到前期的需求工程（Requirements Engineering）中。在动用 AI 编写代码之前，团队应该花两到三周时间撰写极其严谨详尽的 Markdown（`.md`）需求文档，给模型划定严格的逻辑边界与确定性循环；进入开发阶段后，甚至可以构建“双模型闭环”：由 Codex 等工具负责编写代码生成 Pull Request，再交由 Claude 对照前期需求文档进行代码审查（PR Review）。 Michael Caiola 最后提醒开发者要警惕“无脑复制粘贴”的伪提效陷阱。大模型的真正价值在于双向互动中的学习与技能沉淀（当代码报错时深入追问为什么出错）。他举了一个令人啼笑皆非的反面案例：团队曾要求一名程序员为运动障碍视频分类器编写代码，在遇到报错后，该程序员直接把 Michael 私聊询问的“昨天我们讨论的方案试过了吗？”原封不动粘贴进 Claude，随后又把 Claude 回复的“我不知道你昨天说了什么”原样复制回工作群——这种对业务角色与底层代码完全缺乏理解的盲目依赖，正是企业推行 AI 辅助开发时最需警惕的风险。 ### [51:37](https://www.youtube.com/watch?v=n3OYFFujUgg&t=3097s) 警惕“扮演”领域专家的虚假繁荣与黑盒风险 讨论接近尾声时，讲者针对当前 AI 普及过程中出现的一种浮躁现象表达了深切担忧。他指出，随着大模型能够极其迅速地生成华丽且看似专业的回答，行业里涌现出大量实际上缺乏深厚功底的人，正在借此“扮演或假装成领域专家”（cosplaying as subject matter experts）。这种由 AI 工具带来的表面效率和光鲜结果往往掩盖了专业深度的缺失，导致许多人在未真正掌握核心原理的情况下，误以为自己已经掌握了相应领域的话语权。 真正的专业能力在于能够深入底层逻辑，明确系统在何种边界条件下会失效、为何会出错。讲者打比方道，如果自己甚至无法向自己的孩子解释某项技术具体为何能够生效，而仅仅是将其视作一种玄妙的“魔法”，那么依赖这种黑盒输出在本质上就与“掷骰子碰运气”毫无区别。讲者最后结合会议现场幽默地打趣称，鉴于大家此刻正身处赌场（Casino），碰运气摇骰子倒也算应景，但这无疑深刻揭示了盲目依赖黑盒 AI 所带来的不可控风险与理解鸿沟。 - **伪专家的泛滥**：AI 生成的快速、光鲜回答让非专业人士极易产生专业能力的错觉，但其缺乏对系统底层逻辑与失效边界的实质认知。 - **黑盒依赖与掷骰子机制**：无法合理解释与验证的 AI 输出等同于“黑盒魔法”，将其直接投入应用与赌博碰运气无异。 ## Notable quotes > [07:07](https://www.youtube.com/watch?v=n3OYFFujUgg&t=427s) 我们总是主张先构建最基础的基线模型。如果我们连一个哪怕做得很烂的传统机器学习模型（a shitty machine learning model）都打不过，那为什么还要强行套上复杂的 AI？ — Michael Caiola > [24:43](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1483s) 造火箭的大厂绝不会停下脚步，他们只会造出越来越快的星际飞船；但他们不会告诉你究竟是该开一辆车去杂货店，还是造一艘巨型火箭去商场。 — Pedro Alves > [21:59](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1319s) 聚焦自身拥有的存量数据，而不是紧盯模型本身。 — Ravi Ummadisetti > [26:27](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1587s) 如果你现在不是 Google、Meta 或 Anthropic，你已经输掉了基础通用模型的战争。 — Michael Caiola > [29:33](https://www.youtube.com/watch?v=n3OYFFujUgg&t=1773s) 我目前还没有为 Agent 的过度炒作买单。 — Michael Caiola > [45:59](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2759s) 这种病我亲自治过 5 例，这篇论文完全是胡扯。 — Michael Caiola > [47:55](https://www.youtube.com/watch?v=n3OYFFujUgg&t=2875s) Vibe Coding 的核心价值不在于“技能托底”（Upskilling，试图把平庸的工程师变成卓越专家），而在于“产能倍增”（Outskilling，让一名卓越的工程师借助 AI 拥有 10 倍的工程输出）。 — Pedro Alves > [51:37](https://www.youtube.com/watch?v=n3OYFFujUgg&t=3097s) 扮演或假装成领域专家（cosplaying as subject matter experts）。 — Michael Caiola ## Mentioned Jason Hiner, The Deep View, Pedro Alves, Thoth AI, Ravi Ummadisetti, Toyota North America, Toyota, Michael Caiola, Credence, James OKane, Arcova, Boeing Research and Technology, FDA, Toyota GPT, OpenAI, Anthropic, Claude, Claude Code, Google, Meta, DGX Spark, Mindstone, Gearpal, Lang, Codex, Vision Transformer ## Video description Alan Brown - Senior Research Scientist, Boeing Research and Technology Michael Caiola - Senior AI Research Scientist, Credence James OKane - Director, Human AI Services, Arcova Ravi Ummadisetti - Head of Agentic AI & Product Research, Toyota North America Pedro Alves - CTO, Thoth AI Jason Hiner - Editor-in-Chief, The Deep View AI research is moving at unprecedented speed, with advances in large language models, multimodal systems, world models, and agent architectures reshaping what’s possible. Yet key challenges remain unsolved: reasoning reliability, robustness under distributional shift, interpretability, and efficiency at scale. This panel convenes leading AI research heads from industry and academia to reflect on the current state of the field, the breakthroughs that are shaping near-term industry adoption, and the fundamental questions that remain open. Panelists will discuss how cutting-edge research is guiding enterprise applications, where academia and industry must collaborate more deeply, and what innovations might define the next wave—from scalable alignment methods to architectures beyond transformers. Attendees will gain an insider’s perspective on how today’s research frontiers are steering the future of AI in both science and industry. [Source](https://www.youtube.com/watch?v=n3OYFFujUgg)

[7:07] Michael Caiola: 我们总是主张先构建最基础的基线模型。如果我们连一个哪怕做得很烂的传统机器学习模型（a shitty machine learning model）都打不过，那为什么还要强行套上复杂的 AI？ [24:43] Pedro Alves: 造火箭的大厂绝不会停下脚步，他们只会造出越来越快的星际飞船；但他们不会告诉你究竟是该开一辆车去杂货店，还是造一艘巨型火箭去商场。 [21:59] Ravi Ummadisetti: 聚焦自身拥有的存量数据，而不是紧盯模型本身。 [26:27] Michael Caiola: 如果你现在不是 Google、Meta 或 Anthropic，你已经输掉了基础通用模型的战争。 [29:33] Michael Caiola: 我目前还没有为 Agent 的过度炒作买单。 [45:59] Michael Caiola: 这种病我亲自治过 5 例，这篇论文完全是胡扯。 [47:55] Pedro Alves: Vibe Coding 的核心价值不在于“技能托底”（Upskilling，试图把平庸的工程师变成卓越专家），而在于“产能倍增”（Outskilling，让一名卓越的工程师借助 AI 拥有 10 倍的工程输出）。 [51:37] Michael Caiola: 扮演或假装成领域专家（cosplaying as subject matter experts）。

[5:27] James OKane: human AI systems... are not stable [7:07] Michael Caiola: everyone wants AI, but everyone forgets AI is a huge thing... it's not just LLMs. [7:07] Michael Caiola: Like I always say... you want to build the most baseline model first, right? If we can't beat a shitty machine learning model, right? Then why are we putting AI on it? Excuse my language... but that's the first step, right? And then can we build deep learning on top of that? And does it do better? How much better does it do, right? Versus the compute we're doing here versus the black-box part we're going to attach to this? [11:20] Pedro Alves: rhymes with Table 5 [17:34] probably going to poison the earth for that [24:42] Pedro Alves: the next rocket ship that's faster and faster [27:26] Michael Caiola: If you're not Google, Meta, or Anthropic right now, you've lost the fight for a foundational model. [36:56] Michael Caiola: the business is selling you tokens. [36:56] Are you sure you are correct? [47:55] I don't know what you said yesterday [51:37] pretending or cosplaying as subject matter experts.

[0:04] Uh I'm Jason Heiner with Deep View and we have a  a wonderful panel here to talk about AI research. [0:12] uh you know so much is happening uh in AI and the  research world uh often you know moves at a a very [0:24] steady scientific pace right but everything that's  happening in AI is powered by the research and so [0:32] this is a a really unique opportunity for you to  get to hear from some of the researchers behind [0:38] the scenes who are working on the innovations  that are powering all of the innovations that [0:45] are unfolding before us and driving forward uh  both the technology industry um and society at [0:54] this incredible pace that we're all very very  aware of right so we're talking about everything [0:59] from LLMs world models multimodal uh and all of  the advances uh as well as the challenges that [1:09] uh we're now facing right there are challenges  challenges uh that uh have to do with the ways [1:16] that we can observe models, the ways we understand  them, the ways it can overcome the challenges that [1:22] uh these models face. And so the researchers uh  today will unfold some of that for you and help [1:30] us all think about ways that we should understand  the technology and understand the course of where [1:38] it's going both the science and the industry. And  so this is uh a really special opportunity to get [1:45] to hear from um a panel of amazing experts. So,  I'm going to let them introduce themselves, talk a [1:52] little bit about uh you know, who they are, their  background, and where they're from. So, Pedro, go [1:58] ahead. Pedro Alves. I uh been doing AI and machine  learning for about uh 24 25 years. I started off [2:06] in genomics and I moved around to computer vision  and um social networking, Wall Street trading and [2:14] then landed where I am now which is I'm working  with LLMs specifically LLM orchestration LLM [2:20] routing and then on the robotics side how do  you generate the best data sets for robotics? So [2:27] using techniques like active learning where you're  trying to feed the errors of the robot back into [2:32] the data set in order to get a better data for  you know embodied AI. So cool. Uh hey everyone [2:40] I'm I'm Rabi from Toyota. I head the agentic AI  and product research. My background is I started [2:46] as a developer and then I've been working in  industry for almost like 15 years just focusing on [2:53] uh building AI for the real business problems. So  I lead the I lead the uh agentic AI and product [3:01] research at Toyota where I'm responsible for  every AI agents that's creating across Toyota [3:06] from manufacturing supply chain R&D as well as all  the divisions. So we are responsible to make sure [3:12] that all of them is like from an uh from an agent  research standpoint as well as from the production [3:19] productionizing agent standpoint. So we need to  be accountable for how does each agent flow and [3:24] all of that. So I would cover all of them in this  session. Great to meet you all. Hey everybody. Uh [3:32] Mike Kayola from Credence. I'm a senior biomedical  AI engineer over there. Um coming very much from [3:37] the hard scientist side. uh doctorates in  mathematics but then spent a bunch of time [3:41] doing neuroscience both in uh wet primate labs  and then um more on the human translatable spots [3:48] with FDA. Um spend decent amount of time in the  federal government. Um mostly kind of fill in the [3:53] role of can we put experimentation towards AI that  gap that seemed to just open up all of a sudden [3:59] since we all forgot how to do experimentation um  after machine learning and magic AI came out. Um [4:04] so yeah mostly focus on being riskaware for uh  customers. Hey, good morning everybody. Uh James [4:11] Okaine. Uh my research background is primarily uh  social psychology. I focused on crime prevention [4:18] and uh human information processing. Uh my tech  my background in industry about 25 plus years [4:26] focused on technology risk cyber security uh  either in a consulting capacity or operations [4:32] and even research as well. So um currently work  at AROVA uh we provide cyber security services. [4:40] One of the key area we're focused on right now  is uh nextgen workforce design and engineering [4:46] uh for the uh the workforce. Um I'm also a  co-founder of a preede startup where we're [4:52] focused on human machine communication. So thank  you. Very good. So I have some questions for our [4:59] panelists. We'll we'll go through and share uh  some information. obviously very excited to to [5:04] talk to all of these researchers and get this uh  really interesting look behind the scenes. We'll [5:09] have some time at the end for Q&A as well so that  um you all can answer you can ask some questions [5:15] for for them to answer. Uh so as we go through  this if you have questions that arise um keep [5:21] them in mind and at the end we'll we'll leave some  time uh for your questions as well. All right. So, [5:27] let's start by talking uh a little bit about,  you know, one AI research initiative that you're [5:35] working on now or recently completed. Um, you  know, what's the core problem that you're trying [5:42] to solve uh in the real world application? And  why don't we start in reverse order uh this time? [5:51] Okay. Yeah. One of the one of the problems  we're looking at is um the the notion of [5:57] human AI systems and they're they're not stable.  Okay, they're they're difficult to predict. Um [6:04] so we're building um some simulation models that  really focus on um creating synthetic data and uh [6:12] what we're focused on currently is LLM models  as planners. Um the LLM will build um we'll [6:19] start doing task assignment to uh break down the  tasks task de composition. We'll assign the task [6:26] to human or machine. Um and then assemble teams  whether those seams are human to human, human to [6:33] machine, human to AI or machine to machine. Um  and then we build like a controlled environment [6:39] uh that has number one security controls because  we're a security firm but also look at the [6:44] um domain ontologies that can be used as a  harness around that and also some decision logic [6:50] layers. Uh so it's a way that we can start  to simulate human machine teaming and then [6:56] uh build mathematical and computational models  that we may not have thought of that we can then [7:01] go further test. So a little bit of what's top of  mind currently. Cool. Uh yeah, I'll talk about uh [7:07] some biomedical stuff then real fast. Uh the  biggest thing we're running into right now is [7:11] that everyone wants AI, but everyone forgets AI  is a huge thing. Um it's not just LLMs. Uh so [7:17] especially in the medical field where they need  to have this like insight to what's happening and [7:21] explanability. Collision clinician is not going to  use your tool if it's a black box and there's no [7:25] like actual clinical results behind it, right?  Um so we're building more like deep learning [7:29] models right now for like multimodal ingestion of  data. So uh we have one right now for pediatric uh [7:35] cognitive health um heart disease right where we  have echo cardiograms from from children who have [7:39] diseases and they have different genetic diagnoses  from the clinicians u but to actually get those [7:44] genetic diagnosis tested it requires either  a lengthy or costly genetic test essentially [7:49] right and they can treat them differently based on  the genetic tests. So uh can we do with just their [7:53] health record just their echo cardiograms and  fusing those together can we get to a spot where [7:57] we can give that um classification for example  of genetic diagnosis and can we go one step [8:02] further and can we give a short list rank order of  probabilities of diagnosis and now from that can [8:06] we go here's the test you probably should do first  right uh and this will give you like this will [8:11] cross off the first 10 or the first 30 or whatever  it is right and can that then get you to a spot [8:14] where we do more cost-effective or even better  uh patient outcomes for patients for example and [8:19] that's all done with like starting with machine  learning models, right? Um like I always say [8:24] like we you want to build the most baseline model  first, right? If we can't beat a shitty machine [8:30] learning model, right? Then why are we putting AI  on it? Uh excuse my language. Uh [laughter] but [8:34] like that's the first step, right? And then can  we build deep learn on top of that? And does it do [8:37] better? How much better does it do, right? Verse  the compute we're doing here versus the blackbox [8:41] part we're going to attach to this, right? And  then can we build on top of that? Right? If I [8:44] have a lot of people now who just wants to throw  LLMs at it and it might be like quick and flashy, [8:48] but the question is like how do we know it breaks?  how we know what's working all that kind of stuff. [8:52] So we build from the ground up. Cool. So I think  from an productionizing standpoint I would talk [8:58] about I mean we have a lot of use cases but be  interesting one is paints. Uh if you look at [9:04] Toyota cars in the and you like every day I think  the colors that you see on the on the car is all [9:10] we we create all of them. So basically from the  scratch we design them. We create all of them. [9:15] Basically it's all our patent uh colors. Today  I think if you want to design a design a color [9:21] basically if our marketing marketing team uh comes  and give to our research team saying we want this [9:26] color. So our research team goes and does a lot  of research in trying to uh produce that color on [9:31] the car. The the amount of research that we spend  is almost approximately 5 years just to design a [9:38] color mean like create a new color. The reason why  it's 5 years is we thorough a lot of we do a lot [9:43] of thorough testing meaning for one year we put  inside the snow regions we put inside the sunlight [9:49] we we make sure that the paint is not getting  peeled off we make sure that uh the all the [9:54] chemicals are good so in the first year itself we  will try to fix our formula and then the next four [9:59] years we we do a lot of testing analysis so what  we are trying to do is uh we at enterprise AI we [10:05] are trying to uh pitch in and then trying to take  that data that we have inside like the create for [10:12] creating formula we try to create a simulation and  also we are looking at the real robotics uh like I [10:17] would say robots that produces colors on the paint  we take that data as well and then we're trying to [10:22] learn from the backwards basically we are trying  to create our own harness engineering uh for our [10:29] own paint so that we can able to give this to  these new models like fable 5 to figure out uh [10:36] what are the some of these uh trends that you that  they look at that and we need to focus on that's [10:42] I think um uh to go more deeper we are trying  to create our mo our our harness for all our [10:50] uh domain knowledge from a paint standpoint you  look at uh how the corrosion is also coming out [10:56] from these particular paints as well so uh from an  end to end aspect I think this is one of the cool [11:00] problem that we have solved uh that almost like  they say the language models are really capable [11:04] enough basically you should prod you should  give them proper tools uh and then proper I [11:10] would say skills so that it can able to identify  some of these cool research that uh we are doing [11:16] it. So I would say that's a cool problem that we  have solved so far. I mean there are so many but [11:20] that's one of the one. So um understanding what  you can actually get out of LLMs today I think [11:28] is a is a really interesting area of research.  So if you look at some of the pay attention to [11:34] what the research is coming out and you see that  in some benchmarks the cheaper LLMs actually beat [11:40] not in cost but in accuracy the expensive LLMs  or when you see a paper come out that says hey [11:45] if you just ask the exact same LLM that gave you  the answer are you sure it increases its accuracy [11:52] right so it's missing some step and then when you  see a new LLM come out the name of this new LLM [11:58] rhymes with table five and 2 weeks after that LLM  came out, some other researchers said, hey, if you [12:04] just ensemble two cheaper LLMs, we beat table 5,  right? So, you start putting that together, you [12:11] understand, okay, there's an issue here with query  decomposition, there's an issue here with perhaps [12:16] um LLM routing. There's potentially an issue  here with the self uh uh reflection of an LLM, [12:23] right? Which brings you to the research of  are we really using the LLMs properly? So, [12:29] can you train a LLM orchestrator, right, or an LLM  router or an LLM verifier that actually knows how [12:36] to use the tools? Because the the analogy here  is, you know, you build this new LLM is like a [12:41] Ferrari and nobody figured out that there's like  gearbox and you're stuck in first gear and you [12:46] think the maximum speed is there 30 miles an  hour and you realize, oh, wait a minute, oh, [12:50] it can go 100 miles an hour. And so that research,  right, of orchestration, I think is the analogy [12:54] of the the gearbox of figuring out the car that  you already built can go five times faster than [12:59] you think you can. Love that analogy. Uh let's  talk about one of the elephants in the room, [13:07] scaling loss. Uh you know, one the whole current  AI revolution is based on the idea that if we [13:15] throw more compute at problems, things get better.  uh but we're starting to hit some of the limits [13:21] and the challenges of both data and power. And so  what are some of the you know architectural and [13:33] um other uh breakthroughs that you all are finding  that can address some of the challenges with with [13:40] scaling laws? Anybody want to take that one  first? Sure. I can go. uh from an enterprise [13:46] standpoint I think a lot of developers right now  I think basically you throw a question a prompt [13:51] to a cloud and then say just build this app it's  just going building the app but I think there's a [13:56] lot of tokens consumption for the for uh if you go  with cloud it's a lot more consumption than codex [14:02] codeex does a little bit less because the cost as  well so we are also from an enterprise standpoint [14:08] uh we are also making sure that we have uh  basically stopped kept a limit limit for [14:14] every developer a soft limit not a hard limit  but making sure that uh this is the task that [14:20] you have performed and this is how this is the  how much cost that it consumed. So we have also [14:25] built our own I would say uh Toyota GPT uh which  is uh which is like a chat GPT version of it. Um [14:34] we use our own existing software to build our own  functionalities as well. to be honest we also work [14:40] with a lot of the alpha alpha team member alpha  products from open AI and cloud anthropic too [14:45] uh to identify some of these new features  that's coming out that that can help enterprise [14:50] uh solve this big uh scaling problem and on the  other hand uh my team is also researching in lot [14:55] of these DJX sparks uh we are trying to cluster  some of the DJ sparks and then use that as an LLM [15:02] like inference call so that rather than going to a  for example if I'm doing some research on a coding [15:08] agent or let's say if I'm trying to uh build my  own hardness engineering and then I try to ask go [15:14] and do this solve this problem giving giving all  the tools to the coding agents like codex or the [15:20] the harness of codex. So we are trying to use  these inference models by combining maybe like [15:26] eight G eight DX sparks and using that as an  inference and giving that to LM so that rather [15:32] than spending lot of cost in Fable 5 or um new  models like SOL 5.6 so we're using these models to [15:41] uh do these research and then see how is the  comparison between the if I go with table five [15:46] versus this one. So finally I think uh uh I would  say the lot of I would say every developer now in [15:52] my team has their own DJ spark so that they  don't need to spend a lot of time and then [15:56] uh spin up there but yeah I think it's it's not uh  yet solved from an full enterprise standpoint we [16:02] are still working on researching a lot of these uh  existing hardware too from from a data perspective [16:09] at least in robotics there's this massive  rush of companies and money being thrown at it [16:15] um to try to solve the problem of building robots  that can do hand manipulation tasks really well [16:21] and you know it was similar when when people were  trying to get data for computer vision I remember [16:26] building a model that the first million images  I got labeled got me to 90% the next million [16:32] 95 and it took me from the second million to the  10th million to go from 90 uh 95 to 96 right the [16:39] diminishing returns and then I started researching  active learning which is don't just grab data you [16:46] actually look at the decision boundary of the  model you built. Where's in the area of the [16:50] decision boundary that the model is most confused  about? And you get that data and the next 50,000 [16:56] images were more beneficial to the model than  that second to 10 million. So the 8 million [17:02] images benefited the model less than the 50,000  images using active learning. So, you're talking [17:07] about like scaling and with robotics, I believe  it's the same thing. And that's why we're doing [17:12] some research in uh in robotics specifically for  active learning because some of these companies [17:17] are buying 50,000 hours a month and they pay $100  sometimes an hour. You're talking about, you know, [17:23] tens or dozens of millions of just buying data.  Yeah. And it's just stalling because they're not [17:28] trying to get smart data. They're just buying  volume bulk. Yeah. That's you go back then. [17:34] I'll go. Okay. That's a good point. I think uh I  think there the general question of scalability is [17:38] uh worth discussing in general I guess right um do  we want to even do we build a bigger data center [17:44] to get a better LLM right is like a general  question of capabilities right if you asked me [17:48] last year I would say we were probably plateauing  a little bit I've been surprised this year with [17:51] some of the coin breakthroughs and stuff um some  of the models have been better so I I'll take that [17:55] back for the second now but uh the question is  going to be like a bigger even if a bigger data [18:00] center makes a better LLM it just makes a better  general component to a thing right Um, a lot of [18:06] people's uses aren't general LLM purposes, right?  You have a specific use case, you have a specific [18:11] end user who has a certain risk profile and they  want to use it for decision-m or support or maybe [18:17] as a full automatic production pipeline. Those  are different things, right? And maybe a general [18:21] LLM with better data capability might plug into  that a little better. To Pedra's point, maybe [18:25] it's only 1% better, right? And is that going to  help you here? or maybe we should be going closer [18:30] to maybe small language models that are better  for your particular use case that's cheaper, [18:33] it's more cost-effective, for example, right?  Um this different questions that we should be [18:37] asking like they want to build data centers and  try this. Um they're probably going to poison [18:42] the earth for that, but sure. Um it might give  us some capability for it, but I think there's [18:47] there's other components that we could do. I  don't want to even advocate for fine tune and [18:50] stuff since you get into a bottleneck of like  now you're stuck in that static model and then [18:53] you have to throw that away again. But there are  post- drain things you can do or take things in [18:58] smaller components and build modular systems for  example. If I spend all my time working on Fable [19:04] 5 or whatever the the latest and greatest is  for example, right? And then Fable 6 drops, [19:08] right? Am I in a spot now where I need to  rebuild my entire pipeline or have I been [19:12] building this module the entire time so I can  put a new one in and test and how it evaluates? [19:16] I think scalability is one of the questions that  we should be asking. There's like a lot of things [19:19] that at by the time we get to downstream for  example like the end user isn't caring about [19:25] that general LM capability as much as like for  my task I want to work better or worse right [19:32] yeah I think what we're looking at at least  in the context of human machine communication [19:37] um training on the edge you know training data for  the edge um getting much more tiny um and focused [19:45] um so when um what we have is a long we're do this  is an open source project um called Mindstone just [19:53] kind of got started and uh but it's really focused  on a longunning LLM that is learning through [20:00] experience and then the idea here is you could  potentially use that as a small language model um [20:05] a if you think of a knowledge distillation between  a uh you know teacher student paradigm this would [20:11] be a student that could do some federated learning  with other agents but also on the device so that's [20:17] um it's called Mindstone. Um we're just kind of  getting started on there. That's that's one piece. [20:22] The other is when we think um about human machine  communication or just uh in general um in the in [20:28] the context of the the preceded we're working on  it's um we're focused on uh graph neural networks [20:35] and and um and some architectural primitives  like a feed forward loop. Okay, very useful [20:41] um you know triadic data structure in a sense.  Um so we're looking at um building um semiotic [20:47] encoders. Uh I won't get into semiotics right  now but um it's a way to kind of understand how [20:54] either a human or machine interprets its world um  through signs and so um getting very again very [21:01] uh minimalist u and then developing a shared model  between let's say two agents so they can actually [21:07] communicate. So we're we're I'm more on the like  kind of small end, you know. Uh but so when we [21:13] think about scaling, that's just like a whole  another direction there. Can I add real fast to [21:18] that? It's a great point. I think there's with a  lack of data, especially like in the medical field [21:22] and some other fields that have like higher risk  data that isn't usable across things like having [21:26] a baseline piece of data, right? You see this a  lot with like the physics and spiral neural nets [21:30] that came out, right? Or like semantic loss terms  on on error pieces here, right? If you have an [21:34] understanding of the system beforehand, you can  get away with not doing a billion N, right? You [21:39] maybe have a thousand N, but those rules are in  place to kind of help correlate things together, [21:43] which I think is Yeah, James, I'm glad you  brought up small language models and Pedro, [21:48] I'm glad you brought up local processing, too,  because as I as I talk to people in industry, um, [21:54] you know, people that are trying to apply these  things, uh, especially over the last few months, [21:59] a number of problems come up over and over again.  speed, cost, um you know, efficiency or or ROI, [22:08] um and and then also, uh the issue of  sort of data sovereignty and and you know, [22:15] concerns about privacy and and hallucinations.  Sorry, that's a lot of that's a lot of things, [22:20] but they all those concerns come up those same  concerns come up over and over again, right? And [22:25] so I I would love to to ask your all's perspective  is is part of the the industry's answer evolving [22:33] to not just we're going to keep building bigger  models with more parameters and we're going to [22:39] throw more compute at them and that's going to  solve all the problems to more that we're going [22:44] to use SLMs domain specific models. Um Ravi you  mentioned you you all are building your own model [22:52] you know that that's also I I guess you could call  domain specific model um and then we're going to [22:57] use also local compute you know because then we  have sort of data sovereignty um it can be it can [23:03] be faster it can be that so is that combination  of smaller more specific models and ondevice [23:10] processing or local processing even you know onrem  are those part of the solution that's emerging to [23:18] to these challenges. I think uh from an enterprise  standpoint, I would say you should focus on your [23:26] existing data that you have rather than focusing  on the models because the models will improve day [23:31] every single day. Basically, they're training more  on how the tool calling is happening. How does [23:37] uh research, how does the search needs to  happen? If you have an uh if you go to cloud [23:42] coder and say the reasoning very high it's it's  trying to basically do a lot of research ahead [23:48] of time basically good have a good plan before it  executes. So that particular execution of the plan [23:54] is very important. So for that particular plan  we should feed our own data as a tools meaning [23:59] we should uh expose our existing data as a tools  to these existing models so that the amount of [24:07] accuracy or the confidence in executing the code  is also high uh when when you're building any any [24:15] any product right so I think uh from an enterprise  standpoint I think we should uh focus more on [24:21] um looking at our own data and then focus more  on how how I can able to enable these existing [24:27] models which are like damn good at doing all of  these things so that it can able to improve more [24:32] basically it's not just like it's not just  inference but it's also about the uh I would [24:37] say harness around it is important so the next  thing the next big thing will be more of harness [24:42] engineering on your own domain specific models I  mean the the you know the big companies that are [24:49] putting out these models it's not their job or  business to tell you what to do with it They're [24:54] they're never going to stop building the next  rocket ship that's faster and faster, right? It's [25:00] your job to figure out, do you need an engine to  take you to the grocery store or some giant rocket [25:05] to take you to the mall? They're not going to tell  you that and they're not going to stop building a [25:09] bigger and a faster rocket, right? So then, you  know, in order to take that into, like you said, [25:15] ROI, there's two components you need. One is which  type of engine that these companies are building [25:21] do you need? Do you need an SML LLM or SML SLM?  LLM. Um, and then once you choose the one, right, [25:31] what are you going to do with it, right? Are you  going to drive the car to get groceries or are you [25:35] going to go try to mine unoptanium from the moon,  right? And and that translates to figuring out, [25:41] like you were mentioning, within your company,  what's the data that you have, what do you want [25:44] to get out of it, right? and how do you build  something on top of those tools that's actually [25:49] going to move the needle for your company,  right? I think the uh LLM orchestration uh type [25:56] of research I think is helps you um automate that  aspect of figuring out you know to do something [26:03] simple. It's going to automatically tell you and  understand you don't need this LLM, you need this [26:07] one and how to like orchestrate the the different  tools together. Uh I don't think a company needs [26:13] to figure out how to how to build that. I think  companies like Thoth, like us, we're building [26:17] that uh software infrastructure. I think what then  remains to the to the final end user is to just [26:24] figure out what are you going to do with it. Yeah,  I think a lot of our customers are um Fortune 100 [26:30] highly regulated. So, we're starting to see a  lot of small language models coming in-house [26:35] um very specific domains that are being built.  um and if they really need to call a large model, [26:41] they'll they'll combine the two, but it's for it's  rarely rare um for the problems they're trying to [26:47] solve. So, I think uh we're seeing a little of  that ours perspective. Um from a human systems [26:52] integration standpoint, if you think about being  have a human in a loop, we're in our infancy at [26:57] this point. We're just learning what that means.  So that's another kind of frontier that will help [27:03] kind of shrink this this problem of scaling  a little bit and get a little more specified [27:07] particularly uh if you've got domain experts kind  of validating. Um but it it's still that's very [27:12] difficult to do right now and the validation costs  are pretty high and so the output is cheaper than [27:17] the validation. So typically you're just getting  a rubber stamp when the human takes a look at it. [27:22] So again we're still in our infancy there but  that might be something that emerges over time [27:26] particularly in heavily regulated areas. That's  a good point. I think uh we talk more about that [27:31] later if you want. But the Yeah, I think having  the verification step is like super important [27:35] going forward. And uh it's one of the bigger gaps  that I think we're facing. Um can you I think as [27:41] a developer your responsibility is both to put  benchmarks out to show how your model's working [27:45] well, right? In a general sense, right? Um but  also to give an easy verification step to humans, [27:50] right? Um if it's going to take offline audit  to go through your LM to see if you're right or [27:54] wrong, then you're not verifying on the fly or at  inference time, right? and then no one's going to [27:59] trust it, no one's going to use it, and it's just  going to fall apart. So, yeah, I think that's like [28:02] a huge thing that we're working on too now. Um,  but to answer your question, sorry. Um, I think [28:06] yeah, if you're not Google Meta or Anthropic right  now, you've lost the the fight for a foundational [28:10] model you're going to build, right? So, that's  out the window. Uh, they're going to keep building [28:13] them. They have the money for it. They have the  money behind them to do it, right? Um, and no [28:17] one's going to stop them until that's something  that happens, right? Um, so that's definitely a [28:22] path forward for those companies. But if you're  a smaller company, um your options really are to [28:26] buy a GPU for half a million dollars, right, on  the small end, right? Or to buy cloud compute. [28:30] Um and I think we're now in the spot of wild  west where there's enough directors who say [28:34] we want AI no matter what the cost. And you're  seeing news articles come out being like, oh, [28:38] they spent a half billion dollars in tokens today.  Um that's not good for our company. And that will [28:42] happen more and more for a little while until the  the top guys like kind of wake up and or anthropic [28:47] raises their cost by a little bit, right? Because  once you're fully in their ecosystem, right? uh [28:51] there's only reason they put the price up a little  higher, right? Um so I think yes they if you have [28:57] if you're a big enough company to buy a GPU, you  should be buying them now just to get some spot [29:00] for that, right? The downside of that is that you  need to maintain those GPUs and they update every [29:04] year and and Nvidia is going to promote a new  trillion dollar one next year, right? And so on [29:08] from there so like you're fighting that battle if  you're a smaller company who can never afford that [29:12] GPU, right? You're in the spot of like how much  cloud compute do I want to do? And then if not, [29:16] do I want to do the small language model route? Do  I want to train my own models and do fine-tuning? [29:19] And then if I do that, do I want to maintain  those? Do I want to serve those in house? Uh [29:23] how often do I update those? Is there a team  behind that? Right. And a lot of teams, a lot of [29:27] companies are like limping into this process now.  And the only thing they can do is get a clog code [29:31] subscription for now. Very good. All right, Pedro,  you mentioned um one of the magic words of 2026, [29:38] harness. Um [laughter] when uh I want to ask the  audience before January 1st 2026 who had heard [29:48] of the phrase agenic harness? Yeah, me neither. So  like but now it's like the hot buzzword, right? So [29:57] uh agent harnesses are the ways that these models  are finding a lot more purposefulness, a lot more [30:06] usefulness. And so how much of the work that you  all are doing in you know your own research uh is [30:16] shifting toward uh aenic systems uh just in the  past six months because we've seen the industry [30:24] really turn its collective eye you know toward  uh toward you know agenic systems in you know [30:30] such a a big way. Um what's been the difference  and what does that look like for you all now? [30:37] I'll I'll start because I have a counter, I  guess. And you guys can all disagree with me [30:41] in a second. Uh I have not bought the hype on  agents yet, unfortunately. Um I I there is a [30:46] spot for them. It makes sense. I use it in like  the development workflow and stuff like that, [30:51] right? But in production, uh I have not been sold  that agents are the future. Um at least not in the [30:56] iteration we're here, right? And most of it falls  down to verification still, right? Uh you have a [31:01] thousand agents now, each with some error rate,  right? And that doesn't go away. That multiplies [31:05] all the way across, right? And if I'm an end user,  especially in these high-risisk domains, if you're [31:09] on the defense side, the medical side, right,  and I now throw you a black box full of agents, [31:12] you were already not using the black box the first  time because you couldn't see into it now. And now [31:15] you have a thousand black box linked up together  and the errors don't go away. And if it breaks [31:20] or it fails and you don't know how it's going to  fail, right, that's an issue. And if you need to [31:24] debug that, that's also like a crazy system to go  through, right? Um, so I think there's probably a [31:29] purpose for them in the future, I think, for these  production things. But until we start building [31:33] like these test and evaluation harnesses for each  of these guys, for example, right? And be like, [31:37] I know how this component works by itself. I know  how it works in pair with the next person kind of [31:41] deal, which is a bigger system. It's such a hard  question, right? It's not like an easy thing you [31:44] can do. Um, then they're not reliable. And I would  advocate you don't use them in production for [31:49] that. Is there spots where you can do it like I  think in development or in a decision like support [31:54] aid tool. Yeah. having tools connected to things  makes them better. But in the like human machine [31:59] pipeline, I know people were saying human loop was  the big word last year, right? Or two years ago, [32:03] right? And now I keep people saying like human  um near the loop or off the loop a little bit, [32:08] whatever the next term is going to be, right? I  think we're not there yet. And I think people are [32:11] pushing that now because it's the next thing,  but uh I don't think agents, at least in this [32:16] V1 iteration of it, are the current solution.  Thanks for pushing back, Mike. Maybe I should [32:22] debate with you now. [laughter] I think I I agree  50%, I agree, I disagree 50% to be honest. Um the [32:29] thing is I mean maybe uh what we are working on  is for example for me is uh what technology they [32:36] build is useless because if the end user is not  using it there's no point in building what you're [32:40] working on. So the end user satisfaction is very  important when you're building any product. So [32:46] what we are working on is we are trying to create  our own evals of agents. Basically if you look at [32:53] uh lang runtime engine uh it runs on all your  traces and tries to identify the problems that [32:59] sits inside your traces. Meaning whenever you  let's say for example we are working on a use [33:04] case called gearpaul where a manufacturing floor  uh if a manufacturing floor got stopped in Toyota [33:10] we wouldn't make cars. So what we have done is  we built something called gearal agent where [33:14] they go and talk to the gear agent asking for uh  could I fix could you fix could you help me to [33:19] fix this problem. So immediately gpaul goes to  the terabytes of data in the vector stores and [33:24] pulls information and then it uses all its tools  and gives back the response but the response that [33:29] it came is really quality or not is what we  are sitting with the end users like our team [33:34] is really going back and sitting with end users  to see about okay if you don't like the response [33:39] give a thumbs down and give some feedback. So  that's one thing but there's also other other [33:43] research that's happening is the agents itself  are automatically seeing the traces and we have [33:50] done something called unstructured to uh skills  meaning the entire unstructured data we try to use [33:56] uh some graph based approach to convert to a  skills and use that as uh I would say a fe use [34:04] that as a data points to your eval agents to see  uh these are the traces and these are this is my [34:10] data data source and how can I able to improve  my agents automatically. So basically it gives [34:15] you a recommendation in the background saying  uh I think the uh I would say the vector store [34:23] that you're or a retrieval that you're trying to  make is not correct maybe you should change your [34:28] retrieval strategy here. So that is where we are  trying to work on more of I would say the engine [34:32] standpoint the uh we're using lang to be honest in  the background we're using lang engine to identify [34:37] all of these root causes and try again going back  to the customer and seeing whether the engine [34:43] whatever the engine recommended is really correct  or not and we're trying to do directly talk to the [34:47] customer who who is using meaning the engineer who  is using on the ground to solve whether really is [34:52] this really a problem or not then when we spoke  to the engineer he said yes it's a problem so we [34:57] identified all of that and then also So uh looks  at every single use case and then trying to look [35:03] at I would say a graph based approach to fix these  problems. Um yeah that's where I think uh half of [35:08] half of it I think I would say what whatever  you mentioned is really right. I think some of [35:12] them we are still I would say maybe ahead on the  game to be honest. I'll take two perspectives on [35:18] the agent real quick. Um one is yeah more tactical  and practical. um some of my colleagues uh they've [35:24] built a harness uh 35 agents running um over half  are either verifiers, testers, um adversaries uh [35:33] so it's but even that doesn't go into the software  pipeline. This is a prototyping way of getting [35:39] a head start as far as development. So we're  kind of seeing you know that is very useful. [35:43] Um the other is more strategic and um it's it's  more personal. It's it's allowed someone like a [35:48] social scientist like me to take concepts, social  concepts, and then get those engineered into like [35:55] an enigic system. I'll give a quick example. Um,  in criminology, we have this notion that we call [36:01] collective efficacy. You think of a neighborhood  and its crime rate. Uh, the collective efficacy [36:06] of a neighborhood is their the neighborhood's  ability to, you know, defend against crime [36:12] um and to enforce social norms and order. um  you can take that same analogy and say, okay, [36:18] if we're looking at a multi- aent system, can we  get it to communicate with itself to defend itself [36:24] in a sense? So again, we're taking more of a  security perspective, but can a multi- aent system [36:31] practice uh from a social standpoint this notion  of collective efficacy? again a way out there idea [36:37] but you know the ability to take these concepts  and now let's say build some prototypes and let's [36:43] start to experiment a little bit particularly  in the security space so um but yeah as far as [36:48] a highly regulated uh industry working on those um  agents are still I I go with Mike there the uh the [36:56] point about agent communication is I think very  important the three of the areas we're looking [37:01] at one is hallucination detection which is its  own field is fascinating Um when you know LLM's [37:07] hallucinate you can actually see different firing  patterns in the neurons within the neural networks [37:13] um which would be similar to a human being  hallucinating. Uh you can also see sometimes [37:19] different patterns in the language usage. So the  actual verbiage that comes out uh in the pattern [37:24] of uh usage of different words um and and that's  one area to help on the hallucination side. on the [37:32] agentic side, the the communication point is is  super strong because I mean in machine learning, [37:36] right, the theory of ensembles like you know  intelligence of the crowds, it works when you [37:41] have multiple people that are average intelligence  and you combine their you know predictive power, [37:46] you get something better than anybody within the  crowd. And that's the thing with agents. All these [37:51] papers coming out are showing that as long as  you tell the agents how to communicate with [37:54] each other properly, that's when you get two LLMs  that are inferior beating the latest and greatest [38:01] LLM. So I think a lot more research needs to be  done in that space of how can the agents better [38:07] communicate so that they could do their job better  and not just have one central command saying here [38:13] I got this task from the user. I'm going to  send it to 500 agents and then they're going [38:17] to spit back an answer and you know 3 million  tokens later you realize they all did the wrong [38:22] thing. I mean obviously the people building that  they don't care. They want you to spend tokens, [38:27] right? So that's right. Yeah. The business is  selling you tokens. Exactly. I I agree with that. [38:32] I think there's Yes, there's there's definitely  spots for um especially development agents makes [38:36] sense there. I think the ensemble approach here  is a good way to get past some of those things, [38:40] right? Uh, and I've been saying like errorrone or  accuracy for some of these models. I think there's [38:45] another thing talking about is the variability,  right? If I ask that same ensemble four times, [38:49] a hundred times, right? Are we getting the same  answer? Right? And there's a knowledge of the [38:53] crowd part to this. I I agree. But like if I do it  that one time, I'm not going to test it 100 times [38:58] for my use case, right? I'm going to test it one  time when I use that inference. Is it going to be [39:02] the right time or not? And I think that amount of  variability needs to be translated to the end user [39:06] so they know like, hey, I know this is going to  break one out of 100 times or one out of 10 times, [39:10] right? and that end user be like okay I will  trust this with that capa capacity and mental [39:14] model for example right and that's how they can  go forward with that there's like a maybe human [39:18] machine kind of like teaming for this still yeah  some of those papers are showing if you ask the [39:22] same LLM three times and you ask it to average its  own three answers you get a better answer than you [39:28] asking it just once so yes the more times you  ask it it cost you more but you'll get a better [39:33] answer because if you decrease the variability  but yeah I mean it's either that or you have the [39:38] spot where you like ask it to do your question  again are you correct or something like that, [39:41] right? And it could have been correct, but now  that you've given it the idea that it's wrong, [39:44] it's going to recalculate and cost you either  more compute or now be wrong because it's hedge [39:48] against it bet against this other one, right?  You have this like Monty Hall problem of like [39:51] I need to choose another door real fast. It's a  better probability than me stay in here. Yeah. [39:55] Yeah. I love that part of the conversation which  gets me to my last question and then we'll offer [39:59] the audience some some opportunities which is you  all as re researchers you have line of sight into [40:06] what's coming next. things like ensembles,  things like you know um querying the model, [40:12] asking if it's right, like all of these things are  aspects of that that are not broadly, you know, [40:17] in in use. So when you think about um the horizon  of what's coming next, what are the things for for [40:25] the rest of us, you know, in the room as we're  thinking about um how we understand uh AI and [40:32] understand the the course um of uh the trajectory,  you know, of AI. What are the the things that we [40:39] should be thinking about, you know, in addition to  some of these things or you can talk a little bit [40:43] more about them in terms of what's next? you know  what are the problems that you're thinking about [40:47] and the capabilities that are that are coming. I  think uh from an enterprise standpoint I would say [40:54] uh since we look at uh across the supply chain  R&D manufacturing all of things I think we should [41:00] look at uh creating your own um tools skills  I mean I would say I would name harness as a [41:09] combination of skills plus tools plus some all  of the MD files and all of that. So you should [41:15] prepare your own set of these uh I would say  data to feed to these models the existing right [41:22] now the data for example if you look at the  existing databases that you have it's all in [41:26] still not yet consumable you can't just do a text  to SQL on the big database and then say wait for [41:32] like 5 minutes just to get the response so it's  it's a pain so I think you should prepare that [41:37] ground for these agents to give that and that's  one I think the other one is also I think from [41:44] an individual standpoint, I think you all  should create your own agent. Basically, [41:48] let's say if you're working on for example venture  capital or if you're working on some automotive [41:53] or if you're working on healthcare, you should all  prepare your own agents to feed back the knowledge [41:59] uh from these white papers or patents back to you  every single day. Basically you should prepare [42:04] an Asian autom automation that feedbacks the  knowledge that's happening inside these research [42:10] and then gives back the response back so that you  can able to uh look at these new things that's [42:16] solving rather than you know just say always uh  spin up uh very high in reasoning in the cloud [42:22] code and say you figure out I think that's where I  I don't think that will work because uh if you say [42:29] you know simple question that will still very high  is not required because it's going to consume a [42:34] lot of tokens. So, I think that game is very still  important uh from a human aspect standpoint. Well, [42:39] that tip. Yeah, I I think it's actually a  conflict between my research mind and pragmatic [42:46] mind. I think the on the research side, there's  a lot of these exciting things I'm working on, [42:49] but for for answering your question of like what's  next, I think it's it's a lot less exciting from a [42:54] research perspective. It's just so more pragmatic.  It's just there's so much we can do with the tools [42:59] out there. just do something practical  which takes coordination, not technology, [43:04] coordination like within a company like assigning  a team that's going to be responsible for this and [43:08] making sure that they talk to whatever team is  going to use it. Right now companies are just [43:12] throwing money and technology at it and saying  here spend a million dollars a month on tokens [43:16] and you end up with a company where 10 different  teams are spending tokens doing the exact same [43:21] thing when if they were just talking to each  other you'd see that they could build a tool [43:25] together that you know would fix the problem for  the company. So I think there's a human element of [43:30] knowing how to get ROI out of AI and that's a lot  more on the pragmatic side than the the research [43:34] side. I think there's a big move there that  needs to happen within companies of just getting [43:40] pragmatic foot forward. Yes. Yeah. I I'll speak  just real quick because we I think we're out of [43:45] time. Um I think there's things to watch and the  that might change, right? The fusion models are [43:50] seeming to be a little better now in tech stuff.  Um so it's worth keeping an eye on stuff like [43:53] that. There's some other recent models that might  be LM's out but LM's had a huge head start. they [43:57] had the data science behind them and had the data  there. I think as an end user and a company though [44:00] for example, right? That's not my biggest concern  for next year. I think Pedro's kind of right here [44:04] where there's a coordination effort, right? Um  there are companies now waking up that we can't [44:09] just give claw code to one guy who's a developer  in the company and let him go like crazy with it, [44:13] right? We need a team that's going to evaluate  things and capabilities and stuff like that, [44:16] right? And that's a new awareness I think that  people are trying to wake up to. Um so I think [44:19] there's going to be a year of like building  out those teams that existed 10 years prior, [44:23] right? Which is now back to AI again. Um  right full cycle. Um and there's other [44:27] um cache reasons especially in like the defense  and the medical side where um they the customer [44:33] doesn't realize that they also need to present  data and ground truth for a lot of these things [44:36] for verification right I can't not know your data  and then also give you a benchmark on how well our [44:40] tool does on it for example right so there's now  I think more coordinated efforts or people waking [44:44] up to at least being like I need to get my data  together and figure out how I how I could even [44:49] if you're a customer and you can't fathom how you  would verify a tool is correct or incorrect right [44:54] then we're at a hard there's no way you could  either evaluate it, right? And I can't guess for [44:58] you because I'm not the end user with the data,  right? So I think there's more awareness of that [45:02] now or my hope optimistically there's more hope  for that now and hopefully we see more of like [45:08] hey we need to build this tool but also here's our  evaluation set here's a ground truth uh go for it. [45:14] Just final note um uh human machine communication  that's not going away. It's only going to get [45:19] better. Um so I think that's just kind of  stable. Um, one thing I would note on that is, [45:25] you know, everyone's worried about, hey, if  the LLM's hit a ceiling or not and, you know, [45:29] we're running out of data. Um, no, I don't think  so. I think I fall on the side of, you know, [45:33] more there's going to be more synthetic data.  It's going to get curated better. Um, and let the [45:37] machine do what the machine does really well. Um,  let it discover patterns that we would never come [45:42] up with and then have the human review it. That's  going to help with the augmentation of the human [45:47] and human performance down the road. So, I think  that's a frontier that just stays for a while. [45:52] Yeah, thank you. I've noticed our clock changed  and we are out of time or we all right. Let's [45:59] do maybe one question for audience. Couple minute  five minutes. Okay, let's see. Does anybody in the [46:04] audience have a question? There we go. Yeah. So,  you mentioned explaining the outputs of a model [46:09] to physicians or people in the medical field. What  does that look like in a practical sense? Because [46:14] I've had trouble even explaining like linear  regression or decision trees to physicians. So [46:18] like with an LLM or something more complicated,  how would you go about doing that? Uh you need [46:23] to present uh Yeah, you're correct. But first  of all, yeah, I I know a lot of clinicians who [46:28] um I have stories of clinicians who don't trust  normal medical models, right? Um I I I sit next [46:35] to a clinician who is like we're reading a  journal together about some disease and how [46:39] they're doing this new clinical study for it. He's  like, I treated that disease five times. That's [46:44] [ __ ] I'm like, well, this is trusted on 10,000  people, right? So that it's a difference like of [46:48] awareness of like that bias. So you're finding  that no matter what the clinician uh they've been [46:52] trained in whatever capacity they were, they've  seen their own experience. They're already in that [46:56] kind of bubble. Um so you're finding that and then  on top of that you have this LM blackbox thing, [47:00] right? So the only way to do this would be  to give them insight to it, right? So like [47:04] um I'll give you the the echo cardiogram project  we're working on now, right? Um if we can show [47:09] with a test and evaluation set for example that we  can match to some accuracy that that's one part, [47:14] right? That's that's like the lowest bar of entry  for Next part would be like can we now define [47:18] an embed in space to show the most similarity  patients. Right? So now if I go oh you have this [47:22] particular gen genetic diagnosis can I show that  in a cluster in space I'll make a nice visual for [47:27] them for example but like the all the ones around  it are also the same one. Right? So there's now a [47:31] little more buy in from that end right if I can go  even further and go like with an echo cardiogram [47:36] this panel this patch of the um from the image  transformer right is the one given us this right [47:41] and that matches with your particular patch right  there's more buying on that too right. So, it's [47:45] kind of like a working backwards of like [snorts]  this is right, this is wrong. Here's why we think [47:49] it's right or wrong, and here's how hopefully it  agrees with your bias or not, if that makes sense. [47:55] All right, we have another we have a  question up here. Vibe code um a you know [48:03] uh machine learning model, but like where do you  but I I'm at a small medium business. I also can't [48:10] hire a team to build a a system. So like do you  use use the LLM to build the deterministic part [48:18] or what? Yes. Like I mean I I've personally found  that you know when when coding vibe coding I think [48:30] it was a big promise of upskilling people and I  think that's probably not the right way to use [48:37] it. I think you should think of it as outskilling.  So like instead of turning a medium or mediocre [48:43] person into an exceptional person, I would rather  turn an exceptional person into 10 by leveraging [48:50] uh AI vibe coding. So I think that's a safer bet  with uh especially in a small company. So if you [48:56] have one person that isn't that skill and you  say we needed 10, great. I think VIP coding and [49:02] uh trusting an LLM to do that because I mean  I've tried so many projects where there's an [49:08] obvious logistical like a logical next step of  like turning this paper into something new and [49:14] it just can't get there no matter how much you  guide it. But once you say it and then it's like [49:19] oh yeah that's obvious now I can build that  for you. But it's not going to get that that [49:24] uh creativity spark usually. It's uh it's pretty  bland when it's trying to be creative. So, [49:30] um, if you have a pretty good idea of what  you need to build, it's fantastic. Uh, [49:35] if you're wanting it to give you the the spark,  I think you're going to be disappointed. To add [49:40] more to him, uh, maybe you should spend more  time and doing your requirements as well as [49:46] work and creating those I would say MD files like  proper detail requirements that they should not go [49:53] out of this particular loop. That's when I think  you should more focus on right rather than doing [49:58] wipe coding you should more focus on the human  efforts of doing these requirements of course [50:02] you can use clot to generate this as well but  still you need to verify all of these detailed [50:07] requirements maybe I would say you take some like  two weeks or 3 weeks just put your full efforts [50:12] in just doing those proper requirements so that  I think you can just then automate all of your [50:18] uh code generation all of that but again make sure  to maybe every single every single PR uh let's say [50:23] if codex writes the code and ask clot to checks  the requirements and reviews a PR too so that it [50:28] can be a helper for you to for your startup to go  full speed. Yeah, I think the biggest advantage [50:35] of LMS in this kind context would be the back and  forth interaction. So I think there is actually a [50:39] ability to upskill insecapacity, right? If you're  playing with this and it breaks and you're like, [50:43] "Oh, just fix the problem." That's one thing,  right? And that's a fear I have that people [50:47] are doing at the moment. But if it's like, "Why  did this break? Let me understand why it break." [50:51] And you take the time to like figure out why it's  not happening, that's a learning moment for you, [50:55] right? Especially a single developer or small  team, right? And you can use that to build up [50:58] your own skills a little bit. Uh the fear I have  though is that we have a lot of people who don't [51:02] have expertise in that domain, right? who are like  I'll give you an example of someone the other day [51:06] we asked a a programmer to code up like a a v  video classifier for someone with motion disease [51:14] u movement disorder right and I said okay why  is this not working and he just took why is it [51:19] not working from my chat box plugged into claude  and said why is it not working right the question [51:24] was actually um did you try what we tried yes  we talked about yesterday and claude responded [51:28] back I don't know what you said yesterday and he  copy paste that back to me right so that's like [51:32] a lack of understanding of both what his role  was there, right? And lack of knowledge to even [51:37] know how to answer that question, right? So I my  fear is that we have a lot of people who are not [51:41] subject subject matter experts who are pretending  or cosplaying as subject matter experts because [51:46] they can do something real fast and it's a flashy  response which is great but if you take the time [51:50] to like go through and see how it breaks and  figure out why if I can't explain to my kid why [51:54] this is working for example right I might just  understand if just magic then maybe it works. [52:00] to roll a dice that point, right? And  we're at casino, so maybe that makes sense.