PRAXIST: The World's First Autonomous Research System Delivering Validated, Enterprise-Ready Results

PRAXIST：全球首个提供经验证、企业级就绪成果的自主研究系统 William Chen 介绍了 PRAXIST，称其为人工智能第三次演进浪潮——“发现与优化时代”的代表性系统，该系统超越了静态任务执行，迈向自主闭环的实证研究。在涵盖石油路径规划、机器人 SLAM 算法以及火箭推进着陆等实际企业部署中，PRAXIST 通过在夜间运行数百次验证实验，展现出相比人类基准显著的性能提升。演讲详细介绍了 PRAXIST 的线性 DAG（Linear DAG）架构，该架构利用专门的“创新者”（Innovator）、“探索者”（Explorer）和“证伪者”（Falsifier）角色来保留局部洞见，而非像传统树状系统那样丢弃种子，从而达成了 72% 的 SOTA 基准率。Chen 在最后的互动问答中详细说明了企业团队如何集成既有仿真器或从零开始定义形式化评估指标。 人工智能正迈入第三次浪潮——“发现与优化时代”，以 PRAXIST 为代表的自主研究系统已超越创意生成和任务执行，能够自主设计、测试并交付经过验证的解决方案。 (0:04) PRAXIST 的运作模式宛如一支自主研究与咨询团队，不仅能提出分析建议，还能直接实施、测试并交付可投产的优化方案。 (0:04) 通过持续运行，PRAXIST 能够探索人类工程团队难以实际测试的超过 100,000 条不同操作路径，并随着约束条件和市场状况的变化动态重新校准模型。 (3:26) 在涉及 136 个国家的全球石油物流部署中，PRAXIST 在数天内将基准效率提升从 13% 加速至 29%，其运行效率约为人类的 3,000 倍。 (3:26) 在机器人与空间计算领域，PRAXIST 在无需结构性改动的情况下，于五天内将 SLAM 室外导航累积误差从人类团队耗时 6 个月达成的 6.4 米基准降低至 0.9 米。 (7:36)

William Chen, co-founder of Sapient Intelligence, argues that enterprise artificial intelligence is shifting from generative assistance and task execution to autonomous research and optimization. He claims systems like PRAXIST can autonomously design, test, and implement operational solutions at scales human teams cannot match. The Transition to Autonomous Enterprise Discovery (0:04) Chen outlines three evolutionary waves in enterprise AI: the Generation Era of large language models drafting ideas; the Execution Era of tool-using agents handling repeatable tasks; and an emerging Discovery and Optimization Era. Rather than executing static code, systems in this third phase ingest high-level business goals, physical constraints, and metrics, then independently run long-horizon experiments and learn from intermediate failures. Chen notes this shift is gaining industry-wide traction, citing Jeff Dean's departure from Google DeepMind to establish Discovery Loop. Under this model, PRAXIST is framed as an autonomous research and analytical unit that directly implements its findings. Chen argues that because business variables and market constraints fluctuate continuously, autonomous research engines must operate around the clock, testing hundreds of thousands of candidate pathways that human teams lack the time to explore. Empirical Benchmarks in Logistics, Robotics, and Aerospace (3:26) Sapient Intelligence reports deploying PRAXIST across more than 30 enterprises and 100 technical tasks, claiming it functions at roughly 3,000 times human efficiency. In a petroleum logistics network spanning 136 countries, internal human teams had spent months reaching a 13% efficiency gain; PRAXIST reportedly expanded that improvement to 29% within days after ingesting baseline code and system constraints. In spatial computing, PRAXIST reduced a robotics SLAM error from a 6-month human baseline of 6.4 meters down to 0.9 meters in five days. Chen also highlights a benchmark with a commercial aerospace enterprise testing rocket propulsive landing under severe crosswinds and offsets. Iterating through 60 to 75 generations on an 8-GPU server over three days, PRAXIST improved the safe landing rate from a 70% published baseline to 90%, incurring roughly $1,000 in cloud token costs. Chen states that the system averages 300 to 500 validated research findings per night, with several resulting papers submitted to or accepted by academic venues like Nature and major machine learning conferences. Linear DAG Architecture and Resource Efficiency (12:48) Chen contrasts PRAXIST with single-trace coding agents and academic systems like Sakana AI's The AI Scientist. Single-trace agents execute discrete assignments without iterative state memory, while tree-based search architectures prune discarded branches entirely. Chen contends that tree pruning discards valuable structural insights that may have failed solely due to un-tuned hyperparameters. PRAXIST instead connects candidate solutions through a Directed Acyclic Graph (DAG) backed by shared memory, coordinating specialized roles—an Innovator, an Explorer, and a Falsifier—to distill partial successes and ablations into subsequent generations. Across 100 benchmarked tasks, Chen claims PRAXIST achieved a 72% state-of-the-art level rate, running on local bare-metal hardware with near-zero token costs. Under identical benchmarks, Claude Code reached a 45% state-of-the-art rate while incurring $48,000 to $50,000 in token fees, and Baidu's Faro Agent reached 38%. To address enterprise security requirements, the platform supports deployments across isolated virtual clouds and air-gapped on-premise infrastructure. Integration Workflows and Simulation Dependencies (17:11) Addressing integration into legacy environments, Chen explains that Sapient Intelligence employs Forward Deployed Engineering and API adapters to interface with engineering software such as ANSYS and SolidWorks. Engineers can deploy baseline models directly into the pipeline, allowing the system to tune control loops and parameters dynamically. For non-standard domains like novel material synthesis or non-linear chemistry lacking pre-existing simulation software, Chen clarifies that while a pre-configured simulator is the most reliable ground-truth baseline, it is not strictly required. Enterprises can instead supply formal evaluation metrics, enabling PRAXIST to scaffold the verification pipeline autonomously, though custom simulator builds introduce trade-offs in cost, execution speed, and fidelity. Chen presents autonomous research engines as a viable replacement for manual operational optimization, though PRAXIST's performance remains contingent on well-formulated evaluation metrics or high-fidelity simulation environments. Enterprises operating in novel physical domains must balance the cost and trade-offs of developing bespoke simulation harnesses against relying on autonomously generated evaluation frameworks.

Sapient Intelligence 联合创始人 William Chen 提出了名为 PRAXIST 的自主研究系统，主张企业级人工智能正从内容生成和工作流执行，迈向能够自主设计长周期实验并交付验证成果的发现与优化阶段。 企业级人工智能向自主研究范式演进 (0:04) Chen 将现代人工智能划分为三个演化阶段：首先是以大语言模型为核心的生成时代，专注于产出想法与内容；其次是以 AI 智能体为主导的执行时代，涵盖 Codex、Claude Code、Microsoft Copilot Studio 及 Agentforce 等自动化工具，用于执行既定任务；第三个阶段则是以 PRAXIST 为代表的发现与优化时代。该系统超越了静态代码编写，能够接收高级业务目标、环境约束与成功指标，自主设计长周期实验并从失败中提取有效方案。Chen 援引 Jeff Dean 离开 Google DeepMind 创办 Discovery Loop 的行业动态，佐证这一范式转变正成为行业趋势，并将 PRAXIST 比作集成了研究与落地交付能力的自主分析团队。 工业级复杂场景与深科技领域的效果验证 (3:26) PRAXIST 针对企业运营决策的持续优化展开设计，支持全天候探索多达 100,000 条人工团队难以遍历的优化路径，并可根据动态约束进行实时重调。在一家覆盖 136 个国家的能源物流企业案例中，内部 5 人研发团队历时数月将石油管网及海运路线效率提升 13%，而 PRAXIST 基于现有代码库与约束在数日内将效率提升至 29%，号称运行效率达到人类水平的约 3,000 倍。在室外机器人 SLAM 算法优化中，该系统将 10 人团队研发 6 个月得到的 6.4 米累计误差，在 5 天内缩减至 0.9 米。 在商业航天火箭推进着陆基准测试中，人类团队历时 6 个月在 2 公里高度、1,500 米横向偏移与强侧风环境下取得了 70% 的安全着陆率（已发表于 IEEE 论文）。PRAXIST 在相同的对抗性仿真环境中运行 3 天、经历 60 至 75 代迭代，将安全着陆率提升至 90%。该任务在 8 卡 GPU 服务器上运行，每晚生成约 300 至 500 项经过验证的研究成果，云端 Token 成本约为 1,000 美元。Chen 表示，系统产生的三项突破已被《自然》（Nature）或顶级机器学习会议接收或正在审稿。 线性有向无环图与多角色协同架构优势 (12:48) Chen 将 PRAXIST 与现有编程智能体及学术系统进行了架构对比。他指出，Codex 和 Claude Code 属于目标驱动的单轨线性智能体，适合离散型代码修复与网站搭建，但缺乏长周期研究所需的有状态探索机制；Sakana AI 的 The AI Scientist 采用基于树的搜索机制，在评估候选方案时会修剪掉非最优种子，这可能丢弃由于超参数未微调而掩盖了优异架构洞见的潜在方案。PRAXIST 则采用基于共享内存库的线性有向无环图（Linear DAG）架构，通过创新者（Innovator）、探索者（Explorer）和证伪者（Falsifier）三个子智能体角色，审问每次试验并将局部成功与结构性经验提炼入共享知识库。 在 100 项基准任务的评测中，部署于本地或裸机服务器的 PRAXIST 实现了 72% 的业内领先水平（SOTA）达成率，且 Token 成本近乎为零；而在同等测试条件下，Claude Code 的 SOTA 达成率为 45%，并产生了约 48,000 至 50,000 美元的 API 费用，百度基于 Gemini Pro 后端的 Faro 智能体达成率为 38%。在部署模式上，PRAXIST 支持共享集群、独立虚拟云、裸机以及完全物理隔离的本地环境。 企业仿真生态集成与无仿真环境适应方案 (17:11) 针对企业既有软件生态的接入问题，Chen 介绍 Sapient Intelligence 通过前向部署工程（FDE）与 API 适配器实现对接，支持原生集成 ANSYS、SolidWorks 等工业工程软件及内部数据分析流水线。工程师可将基准代码直接置于代码库中，由系统动态优化参数与控制闭环。 面对新材料研发或非线性化学等缺乏现成仿真软件的领域，Chen 指出拥有完备的仿真环境虽然是使智能体严格遵循物理限制的最稳健路径，但并非硬性前置条件。企业可以直接提供目标评估指标与清晰的形式化定义，PRAXIST 能够据此自主构建评估测试框架与执行工作流。对于选择事后从零构建独立仿真器的企业，Chen 提醒不同实现方式可能在执行成本、系统性能及保真度上带来相应权衡。 Chen 将 PRAXIST 呈现为解决高难度工程与业务建模的自主闭环优化方案，展示了其在降低 Token 成本与提升指标上的优势。然而，在缺乏成熟物理仿真器的前沿探索领域，自主构建的评估框架与现实物理约束之间能保持多高的真实保真度，仍需在更多工业实践中经受检验。

## Background 该内容由科技频道 Ai4 于 2026 年 8 月 20 日发布在 YouTube 平台。视频由 Sapient Intelligence 联合创始人 William Chen 主讲，介绍了全球首个面向企业级交付的开源自主研究系统 PRAXIST，展示其如何以极低成本和极少人工干预自主完成实验设计、模型微调与 SOTA 级科研突破。 ## TL;DR 演讲阐述了 AI 技术从“生成时代”和“执行时代”向“发现与优化”自主研究时代的跨越，并正式推出了 PRAXIST 系统。该系统通过独创的线性有向无环图（Linear DAG）与多智能体共享记忆架构，有效克服了传统树状搜索容易过早收窄的缺陷，实现了高效的全局实验探索。在跨国能源调度、火箭回收控制和机器人 SLAM 等实际工业落地中，PRAXIST 在数天内便达成了远超人类专家团队数月研发成果的指标，并大幅降低了计算成本。系统支持指标驱动评估与本地私有化部署，为工业和前沿科学探索提供了端到端的百倍加速方案。 ## Atmosphere & style 演讲展现出充满自信的前瞻科技视野与极具说服力的工程底色，开场以幽默的深呼吸互动迅速抓牢观众注意力。随后通过严密的架构演进逻辑、详实的多系统横向评测以及令人震撼的真实工业案例，构筑出既严谨前沿又高度务实的极客叙事氛围。结尾的问答互动从容切中工程与行业落地痛点，给观众带来一场见证科研范式被彻底重构的沉浸式体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=4s) AI 技术正在从大模型内容生成（第一阶段）与工具执行智能体（第二阶段），跨入以 PRAXIST 为代表的“发现与优化（Discover & Optimize）”自主研究时代。 - [00:50](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=50s) PRAXIST 专注于深刻理解业务目标、数据、约束与指标，能够自主构思并运行长周期实验（Long-horizon Experiments），直接交付经过严格消融验证的生产级代码与模型。 - [03:26](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=206s) 只要商业与工程决策可建模且可量化测量（如物流调度、火箭制导、排产调度等），PRAXIST 便能以相当于人类 3,000 倍的效率 7×24 小时进行自主闭环优化。 - [04:00](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=240s) 在大型跨国能源运输调度案例中，PRAXIST 在短短数天内将整体运输与成本效率提升至 29%，显著超越了原 5 人专家团队耗时数月达成的 13% 优化成果。 - [07:36](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=456s) 在火箭回收控制算法案例中，系统在 8-GPU 服务器上运行 3 天、自主迭代 60 至 75 代，将安全着陆率从人类团队研发 6 个月达成的 70% 提升至 90%。 - [08:40](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=520s) PRAXIST 每晚可产生 300 到 500 个有效科研发现，大幅压降了室外机器人 SLAM 误差，且已有 3 项系统深度参与的科研成果被投递至《Nature》或顶级机器学习会议。 - [12:48](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=768s) 架构上摒弃了传统丢弃非冠军种子的树状演进，独创“线性 DAG（有向无环图）+ 共享记忆空间”，全面提炼所有测试种子的经验，避免过早收窄探索路径。 - [13:40](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=820s) 系统内部设立了证伪者（Falsifier）、探索者（Explorer）和创新者（Innovator）等多角色智能体，将实验探索沉淀为全局知识图谱以指导下一代演进。 - [14:40](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=880s) PRAXIST 取得了约 72% 的 SOTA 达成率（远高于行业同类工具），并提供共享集群、隔离云以及边际 Token 成本为零的本地私有化部署方案。 - [17:11](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=1031s) 支持通过编程接口无缝对接 Ansys、SolidWorks 等专业工程仿真与分析工具，亦可直接将算法工程师现有的基线代码（Baseline）作为输入展开长周期调优。 - [18:57](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=1137s) 针对缺乏现成仿真模拟器的研发场景，用户只需提供明确的评估指标（Evaluation Metric）定义，PRAXIST 即可自主构建评估与验证流程完成闭环研究。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=4s) 迈入自主研究与优化时代：从大模型到 PRAXIST Sapient Intelligence 联合创始人 William Chen 登台开场。Sapient Intelligence 是一家专注于基础模型与智能体治理（Agent Harness）的前沿科技企业，在新加坡、帕洛阿尔托（Palo Alto）和中国设有分支机构。在正式切入主题前，William 幽默地邀请全场观众一同深呼吸，因为接下来展示的内容将“彻底重塑大家对技术可能性的认知”。他正式推出了全球首个面向企业级交付与验证的自主研究系统——PRAXIST，旨在通过端到端的自主科研优化大幅提升企业的商业产出与核心指标。 ```mermaid flowchart LR accTitle: AI 技术演进三阶段 subgraph S1["第一阶段：生成时代"] A["大语言模型 (LLM)<br/>内容生成 / 提供方案与构想"] end subgraph S2["第二阶段：执行时代"] B["执行智能体 (Agents)<br/>工具调用 / 自动化既定工作流"] end subgraph S3["第三阶段：自主研究时代"] C["PRAXIST (Discover & Optimize)<br/>自主设计长周期实验 / 验证并交付 SOTA"] end A --> B --> C ``` William 回顾了人工智能的发展路径：从以大语言模型为代表、擅长生成内容与发散想法的“生成时代（Generation Era）”，演进到以 Claude Code、Microsoft Copilot Studio、Agentforce、OpenClaw 等为代表、擅长调用工具完成特定重复性工作流的“执行时代（Execution Era）”。而当下技术已经迈入“发现与优化（Discover & Optimize）”的自主研究时代。正如近期 Google DeepMind 首席科学家 Jeff Dean 离职创立 Discovery Loop 投身该赛道一样，PRAXIST 专注于在深刻理解业务目标、数据、约束条件及衡量指标的基础上，自主构思并运行长周期实验（Long-horizon Experiments），从反馈中持续迭代，输出经过严格验证的落地成果。 - 商业本质由无数决策构成，只要该决策可建模、可量化测量，PRAXIST 便能对其进行持续优化。 - PRAXIST 扮演的角色不仅是提供咨询建议的“麦肯锡或 BCG 顾问团队”，更是一支自带顶级研究员、能直接动手实现并交付最优代码与模型的工程科研团队。 --- ### [03:26](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=206s) 商业与工程决策优化：能源运输网络真实案例 企业每天都在面临海量复杂的决策难题：从日常运营端的库存水位设定、物流路径规划、资本配置、排产调度，到前沿工程研发端的机器人控制策略、火箭制导算法、图像识别、新材料分子设计以及制造工艺调优。传统团队受限于人力与时间，难以穷尽所有可能方案，而 PRAXIST 可以 7×24 小时无休运转，探索比人类团队多出成千上万倍的潜在路径，并随着外部市场与风险动态持续调整模型。 ```mermaid flowchart TD accTitle: PRAXIST 自主优化工作流 subgraph Input["1. 任务输入"] D["业务数据与基线代码 (Baseline)"] K["KPI、评估指标与系统约束"] end subgraph PRAXIST["2. PRAXIST 多智能体闭环研究"] E["多智能体协作组装<br/>(相当于人类 3000 倍效率)"] F["实验设计与仿真运行"] G["消融实验 (Ablation) 与结果验证"] E --> F --> G --> E end subgraph Output["3. 企业级交付"] H["经过严格验证的 SOTA 生产级模型"] end D --> E K --> E G --> H ``` William 分享了一家业务遍布 136 个国家的大型跨国能源运输公司的真实落地案例。该企业需要优化极其复杂的石油管道与航运调度路线。用户只需将现有基线（Baseline）数据与代码放入文件夹，设定好 KPI 目标与系统物理约束交付给系统，PRAXIST 便会组建具备“人类 3,000 倍效率”的多智能体系统自主展开闭环研发——搭建仿真环境、自动运行测试、通过消融实验（Ablation Study）验证结果并推进下一轮迭代。 - 该能源企业原先由 5 名专家组成的研发团队历时数月将成本与运输效率优化了 13%，已为企业带来数亿美元的效益提升。 - 引入 PRAXIST 后，系统在短短数天内将整体效率提升幅度拉升至 29%。 - 传统需要研究团队耗费数月甚至数年的建模、调优、消融与验证全流程，被 PRAXIST 压缩至数天到一两周内完成。 --- ### [07:36](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=456s) 跨行业落地成果与火箭回收控制案例 PRAXIST 已经在 30 余家行业头部企业中落地了超过 100 项复杂任务，覆盖量化交易（提高收益率并显著压降回撤）、核聚变控制、医疗影像识别及自动驾驶等多个硬核领域。William 强调，PRAXIST 的定位绝非取代人类研究员，而是作为将工作流效率放大百倍的得力工具，将专家从重复的模型调优中解放出来，专注于构思更高维度的科学猜想。 ```mermaid flowchart LR accTitle: 商业火箭着陆算法优化对比 subgraph Human["人类研究团队 (历时 6 个月)"] H1["IEEE 论文成果"] H2["安全着陆率: 70%"] H1 --> H2 end subgraph PRAXIST["PRAXIST (8-GPU 服务器 / 运行 3 天)"] P1["迭代 60~75 代 / 查阅最新论文"] P2["安全着陆率: 90%<br/>姿态更平稳 / 触地速度更低"] P1 --> P2 end ``` 在商业航天领域的一项火箭回收算法优化任务中，面临着 2 公里高空回收、1,500 米水平初始偏移以及复杂强侧风等严苛工况。原合作方科研团队耗时 6 个月达成了 70% 的安全着陆率并发表了 IEEE 论文。而接入配备 8 块 GPU 的 PRAXIST 系统后，系统在无人干预下自动查阅最新前沿文献，自主迭代了 60 至 75 代： - 系统每晚能够产生 300 到 500 个有效发现，每个发现的体量相当于一名普通博士生一周的研究工作量。 - 仅用时 3 天，PRAXIST 便将火箭安全着陆率从 70% 提升至 90%，同时大幅降低了触地误差与倾斜姿态。 - 整个任务若使用云端大模型，Token 消耗成本仅约 1,000 美元；若采用本地离线部署则边际 Token 成本为零。 - 室外机器人 SLAM 算法案例：某 10 人团队研发 6 个月达成 6.4 米累积误差，PRAXIST 运行 5 天后将误差降至 0.9 米，且方案完全即插即用。 - 截至目前，已有 3 项由 PRAXIST 深度参与并完成的科研成果被投递至《Nature》或顶级机器学习会议。 --- ### [12:48](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=768s) 架构差异：线性 DAG 演进机制与行业对比 William 深入拆解了 PRAXIST 与其他 AI 系统的底层技术差异。像 Claude Code、CodeX 这类工具本质上是面向具体目标的线性追踪（Linear Trace），在编写独立应用、修补 Bug 或搭建网站等确定性任务上表现出色，但在长周期科研探索上存在天花板；而以 Sakana AI 研发的 AI Scientist 为代表的系统则采用树状结构（Tree Structure）进行演化，每代仅筛选胜出的单一冠军种子（Champion Seed）繁衍，其余种子全盘丢弃。这种方式的弊端在于许多被淘汰的种子可能仅是由于超参数未对齐或训练轮数（Epoch）不足，直接舍弃会导致研究路径迅速收窄。 ```mermaid flowchart TD accTitle: PRAXIST 线性 DAG 与多角色共享记忆架构 subgraph Generation["当前代候选方案 (Seeds)"] S1["种子 A"] S2["种子 B"] S3["种子 C"] end subgraph Agents["多角色智能体分析与提炼"] F["证伪者 (Falsifier)"] E["探索者 (Explorer)"] I["创新者 (Innovator)"] end subgraph Memory["共享记忆空间 (Shared Memory)"] M["知识图谱与沉淀经验"] end subgraph NextGen["下一代演进 (Next Gen)"] N["更高胜率的 SOTA 迭代方案"] end S1 & S2 & S3 --> F & E & I F & E & I --> M M --> N ``` 针对这一瓶颈，PRAXIST 独创了“线性有向无环图（Linear DAG）”架构与共享记忆空间（Shared Memory）。系统内部设立了证伪者（Falsifier）、探索者（Explorer）与创新者（Innovator）等专门智能体角色，全面提取所有测试种子中的有效经验并沉淀为全局知识，再统一赋能下一代演化，实现了极高效率的全局搜索与突破。 | 系统 / 方案 | SOTA 达成率 | Token 综合成本 | 核心演进架构 | |:--- |:--- |:--- |:--- | | **PRAXIST** | **~72%** | **$0** (本地私有化部署) / 极低云端成本 | **线性 DAG (Linear DAG) + 共享记忆机制** | | **Claude Code** | ~45% | 约 $48,000 - $50,000 (同等任务基准) | 目标导向线性追踪 (Linear Trace) | | **百度 Faro Agent** | ~38% | 较高 (基于 Gemini 3.1 Pro) | 传统 Agent 工作流 | 在企业部署与数据安全层面，PRAXIST 提供了极致的灵活性： - **共享集群模式**：成本效益最高，适合轻量级快速接入。 - **隔离云 / 裸金属服务器**：兼顾算力性能与网络隔离。 - **本地私有化部署（On-premise）**：模型与研究闭环系统全部部署在企业内部机房，实现完全的数据控制与合规治理。 --- ### [17:11](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=1031s) 现场问答：工具链对接与新产品研发挑战 演讲结束后进入现场提问环节。观众 Miguel 率先提问，询问 PRAXIST 如何与企业现有的专业工具、仿真环境及工程软件进行对接。 William 解释道，针对不同企业的异构系统，Sapient Intelligence 会根据需求派遣前向部署工程师（FDE）提供定制支持，但本质上对接是在智能体交互层面完成的。只要现存工具具备可编程或可调用的操作接口（例如工程仿真领域的 Ansys、SolidWorks，或是各类金融分析工具），PRAXIST 都能无缝挂载。如果企业内部已有算法工程师或研究员写好了基线代码，也可以像使用高级版 Claude Code 一样，直接将代码基线作为输入，交由系统展开长周期自动化调优。 随后，另一位来自新产品研发（NPD）与制造领域的观众提出了深入探讨：在许多前沿材料研发场景中，由于涉及复杂的非线性化学反应与物理方程，行业内甚至尚未建立成熟的仿真软件。在缺乏现成仿真器验证方案可行性的情况下，系统如何贯穿从科研构想到实际测试的闭环流程？（讨论在此处展开延伸）。 ### [18:57](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=1137s) 问答环节：领域知识注入与仿真模拟器的接入方式 在现场问答环节中，有观众提问探讨了在实际落地 PRAXIST 时，企业是否需要将自身的领域知识（Domain Knowledge）深入转化为一个专门的仿真模拟器（Simulator），以便后续接入该自动化研究系统。 Sapient Intelligence 联合创始人 William Chen 对此回应表示，这在很大程度上取决于具体的业务场景。如果企业自身没有现成的模拟器，通常只需提供清晰的评估指标（Evaluation Metric）定义，PRAXIST 系统就能够自主地为用户构建相应的评估与验证流程。当然，如果用户已经具备成熟的仿真模拟器，这无疑是最理想的状态，因为系统可以直接且严格地按照模拟器的规则与环境进行实验迭代与验证。如果完全缺乏模拟环境，虽然系统支持从头构建，但在性能成本或最终效果上可能会存在一定的差异与权衡。 - **指标驱动替代模拟器**：在缺乏完整仿真环境时，用户只需明确业务或算法的评估指标定义，系统即可自动生成评估机制。 - **现成模拟器的优势**：若已有模拟器，系统能够实现端到端严格对齐模拟环境的闭环探索。 - **成本与性能权衡**：从零构建模拟环境可能会引入额外的算力与研发成本，并在模拟精度与实际效果之间产生一定差异。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=4s) 彻底重塑大家对技术可能性的认知 — William Chen > [03:26](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=206s) 人类 3,000 倍效率 — William Chen > [07:36](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=456s) 系统每晚能够产生 300 到 500 个有效发现，每个发现的体量相当于一名普通博士生一周的研究工作量。 — William Chen > [07:36](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=456s) 仅用时 3 天，PRAXIST 便将火箭安全着陆率从 70% 提升至 90% — William Chen > [07:36](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=456s) 整个任务若使用云端大模型，Token 消耗成本仅约 1,000 美元；若采用本地离线部署则边际 Token 成本为零。 — William Chen > [07:36](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=456s) 某 10 人团队研发 6 个月达成 6.4 米累积误差，PRAXIST 运行 5 天后将误差降至 0.9 米，且方案完全即插即用。 — William Chen > [00:04](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=4s) 只要该决策可建模、可量化测量，PRAXIST 便能对其进行持续优化。 — William Chen > [12:48](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=768s) 系统内部设立了证伪者（Falsifier）、探索者（Explorer）与创新者（Innovator）等专门智能体角色，全面提取所有测试种子中的有效经验并沉淀为全局知识 — William Chen > [18:57](https://www.youtube.com/watch?v=o2wMJRLAQlc&t=1137s) 如果企业自身没有现成的模拟器，通常只需提供清晰的评估指标（Evaluation Metric）定义，PRAXIST 系统就能够自主地为用户构建相应的评估与验证流程。 — William Chen ## Mentioned William Chen, Sapient Intelligence, PRAXIST, Claude Code, Microsoft Copilot Studio, Agentforce, OpenClaw, Jeff Dean, Google DeepMind, Discovery Loop, Sakana AI, AI Scientist, 百度 Faro Agent, Gemini 3.1 Pro, CodeX, Ansys, SolidWorks, Nature, IEEE, MLE-bench ## Video description William Chen - Co-founder, Sapient Intelligence Pte. Ltd. Tired of repetitive model tuning or running out of new ideas to explore? Free yourself from the tedious side of research. PRAXIST is the world's first production-ready, open-source autonomous research system. It handles the repetitive heavy lifting while you focus on exploring ideas, developing insights, and tackling higher-level research questions. In this session, Sapient Intelligence will introduce a new research paradigm—one in which just a few lines of instruction can autonomously turn a baseline into an above-SOTA result. At a fraction of the cost of Claude Code, PRAXIST can research, write code, design and configure experiments and simulations, analyze results, refine iterations and tune models … all with minimal human intervention. This fundamentally transforms both the workflow and economics of enterprise R&D. Today, PRAXIST supports more than 50 task types across over 30 industries and achieves leading performance on agentic research benchmarks such as MLE-bench. In this session, you will discover what a 100×-accelerated frontier research workflow feels like, including: - How PRAXIST operates - The technology behind its autonomous research capabilities - The latest prompting techniques for AI-driven research - State-of-the-art results produced autonomously by PRAXIST The next generation of leading researchers will not simply use AI. They will know how to turn it into a decisive research advantage. [Source](https://www.youtube.com/watch?v=o2wMJRLAQlc)

[0:04] William Chen: 彻底重塑大家对技术可能性的认知 [3:26] William Chen: 人类 3,000 倍效率 [7:36] William Chen: 系统每晚能够产生 300 到 500 个有效发现，每个发现的体量相当于一名普通博士生一周的研究工作量。 [7:36] William Chen: 仅用时 3 天，PRAXIST 便将火箭安全着陆率从 70% 提升至 90% [7:36] William Chen: 整个任务若使用云端大模型，Token 消耗成本仅约 1,000 美元；若采用本地离线部署则边际 Token 成本为零。 [7:36] William Chen: 某 10 人团队研发 6 个月达成 6.4 米累积误差，PRAXIST 运行 5 天后将误差降至 0.9 米，且方案完全即插即用。 [0:04] William Chen: 只要该决策可建模、可量化测量，PRAXIST 便能对其进行持续优化。 [12:48] William Chen: 系统内部设立了证伪者（Falsifier）、探索者（Explorer）与创新者（Innovator）等专门智能体角色，全面提取所有测试种子中的有效经验并沉淀为全局知识 [18:57] William Chen: 如果企业自身没有现成的模拟器，通常只需提供清晰的评估指标（Evaluation Metric）定义，PRAXIST 系统就能够自主地为用户构建相应的评估与验证流程。

[0:04] William Chen: Let's do breathe in and breathe out. [0:04] William Chen: You're basically getting a McKinsey or a BCG consultant team with researchers—so they not only give you advice, but they will actually implement and deliver those changes for you. [18:57] William Chen: strictly follow what's on the simulator

[0:04] Welcome to our session. So my name is  William Chen. I'm the co-founder at [0:07] Sapian Intelligence. We are a foundation  model and agent harness company based in [0:12] uh with offices in Singapore, Palo Alto  and China as well. Um so before we start, [0:17] I would like to invite everyone to take a deep  breath with me. Let's do breathe in and breathe [0:24] out. All right? because um what you're going to  see is we're going to introduce you to something [0:31] that's going to completely redefine what you  think was was possible and uh we're going to [0:36] introduce you to practice. It's an autonomous  research system that can optimize and delivers [0:42] validated enterprise ready results and that can  significantly improve your business outcome. [0:51] So, uh a lot of you might be new to the concept  of autonomous research. uh you know we came from [0:57] the era of generation right foundation models  large angry models uh to an era of uh execution [1:04] agents and then now we have arrived to the era  of autonomous research and optimization or we [1:10] call discover and optimize. Um so you know large  language models they are really good at generating [1:17] answers and possibilities. they can produce uh  content to answer questions and propose ideas [1:23] uh strategies and plans based on the content you  provided. And then there comes agents. I think as [1:29] a matter of fact they they probably make up of uh  90% of the conference exhib ex exhibitors right [1:34] right here at the conference. Um they're really  good at executing a defined work right. They [1:39] use tools to complete assigned task and automate  repeatable work uh workflows and across enterprise [1:45] systems. uh a lot of familiar faces like codeex  cloud code uh Microsoft copilot studio agent force [1:54] open claw you name it but right now we have  arrived at the age of discovery and optimize [2:01] uh and that's what practis does autonomous  research and optimization for those of you [2:05] that follows the news right um Jeff Dean the chief  scientist of Google deepmind just announced his [2:12] u exit from Google to join to to go to his own  startup called discovery loop which is practically [2:18] practically doing the same thing. Uh so practiced  we can discover validate and improve. Practice [2:25] understands your objectives data constraints and  success metrics and autonomously designs and runs [2:31] long horizon experiments learn from the results  and delivers validated improvements and solutions. [2:37] So basically you know large angle models they can  generate agents they can execute but right now we [2:41] got uh autonomous research systems like practice  they can discover what works and helps you build [2:47] it. So in short uh you know all businesses they  are made up of decisions um and practice can [2:53] optimize every single decision for you as long  it's measurable. uh so practices and autonomous [2:58] research analyst team and research team that can  understand your KPIs your constraints data and [3:04] codes even there there is and then designs runs  and val validate experiments to improve business [3:10] outcomes at a state-of-the-art level so you're  basically you're getting a McKenzie consultant [3:15] team or a McKenzie or a BCG consultant team with  researchers so they don't they not only give [3:21] you advice but they will actually implement and  deliver for you to to implement those changes for [3:26] you. So, [snorts] you probably wake up to these  questions or your company probably wake up with [3:30] these questions every day. Um, how much inventory  should we keep? Um, what roots should we should we [3:36] take? How should we allocate capital? You know,  what production schedule performs best? Uh, [3:41] how should we model or or control how should we  model a control policy respond? Or if you're on [3:46] a tech side, if your company is an engineering  or R&D side, you're probably wondering how can [3:51] I make a best performing robot control system?  How can I make a best performing rocket guidance, [3:56] image recognition, materials design, manufacturing  processes or autonomous research, autonomous [4:01] systems? Um, so you probably have a team for these  already. You probably have a system or at least [4:06] you have a model for these. You better be modeling  your inventory and sales, right? Uh, it doesn't [4:12] uh you're definitely behind the game, but you it  the question you're probably asking wondering is, [4:17] is there room for improvement? And the answer  is heck yeah. Um the teams using practice they [4:24] can probably try a 100 thousand times different  methods and pathways than you know teams without [4:30] using practice they would ever imagined. Um AIS  they can do practice run can run 24/7 they can try [4:36] so many different stuff uh on on a ground basis.  Uh so as and also as demand markets constraints [4:42] and risks change these decisions must change too.  Uh it's it's an active it's a dynamic system. So [4:48] if a decision or a research that can be modeled  and measured, practice can help can help you [4:52] to model your business research and continuously  optimize it. So this is a real world case. We're [4:59] going to use this uh as a uh this is a real case  with a we're working with a company a very big [5:05] energy transportation company. So their request is  basically to optimize the petroleum transportation [5:10] route. They have the pipelines, they have the  shipping across to uh 136 countries. Um and uh [5:18] for for the codeex or cloud code users, this might  look very familiar. You just put your data uh if [5:23] you have a current solution or you have a current  baseline, just put it in there in a folder in a [5:28] directory uh and then define your KPI and goal  uh or your uh evaluation metrics and also your [5:34] system constraints. Give it to PRAIS to hand over.  Uh practice will utilize a multi- aent system. It [5:40] will assemble a world-class team that'll work at  about a 3,000 times uh times the human efficiency. [5:46] So actually it will produce a lot of results and  experiments. It will actually run the experiments [5:50] for you. Um so conductments uh set up simulations  and uh iterates the solutions and then it will go [5:58] to verify and validate the results through  ablation studies and prepare for the next [6:01] iteration. So you just you just treat them like  your own AI researcher, research scientist or a [6:07] analyst. uh you give it what you gave to a real  person. It's like a server or a laptop with with [6:15] GPU preferred. Um and then it will give you a  thoroughly optimized state-of-the-art model for [6:21] your given petroleum transportation network. Um  so before introducing practice into the system, [6:27] um the team that was at the company, it's a  it's a five it's a five people team. uh they [6:32] they actually did a 13% increase in the cost and  efficiency. Uh that's a big deal. It's already [6:40] at the hundreds of millions level in improvements  because you know it's metro is very large. Uh but [6:46] after introducing practice we were able to bump  that into a 29%. That means actually makes a lot [6:51] of changes uh for that and that's within a matter  of days. So in a normal research routine, right, [6:57] you have an average research or a analyst group  that will that will come in and then take over set [7:04] uh set up the environment do the research um and  set up simulations uh experiments and tune your [7:11] model uh iterations abations verification  validation that it's going to take about at [7:16] least a few months to do so uh whether in in  whatever industry. Um so practis can do that [7:22] all in a closed loop system and the entire process  only takes about just a few days up to a week or [7:29] two. Um and it will actually generate sotai level  state-of-the-art level productions. [snorts] Oh [7:36] got plenty of time. So um so far we have  uh collaborated with uh around 30 different [7:42] enterprises about across industries. Uh we have  done about uh more than a 100 different tasks [7:47] across different industries. Uh we have quant  trading. We were able to bump that. We're actually [7:51] working with some of the largest quant trading  companies. Uh they're they're live right now. [7:56] Uh we have to bump their uh returns and minimize  their draw downs. Um and also the second one is a [8:04] uh a outdoor robotics company we're working with.  So they're trying to make a uh slam algorithm for [8:10] a robot. They have a 10 people team, 10 person  team that worked on this subject for about six [8:16] months. and they uh they they came up with a  system that's uh 6.4 meters uh accumulate error [8:23] which is not bad. It's already at uh sly level.  So we put that entire system baseline uh into [8:29] practice and ran for about 5 days. Uh and we gave  them back a.9 meter mcumid error solution. Uh and [8:37] it just it just can be plug and play. Uh you don't  need any modification. Just put it in their system [8:43] and run. Um and then the third one's uh uh rocket  landing algorithm we developed for a commercial [8:48] aerospace company. Uh we can do nuclear fusion.  Uh so these are more tech related because you know [8:54] I show the muscle. Uh a lot of thing other things  we can do in terms of uh practice. We got a lot of [9:00] feedbacks from uh our our collaborators. Practice  can 100 100x my workflow efficiency allowing me [9:07] to explore more higher concepts. So we're not  replacing people. uh practice is a tool for um [9:13] you know analysts or researchers that can actually  help you to uh do all those experiments run those [9:18] experiments autonomously. uh and you know some of  for some of the companies app implementation they [9:23] reached state-of-the-art results that our  competitors would never have imagined and [9:27] within just days or weeks uh things like that  uh and then uh yeah worthy to mention we got [9:34] uh three of the works that practice actually took  place on is already uh submitted to nature or ML [9:41] is actually you will be able to see in the uh  actually it already took place so uh yeah so we [9:48] can we actually transformed from so many different  other cases. Uh whether it's a business decision [9:53] or a tech decision. Uh you you probably see a  lot of a medical on there. Uh there a lot of [9:58] u a lot of medical uh image recognition or uh  system optimization problems out there. Um you're [10:07] going to see so uh let's do an interesting case  study. So this is the case that we were doing with [10:12] a um with a commercial uh commercial aerospace  company. Um so the task is basically they handed [10:20] over a uh simulator where their researchers have  been running very uh on that for at least a uh six [10:27] months and they they got to 70% safe landing rate  on that algorithm which is not bad very impressive [10:34] it's already state-of-the-art and they published  that on a uh IE paper um so the challenge is you [10:42] go we got two kilometers in height and then within  that two kilometers we got to save the rocket [10:47] uh that which had a 1,500 meter horizontal u  offset. So and and on top of that you got all the [10:54] crosswind you got all the like uh very adversary  adversarial uh conditions um and we give it a [11:02] u we define a success metric. We we want to  maximize the landing rate obviously safe landing [11:07] rate and we also want to minimize the landing  error. We want to make the landing more gentle [11:12] the touchdown speed and tilt. Um so and we put  that into the system and then just let practice [11:17] take over. Uh practice will automatically we gave  practis a 8GPU server so it will actually just set [11:23] up the all autonomously with our human in the  loop set up in the environment. And worthy to [11:28] mention practis can uh can produce about 300 to  500 findings per night and each of the findings [11:35] is pretty much equivalent to a uh week worth  of work done by an average PhD candidate. So [11:42] uh it automatically iterate this system for  about 60 75 generations uh and it return us a [11:50] uh save rate of 90% uh landing rate. Uh and uh  before that the best they got they ever got into [11:57] was uh 70% that's by uh the human researchers. Um  so there it's going to try a lot of things. It's [12:03] going to look up the new latest papers on this.  It's going to look look up everything that's out [12:07] there and it's going to implement those studies  and um and results and it only takes three days [12:13] and uh we we were using a cloud-based ser uh model  uh and we can also you know deploy in the locally [12:19] that wouldn't cost any other token charges but uh  a cloud-based model it takes about $1,000 to run [12:26] this task. Um oh yeah this is another task that we  built for a cargo delivery uh company. So they're [12:32] trying to optimize the cargo delivery. this only  ran for about just a few hours and you get entire [12:38] system not just a solution you get a system  that will actually fit into your needs uh and [12:42] that will work every time and with practice you  can even dynamically optimize the system as well [12:48] um yeah so the differences between uh practice and  uh codeex and cloud code and you know probably you [12:55] guys have heard of sakana ai they're a they're  one of the uh they are one of the inventors of AI [13:01] scientist back in 2021 um So it's basically how we  evolve between generations. So uh cloud cloud code [13:10] and codeex are more goal oriented. It's more like  a linear trace. They're really good at you know a [13:15] lot of task more general tasks like building  you a uh app or uh fixing some of the bugs [13:21] um or like building a website. Um so it's they're  very good. We're not we're not saying they're [13:28] bad and you know but on ter in terms of uh top  analysis [clears throat] and research yield uh [13:33] we we do perform significantly better. Uh Sakana  AI they were one of the inventors of uh scientist [13:38] AI scientist. Um but their science their system  evolves on a tree structure. So they will try I [13:45] don't know like hundred different stuff and they  will pick the champion seed that's in there and [13:49] they will evolve based on the champion seeds into  the next generation. Uh it's a very good working [13:54] system, but why it's underperforming is because  uh they're throwing away all the other seeds [14:00] out there. That's not a champion seed. And the  problem with that is, you know, a seed might not [14:04] be performing well because uh it might not be just  be it's a bad solution. Uh it it might just be it [14:10] has the wrong parameter set or it's not trained  to enough epoch or or it's not tuned enough. Um so [14:18] there are a lot of things that it's throwing away  and it's it's very narrowing down on the path of [14:23] uh research. Um so that's why they're not really  um performing that well overall in terms of the [14:29] deliveries. Um so they're throwing away a lot of  stuff. So we use something called a linear DAG [14:35] uh which will utilize everything from the seeds.  We will actually have a shared memory and we [14:40] have a falsifier and a exploer and a um and an  innovator uh character that will actually take [14:46] what's good in all the seeds and then put that  in a uh uh shared memory to transfer them into [14:54] uh knowledge or experience and then take that  entirely into the next generation. So that evolves [15:00] uh with greater productivity and with greater  uh efficiency and effectiveness as well. [15:07] Um so across the 100 different tasks that we  have tried uh and on a lot of the benchmarks [15:13] which we well going to release in uh very soon  um we can we hit about a 72% soda level rate [15:22] uh with a cost of depending on how you want to  deploy the system. If you want to deploy on your [15:27] own server, we can do that uh with the model. Um  [snorts] uh that's going to cost you practically, [15:33] you know, $0 as in terms of the token  cost. Uh if you were to use cloud code, [15:39] uh we got a under the same uh environments, we got  about 45% in in all those tasks. Uh and it's going [15:46] to cost it costed us at least about uh 48,000  about 50,000 in token fee uh in total. And then [15:55] uh another famous uh it's called the faro agent.  It's by a Chinese company called BU. They they [16:00] claim to be they're good on these kind of tasks  as well. They only got about 38%. I think it's [16:05] a little too light but yeah but we don't know  the cost they're using Gemini 3.1 Pro. So which [16:11] is probably going to be on the expensive side as  well. Um so in terms of security and u governance [16:19] uh we we have a lot of freedom on the system. If  your company wants it to be on a shared cluster, [16:25] we can do that. It's more it's the most probably  the most cost efficient. Uh and we can deploy this [16:30] on a isolated cloud or on a bare metal. And if you  want, you know, maximize control, uh we can deploy [16:36] this entire system with a model uh on a on-remise  server in your uh in your on location or locally [16:44] deployed. Um yeah, and uh that's basically it.  Uh so we're going to release this very soon to [16:50] the general public and we're doing a lot of uh  collaborations with the enterprise already uh [16:56] some of the largest uh manufacturers uh R&D or  business or shipping companies uh already. So [17:02] if you're interested you can contact us at  the email or find us at booth uh 15641564. [17:11] I think we have a few minutes. That's  design. Any questions? We're going to [17:15] take some questions. All right. I'm  going to just come down. Yes. So, [17:19] how do you connect to to the tools and how do you  connect to the simulators uh environments that the [17:25] companies or teams have already? Yeah. Uh that's  a very good question. Uh thank you Miguel. Um so, [17:33] so basically what we will do is we will uh  do a FDE or uh depending on the need of the [17:38] company. It's more like a case by case because  you know everyone has different tool sets. [17:42] uh but generally it's more on the agentic level  as long as you have uh the tool can be easily [17:48] manipulated like like for example on engineering  side you have answers you have solid works uh we [17:53] can actually we can definitely connect to those  uh on uh a lot of analyst tools uh we can connect [17:58] to those as well uh so basically if you have a  current system that's working we can definitely [18:02] just you know plug into that right into that for  you uh and if you have a baseline if you have a [18:08] you know human programmer or like a analyst that's  that's writing the code for you. Uh he he can just [18:14] use this as a cloud code or codeex but for long  horizon research analysis. Uh they can literally [18:19] just put in their baseline into the system and  it will dynamically optimize it for them. Uh [18:24] basically like that. Yes. So my question is my my  question is I work in new product development for [18:30] R&D and manufacturing. A lot of times we don't  have the simulation software for new product [18:36] development because it just hasn't been developed  yet. We're trying to understand what new material [18:39] we can introduce and some of those are like  complex chemical nonlinear equations and nonlinear [18:45] simulations. So when it comes to like the end  process of you've gotten an idea come through [18:49] the research task all the way to the testing  phase of whether that's even feasible. How do [18:54] you guys incorporate that? Is that something that  comes from your end or is it that we are expected [18:57] to bring our domain knowledge to the point of  building a simulator that you can then connect? [19:02] Thank you. Uh that's a very good question. So um  again it really depends on a case. I I I don't [19:07] have enough information on that but usually you  can just define a uh evaluation metric. Uh just [19:14] give us the definitions and we should be able to  the the system should anonymously build that for [19:19] you. If you don't have a simulation if you have a  simulation itself it will be the best uh because [19:23] you'll just ex you know strictly follow what's on  the simulator. Uh but if not uh you can definitely [19:29] build it but you know it might there might be cost  of performance uh on that or like differences.