PNNL Benchmarks Show Human-AI Networks Outperform Solo Frontier Models

Networked Intelligence: Scaling Human-AI Networks with Active Context Graphs With PNNL — Ai4

Ai4

Standalone frontier models stumble on ambiguous discovery tasks because high-consequence research requires synchronized context graphs and human verification to prevent compounding errors.

AI-assisted summary of Networked Intelligence: Scaling Human-AI Networks with Active Context Graphs With PNNL — original source. Check the original for full context.

Source published: 2026-08-17

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 17, 2026, this presentation features Sutanay Choudhury, Chief Scientist at Pacific Northwest National Laboratory (PNNL). According to the video description, the talk argues that scaling individual model context windows is hitting a plateau, introducing the 'Mycelium' runtime architecture and its Active Context Graph to coordinate distributed human-AI scientific campaigns.

## TL;DR Sutanay Choudhury argues that solving complex, high-consequence enterprise and scientific problems requires shifting focus from isolated model scaling to networked intelligence among distributed human-AI teams. To overcome the limits of standalone autonomous reasoners, PNNL developed Mycelium, an orchestration system built around an Active Context Graph that tracks hierarchical computational provenance and routes findings via dynamic push and pull mechanisms. The architecture implements strict epistemic weighting, sprint-aware goal alignment, and human approval gates to eliminate autonomous error compounding. Empirical evaluations demonstrate that while autonomous frontier models excel at narrow, well-defined prompts, human-AI networked intelligence significantly outperforms them in ambiguous, open-ended scientific discovery, prompting an upcoming open-source release.

## Atmosphere & style The presentation carries an authoritative, pragmatic, and intellectually rigorous tone rooted in Choudhury's decades of national laboratory experience across high-stakes domains like cybersecurity and biology. Rejecting uncritical frontier model hype, the delivery methodically unpacks the practical failure modes of solo autonomous agents with clear architectural logic and measured skepticism. It feels both grounded in messy day-to-day scientific reality and forward-looking, building from distributed systems mechanics to an inspiring closing call to dismantle organizational silos and 'let the mycelium spread.'

## Key takeaways - [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) Individual model scaling is insufficient for high-consequence domains like biology, cybersecurity, and national security, which fundamentally rely on distributed teams with tacit domain knowledge. - [03:44](https://www.youtube.com/watch?v=aG8foA_qXgY&t=224s) Networked intelligence differs from workspace over-broadcasting by focusing on high-precision routing to deliver updates specifically to the people or agents whose decisions they impact. - [05:14](https://www.youtube.com/watch?v=aG8foA_qXgY&t=314s) The Mycelium runtime replaces single-user chat setups with an Active Context Graph that serves as a hierarchical provenance index linking chat transcripts, datasets, and background computations. - [08:43](https://www.youtube.com/watch?v=aG8foA_qXgY&t=523s) Mycelium utilizes frontier models to infer data schemas from natural language descriptions and orchestrates fault-tolerant computational DAGs recorded dynamically as Jupyter notebooks. - [08:43](https://www.youtube.com/watch?v=aG8foA_qXgY&t=523s) Contextual updates are shared across multidisciplinary collaborators via dual mechanisms: proactive push recommendations and direct graph pull queries. - [12:47](https://www.youtube.com/watch?v=aG8foA_qXgY&t=767s) To prevent error compounding and noise drift, unconstrained autonomy is restricted: no autonomous agent finding is committed to the shared context graph without human verification and approval. - [12:47](https://www.youtube.com/watch?v=aG8foA_qXgY&t=767s) Effective human-AI collaboration requires epistemic weighting to calibrate the evidential influence of findings based on collaborator expertise and seniority. - [12:47](https://www.youtube.com/watch?v=aG8foA_qXgY&t=767s) Agent behavior and recommendations must dynamically adapt to the project lifecycle, transitioning from early data exploration and quality control to late-stage experimental synthesis. - [17:30](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1050s) Empirical evaluations demonstrate that fully autonomous frontier models yield polarized results on ambiguous discovery tasks, whereas networked human-AI execution maintains consistent breadth and depth. - [19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) The Active Context Graph runtime is domain-agnostic—applicable across science, finance, and medicine—and PNNL is preparing an open-source release of the platform.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) Beyond Model Scaling: The Case for Networked Intelligence in Complex Domains

Over the last decade, artificial intelligence progress has been almost exclusively driven by scaling individual models—expanding context windows, lengthening autonomous execution horizons, and pushing raw model capabilities. However, Sutanay Choudhury challenges the premise that handing an ultra-capable frontier model all of an organization's datasets will magically yield answers that satisfy rigorous regulatory, legal, or fiduciary standards: "we believe the answer to that question is largely no." Drawing from his ten to twenty years of experience at Pacific Northwest National Laboratory (PNNL) across high-consequence domains like cybersecurity, power grid resilience, chemistry, and biology, he argues that critical scientific and national security challenges are fundamentally never solved by a single, isolated reasoner.

Real-world enterprise problems require distributed teams holding distinct disciplinary priors, tacit knowledge, and domain-trained intuitions. Furthermore, work in these domains does not operate like standard software development sprints; it progresses asynchronously in response to dynamic events, such as an emerging cyber intrusion or the sudden completion of an experimental run. When events occur, human experts must synthesize new findings, consult collaborators, and navigate subsequent decision points. Complementing the model scaling coming out of frontier AI labs requires an orthogonal dimension of scaling: building human-AI collaborative networks that scale the connections between people, agents, and instruments.

- **Limits of Solo Reasoners:** Frontier models fed complete data lakes cannot substitute for the diverse domain expertise, accountability, and regulatory compliance demanded in high-stakes fields. - **Asynchronous Team Workflows:** Mission-driven scientific research proceeds through event-driven, distributed interactions rather than isolated, linear sprints.

[03:44](https://www.youtube.com/watch?v=aG8foA_qXgY&t=224s) Defining Networked Intelligence and the "Mycelium" Substrate

Choudhury defines *Networked Intelligence* by contrasting it with common industry approaches. It is not merely giving an AI model all available data, nor is it the "over-broadcasting" seen in shared project workspaces like Anthropic Claude Projects or ChatGPT Team workspaces, where every update is dumped on every participant. Instead, networked intelligence centers on intelligent, high-precision routing: determining how a change in data—whether a cyber incident alert, a new scientific dataset, or an updated patient record—can reach the exact person or agent whose decisions it will influence.

```mermaid flowchart TD accTitle: Mycelium Active Context Graph Architecture accDescription: Flow diagram showing how heterogeneous tools, compute, and user interactions synchronize through the Active Context Graph.

subgraph DataCompute["Heterogeneous Compute & Data Layer"] D1["Raw Scientific Datasets"] D2["Dynamic Jupyter Notebooks"] D3["GPU Clusters & Simulations"] end

subgraph RuntimeEngine["Mycelium Orchestration Layer"] MCP["MCP Server & Tooling"] ACG["Active Context Graph\n(Provenance & Routing Index)"] end

subgraph UserAgents["Multi-Principal Collaboration"] U1["Jeff\n(Proteomics Expert)"] U2["Lumi\n(Mechanistic Biologist)"] AG["Autonomous Background Agents"] end

D1 --> MCP D2 --> ACG D3 --> MCP MCP --> ACG U1 <--> MCP U2 <--> MCP ACG <--> AG ACG -->|Push / Pull Recommendations| U1 ACG -->|Push / Pull Recommendations| U2 ```

The system is named **Mycelium** after the underground fungal network that weaves beneath soil, linking plant root systems together to distribute nutrients and biochemical signals across an entire ecosystem. In the same manner, Mycelium functions as a shared knowledge substrate for multidisciplinary teams, maintaining organizational memory and coordinating human-agent interaction.

[05:14](https://www.youtube.com/watch?v=aG8foA_qXgY&t=314s) Architectural Shift: From Single-User MCP to Multi-Principal Context Graphs

To demonstrate the runtime architecture, Choudhury presents a "hello world" collaboration scenario between two scientists: Jeff, a proteomics specialist studying how protein expression varies across experimental conditions, and Lumi, a mechanistic biologist focused on identifying the specific biological pathways triggered by those conditions. In a conventional single-user setup, Jeff interacts with a chat client connected to a Model Context Protocol (MCP) server, which accesses heterogeneous tools, datasets, and high-performance GPU clusters for simulations.

The coordination challenge begins when Lumi joins the session wanting to investigate the same biological system by picking up precisely where Jeff left off. The architecture introduces an **Active Context Graph** to bridge their workflows. Rather than functioning as a standard vector database or static graph database, the context graph acts as a hierarchical provenance index that binds conversational logs, execution traces, background computations, and emerging hypotheses together. When assessing any scientific conclusion or decision, the graph links the human dialogue directly to the underlying computational history.

- **Heterogeneous Scientific Backends:** Integrates disparate data formats, analytical tools, and GPU/EC2 computation clusters through unified MCP interfaces. - **Hierarchical Provenance Indexing:** Couples chat transcripts directly with background computational scripts, data lineage, and decision paths.

[08:43](https://www.youtube.com/watch?v=aG8foA_qXgY&t=523s) Fault-Tolerant DAG Workflows and Dynamic Notebook Provenance

Scientific data is notoriously heterogeneous and messy, making it impractical to force researchers to format inputs into rigid tabular schemas or strict JSON structures. Taking advantage of model scaling, Mycelium uses frontier models (such as Claude Opus-class LLMs) to infer data schemas directly from natural language explanations—asking the user to describe the data just as they would to a colleague—and caches these schemas for future interactions.

Behind a straightforward user prompt such as "Show me what are the top proteins in this particular data set," the system orchestrates a complex directed acyclic graph (DAG) of computations. It identifies target datasets (e.g., glycolysis pathways), selects the appropriate tables and columns, and schedules workflows ranging from simple histograms to spectral and K-means clustering across distributed EC2 instances. Because distributed execution can fail, the runtime incorporates fault-tolerant checkpointing to resume execution seamlessly. Every computational action is dynamically recorded as a Jupyter notebook tied directly to the context graph via unique identifiers.

Information is shared across the network via two core modes: - **Push Mechanism:** Proactive recommendations surfaced to a user (e.g., notifying Lumi upon login of relevant findings generated by Jeff). - **Pull Mechanism:** Direct graph queries formulated by the user (e.g., Lumi asking the context graph to summarize Jeff's recent analyses).

[12:47](https://www.youtube.com/watch?v=aG8foA_qXgY&t=767s) Calibrating Autonomy, Error Compounding, and Epistemic Weighting

While background agents can infer user intent and spawn autonomous computational investigations after researchers log off, Choudhury notes that unconstrained autonomy quickly leads to user frustration. To prevent noise compounding and unverified drift, Mycelium enforces a strict policy: **no autonomous agent finding is published to the shared context graph without explicit human review and approval**. Allowing unverified autonomous computations into the shared graph risks compounding errors down the line, making original sources of failure impossible to isolate.

```mermaid flowchart LR accTitle: Sprint Lifecycle and Epistemic Context Graph Flow accDescription: Flow showing the evolution of collaborative tasks during a research sprint alongside epistemic weighting.

subgraph SprintLifecycle["Sprint Temporal Evolution"] S1["Early Sprint:\nData Exploration & QC"] S2["Mid Sprint:\nDeep Mechanistic Analysis"] S3["Late Sprint:\nExperimental Proposals"] S1 --> S2 --> S3 end

subgraph Governance["Networked Intelligence Guardrails"] EW["Epistemic Weighting:\nIntern vs. Senior Fellow"] HITL["Human Approval Gate:\nNo Unchecked Publishing"] end

SprintLifecycle -.-> Governance ```

Choudhury breaks down three foundational mechanisms for amplifying networked intelligence: 1. **Network-Informed Recommendations with Epistemic Weighting:** Recommending paths based on team actions while addressing "epistemic weighting"—calibrating the influence of findings from an enthusiastic high school intern versus a veteran, conservative laboratory fellow. 2. **Intent-Driven Background Investigation:** Allowing agents to run autonomous computations based on shared team interests while remaining constrained by human verification gates. 3. **Sprint-Aware Global Goal Alignment:** Dynamically adapting agent assistance to the team's project stage—shifting from data exploration and quality control early in a sprint, to deep scientific analysis in the middle, to experimental synthesis and decision-making near deadlines.

[17:30](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1050s) Empirical Evaluation: Autonomous Frontier Models vs. Human-AI Networks

To evaluate the hypothesis that individual model scaling is insufficient for open-ended scientific campaigns, Choudhury presents benchmark data comparing fully autonomous frontier models (Claude Opus 4.8 executed with maximum parallelism, compute resources, and autonomy) directly against the human-AI Mycelium system.

| Metric / Execution Paradigm | Fully Autonomous Opus 4.8 (High Autonomy & Parallelism) | Mycelium Networked Intelligence (Human-AI Network) | |:--- |:--- |:--- | | **Task Execution Profile** | High performance (scores of 4 and 5) on narrowly scoped, highly specific tasks. | Consistent breadth and depth across both open-ended and specialized scientific queries. | | **Result Distribution** | High sparsity, characterized by polarization (numerous 4s/5s interspersed with 0s and 1s). | Robust, well-distributed scoring across exploratory, analytical, and synthesis tasks. | | **Exploration & Serendipity** | Struggles when problem framing is ambiguous or initial search directions are unconstrained. | Combines algorithmic execution with human domain intuition, lateral reasoning, and serendipitous discovery. |

The empirical results show that while standalone frontier models execute tightly defined tasks exceptionally well, they stumble on ill-defined discovery problems where the starting point is ambiguous, such as dissecting a novel biological mechanism or triaging an uncharacterized cyber intrusion. Combining human creativity, domain expertise, and serendipity with AI-driven computational infrastructure provides the optimal balance of breadth and depth. PNNL is actively expanding these capabilities across five to six biological pilots and scaling multi-agent coordination across distributed data centers.

[19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) Expanding Beyond Biology: Enterprise Intelligence and Open-Sourcing Mycelium

Reflecting on the broader implications of the Mycelium framework, Sutanay Choudhury brings the presentation to a close by offering three overarching conclusions. While the demonstrated empirical breakthroughs were grounded in biological scientific campaigns, the underlying runtime architecture and Active Context Graph design are far from domain-specific. The model is built to generalize across complex, high-stakes environments—ranging from fundamental science and national security to enterprise domains such as finance and medicine. Choudhury emphasizes that the team is actively eager to connect and collaborate with practitioners tackling similarly intractable challenges in these sectors, where critical insights remain trapped in silos.

Achieving this vision extends well beyond merely constructing an orchestration harness. The deeper challenge lies in understanding how to harvest the fast-moving insights generated during scientific sprints, and how to methodically link those disparate "capsules of knowledge" to cultivate true enterprise-grade intelligence. To accelerate adoption and invite community experimentation across industry and academia, the team is nearing a full open-source release of the platform. Closing on an evocative note borrowed directly from his biological research collaborators, Choudhury leaves the audience with an invitation to join the networked movement: "let the mycelium spread."

```mermaid flowchart TD accTitle: Cross-Domain Mycelium Architecture subgraph Core ["Active Context Layer"] C1["Capsules of Knowledge"] C2["Active Context Graph"] C1 --> C2 end subgraph Domains ["Application Domains"] D1["Biology & Science"] D2["National Security"] D3["Finance"] D4["Medicine"] end C2 --> D1 C2 --> D2 C2 --> D3 C2 --> D4 ```

- **Cross-Domain Generalization:** The architecture applies directly to science, security, finance, and medicine, bridging domain-specific tacit knowledge. - **Enterprise-Grade Scaling:** The objective shifts focus from simple test harnesses to synthesizing synchronized "capsules of knowledge" across distributed teams. - **Open-Source Availability:** Pacific Northwest National Laboratory is preparing an open-source release of the Mycelium tooling for public experimentation and development.

## Notable quotes > The current paradigm of AI scaling relying on massive context windows and long-horizon agentic execution is rapidly approaching a plateau. — Sutanay Choudhury

> Challenging scientific problems are rarely solved by one isolated reasoner; they are solved by distributed teams carrying different priors, tacit knowledge, and domain-trained intuitions. — Sutanay Choudhury

> The open problem is therefore not just how to scale individual models, but how to cultivate networked intelligence: scaling the connections between humans and AI systems so that a hypothesis produced in one context reaches the exact person, agent, instrument, or robot that can act on it. — Sutanay Choudhury

> [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) we believe the answer to that question is largely no. — Sutanay Choudhury

> [08:43](https://www.youtube.com/watch?v=aG8foA_qXgY&t=523s) Show me what are the top proteins in this particular data set — Sutanay Choudhury

> [12:47](https://www.youtube.com/watch?v=aG8foA_qXgY&t=767s) no autonomous agent finding is published to the shared context graph without explicit human review and approval — Sutanay Choudhury

> [17:30](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1050s) Combining human creativity, domain expertise, and serendipity with AI-driven computational infrastructure provides the optimal balance of breadth and depth. — Sutanay Choudhury

> [19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) capsules of knowledge — Sutanay Choudhury

> [19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) let the mycelium spread. — Sutanay Choudhury

## Mentioned Sutanay Choudhury, Pacific Northwest National Laboratory (PNNL), Ai4, Mycelium, Active Context Graph, Anthropic, Claude Projects, ChatGPT Team, Model Context Protocol (MCP), Claude Opus 4.8, Jupyter Notebook, Amazon EC2

## Video description Sutanay Choudhury - Chief Scientist, Pacific Northwest National Laboratory

The current paradigm of AI scaling relying on massive context windows and long-horizon agentic execution is rapidly approaching a plateau. Challenging scientific problems are rarely solved by one isolated reasoner; they are solved by distributed teams carrying different priors, tacit knowledge, and domain-trained intuitions. The open problem is therefore not just how to scale individual models, but how to cultivate networked intelligence: scaling the connections between humans and AI systems so that a hypothesis produced in one context reaches the exact person, agent, instrument, or robot that can act on it. The talk will introduce Mycelium, a runtime architecture using an Active Context Graph to connect isolated users, agents, and instruments into a synchronized network. I will detail empirical success stories showing how active context propagation fundamentally altered the trajectory of a live scientific campaign. Join this session to see a direct contrast between networked execution and standalone autonomous models, providing actionable evidence of a completely new axis for AI scaling.

[Source](https://www.youtube.com/watch?v=aG8foA_qXgY)

中文简报

PNNL基准测试表明人机网络表现超越单体前沿模型单体前沿模型在模糊探索任务中面临瓶颈，因为高风险研究需要同步上下文图谱与人工验证以防止错误累积。 Sutanay Choudhury 认为，要解决复杂且具有重大影响的企业与科学问题，必须将重心从孤立的模型扩展转向分布式人机团队之间的网络化智能。为了克服独立自主推理模型的局限，PNNL 开发了编排系统 Mycelium，该系统围绕“活跃上下文图”（Active Context Graph）构建，可追踪分层计算溯源，并通过动态推送和拉取机制传递发现。该架构实施了严格的认识论加权、敏捷冲刺感知的目标对齐以及人工审批关卡，以消除自主错误累积。实证评估表明，尽管自主前沿模型在狭窄且定义明确的提示词任务中表现优异，但在模糊、开放式的科学探索中，人机网络化智能的表现显著胜过单体模型，这也促成了该系统的即将开源发布。对于生物学、网络安全和国家安全等从根本上依赖具备隐性领域知识的分布式团队的高风险领域，仅靠单一模型的规模扩展是不够的。 (0:04) 网络化智能与工作区的信息过度广播不同，它侧重于高精度路由，专门将更新推送给决策会受到影响的人员或智能体。 (3:44) Mycelium 运行时取代了单用户对话机制，采用活跃上下文图作为分层溯源索引，将聊天记录、数据集和后台计算连接起来。 (5:14) Mycelium 利用前沿模型从自然语言描述中推断数据模式，并编排以 Jupyter notebook 形式动态记录的容错计算 DAG。 (8:43) 跨学科协作者之间通过双重机制共享上下文更新：主动推送推荐与直接图谱拉取查询。 (8:43)

Reading summary

Sutanay Choudhury of Pacific Northwest National Laboratory argues that scaling individual frontier AI models cannot resolve complex challenges in high-consequence fields. Instead, mission-driven domains require networked intelligence systems that connect distributed human experts, autonomous agents, and computational backends through active context graphs. Limits of Solo Models in Complex Domains (0:04) Progress in artificial intelligence has focused on scaling individual model parameters, context windows, and autonomous task horizons. However, Choudhury contends that feeding entire enterprise datasets into a frontier model will not satisfy the legal, fiduciary, and regulatory standards required in high-stakes sectors like power grid resilience, national security, chemistry, and biology. Real-world scientific research relies on asynchronous, event-driven workflows carried out by teams with divergent disciplinary training and tacit knowledge, rather than isolated, linear sprints. Because no single reasoner possesses the breadth to solve these multidisciplinary problems, AI deployment must shift toward an orthogonal dimension: scaling the interactive fabric between people, algorithms, and scientific instruments. Mycelium Architecture and the Active Context Graph (3:44) To coordinate distributed discovery, PNNL developed Mycelium, an orchestration system named after fungal root networks. Rather than over-broadcasting every project update to all users—a common failure mode in shared AI workspaces—Mycelium routes dynamic updates exclusively to the specific collaborators or agents whose decisions rely on that data. At the center of this runtime is the Active Context Graph. Operating across Model Context Protocol interfaces, the graph acts as a hierarchical provenance index rather than a static database. It links conversational logs directly to underlying computations, distributed execution directed acyclic graphs (DAGs), dynamic Jupyter notebooks, and datasets. Researchers can engage the network proactively through pull queries or receive push recommendations as collaborators generate new findings. Governing Autonomy and Calibrating Epistemic Influence (12:47) While autonomous agents within Mycelium can infer user intent and execute computational tasks in the background, Choudhury cautions that unconstrained autonomy rapidly introduces noise and error compounding. To preserve reliability, Mycelium mandates a strict human-in-the-loop gate: no autonomous agent finding is committed to the shared context graph without explicit human review and approval. Additionally, the framework integrates epistemic weighting and sprint-aware goal alignment. Epistemic weighting calibrates the influence of inputs across varied team members—balancing proposals from junior interns against veteran laboratory fellows. Sprint-aware alignment dynamically alters the nature of agent support as projects progress, shifting from data exploration and quality control early in a sprint to deep mechanistic analysis and synthesis near deadlines. Empirical Performance: Autonomous Agents Versus Networked Teams (17:30) PNNL benchmarked the human-AI Mycelium system against a standalone frontier model (Claude Opus 4.8) configured with maximum compute, parallelism, and autonomy. The evaluation revealed that while solo frontier models excelled on tightly specified tasks, their performance was highly polarized—producing top scores alongside complete failures when facing ambiguous or unconstrained problem definitions. In contrast, the Mycelium human-AI network maintained consistent depth and breadth across exploratory, analytical, and synthesis workflows. Choudhury attributes this robustness to the combination of algorithmic compute and human lateral reasoning, which successfully guides search trajectories when starting points are ill-defined. PNNL is currently deploying the Mycelium framework across five to six biological pilots and preparing an open-source release to encourage broader evaluation. Choudhury emphasizes that the core architecture is domain-agnostic, designed to generalize across medicine, finance, and national security, though the broader challenge remains methodically capturing fast-moving sprint discoveries and turning them into synchronized institutional knowledge.

精炼阅读摘要

太平洋西北国家实验室（PNNL）的 Sutanay Choudhury 指出，在网络安全、电网韧性与生物学等高风险专业领域，仅靠扩大单体前沿模型的规模和灌注全量数据无法满足严格的监管与信托决策标准；解决复杂现实问题必须转向“网络化智能”（Networked Intelligence），通过动态上下文图谱连接分布式人类专家、自主代理与异构算力工具。单体模型局限与网络化智能定位 (0:04) Choudhury 结合其在 PNNL 跨多个关键技术领域的经验提出，过去十年人工智能的进展几乎完全由单体模型的扩展驱动（如扩大上下文窗口、延长自主执行周期），但在应对高风险任务时，单一隔离的推理系统无法替代具备不同先验背景、隐性经验与专业直觉的分布式团队。这类高精尖领域的任务推进并非标准软件开发的线性迭代，而是由动态事件驱动的异步协作过程（例如突发网络入侵事件或实验阶段性完成），需要专家在动态演变中持续综合新发现并协同研判。因此，网络化智能并非单纯向大模型提供全量企业数据，亦非 Anthropic Claude Projects 或 ChatGPT Team 等团队工作区中常见的全员无差别广播；其本质在于高精度的智能路由机制，即在数据、警报或研究进展发生变动时，准确将其定向推送至决策直接受其影响的具体专家或软件代理手中。 Mycelium架构与动态上下文图谱 (5:14) PNNL 开发的 Mycelium 平台以自然界中连接植物根系并调配生化信号的真菌菌丝网络命名，旨在作为多学科团队的共享知识基质。在传统的单用户协作中，研究人员（如蛋白质组学专家 Jeff）通常使用聊天界面连接模型上下文协议（MCP）服务器，以此调度异构数据集、分析工具与高性能 GPU 仿真集群。但当另一位专家（如机制生物学家 Lumi）加入并希望接续探索同一生物系统时，现有的单用户工具链便会出现协同断层。为此，Mycelium 引入了“动态上下文图谱”（Active Context Graph）。该系统并非传统的向量数据库或静态图数据库，而是一个分层的源流索引（Provenance Index），能够将对话记录、执行轨迹、后台计算任务与正在演进的科学假设紧密绑定。任何推断和结论均可直接追溯到底层的计算历史与讨论链条，并支持系统的主动推荐（Push）以及用户的定向图谱查询（Pull）。容错工作流、认知权重与自主性治理 (8:43) 针对科学数据高度混乱、难以强行转化为标准表格或结构化 JSON 的现实，Mycelium 利用 Claude Opus 等前沿大语言模型直接从用户的自然语言描述中推断数据模式，并对其建立缓存。依托这一机制，简单的自然语言指令即可触发有向无环图（DAG）工作流，在底层自动完成数据筛选、直方图分析或跨 EC2 实例的聚类计算；若分布式任务失败，其内置的检查点容错机制可确保任务无缝恢复，且所有步骤均以动态 Jupyter Notebook 形式记录并关联到上下文图谱。针对后台代理自主性的蔓延，Choudhury 强调无约束的自主探索极易带来噪音累积与用户抵触。Mycelium 建立了严格的治理规则：任何自主代理生成的推断成果，未经人类专家的显式审核与批准，严禁合并至共享上下文图谱，以防止错误发生连锁复合。此外，系统引入了“认知权重”（Epistemic Weighting）机制以区分不同资历成员（如实习生与资深研究员）的影响力，并能根据项目冲刺阶段（从早期数据质控到中期深度机制分析，再到后期的实验决策）自适应调整代理的辅助重点。实证对比：全自主模型与人机协同网络 (17:30) PNNL 将处于最高并行度与完全自主状态下的单体前沿模型（Claude Opus 4.8）与基于 Mycelium 的人机协同系统进行了基准对比测试。结果显示，单体前沿模型在边界清晰、定义明确的特定子任务上表现优异（评测得分集中在 4 分和 5 分），但在面对定义模糊、起点不确定的开放式科学发现问题时，表现出极高的结果离散性与两极分化，频繁出现 0 分与 1 分。相比之下，Mycelium 人机协同系统在开放式探索与深度专门分析中均表现出稳定的深度与广度。该网络通过将 AI 的高通量计算执行力与人类专家的领域直觉、横向推理及偶然性发现（Serendipity）相结合，补足了前沿模型在探索阶段的缺陷。PNNL 目前正在 5 至 6 个生物学试点中推进该方案，并计划在分布式数据中心中扩大多代理协同规模。 Choudhury 总结认为，Mycelium 及其动态上下文图谱架构不仅适用于生物学探索，同样能够泛化至国家安全、金融和医疗等存在严重信息孤岛的高风险领域。其深层挑战不仅在于构建调度工具，更在于如何持续提取科研冲刺中快速产生的高价值见解，并将其沉淀为有机连接的“知识胶囊”（Capsules of Knowledge）以构建真正的企业级智能；目前研究团队正准备将 Mycelium 平台进行开源发布，以促进行业与学术界的进一步验证与协作。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 17 日发布在 YouTube 平台。视频由来自太平洋西北国家实验室（PNNL）的首席科学家 Sutanay Choudhury 主讲，页面简介指出，随着单体 AI 模型在超大上下文与长周期推理上的扩展逐渐触及瓶颈，演讲将介绍通过 Active Context Graph 连接人、Agent 与仪器的 Mycelium 运行时架构，探索人机网络智能这一全新的 AI 扩展维度。 ## TL;DR 演讲指出，面对现实中的复杂科学与安全难题，单纯依靠独立单体 AI 模型的单向扩展已无法满足多领域协作与合规决策需求，亟需构建人机协同的网络智能。为此，团队研发了名为 Mycelium 的运行时系统，利用动态上下文图谱（Active Context Graph）实现跨学科专家与 AI 之间全链路溯源的知识连接与精准智能路由。实证对比表明，完全自主的独立前沿 Agent 容易陷入认知盲区和局部死胡同，而 Mycelium 人机协作网络兼具人类直觉的广度与机器计算的深度。该架构不仅已在生物学试点中取得显著成效，未来还将拓展至国家安全、金融与现代医学等广泛领域并逐步开源。 ## Atmosphere & style 整场演讲呈现出严谨、前瞻且极具启发性的学术与工业应用氛围。主讲人以清晰从容的语调抽丝剥茧，将抽象的架构设计与生动的生物学协同案例紧密结合，叙事逻辑严密且富有说服力。实证对比环节展现了对独立大模型局限性的客观审视，结尾处则充满对开源与跨领域协作的澎湃热情，为听众传递出一种见证下一代 AI 协作范式正在破土而出的探索感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) 单体模型扩展（超大上下文、长周期 Agent）正面临边际效益递减，难以独立解决依赖多领域先验与合规闭环的现实复杂难题。 - [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) 提出与参数扩展正交的新维度——网络智能（Networked Intelligence），通过构建高效的人机协同网络弥补单体模型的局限。 - [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) 复杂科研与安全任务本质上高度异步且由事件和假设驱动，依赖跨领域团队在突发事件或新数据产生时迅速响应与迭代。 - [03:44](https://www.youtube.com/watch?v=aG8foA_qXgY&t=224s) 网络智能的核心在于精准的智能路由，而非全局无差别广播项目记忆，从而避免严重的认知过载与信噪比失衡。 - [03:44](https://www.youtube.com/watch?v=aG8foA_qXgY&t=224s) PNNL 推出的运行时架构命名为 Mycelium（菌丝体），旨在像地下真菌网络一样为人机协作团队提供互联的知识基质。 - [05:03](https://www.youtube.com/watch?v=aG8foA_qXgY&t=303s) Active Context Graph 超越传统向量库或知识图谱，作为层次化记忆中枢实现异构数据、后台计算与决策链条的全链路溯源。 - [05:03](https://www.youtube.com/watch?v=aG8foA_qXgY&t=303s) 系统利用前沿大模型实现自适应 Schema 推断，并自动将自然语言查询编译为具备检查点（Checkpoint）容错能力的计算 DAG 与可复现的 Jupyter 笔记本。 - [11:48](https://www.youtube.com/watch?v=aG8foA_qXgY&t=708s) 信息交互兼具 Pull（主动拉取查询）与 Push（基于上下文图谱在成员上线时的主动智能推荐）两种模式，放大团队网络效应。 - [11:48](https://www.youtube.com/watch?v=aG8foA_qXgY&t=708s) 为防止错误扩散导致图谱发生噪声复合（Noise Compounding），离线自主 Agent 的探索结果在未经用户人工审查批准前严禁直接写入 Context Graph。 - [11:48](https://www.youtube.com/watch?v=aG8foA_qXgY&t=708s) 认知加权（Epistemic Weighting）是网络协同的关键挑战，系统需合理平衡活跃初级人员与资深专家在全局推荐中的知识权重。 - [16:49](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1009s) 实证基准测试表明：最高配置的独立自主 Agent 虽在局部聚焦任务表现优异，但全局视野稀疏且易陷入死胡同；人机网络则兼顾了探索深度与广度。 - [19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) Mycelium 架构具有高度通用性，不仅适用于生物学研究，还可拓展至国家安全、金融决策和现代医学等高度协同领域，且目前正处于开源收尾阶段。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) 探索 AI 扩展的新维度：从单体模型到人机网络智能在过去五到十年间，人工智能的绝大部分飞跃都是由“单体模型扩展”（Individual Model Scaling）驱动的。学术界与前沿实验室的核心焦点始终是如何构建更大的上下文窗口（Context Windows）、延长长周期规划（Long Horizon）能力，以及提升单个模型应对复杂任务的能力。然而，面对企业和国家实验室必须解决的复杂现实难题，即便拥有能力极强的顶尖模型并赋予其访问所有数据集的权限，它也无法直接给出符合监管合规或法律标准的最终答案。为了弥补这一鸿沟，研究人员提出了一条完全正交的扩展维度：构建人与 AI 协同的网络系统（Networked Intelligence）。前沿实验室致力于推进模型本身的参数与推理能力，而网络智能则是通过构建高效的人机协作网络来互补这种单体模型扩展。 ```mermaid flowchart TD accTitle: 单体模型扩展与网络智能扩展对比 subgraph S1["传统单体模型扩展路线"] A1["超大上下文窗口"] --> B1["长周期单体 Agent"] B1 --> C1["独立推理但缺乏多领域先验与合规闭环"] end subgraph S2["网络智能扩展路线 (Mycelium)"] A2["异构人类专家网络"] --> B2["Active Context Graph 动态路由"] B2 --> C2["人机协同决策与全局目标对齐"] end ``` - 现实世界的复杂问题（网络安全、电网调度、医学诊断、化学与生物发现等）极少能由单一领域的专家或单一推理器独立解决。 - 复杂科学与安全领域的工作模式并非像软件工程那样按部就班地进行 Sprint 冲刺，而是由大型团队以高度异步（Asynchronous）的方式推进。 - 现实任务依赖于事件驱动与假设迭代：一旦发生网络攻击或新实验得出数据，团队成员需要迅速响应，与跨领域同事交换假设并推进到下一个决策点。 --- ### [03:44](https://www.youtube.com/watch?v=aG8foA_qXgY&t=224s) 什么是网络智能与 Mycelium 系统的命名由来网络智能的核心绝非单纯将全部数据塞给某个模型，也不是像 Claude Project Memory 或 ChatGPT 团队工作区那样简单地将整个项目记忆无差别广播给所有人——因为全局广播会导致严重的信噪比失衡与认知过载。网络智能的真正精髓在于“智能路由”（Intelligent Routing）：当发生数据变动、收到新实验结果、检测到入侵事件或病患指标更新时，系统能精确判定其价值，并将该信息推送到最需要它的人员、Agent 或仪器面前，以驱动下一步决策。该运行时架构被命名为 **Mycelium（菌丝体）**，灵感来源于其早期落地的生物学团队。生物学家形象地将其比作土壤深处蔓延的真菌菌丝网络，它在地下默默连接着各方生态，为整个人机协作团队筑起一层坚实的“知识基质”（Knowledge Substrate）。 --- ### [05:03](https://www.youtube.com/watch?v=aG8foA_qXgY&t=303s) 运行时架构：从单用户 MCP 到基于上下文图谱的多人网络在传统单用户模式下，系统架构通常基于类似模型上下文协议（Model Context Protocol, MCP）的交互方式：用户的浏览器向 MCP 服务器发送请求，服务器连接底层异构数据源与异构算力（例如需要独立 GPU 集群支持的高性能仿真任务）。但在多人跨学科科研场景中，工作流展现出截然不同的需求。以生物学领域的协同场景为例： 1. **科学家 Jeff**：专注于蛋白质组学（Proteomics），擅长分析不同实验条件下蛋白质的表达差异。 2. **科学家 Lumi**：专注于生物通路的机制机理（Mechanistic Understanding），探究特定条件如何触发下游生物通路。当 Jeff 针对某批蛋白质数据发起分析并提出假设后，Lumi 希望无缝接手并在该假设的基础上继续深挖。为了让 Jeff 探索过程中沉淀的知识与假设有效指导并赋能 Lumi 的工作，Mycelium 在中间引入了**动态上下文图谱（Active Context Graph）**。 ```mermaid flowchart LR accTitle: Mycelium 动态上下文图谱运行时架构 subgraph Users["科研人员终端"] J["Jeff (蛋白质组学专家)"] L["Lumi (通路机理专家)"] end subgraph Core["Mycelium 核心层"] CG[("Active Context Graph\n上下文图谱")] Router["智能路由与推荐引擎"] end subgraph Backend["异构计算与数据底座"] MCP["MCP 服务与工具链"] JNB["动态 Jupyter 笔记本 (数据溯源)"] Compute["异构算力 (EC2 / GPU 集群)"] end J -->|"发起提问与假设"| MCP MCP -->|"生成计算流"| JNB JNB -->|"记录算力与状态"| Compute JNB -.->|"绑定 ID 追踪"| CG CG -->|"语义解析与路由"| Router Router -->|"Push / Pull 上下文分发"| L ``` - **超越传统知识库/向量库**：Context Graph 并不是普通的知识图谱、向量数据库或图数据库，而是一个层次化（Hierarchical）的系统记忆中枢。 - **全链路溯源（Provenance）**：它不仅捕捉聊天记录，还将背后触发的全部后台计算、异构数据集、决策链条统一编排，作为高阶索引挂载到各类业务决策插件与 API 上。 - **自适应 Schema 推断**：科学数据极其杂乱且非结构化。系统借助前沿模型（如 Claude Opus 系列）的理解能力，在用户仅用自然语言描述数据时动态推断 Schema 并建立缓存，无需强行限制用户上传固定格式的 JSON 或表格。 - **容错计算（Fault-Tolerant DAG）**：系统自动将自然语言查询解析为包含筛选、直方图、K-Means 或谱聚类（Spectral Clustering）的有向无环图工作流；每一项分布在 EC2 或 GPU 上的计算均记录 Checkpoint，以支持中断后的无缝恢复。 - **代码级可复现**：所有的计算轨迹与分析逻辑均以动态生成的 Jupyter Notebook 形式记录，并与 Context Graph 中的节点 ID 严格绑定。 --- ### [11:48](https://www.youtube.com/watch?v=aG8foA_qXgY&t=708s) 网络效应放大机制：主动推荐、受控自主性与认知加权当多位用户介入同一协作空间时，网络效应开始显现。Mycelium 提供了两种信息交互模式： - **Pull（主动拉取）**：Lumi 主动询问系统：“Jeff 最近有什么新发现？帮我总结一下”，系统基于 Context Graph 查询并返回对应上下文。 - **Push（智能推送/推荐）**：系统感知到 Jeff 的最新工作结论对 Lumi 当前的研究方向具有潜在价值，在 Lumi 上线时主动推送提示：“Jeff 刚刚在分析这批数据，你是否需要查看？” 为了进一步放大网络效应，Mycelium 探索了具备长周期探索能力的自主 Agent，但同时也对其潜在风险进行了严格校准： | 机制维度 | 运作逻辑与设计考量 | 关键治理策略 / 面临挑战 | |:--- |:--- |:--- | | **离线自主探索 (Autonomous Investigation)** | Agent 理解用户的探索意图，在用户下线后在后台自主运行推演与验证计算，待用户下次上线时呈现分析结果。 | **严格把控自主权**：未经用户审查批准的自主探索结果**严禁直接发布至 Context Graph**，防止单一早期错误在全网扩散，导致图谱遭遇严重的噪声复合（Noise Compounding）。 | | **网络级协同理解** | Agent 结合 Jeff 与 Lumi 的共同兴趣与跨学科背景，在全员离线时自动派发符合团队整体诉求的计算任务。 | 必须保持高度可解释性，避免过度自主导致科研人员产生抵触与挫败感。 | | **认知加权 (Epistemic Weighting)** | 系统需要评估网络中不同成员的输入在指导全局推荐时的权重。 | **核心挑战**：如何权衡一位极度活跃但资历尚浅的实习生提出的激进假设，与一位经验丰富但行事保守的资深实验室院士（Lab Fellow）的判断。 | | **阶段性目标自适应 (Sprint Alignment)** | 系统推荐与后台任务随团队所处的项目生命周期动态演变。 | **Sprint 周期演进**：Sprint 初期侧重宽泛的数据探索与质检，中期聚焦深层次科学分析，后期则全面收敛至实验设计与决策闭环。 | --- ### [16:49](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1009s) 实证对比：人机网络智能 vs. 独立自主模型通过可视化两天内早、中、晚三个时段的 Context Graph 演变，研究人员直观展示了团队在 Sprint 周期内从数据质检、深度科学分析到最终锁定实验方案的动态迁移过程。针对“单靠顶尖大模型赋予全部数据即可解决问题”的假设，团队开展了详尽的基准测试，将 Mycelium 人机协作网络（Column A）与在最高自主度、最大算力资源及完全并行设置下运行的 Opus 4.8 独立模型（Columns B & C）进行了全面对照评估： ```mermaid flowchart TD accTitle: 独立 Agent 与人机协作网络在复杂科学任务中的表现对比 subgraph Standalone["独立自主 Agent (Opus 4.8 单独执行)"] B1["特定聚焦任务得分极高 (出现较多 4-5 分)"] B2["全局视野极度稀疏 (存在大量 0-1 分盲区)"] B1 & B2 --> B3["缺乏探索广度，容易陷入局部死胡同"] end subgraph Networked["Mycelium 人机网络智能"] M1["人类专家：提供直觉、发散创意与意外发现 (Serendipity)"] M2["AI 代理：提供高效计算、自动化执行与后台推演"] M1 & M2 --> M3["兼具广度与深度，全链路无认知死角"] end ``` - **单体模型的局限**：前沿独立 Agent 虽然在高度受限、目标明确的单点任务上表现出色（评测中拿下不少 4 分甚至 5 分的高分），但整体表现极度稀疏，存在大量 0 分和 1 分的盲区。面对定义模糊、起点未知的开放式探索（如前沿生物学发现或突发网络入侵溯源），单纯的自动化模型极易迷失方向。 - **人机网络的优势**：人类专家的直觉、跨学科经验与捕捉意外发现（Serendipity）的创造力，与 AI 的高并发执行力在 Context Graph 的连接下形成合力，在兼顾探索深度的同时提供了无死角的广度。 - 目前该项目已在生物学领域推进了 5 到 6 个 Pilot 试点，并正逐步拓展至跨多个数据中心、调度多智能体协同运行的大规模科学发现任务中。 ### [19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) 跨领域拓展与开源展望：让“菌丝体”网络智能全面蔓延在演讲的结尾，苏塔奈·乔杜里（Sutanay Choudhury）提炼出了三项核心结论与展望。他首先指出，尽管本场演讲所展示的实证结果主要聚焦于生物学领域，但基于动态上下文图（Active Context Graph）的 Mycelium 架构设计绝非局限于单一学科。这种协同范式不仅适用于广泛的基础科学研究，更能够横向迁移至国家安全（Security）、金融（Finance）以及现代医学（Medicine）等复杂领域。他真诚地呼吁并期待与那些在各自行业中同样面临高难度复杂协同挑战的研究者和从业者展开对话。乔杜里进一步强调，构建这套系统的意义远不止于搭建一个智能体运行载体或自动化测试测试桩（Harness）。更深层的价值在于重新思考知识的流动方式：如何在密集的冲刺式研发（Sprints）中捕获那些零散沉淀的隐性认知，并将这些孤立的“知识胶囊”（Capsules of Knowledge）有效联结、激活，从而自底向上构建出真正具备企业级水准的系统级网络智能（Enterprise Grade Intelligence）。最后，他透露团队目前已非常接近将 Mycelium 项目正式对外开源，以便让开源社区中的每一位研究人员和开发者都能亲自上手体验和评估。在致谢观众时，乔杜里借用了生物学合作者们最喜欢的一句行话作为收尾，充满激情地呼吁道：“让菌丝体（Mycelium）全面蔓延扩散开来吧！”（"Let the mycelium spread."）。 - **多领域适用性**：Mycelium 架构不仅服务于生物学实验，还将拓展至广泛的科研探索、安全情报分析、金融量化决策与精准医疗等领域。 - **知识封装与联结**：核心技术追求不单是工具链封装，而是将各个工作节点和冲刺周期中生成的“知识胶囊”打通，凝炼为组织层面的企业级智能。 - **开源计划**：项目已处于开源准备的最后收尾阶段，近期将面向全球开发者与科研机构开放代码库。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=aG8foA_qXgY&t=4s) 前沿实验室致力于推进模型本身的参数与推理能力，而网络智能则是通过构建高效的人机协作网络来互补这种单体模型扩展。 — Sutanay Choudhury > [03:44](https://www.youtube.com/watch?v=aG8foA_qXgY&t=224s) 网络智能的核心绝非单纯将全部数据塞给某个模型，也不是像 Claude Project Memory 或 ChatGPT 团队工作区那样简单地将整个项目记忆无差别广播给所有人——因为全局广播会导致严重的信噪比失衡与认知过载。 — Sutanay Choudhury > [03:44](https://www.youtube.com/watch?v=aG8foA_qXgY&t=224s) 网络智能的真正精髓在于“智能路由”：当发生数据变动、收到新实验结果、检测到入侵事件或病患指标更新时，系统能精确判定其价值，并将该信息推送到最需要它的人员、Agent 或仪器面前，以驱动下一步决策。 — Sutanay Choudhury > [05:03](https://www.youtube.com/watch?v=aG8foA_qXgY&t=303s) Context Graph 并不是普通的知识图谱、向量数据库或图数据库，而是一个层次化的系统记忆中枢。 — Sutanay Choudhury > [11:48](https://www.youtube.com/watch?v=aG8foA_qXgY&t=708s) 未经用户审查批准的自主探索结果严禁直接发布至 Context Graph，防止单一早期错误在全网扩散，导致图谱遭遇严重的噪声复合。 — Sutanay Choudhury > [16:49](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1009s) 前沿独立 Agent 虽然在高度受限、目标明确的单点任务上表现出色，但整体表现极度稀疏，存在大量 0 分和 1 分的盲区。面对定义模糊、起点未知的开放式探索，单纯的自动化模型极易迷失方向。 — Sutanay Choudhury > [16:49](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1009s) 人类专家的直觉、跨学科经验与捕捉意外发现的创造力，与 AI 的高并发执行力在 Context Graph 的连接下形成合力，在兼顾探索深度的同时提供了无死角的广度。 — Sutanay Choudhury > [19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) 构建这套系统的意义远不止于搭建一个智能体运行载体或自动化测试测试桩。更深层的价值在于重新思考知识的流动方式：如何在密集的冲刺式研发中捕获那些零散沉淀的隐性认知，并将这些孤立的“知识胶囊”有效联结、激活，从而自底向上构建出真正具备企业级水准的系统级网络智能。 — Sutanay Choudhury > [19:33](https://www.youtube.com/watch?v=aG8foA_qXgY&t=1173s) Let the mycelium spread. — Sutanay Choudhury ## Mentioned Sutanay Choudhury, Pacific Northwest National Laboratory (PNNL), Ai4, Mycelium, Active Context Graph, Model Context Protocol (MCP), Claude Opus, Claude Project Memory, ChatGPT, Jupyter Notebook, Amazon EC2 ## Video description Sutanay Choudhury - Chief Scientist, Pacific Northwest National Laboratory The current paradigm of AI scaling relying on massive context windows and long-horizon agentic execution is rapidly approaching a plateau. Challenging scientific problems are rarely solved by one isolated reasoner; they are solved by distributed teams carrying different priors, tacit knowledge, and domain-trained intuitions. The open problem is therefore not just how to scale individual models, but how to cultivate networked intelligence: scaling the connections between humans and AI systems so that a hypothesis produced in one context reaches the exact person, agent, instrument, or robot that can act on it. The talk will introduce Mycelium, a runtime architecture using an Active Context Graph to connect isolated users, agents, and instruments into a synchronized network. I will detail empirical success stories showing how active context propagation fundamentally altered the trajectory of a live scientific campaign. Join this session to see a direct contrast between networked execution and standalone autonomous models, providing actionable evidence of a completely new axis for AI scaling. [Source](https://www.youtube.com/watch?v=aG8foA_qXgY)

Source quotes

[0:04] Sutanay Choudhury: 前沿实验室致力于推进模型本身的参数与推理能力，而网络智能则是通过构建高效的人机协作网络来互补这种单体模型扩展。 [3:44] Sutanay Choudhury: 网络智能的核心绝非单纯将全部数据塞给某个模型，也不是像 Claude Project Memory 或 ChatGPT 团队工作区那样简单地将整个项目记忆无差别广播给所有人——因为全局广播会导致严重的信噪比失衡与认知过载。 [3:44] Sutanay Choudhury: 网络智能的真正精髓在于“智能路由”：当发生数据变动、收到新实验结果、检测到入侵事件或病患指标更新时，系统能精确判定其价值，并将该信息推送到最需要它的人员、Agent 或仪器面前，以驱动下一步决策。 [5:03] Sutanay Choudhury: Context Graph 并不是普通的知识图谱、向量数据库或图数据库，而是一个层次化的系统记忆中枢。 [11:48] Sutanay Choudhury: 未经用户审查批准的自主探索结果严禁直接发布至 Context Graph，防止单一早期错误在全网扩散，导致图谱遭遇严重的噪声复合。 [16:49] Sutanay Choudhury: 前沿独立 Agent 虽然在高度受限、目标明确的单点任务上表现出色，但整体表现极度稀疏，存在大量 0 分和 1 分的盲区。面对定义模糊、起点未知的开放式探索，单纯的自动化模型极易迷失方向。 [16:49] Sutanay Choudhury: 人类专家的直觉、跨学科经验与捕捉意外发现的创造力，与 AI 的高并发执行力在 Context Graph 的连接下形成合力，在兼顾探索深度的同时提供了无死角的广度。 [19:33] Sutanay Choudhury: 构建这套系统的意义远不止于搭建一个智能体运行载体或自动化测试测试桩。更深层的价值在于重新思考知识的流动方式：如何在密集的冲刺式研发中捕获那些零散沉淀的隐性认知，并将这些孤立的“知识胶囊”有效联结、激活，从而自底向上构建出真正具备企业级水准的系统级网络智能。 [19:33] Sutanay Choudhury: Let the mycelium spread.

Source quotes

Sutanay Choudhury: The current paradigm of AI scaling relying on massive context windows and long-horizon agentic execution is rapidly approaching a plateau. Sutanay Choudhury: Challenging scientific problems are rarely solved by one isolated reasoner; they are solved by distributed teams carrying different priors, tacit knowledge, and domain-trained intuitions. Sutanay Choudhury: The open problem is therefore not just how to scale individual models, but how to cultivate networked intelligence: scaling the connections between humans and AI systems so that a hypothesis produced in one context reaches the exact person, agent, instrument, or robot that can act on it. [0:04] Sutanay Choudhury: we believe the answer to that question is largely no. [8:43] Sutanay Choudhury: Show me what are the top proteins in this particular data set [12:47] Sutanay Choudhury: no autonomous agent finding is published to the shared context graph without explicit human review and approval [17:30] Sutanay Choudhury: Combining human creativity, domain expertise, and serendipity with AI-driven computational infrastructure provides the optimal balance of breadth and depth. [19:33] Sutanay Choudhury: capsules of knowledge [19:33] Sutanay Choudhury: let the mycelium spread.

Original transcript

[0:04] In the next 20 minutes, I'm going to take you on a tour of a very different branch of scaling. [0:10] Last few years, last 5 years, last 10 years, much of the progress in artificial intelligence has [0:16] been driven by scaling individual models. So, the focus has been how do you scale up larger context? [0:22] How do you build models with larger horizons and the ability to do more capable tasks? Now, let [0:28] me ask you a question. Do you believe that if you have for the complex enterprise problems that we [0:34] must deal with, if you had a very capable model, and if you gave it access to all the data sets, [0:41] it'll come back with the right level of answers that you can just submit or meet your regulatory [0:46] or other legal standards? And we believe the answer to that question is largely no. [0:52] And to address that gap, we want to talk about an alternate dimension of scaling because [1:00] we believe that building human and AI networks and building effective systems for human and AI [1:06] collaboration is the perfect way to complement what we're the model scaling is coming out of [1:12] from the frontier labs and other efforts. So, in the next 20 minutes, I'm going to talk about [1:28] Can you um Look. Okay. So, the next 20 minutes, I'm going to talk about what was the motivation [1:34] that drove us to explore this alternate direction because most of most people, most organizations [1:40] today are relying on model scaling. Why do you want to explore this orthogonal dimension? [1:45] I'm going to going to walk you through the runtime architecture we have built. [1:49] I'll walk you through a case study in biology, how people are using it. I'll share some analytical [1:54] insights that you can use to So, I'm going to share some analytical insight that you can use [1:59] to think about further and how can you build network intelligence for the problems that you [2:05] face in your organizations or your problems. So, I work as I as said in the first slide, [2:12] um I work at Pacific Northwest National Laboratory and over the last few 10 to 20 years, [2:17] I have worked on different kinds of problems in science and security. And the where we come from, [2:24] the this problems, whether you're dealing with cybersecurity, power grid, you're dealing with [2:29] problems in medicine, or problems with chemistry or biology, this problems are rarely solved by [2:36] one background or experts within a single kind of expertise. And that is one of the major reasons I [2:42] believe that the hard problems are not really solved by one reasoner alone. We need people [2:49] with multiple kinds of expertise. And even if you provide an uh all the like I said, all the data [2:55] sets to the AI model and to solve the problem, we just are not ready to take the AI driven work and [3:02] submit for our own either professional obligations or meet our legal or fiduciary obligations. [3:10] And the other thing important to think about is that for a lot of this domains, this these are [3:15] the domains where the work is not done like in software sprints. Work happens asynchronously. [3:22] Work happens asynchronously in big teams. Stuff happens, events happen, you have a hack, [3:28] or you have a new experiment that just gets done, and people have to respond to that. And [3:33] once people respond to that, they exchange ideas with their colleagues or other collaborators, [3:39] and they think about how are you going to respond and go to the next decision point. [3:44] So, the definition of network intelligence, if I were to put this in a succinct fashion, [3:50] is not just about um giving an AI model all the data. It's also not about giving [3:57] a team of people access to a project memory like you can do this in the Anthropic Cloud or [4:04] chat GPT because then you're over broadcasting everything to everybody. How do you find the [4:10] sweet spot of the doing the intelligent routing? So whenever there's a change in data, [4:17] whenever there is a new update that arrives to you, whether it's in the form of a hack, [4:22] a new scientific result, a new patient update, how do you get it to the right people and influence [4:29] them to make decisions? This is our definition or formulation of the network intelligence. [4:36] Why do you call it mycelium? Because our first one of the first use cases we were working on [4:40] was with a team of biologists and they really like to use this example of a fungus that grows [4:46] beneath the soil and really keeps all these networks together and this look at this [4:51] as that we're building a system that essentially creates a knowledge substrate for teams here. In [4:59] the next few slides, I'm going to walk you through our runtime architecture. I'll start with the [5:03] view of how a single user version of that works and then I'm going to introduce the alterations [5:09] we make to make it for in a multiplayer or multi-principle setting. So imagine that we [5:14] have you got two scientists working together in a team and this is our hello world example. [5:20] One works on procuring means that his expertise is really into how does proteins express differently [5:27] under different conditions. The second person doesn't really focus on that kind of problem. [5:32] She focuses on how does one active one condition triggers a certain pathway. She's more into the [5:39] why and the mechanistic understanding of things. So imagine that both of them have a cloud desktop [5:45] and they have they have given access to cloud to all the data. So, Jeff starts asking a question [5:51] about, "Hey, I'm seeing this particular signature in my data. What what do you think about this?" [5:57] Now, this is a familiar this probably is a familiar architecture for a lot of you in [6:01] this room, that you have a question, it goes from your browser, it goes to a MCP server. [6:07] On the MCP server, you have given access to all the different kinds of data and all the different [6:12] kinds of tools. In the case of scientific workloads, this data and are very heterogeneous [6:18] because some of the computations that we might might have to do might require their own GPU [6:22] cluster like running a simulation. But that's besides the point. The main point here is that [6:27] we have different heterogeneous data, we have heterogeneous compute, and we also have [6:32] heterogeneous workloads. But now I bring in Lumi instead of in addition to Jeff, the other user, [6:39] into the play. So, Jeff starts asking questions like I said before, "What's happening with this [6:44] this protein data set?" And now Lumi wants to investigate this and she wants to start where [6:50] her colleague left off. Now, how do how does the knowledge from Jeff's investigation goes to [6:57] impact Lumi's or steer Lumi's work? That means we need to introduce something [7:03] in the architecture I was showing before so that it can hold whatever knowledge or synthesis [7:09] or hypothesis that was coming out of Jeff. And we want to have do some smart intelligent processing [7:16] to determine the routing value of that information from Jeff to Lumi. The way we do this [7:22] is through this architecture of an uh a context graph bridging or holding everything together. [7:30] So, it is not I I don't want you to think about this as just a one more knowledge graph [7:35] or a vector database or a graph database. The the real big shift here in terms of thinking is that [7:42] we're looking as all this data or memory for the AI system in a hierarchical fashion. For us, [7:50] data or knowledge is not just about the chat logs. It's not just about the project memory that cloud [7:56] would give you. For us, we have to trace through the decisions, all the different computations that [8:02] were run in the background, and that that also steered or or influenced these conversations. [8:09] So, when you have to think about a decision, we think about the provenance behind the decision, [8:13] we really have to think about all the conversational data sets and all the [8:18] other computational data sets in unison. So, the context graph really serves as a as a higher level [8:25] index that hooks into all these disparate sources of information, and then on top of this graph, [8:31] now you can ask questions to say, "Hey, get me a particular context." And you can [8:36] implement different kinds of plugin or different kinds of decision-making applications or APIs. [8:43] So, I'm going to walk you through in the next few slides. I'm going to walk you [8:46] through an implementation, a very one of the earliest ones in in the the biology use case, [8:52] and I'm going to call out where we're leveraging the model scaling features, [8:56] because that's really important, but where we're also still going away from pure model scaling [9:02] and thinking about new features to do this amplification of the network intelligence. [9:08] So, let's say with the previous example, Jeff goes to do it goes to his cloud chat, [9:13] asks this question, and that triggers a whole bunch of computations in the background. [9:19] One of the requirements for us is that scientific data is messy, it's extremely heterogeneous. [9:24] We have decided that if you wanted to support a really large number of users, [9:28] we cannot really restrict them to tabular format or a specific JSON format. So, this is a this is a [9:34] classic example of model scaling, where we let the model figure out the schema of a data. It caches [9:41] this information and next time it goes and hits the similar data sources, it uses that knowledge. [9:47] So, basically what you tell users is that if you were to turn around and explain your data [9:52] to some colleague, what did you say them tell tell them? And if you can express that to model [9:59] and today's Opus-like models are quite capable of being able to do that sort of inference. [10:05] The next information and very unique requirement for us has been [10:09] supporting fault-tolerant computations. So, I'll I'll give you an example. [10:15] When some a scientist says, "Show me what are the top proteins in this particular data set." [10:21] In the back behind the scenes, we have to go figure out which data set he's talking about [10:26] because if he's saying, "I care about glycolysis." So, we need to go to the right data set, [10:31] look at the right tab, identify the right columns. All these inferences has to work correctly. [10:37] After that, we might choose to do some simple histograms or somebody might also choose to do a [10:42] K-means or spectral clustering. So, this leads to all kinds of very interesting directed acyclic [10:49] graphs or workflow patterns that we're seeing here. And what is important for us to be able to [10:54] make sure is that each of them as they're getting spawned on different EC2 instances or or GPUs, [11:00] we have to be able to pick up from a failure and then re-execute from a checkpoint and so [11:04] on and so forth. Now, this is again not about network intelligence but just setting the stage [11:09] that we have to do all of this in order for us to be able to get to where we want to be. [11:15] We also have to think about providence I I that I'm mentioned before [11:20] that we track every computation in the form of a dynamically created Jupyter notebook. [11:26] So, our data model is pretty simple. We have the raw data sets of the scientific [11:31] data sets that they're getting computed. We also have the conversational data sets [11:35] with the people with the users from the cloud or chat GPT like interfaces. And then you also [11:40] have these notebooks which are tracing everything back and hooking into the context graph via IDs. [11:48] Now, when a separate user comes into the conversation, now this is where the network [11:54] effect starts kicking kicking in. So, we have two choices. We can either let a user pull information [12:01] from Mycelium's knowledge store or the context graph or this memory structure, [12:05] or it can also push the information. So, if Jeff and Lumi has been working in a team and Jeff has [12:10] already had logged some activities in there, next time Lumi joins, we could pop a message in front [12:16] of her and say, "Hey, Jeff has been working on these things. Should you take a look at this?" [12:20] That's an example of a pull. You have to do a smart implementation of recommendation. [12:26] But Lumi can also say, "Hey, what has what Jeff have been up to? Tell me about it." And that would [12:31] lead to again a call on the context graph. The point is that at this point, I'm just describing [12:37] different ways to do recommendations on top of the context graph and the associated data model. [12:47] Um I'll I'll jump through here a little bit. So, we just don't want this that when you talk about [12:54] network intelligence, it's not just about letting users chat around a common store. [13:00] We want to take advantage of model scaling and acting on in an autonomous fashion is a [13:05] very important part. We also want to What do you want to do is that as users are [13:11] working through a problem in their own chat, our agent understands the intent. [13:16] And sometimes once they log off, the agent would spawn off a background investigation [13:20] based on that. And next time they log come back in, the those results will be presented to them. [13:26] It can also happen that based on the shared understanding of their work needs. For example, [13:32] once the My Selim agent has a good understanding of what Jeff and Lumi's common interest is, it [13:38] can spawn background computations when they have signed off based on their collective work and show [13:45] the results to them next time. But in the what we have also discovered that people don't like to [13:51] from AI researcher perspective is very interesting to see AI models doing all these autonomous [13:56] investigations. But from the user perspective we have seen they get increasingly frustrated. [14:02] So autonomy is something that we have to really really calibrate. How much degree of autonomy [14:07] you want? One of the policies we came up is that we do not publish any autonomous investigation [14:13] into the act the context graph until some user has approved or okayed that. [14:20] And the reason is because there is a very important balance of [14:23] uh the trade-off between if you do a lot of trigger a lot of autonomous computation, [14:29] we has to be reviewed. Because even a simple error getting in the way would essentially trigger [14:35] subsequent amplification of errors and then you have noise compounding to the context graph and [14:41] at some point you have no idea when this error propagation is happening and where it all started. [14:49] So explaining autonomous behavior, this is again what I just the problem I mentioned. And in short, [14:55] so I'm basically saying there are three different ways we think about amplifying the network effect. [15:02] We want to provide recommendations where you are using the the we are letting the network of a user [15:09] influence their activity. And this is subtle because one of the problems that we have seen [15:14] and have not addressed very well is that should we give equal importance to people's [15:19] everyone in the network? So let's say in one one case we have a very enthusiastic intern from [15:25] high school and she's very very excited about the AI-driven discoveries. I found this great result. [15:31] And we have somebody in the same team who's a lab fellow, very senior, very conservative. [15:37] And so, we do want to How do we give equal import Should I give equal importance to the [15:42] what they're saying to guide the recommendation somebody else? So, we call this the problem [15:47] of epistemic waiting and this is something that's a that's a really important for us to [15:52] to address properly going forward. But, letting the the network influence you, letting the [16:00] agents do background investigation, understanding what the network is doing, that's number two. And [16:06] number three is this that what the agent does for you should not be just dictated by the users or [16:12] the network goals. When When you start a sprint, the beginning of the sprint, we could be doing a [16:18] lot of exploratory computations. But, towards the end of the sprint, we suppose we have to propose [16:24] an experiment and and start the experiment next day. So, that even the agent's recommendation [16:30] starts to has to adapt to the goals. So, that this is also a very important aspect of network [16:36] intelligence. The intelligence here is not just about a single user or single user's [16:41] surrounding context, but it also has to factor in the global context and the objective of the team. [16:49] The context graph is a very useful tool for understanding this kind of challenges. [16:53] What I'm showing you here on the screen is a screenshot of a context graph of work two days, [16:58] morning, after morning, afternoon and evening. And it really allows us to see where the work is [17:04] happening. It Each kind of work is is color-coded by uh you know, in a different way. And initially, [17:11] at the beginning of the sprint, people are doing a lot more quality checks, data exploratory [17:16] uh investigations. Towards the middle of the spring, they started doing a lot more [17:20] of scientific deep analysis. Towards the end, they turned more towards [17:24] decision-making, deciding what experiments to wrap up, and so on and so forth. [17:30] We have also done significant number of benchmarking with autonomous agents. So, I [17:35] the the beginning I mentioned the thesis is that that if you just hand over all of this data [17:42] to a very capable model, will it just do magic? And here is an example proof. So, column A is uh [17:50] is an is the execution of our MyCilium system. Column B and C are executions [17:55] from Opus 4.8, run with the highest degree of autonomy, maximum resources, and parallelism. [18:03] So, what you really see is this that these models, the current frontier models, are highly capable. [18:08] But they're really good at doing very focused tasks. What I want you to take away from the table [18:13] is that if you look at the scores, higher is better, the you you see as much as four or five [18:21] or three scores in columns B and C. And you see a a lot of fours and fives in the column B and C. [18:29] And it but you also see a lot of sparsity, a lot of zeros and ones in there. [18:34] So, what it means is that when the AI agents went after some work, they did that very well. [18:40] But the human AI team, their advantage was bringing the breadth and the depth. [18:46] And this is where the if you work on problems such such as in in scientific discovery, [18:51] where the the problem is not very well scoped, or if you're working on trying to [18:57] investigate a cyber hack, where you don't know where to begin, the human creativity and being [19:03] able to go serendipity is really important. And this is why we believe that being able to combine [19:09] the the strengths of the human and the AI system, this is one way to get the best of the both world. [19:17] I'm going to skip this slide, but the quick thing is right now we have we're still work focused on [19:22] biology. We're running about five five six pilots and getting into more ambitious efforts where [19:27] we can coordinate between agents running across multiple data centers. And but towards the end of [19:33] it, I want to finish with three conclusions. So, we started by showing these results in biology, [19:39] but the design is not limited for just one domain. It can work for science, it can work for security, [19:45] it can work for other domains such as finance and medicine. And we want to hear from people who are [19:49] dealing with similar hard problems in those domains. It's not just about building a harness [19:54] either. It's also having conversations about how do you take the knowledge in the sprints and how [20:00] do you really go from connecting these capsules of knowledge to build enterprise grade intelligence. [20:06] So, we'd love to hear from you. We're also getting close to getting releasing this [20:11] work as open source, so everybody can try this out. Thank you. And like our work collaborators [20:17] in biology like to say, I'll finish by saying let the mycelium spread. Thanks for your time.