Meta Deploys Rhythm AI to Cut Project Tracking to Five Minutes

Rhythm - Anticipatory Intelligence for Organizations | Sundararajan Subramanian from Meta — @Scale

@Scale

By automatically synthesizing cross-company data while strictly respecting document access controls, Rhythm eliminates coordination bottlenecks and prevents strategic misalignment before risks materialize.

AI-assisted summary of Rhythm - Anticipatory Intelligence for Organizations | Sundararajan Subramanian from Meta — original source. Check the original for full context.

Source published:

Key ideas

  • Meta's Rhythm AI shifts organizational intelligence from reactive prompt-based LLMs to an anticipatory architecture that autonomously surfaces operational risks and dependencies before team members think to ask. (3:32)
  • To prevent data leaks, Rhythm employs a dual-layer, viewer-aware pipeline that calculates a public baseline and dynamically layers restricted strategic context based on individual viewer access clearances. (12:43)
  • Deployed across more than 200 enterprise projects, the platform has cut weekly status reporting cycles from multi-hour coordination marathons to roughly five minutes of manual review on Monday mornings. (16:17)
  • The platform incorporates a self-learning loop that separates cosmetic edits from critical factual modifications, writing corrected facts into the State Engine as authoritative precedents for future synthesis cycles. (6:23)

Detailed notes

## Background This presentation was published on YouTube by the @Scale channel on October 30, 2026, featuring Sundararajan Subramanian from Meta. No video description or show notes were provided for the talk.

## TL;DR Sundararajan Subramanian introduces Rhythm, an anticipatory intelligence system developed at Meta to solve core organizational failure modes like strategic mutation, decision latency, and cascading cross-team risks. Unlike reactive LLMs that rely on users asking the right questions, Rhythm proactively identifies cross-organizational dependencies and surfaces critical context before blockers derail delivery dates. The architecture combines evidence ingestion, context synthesis, state tracking, and a self-learning loop with rigorous privacy controls—such as viewer-aware dynamic decoration and explicit consent models. Deployed across more than 200 projects, Rhythm automates weekly status reporting rituals and acts as a high-agency partner protecting enterprise vision.

## Atmosphere & style The presentation carries the polished, technical confidence of a modern engineering keynote, blending architectural precision with pragmatic systems thinking. Subramanian delivers an earnest, structured breakdown of workplace friction, making abstract communication breakdowns feel tangible through concrete operational examples like GPU shortages and leadership acquisitions. Watching the talk feels like an authoritative peek under the hood of enterprise AI at scale, marked by a steady, solution-oriented cadence that moves deliberately from systemic organizational pain points to automated resolution.

## Bulletin - [03:32](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=212s) Meta's Rhythm AI shifts organizational intelligence from reactive prompt-based LLMs to an anticipatory architecture that autonomously surfaces operational risks and dependencies before team members think to ask. - [12:43](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=763s) To prevent data leaks, Rhythm employs a dual-layer, viewer-aware pipeline that calculates a public baseline and dynamically layers restricted strategic context based on individual viewer access clearances. - [16:17](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=977s) Deployed across more than 200 enterprise projects, the platform has cut weekly status reporting cycles from multi-hour coordination marathons to roughly five minutes of manual review on Monday mornings. - [06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) The platform incorporates a self-learning loop that separates cosmetic edits from critical factual modifications, writing corrected facts into the State Engine as authoritative precedents for future synthesis cycles.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=5s) Organizational intelligence has evolved through four main phases: manual human ledgers, web-based project tracking tools, reactive LLM query layers, and proactive anticipatory systems. - [00:05](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=5s) Reactive LLMs place an unmanageable cognitive burden on users by requiring them to formulate the right prompt, whereas anticipatory intelligence surfaces critical cross-team context without any prompt. - [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) Modern enterprises suffer from three recurring failure modes: strategic mutation as goals cascade downward, decision latency when escalating technical blockers upward, and undetected cascading cross-team risks. - [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) The core thesis behind Rhythm AI is that an intelligence platform should actively protect an organization's vision rather than merely act as a passive project tracking tool. - [03:32](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=212s) Anticipatory systems eliminate execution bottlenecks by recognizing shared resource dependencies (such as GPU constraints) at the document drafting stage and flagging historical delays immediately. - [06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) Rhythm's core pipeline relies on specialized engines: Evidence Engine for recency conflict resolution, Context Engine for structured synthesis, and State Engine for extracting granular objectives and risks. - [06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) The platform's self-learning loop analyzes human corrections to insight rollups, ignoring cosmetic wording edits while committing validated factual fixes as authoritative organizational precedent. - [09:52](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=592s) Enterprise adoption of organizational AI fails if users suspect cross-team data leaks; systems must enforce original ACLs and guarantee zero audience escalation. - [11:50](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=710s) Cross-synthesis boundary protection ensures that synthetic narratives aggregated from multiple private documents are restricted strictly to the intersection of authorized viewers. - [11:50](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=710s) Rhythm bars ephemeral side chatter from indexing and enforces explicit ingestion consent from project owners before adding candidate sources to the knowledge graph. - [12:43](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=763s) Viewer-aware insights dynamically decorate public status baselines with confidential executive context—such as impending acquisitions—without leaking sensitive information down the chain. - [15:09](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=909s) Fine-grained citation tracking links every generated insight directly back to source artifacts, providing both factual credibility and an audit trail for downstream privacy controls. - [16:17](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=977s) Across more than 200 projects, Rhythm reduced weekly status compilation from Friday afternoon drudgery to a five-minute Monday review. - [16:17](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=977s) Teams use Rhythm's canonical data model to perform strategic gap analyses that highlight discrepancies between long-term executive roadmaps and day-to-day project backlogs.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=5s) The Evolution of Organizational Intelligence

Organizational intelligence has undergone several generational shifts over the history of modern industry. In its earliest form, organizational awareness relied entirely on humans maintaining physical ledgers, manually deriving summaries, and routing status reports to leadership across bureaucratic layers through paper mail. This paradigm eventually transitioned into web-based tooling, where centralized project management platforms, live dashboards, automated reports, and digital communication tools attempted to coordinate complex operations.

```mermaid flowchart LR accTitle: Evolution of Organizational Intelligence accDescr: Shows the progression from manual human ledgers to web tools, reactive LLMs, and finally proactive anticipatory intelligence. A["Manual Era<br/>(Ledgers & paper mail)"] --> B["Web Tooling<br/>(Dashboards & PM software)"] B --> C["LLM Era: Reactive<br/>(User-prompted query layer)"] C --> D["Rhythm: Anticipatory<br/>(Proactive, zero-prompt intelligence)"] ```

The emergence of Large Language Models marked a breakthrough by consolidating fragmented data streams into a single intelligence layer that team members could directly query. However, this model remains fundamentally reactive because it places the burden on the user to formulate the correct prompt. As Sundararajan Subramanian observes: “The proactive or intelligence is where you don't even have to ask the right questions because it's only good when as much as the questions that are being asked.” Proactive organizational intelligence bypasses this limitation by automatically recognizing which critical signals must be surfaced and pushing them to stakeholders before anyone thinks to ask.

[01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) Common Organizational Failure Modes and Rhythm AI

Modern organizations frequently stumble over three systemic failure modes: 1. **Strategic Mutation:** A strategic objective is defined at the executive layer, but as it trickles down through intermediate organizational tiers toward the execution teams, the core intent mutates, generating a damaging gap between leadership strategy and day-to-day execution. 2. **Decision Latency:** In the reverse direction, as teams execute on the ground, emerging operational risks and technical blockers must be escalated upward for executive decisions or strategic guidance. By the time these escalations climb to the appropriate altitude and reach the right decision-maker, valuable time has slipped away, leaving the eventual decision too late to be effective. 3. **Cascading Cross-Team Risk:** Multiple teams simultaneously pursue shared organizational goals, yet a schedule slip, risk, or resource conflict inside a single team can silently derail interdependent teams across the company without anyone connecting the dots until the wider delivery date is compromised.

To address these breakdowns, an organization cannot rely merely on a tracking mechanism. As Sundararajan Subramanian argues: “So what you need is an AI that does not track an org. It has to protect its vision.” Rhythm AI was introduced to supply this anticipatory intelligence layer, preserving cross-functional alignment and resolving hard coordination bottlenecks across disparate teams.

[03:32](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=212s) Reactive AI vs. Anticipatory Intelligence

Anticipatory intelligence operates as an AI system with full organizational awareness, synthesizing cross-company knowledge to deliver critical context at the exact moment a decision or plan is drafted, without requiring an explicit prompt.

To illustrate the difference between reactive and anticipatory paradigms, consider an engineering team drafting an architecture specification for a new project that requires training hardware: - **The Reactive Failure:** The engineers document a dependency on GB300 GPUs to begin model training within one month. If operating under reactive AI, the author would have to explicitly search or ask the system whether GB300 capacity is constrained. Because they do not suspect a shortage, they do not ask. Weeks later, when entering the execution phase, the team runs directly into an allocation bottleneck, discovering too late that earlier teams faced massive delays waiting for the same GPUs. - **The Anticipatory Resolution:** Under anticipatory intelligence, the platform detects the GB300 GPU dependency the moment the draft architecture document is authored. It cross-references current company-wide constraints, flags that earlier teams were blocked on identical hardware, and immediately surfaces the risk to the author alongside actionable mitigation paths or alternative solutions already proven by other teams.

By maintaining deep, cross-linked organizational context, the platform ensures that risks encountered by one team automatically become preventive warnings for every other team across the organization.

[06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) System Architecture and the Self-Learning Loop

Rhythm structures its processing pipeline through an orchestrated series of specialized engines, turning raw company artifacts into cross-organizational intelligence and self-correcting models.

```mermaid flowchart TD accTitle: Rhythm Architecture and Self-Learning Loop accDescr: Pipeline mapping sources through evidence, context, state, and insight generation engines, including a self-learning loop for factual edits. subgraph Ingestion["Ingestion & Contextualization"] S["Artifacts & Sources<br/>(Docs, Dashboards, Sheets)"] --> E["Evidence Engine<br/>(Prioritization & Precedence)"] E --> C["Context Engine<br/>(Ordered Context Synthesis)"] end

subgraph StateAndInsights["State Extraction & Reporting"] C --> St["State Engine<br/>(Objectives, Milestones, Risks)"] St --> I["Insights Generator<br/>(Human-Readable Project & VP Rollups)"] I --> N["Notifications to Project Owners"] end

subgraph Learning["Self-Learning Loop"] N --> Edits{"Human Edits"} Edits -->|"Cosmetic/Style"| Ignore["Discard"] Edits -->|"Critical Facts"| SL["Self-Learning Loop<br/>(Analyze Root Cause)"] SL -->|"Record New Precedent Fact"| St end ```

The end-to-end data pipeline functions through several specialized components: - **Evidence Engine:** Ingests disparate internal artifacts—such as documents, live dashboards, Google Sheets, and Excel workbooks—and resolves conflicting information by ordering them by priority and recency. For example, fresh decision meeting notes automatically supersede conflicting decisions outlined in an older architecture spec. - **Context Engine:** Organizes prioritized evidence into a coherent, structured contextual narrative. - **State Engine:** Deconstructs this context into discrete organizational facts and states, categorizing them into tracked items such as active objectives, project milestones, verified risks, and critical operational updates. - **Insights Generator:** Synthesizes granular state data into human-readable narratives tailored to specific organizational altitudes. It delivers localized project updates for individual teams while generating portfolio rollups for VPs, proactively flagging systemic blockers across entire divisions (such as cross-team training hardware contention). - **Self-Learning Loop:** When generated insight notifications prompt human project owners to submit corrections, the system automatically classifies edits as either cosmetic/language tweaks or critical factual modifications. When a critical factual edit occurs, the loop determines why the State Engine missed the detail and records it as an authoritative new fact within the State Engine, establishing it as precedent for future insight cycles unless overridden by an authoritative new source. - **Integration Layer:** Provides dedicated CLI and API integration hooks that enable external tools and internal workflows to interact with Rhythm's underlying intelligence core.

[09:52](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=592s) Trust at Scale and Information Access Governance

Deploying an anticipatory intelligence layer across large enterprises introduces severe data boundary challenges. As Sundararajan Subramanian emphasizes: “The AI must read everything but it must never surface the information to someone who should not see it.” If employees suspect the system will broadcast sensitive or confidential team artifacts indiscriminately, adoption collapses and users stop feeding authoritative data into the platform.

To build trust at scale, the architecture enforces five foundational principles of information access and security: - **Respect Original ACLs:** Every extracted fact and state item maintains direct cryptographic or metadata linkage back to its source artifact, ensuring generated insights strictly mirror the access control lists (ACLs) of the underlying documents. - **No Audience Escalation:** Context or operational deliberations meant for a restricted audience remain strictly confined to that specific group, preventing automated rollups from bubbling up sensitive discussions to wider audiences without explicit approval or promotion. - **Cross-Synthesis Boundary Protection:** When the AI aggregates disparate facts sourced across multiple private documents with distinct access permissions to synthesize a unified narrative, it must guard against inadvertent exposure. Making a synthesized high-level insight public risks leaking underlying sensitive information that specific consumers lacked authorization to see.

[11:50](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=710s) Privacy Principles: Ephemerality, Consent, and Source Discovery

Maintaining strict boundaries around organizational communication is critical when automated systems synthesize company-wide data. Any synthetic narrative or summary generated by the platform must remain visible solely to the intersection of the authorized audiences of the underlying inputs. Ephemeral chatter must strictly remain ephemeral; one-off chats and transient side conversations should never be indexed or cemented into the core organizational facts.

Extraction operates under explicit permissions. Before any artifact is ingested or analyzed, the designated project owner must grant explicit consent to pull information from it. While Rhythm has the intelligence to scan the workspace and identify potentially relevant documents, repositories, or channels, it does not unilaterally crawl them. As Sundararajan Subramanian emphasizes: “It should not go ahead and collect information from the sources that were not explicitly added but can suggest some of the sources then the project owner should go ahead and explicitly add those sources to be part of the org intelligence here.”

- **Audience Intersection**: Generated narratives are strictly partitioned and visible only to users who hold overlapping access to all source materials referenced. - **Ephemerality Protection**: Transient, informal discussions (such as ad-hoc chat threads) are barred from becoming permanent organizational facts. - **Explicit Ingestion Consent**: Artifacts undergo extraction only after designated project owners give verified approval. - **Suggested Sources**: Rhythm passively surfaces candidate knowledge sources, requiring human-in-the-loop sign-off before incorporating them into the ingestion graph.

---

[12:43](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=763s) Viewer-Aware Insights: Dual-Layer Pipelines and Leadership Context

A common failure mode of strict access-control systems is that they become overly restrictive for executives and leaders who need high-level, confidential context to make informed decisions. If the intelligence engine strips out confidential business context, its synthesized summaries become useless to executive decision-makers. Rhythm solves this dilemma through a mechanism called viewer-aware insights. The system first executes a baseline public pipeline: it gathers unrestricted evidence, establishes a shared organizational state, generates a public insights layer, and commits that baseline version.

```mermaid flowchart TD accTitle: Viewer Aware Insights Pipeline accDescr: Flow diagram showing how public artifacts and restricted artifacts are processed to decorate baseline insights for authorized viewers.

subgraph PublicPipeline["Public Evidence Pipeline"] A["Public Artifacts"] --> B["Evidence Collection"] B --> C["Shared Public State"] C --> D["Baseline Public Insights"] end

subgraph ViewerPipeline["Viewer-Aware Decoration"] E["Restricted Artifacts"] --> F{"Viewer Access Check"} F -->|"Authorized"| G["Privacy-Aware Extraction"] D --> H["Decorated Unified Insights"] G --> H end ```

When a user views an insight, Rhythm dynamically checks the viewer's individual clearance against artifacts marked as restricted during evidence collection. If the viewer possesses access rights to those restricted items, the platform runs those specific artifacts through the privacy-aware pipeline on the fly and decorates the existing public insights with confidential context. Subramanian illustrates this through a project risk scenario: an engineering project faces a major milestone delay that looks catastrophic to the core team, but an executive is secretly pursuing an acquisition that would render the delayed tool redundant. Sundararajan Subramanian explains how this calibration functions in practice: “So if it goes well then I should not risk my entire portfolio. So he has this additional layer of information presented but none of the other team members even a scope to the information that is relevant to them.”

- **Public Baseline**: Rhythm establishes an organization-wide baseline from open artifacts, keeping broad team alignment transparent. - **Just-in-Time Privacy Decoration**: Restricted insights are dynamically layered on top of the public view according to the viewer's explicit permissions. - **Portfolio Risk Calibration**: Leadership gains strategic context—such as unannounced acquisitions or partner discussions—without leaking sensitive operational plans down the reporting chain.

---

[15:09](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=909s) Building Trust through Provenance, Owner Controls, and Structured Data

Enterprise adoption hinges entirely on organizational trust. Rather than asking teams to sanitise their work by withholding sensitive materials, Rhythm is designed to accept all project artifacts while handling granular access control downstream. Every tracked entity is explicitly owner-controlled, meaning project leads maintain definitive say over whether specific facts and milestones are permitted to be surfaced outside their working group.

Artifacts ingested into Rhythm are parsed into a structured entity model where individual tracked items stay linked directly to their originating permissions and privacy constraints. Every insight produced is accompanied by rigorous citations and provenance tracking. Sundararajan Subramanian highlights the dual utility of this architecture: “Then you have this citations or provenance where whenever an insight is generated, each part of this is tied to the exact source where it is taken from which provides credibility to that information. On top of it, it also helps attach the right privacy to each of this information.” If an insight makes an assertion, the user can verify the exact document, task, or meeting note it originated from, while the underlying citation ensures privacy restrictions follow the data wherever it surfaces.

- **Unrestricted Ingestion with Safe Boundaries**: Teams can upload complete documentation without manual redaction, trusting the platform's permissioning engine. - **Structured Tracked Models**: Textual artifacts are transformed into structured models tied to access-control lists (ACLs). - **Fine-Grained Provenance**: Citations link synthetic statements back to primary sources, boosting credibility while preserving provenance-based security boundaries.

---

[16:17](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=977s) Impact: From Weekly Status Drudgery to Engineering Abundance

Over its initial deployment, Rhythm fundamentally altered weekly reporting rituals. Historically, project managers and technical program coordinators spent every Friday afternoon chasing down engineers and cross-functional partners, compiling disparate updates, and manually resolving conflicting status reports so they could publish a digestible update on Monday morning. Rhythm automates that entire coordination loop: by Monday morning, all facts, progress metrics, and risks are pre-compiled and reconciled. Project leads need only spend roughly five minutes making minor adjustments, and those manual edits feed back into the system’s self-learning algorithms.

The efficiency gains have freed teams across Meta from routine operational overhead. Rhythm currently powers more than 200 projects across the enterprise. Furthermore, teams are using Rhythm's canonical data model to build strategic analysis tools that identify capability gaps between stated long-term goals and ongoing initiatives, suggesting new roadmaps and workstreams to bridge the divide. Looking forward, Subramanian sees this as a transition toward true scale: “Rhythm aims to be the high agency teammate that actively protects the organization by surfacing the right information at the right time and also go ahead and proactively take some low-hanging remediation actions and help you achieve a lot more in the engineering pipeline.”

- **Time Savings**: Compiling cross-functional status updates dropped from several hours of manual Friday coordination down to approximately 5 minutes of review. - **Scale**: The system actively supports more than 200 projects across the organization. - **Self-Learning Loop**: Minor human edits on Monday mornings are ingested to continuously fine-tune subsequent automated summaries. - **Strategic Gap Analysis**: Secondary tooling built on Rhythm analyzes project backlogs against company strategy to identify roadmap shortfalls and suggest remediation plans. - **High-Agency Teammates**: The ultimate vision centers on proactive remediation, where anticipatory systems handle administrative toil and surface insights before project risks materialize.

## Notable quotes > [09:52](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=592s) The AI must read everything but it must never surface the information to someone who should not see it. — Sundararajan Subramanian

> [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) So what you need is an AI that does not track an org. It has to protect its vision. — Sundararajan Subramanian

## Mentioned Sundararajan Subramanian, Meta, Rhythm, Rhythm AI, Scale, GB300 GPUs, Google Sheets, Excel

[Source](https://www.youtube.com/watch?v=vnEWk2PQY4E)

中文简报

Meta部署Rhythm AI，将项目跟踪耗时缩短至五分钟 通过在严格遵守文档访问权限控制的同时自动综合全公司数据，Rhythm消除了协调瓶颈，并在风险显现之前防止战略偏离。 Sundararajan Subramanian介绍了Meta开发的先验式智能系统Rhythm，旨在解决战略变异、决策延迟和跨团队级联风险等核心组织失效模式。与依赖用户提出正确问题的反应式大语言模型不同，Rhythm能够主动识别跨组织依赖关系，并在阻碍因素影响交付日期前呈现关键背景信息。该架构结合了证据摄取、背景综合、状态跟踪以及带有严格隐私控制（如面向查看者的动态修饰和明确授权模型）的自学习循环。Rhythm已在200多个项目中部署，实现了每周状态报告惯例的自动化，并作为一个具备高能动性的合作伙伴守护企业愿景。 Meta的Rhythm AI将组织智能从基于提示词的反应式大语言模型转变为先验式架构，能够在团队成员主动提问之前，自主呈现运营风险和依赖关系。 (3:32) 为防止数据泄露，Rhythm采用了面向查看者的双层流水线，可计算出公共基线，并根据个人查看者的访问权限动态叠加受限的战略背景信息。 (12:43) 该平台已部署在200多个企业项目中，将每周的状态报告周期从数小时的跨部门协调拉锯缩减为周一早上大约五分钟的人工审核。 (16:17) 该平台包含一个自学习循环，可将表面修饰与关键事实修改区分开来，并将修正后的事实写入状态引擎（State Engine），作为未来综合周期的权威依据。 (6:23)

Reading summary

Sundararajan Subramanian presents Rhythm, an anticipatory intelligence architecture developed at Meta to eliminate organizational coordination bottlenecks by surfacing cross-team risks and context without requiring explicit prompts. By combining automated state extraction with viewer-aware access controls, the system seeks to protect organizational alignment from strategic mutation and decision latency. From Reactive Prompts to Anticipatory Context (3:32) Organizational coordination historically evolved from physical ledgers and paper reports to centralized web dashboards and project management software. While large language models introduced a consolidated data query layer, Subramanian argues that prompt-based interfaces remain fundamentally reactive because they rely on users knowing the right questions to ask. This dependency exacerbates three common enterprise failure modes: strategic mutation as leadership goals filter down, decision latency as escalating blockers crawl upward, and cascading cross-team risks where silent delays derail interdependent projects. Anticipatory intelligence replaces user-initiated querying with automated risk surfacing across organizational boundaries. For instance, if an engineering team drafts a specification requiring GB300 GPUs without realizing company-wide capacity is constrained, a reactive model yields nothing unless queried. Rhythm instead detects the hardware dependency during authoring, cross-references internal resource bottlenecks, and immediately surfaces historical delays and mitigation pathways encountered by earlier teams. Multi-Engine Pipeline and Self-Learning Feedback (6:23) Rhythm orchestrates unstructured artifacts into structured intelligence through a specialized processing pipeline. An Evidence Engine ingests project documents, spreadsheets, and live dashboards, resolving conflicts through recency and priority ranking—such as letting recent meeting decisions supersede older specifications. A Context Engine organizes this prioritized material, which a State Engine then decomposes into discrete tracked items including active objectives, verified risks, and milestones. From this state layer, an Insights Generator compiles altitude-tailored narratives, producing localized project digests for teams and high-level portfolio rollups for vice presidents. To continuously refine its extraction accuracy, Rhythm incorporates an automated self-learning loop triggered when project owners edit weekly notifications. The system categorizes human changes as either stylistic preferences or critical factual corrections. When critical facts have been missed, the loop conducts a root-cause analysis and records the correction as an authoritative precedent within the State Engine, preventing the omission from recurring in future reporting cycles. Granular Governance and Viewer-Aware Insights (12:43) Because anticipatory systems must ingest wide-ranging company documentation to detect dependencies, enterprise adoption collapses if sensitive information leaks. Rhythm enforces five governance constraints: mirroring original access control lists (ACLs), barring audience escalation, preventing the public synthesis of multiple restricted inputs, maintaining explicit project-owner consent before indexing new sources, and treating transient chat chatter as strictly ephemeral rather than permanent organizational facts. To balance strict security with executive context, the platform implements viewer-aware insight pipelines. Rhythm builds an open baseline layer from public artifacts, then dynamically layers restricted context on top based on an individual viewer's verified credentials. In an operational scenario where a team faces an alarming milestone delay, an authorized leader with access to confidential corporate development data sees that an unannounced acquisition makes the delayed tool redundant, allowing portfolio risk to be recalibrated without disclosing the sensitive deal downward. Operational Scale and Strategic Alignment (16:17) Within Meta, Rhythm currently supports more than 200 projects, replacing manual Friday status collection with pre-reconciled updates that project leads review in roughly five minutes on Monday morning. To ensure credibility, all synthesized claims carry granular citations linking back to original sources, which simultaneously prevents access-control leaks across data boundaries. Beyond status tracking, engineering groups use Rhythm's canonical data model to perform strategic gap analyses, comparing active project roadmaps against stated enterprise priorities to identify missing workstreams. Subramanian outlines an ultimate vision where the platform operates as a high-agency collaborator that not only highlights latent coordination friction, but autonomously executes low-hanging remediation actions across the delivery pipeline. Rhythm demonstrates how anticipatory intelligence can reduce administrative overhead and cross-team blind spots across large engineering enterprises. Its operational success remains tied to disciplined governance, requiring strict owner consent and dynamic permission checking to maintain trust while synthesizing cross-organizational data.

精炼阅读摘要

Meta 的 Sundararajan Subramanian 提出，现代组织在协同中普遍面临战略变异、决策延迟与跨团队连锁风险，而基于提示词的被动大语言模型无法从根本上解决这些痛点；为此 Meta 构建了前瞻性智能系统 Rhythm，能够在无显式提问的情况下自动汇集全公司上下文、预判潜在阻碍并精准推送关键信息。 从被动问答转向主动前瞻智能 (0:05) 组织协同经历了从早期的纸质账本、纸质信件，到基于网页的项目管理工具和实时仪表板，再到当前大语言模型驱动的交互问答层的演化。然而，现有的 LLM 架构本质上仍是被动式的，严重依赖用户提出恰当的问题；若员工未意识到资源短缺或隐藏冲突，系统便无法发出预警。例如，工程团队在架构规划中依赖 GB300 GPU 训练硬件，若未主动检索受限情况，便无法得知早期团队已遭遇配额瓶颈，往往直至执行阶段才爆发严重延误。 Subramanian 指出，现代组织频繁陷入三大系统性故障模式：高层战略在层层下发中偏离初衷的“战略变异”、底层风险向上汇报耗时过长的“决策延迟”，以及单个团队进度滑坡暗中波及依赖方的“跨团队连锁风险”。前瞻性智能（Anticipatory Intelligence）旨在打破这一僵局，其核心理念是系统在员工起草方案或决策的瞬间，主动调取全公司受控知识，在无人提示的情况下识别风险并提供已验证的替代路径，从而主动捍卫组织的整体愿景。 多引擎流水线与事实自学习闭环 (6:23) 为实现全流程自动化，Rhythm 构建了一套结构化的数据处理流水线。证据引擎（Evidence Engine）负责摄取文档、实时仪表板与电子表格等多源材料，并依据优先级与时效性解决冲突（例如最新的决策会议纪要优先于旧的架构规范）；上下文引擎（Context Engine）将证据串联为有序叙事；状态引擎（State Engine）则将上下文提炼为目标、里程碑、风险等离散的组织实体状态；最终由洞察生成器（Insights Generator）针对不同层级（如一线团队进展或副总裁级的业务组合汇总）生成可读摘要。 该架构内嵌了自学习闭环（Self-Learning Loop）以保证长期准确性。当系统向项目负责人推送周报时，若负责人对内容进行修订，系统会自动识别修改类型：纯排版或文风调整会被过滤忽略，而关键事实的纠正会触发根因分析，并将新事实写入状态引擎确立为权威先例。此外，Rhythm 还提供 CLI 和 API 接口，供外部工具和内部工作流进行程序化交互。 严格访问控制与可信数据治理 (9:52) 在全公司范围内推行全知型 AI 面临极为严苛的数据边界挑战。Subramanian 强调，“AI 必须读取所有内容，但绝不能将信息展现给不该看到的人”，否则一旦员工担忧敏感信息泄露，系统的数据摄取基础便会瓦解。为此，Rhythm 确立了严格的访问控制治理准则：所有提取的事实与状态元数据均直接关联原始来源的访问控制列表（ACL），严格遵循原始权限，坚决杜绝在聚合汇报中非授权扩散限定受众的讨论内容。 在隐私与摄取控制方面，系统执行交集受众可见原则，即由多份私密文档综合而成的洞察，仅对同时拥有所有源文件查看权限的人员开放；严禁抓取和持久化瞬态聊天（如即时通讯中的临时闲聊）；所有材料的摄取必须经过项目负责人的显式授权，Rhythm 仅负责被动发现和推荐候选数据源，必须经人工审批后方可加入知识图谱。同时，每条生成的洞察都附带精细的溯源引用，确保结论高度可信且权限随数据流动精准绑定。 查看者感知的双层动态洞察 (12:43) 过于严苛的权限隔离往往会导致高层管理者无法获取全局机密商业背景，从而削弱决策质量。Rhythm 提出了“查看者感知洞察”（Viewer-Aware Insights）机制来解决这一矛盾。系统首先运行一条基础公共流水线，汇总公开证据并沉淀为共享公共状态，生成基线公开洞察；当特定用户访问这些洞察时，系统会实时校验该用户的个人安全许可。 若查看者拥有针对受限材料的权限，系统会在运行时即时调用隐私流水线，将机密背景动态装饰在公共基线之上。Subramanian 举例说明：当工程项目出现严重里程碑延期时，普通团队成员仅能看到公开的技术排期风险；但具有高层权限的管理者能看到一条叠加的战略背景——公司正在秘密筹备一项相关收购，若收购成功该滞后工具将被直接替代。这一机制既防止了未公开收购消息过早泄露，又使管理层能准确校准投资组合的整体风险敞口。 业务效益与高能动性协作愿景 (16:17) 在 Meta 的实际部署中，Rhythm 彻底重构了周报流程。以往项目经理与工程团队需要在每周五耗费数小时逐一催促进度、核对冲突信息并在周一手动整理报告；而在 Rhythm 运行后，周一早晨所有进展、指标和风险均已自动对齐生成，项目负责人通常仅需约 5 分钟完成微调，微调数据反过来持续训练系统的精准度。目前该系统已在 Meta 内部支持超过 200 个项目。 此外，团队开始依托 Rhythm 的规范化数据模型开发更深层的战略差距分析工具，比对长期目标与在研项目之间的能力断层，主动提出补齐短板的路线图建议。Subramanian 将此视为工程组织向规模化演进的关键一步，未来的终极愿景是将系统打造为具有高能动性的虚拟队友，不仅主动预警，还能在工程流水线中自动执行低悬果实的补救措施。 Subramanian 描绘的前瞻性智能展示了组织管理从人工催办向自动化智能协同的转变潜力，但该系统的成功高度依赖于员工对严格权限隔离机制的信任，以及各团队对数据源持续显式授权的意愿。未来如何在扩大自动化修复权限的同时规避误操作风险，并在更广泛的异构工程场景中持续验证其事实纠偏能力，仍是前瞻性智能走向完全自主化需要持续探索的问题。

Detailed notes — zh

## Background 该视频由 YouTube 频道 @Scale 于 2026 年 10 月 30 日发布。页面未提供具体的简介说明，内容主要记录了来自 Meta 的 Sundararajan Subramanian 带来的技术演讲，分享为组织构建前瞻性智能系统 Rhythm 的实践经验。 ## TL;DR Sundararajan Subramanian 探讨了组织智能从传统被动式提问到前瞻性、主动式感知的演进路径，并重点剖析了现代企业面临的战略脱节、决策延迟与风险级联等失效模式。为攻克这些难题，Meta 打造了 Rhythm 系统，通过证据引擎、上下文引擎、状态引擎与自学习闭环，实现跨团队潜在风险的主动预警。为了在多源数据融合与多层级汇总中建立信任，Rhythm 落地了细粒度访问控制、受众交集校验及“感知查看者洞察”架构。目前该系统已在 Meta 内部 200 多个项目中落地，将原本耗时数小时的周报汇总压缩至数分钟，推动组织迈向高能动的“工程丰沛”时代。 ## Atmosphere & style 整场演讲呈现出典型大厂资深技术专家严谨、务实且充满洞见的专业气场。Subramanian 的讲述条理清晰、抽丝剥茧，没有浮夸的营销词藻，而是直击大组织跨部门协作与信息流转中的核心痛点。随着架构图的逐步拆解与具体工程案例的生动呈现，观众能直观感受到其在兼顾系统前瞻推理与严密权限控制之间的深层思考，整场分享透露出一种用扎实工程架构化解复杂组织摩擦的从容与底气。 ## Bulletin - [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) Meta 推出名为 Rhythm 的前瞻性智能（Anticipatory Intelligence）系统，旨在从根本上解决现代组织中战略执行脱节、决策延迟以及跨团队风险级联效应等典型失效模式。 - [12:43](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=763s) Rhythm 采用“感知查看者洞察”（Viewer-Aware Insights）架构，通过公共与保密两条同构数据流水线，在严格遵守 ACL 权限的前提下，为高层管理者动态修饰并呈现全局视角。 - [15:09](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=909s) 该系统在 Meta 内部已覆盖 200 多个在研工程项目，成功将原本耗时数小时的周报状态同步与对齐流程缩减至 5 分钟的人工微调与校对。 - [06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) 系统配备事实自学习闭环，能够将项目负责人对报告的关键事实修改反哺给状态引擎，确立为最高优先级的有效事实，持续提高洞察准确度。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=5s) 组织智能经历了从纸质报表、Web 工具再到 LLM 被动式问答的演进；但 LLM 仍受限于提问质量，前瞻性智能的核心是无需用户提问即可主动感知并呈现实时风险。 - [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) 现代复杂组织普遍存在三大失效模式：战略向下一线执行时脱节、风险向上汇报导致决策滞后，以及跨团队交付排期的负面级联连锁反应。 - [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) Rhythm 的定位不是被动追踪组织的进度工具，而是通过跨团队深度上下文感知来捍卫组织愿景的前瞻性 AI 系统。 - [03:32](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=212s) 以训练硬件依赖为例，前瞻性智能可在架构设计阶段主动比对全组织依赖，提前告警潜在供应短缺，避免项目在执行阶段受阻。 - [06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) Rhythm 架构包含证据引擎、上下文引擎、状态引擎与洞察生成器，支持多源工件的时效排序、事实拆解及跨层级报告生成。 - [06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) 自学习循环（Self-Learning Loop）能自动区分语言微调与关键事实修改，将项目负责人的事实纠正固化为最高优先级的权威状态。 - [09:52](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=592s) 规模化落地的核心瓶颈在于信任，AI 必须读取所有文档，但呈现洞察时绝不能向无权限人员泄露任何未授权信息。 - [11:50](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=710s) 跨事实综合推理可能带来泄密风险，因此多源合成的新叙事只能对相关原始信源受众的“交集”开放，且临时聊天绝不沉淀为核心事实。 - [11:50](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=710s) 系统严格恪守“提取前先授权”原则，只主动推荐潜在相关信源，绝不擅自越权抓取未获负责人许可的工件数据。 - [12:43](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=763s) “感知查看者洞察”架构先生成共享公共洞察，再根据登录高管的安全权限动态叠加保密工件的事实，实现全局风险的精准评估。 - [15:09](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=909s) 系统生成的每条洞察均附带严格的引文与信源溯源（provenance），既增强了结论可信度，也实现了数据隐私策略与具体事实的绑定。 - [15:09](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=909s) Rhythm 沉淀的标准事实库已被用于上层战略工具，能够自主识别当前在研项目与公司既定战略之间的空白缺口并推荐新提案。 - [17:48](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=1068s) 随着行业步入“工程丰沛”时代，具备高能动性（high agency）的 AI 协同系统将在后台化解组织杂音，让团队专注高价值的产品创新。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=5s) 组织智能的演变：从被动查询到前瞻性智能 组织智能（Organizational Intelligence）在科技行业经历了漫长的技术迭代。最初阶段依赖人工记录分类账册、手动汇总报表、发送纸质邮件，并在组织的各层级与领导层之间传递状态报告。随后，行业演进至基于 Web 的工具化时代，项目管理软件、实时数据看板、分析报表及沟通协作平台成为承载组织智能的主要载体。大语言模型（LLM）的出现带来了真正的技术突破，系统能够将海量信息消化并整合为一个统一的智能层，员工可以通过自然语言向该层发起提问以获取所需信息。 然而，当前的 LLM 范式本质上仍是“被动式”（reactive）的，其价值完全取决于提问者能否提出恰当的问题。正如 Sundararajan Subramanian 所指出的：“The proactive org intelligence is where you don't even have to ask the right questions because it's only good when as much as the questions that are being asked.” 真正的演进方向在于构建一种前瞻性、主动式的组织智能——它无需用户主动发起查询，而是能够自我感知并主动将关键问题与信息呈现在决策者面前。 ### [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) 现代组织的常见失效模式与 Rhythm AI 的定位 现代复杂组织在运转过程中普遍存在三种典型的失效模式（Failure Modes）： 1. **战略执行脱节（Strategy Misalignment）**：高层制定的战略在向下渗透至具体执行层的过程中不断发生变异与走样，导致顶层规划与一线实际工作之间出现巨大的鸿沟。 2. **决策延迟（Decision Latency）**：一线执行过程中遇到风险、需要制定决策或寻求指导时，这些诉求在逐级向上汇报的过程中消耗了大量时间，导致决策滞后。“There are some guidance that the team wants to hear and all these ask it travels up to the right altitude and before it reaches the right person a time has passed by and this has caused a decision latency as such.” 3. **风险级联效应（Cascading Risks）**：多个团队往往为了同一个大目标协同工作，单个团队的排期延误或局部风险，极易在知情或不知情的情况下对组织内其他团队的交付产生负面连锁反应，而传统机制很难及时捕获这种跨团队的风险传导。 针对这三大痛点，组织需要的不再仅仅是一个跟踪项目进度的工具，而是一个能够捍卫组织愿景的 AI 系统。Sundararajan Subramanian 强调：“So what you need is an AI that does not track an org. It has to protect its vision.” 为此，Meta 推出了 Rhythm AI，旨在通过前瞻性智能（Anticipatory Intelligence）攻克跨团队协作中的硬核难题。 ### [03:32](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=212s) 前瞻性智能的机制与硬件依赖案例 前瞻性智能本质上是一种具备“完整组织感知力”（complete organizational awareness）的 AI 系统，它不仅全量捕获信息，更能在最关键的时间节点主动推送用户必须了解的事项，摆脱对人工精准提问的依赖。 Sundararajan Subramanian 列举了一个典型的工程研发案例： - 某团队正在为新项目设计架构方案，该架构在训练层面存在强依赖，需要在接下来的 1 个月内获取 GB300 GPU 训练硬件。 - 架构师在撰写设计文档时直接写入了该硬件依赖，但该团队并不知晓，此前已有多个团队因该特定训练硬件缺货而遭遇严重延期。 - 在被动式 AI 模式下，如果架构师没有主动向系统提问（例如：“我们依赖这款硬件是否有供应风险？”），系统便不会告警，团队往往要等到实际执行阶段撞墙受阻，才意识到项目已被卡死。 - 而在前瞻性智能模式下，当架构文档生成的瞬间，系统即可洞察这一依赖关系，主动触发风险预警：“Many teams who are actually dependent on this particular training hardware have been delayed or been blocked. Do you want to consider something else or even better these are all the some suggestions that the way teams have solved these problems.” 前瞻性智能不仅能够跨团队交叉比对深层组织上下文，还能将某一个团队踩过的坑转化为全组织的预防性洞察。 ### [06:23](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=383s) Rhythm 核心架构：从多源证据到自学习循环 为了实现前瞻性智能，Rhythm 构建了一套分层的端到端数据流水线与推理引擎。 ```mermaid flowchart TD accTitle: Rhythm AI Engine Architecture accDescr: 展现 Rhythm 从多源证据摄取、上下文组装、状态切分、洞察聚合到自学习闭环的架构流程。 subgraph Ingestion["数据摄取与上下文组装"] A["多源工件 (Docs / Sheets / Dashboards)"] --> B["证据引擎 (Evidence Engine)"] B -->|"时效排序与优先级仲裁"| C["上下文引擎 (Context Engine)"] end subgraph Processing["状态解析与洞察生成"] C --> D["状态引擎 (State Engine)"] D -->|"事实提取与分类"| E["项目状态 (目标/里程碑/风险)"] E --> F["洞察生成器 (Insights Generator)"] F --> G["多层级报告 (项目洞察 / VP 业务组合汇总)"] end subgraph Feedback["反馈与自学习闭环"] G --> H["项目负责人校正 (事实性修改)"] H -->|"触发自学习"| I["自学习循环 (Self-Learning Loop)"] I -->|"记录为高优先级新事实"| D end ``` - **证据引擎（Evidence Engine）**：统一摄取散落在组织各处的多源工件（文档、Google Sheets、Excel 表格、看板等），并负责仲裁优先级与时效性。例如，早期架构文档虽然存在，但最新决策会议纪要具有更高的时效效力，证据引擎会确保最新记录在事实冲突中占据主导。 - **上下文引擎（Context Engine）**：接收经过排序与权重计算的证据链，按优先级重构出全局上下文。 - **状态引擎（State Engine）**：从上下文中提取具体事实，将其结构化切分为不同的受跟踪项（Tracked Items），包括目标（Objectives）、里程碑（Milestones）、风险（Risks）及重要须知（Things to know）等独立状态。 - **洞察生成器（Insights Generator）**：将上述结构化状态进行聚合，针对不同受众生成人类可读的高维摘要。在一线团队层面提供项目细粒度洞察；在副总裁（VP）层面提供业务组合级汇总（Portfolio Rollup），例如直接预警“您的业务组合下有多个项目因训练硬件受阻”。 - **自学习循环（Self-Learning Loop）**：每次生成洞察后，系统会向项目负责人（Project Owner）发送通知。若用户进行修改，自学习循环会自动将修改分类为“表述/语言微调”（cosmetic/language edits）与“关键事实修改”（critical factual edits）。针对后者，系统会反查状态引擎漏判原因，并将该修正固化为状态引擎内具备最高裁决效力的新事实，直至被更新的工件覆盖。 - **CLI 与平台集成**：系统提供统一的 CLI 接口与集成点，供其他第三方内部系统与工具调用底层能力。 ### [09:52](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=592s) 大规模信任与信息访问控制原则 在前瞻性智能广泛落地的过程中，核心挑战并非技术实现，而是规模化信任（Trust at Scale）。Sundararajan Subramanian 明确指出：“The AI must read everything but it must never surface the information to someone who should not see it.” 如果 AI 系统因为掌握全量文档而将机密信息泄露给无权查看的人员，整个信任体系便会崩溃，团队也将拒绝向其接入核心文档。 为此，系统确立了严格的信息访问安全原则： - **尊重原始访问控制列表（Respect the Original ACL）**：状态引擎中的每一项事实都必须与其原始工件强绑定。当向特定用户呈现状态与洞察时，其可见性必须严格受限于该用户在底层原始工件上所拥有的访问权限。 - **禁止受众越权扩散（No Audience Escalation）**：仅在小范围内讨论和流转的信息，严禁被推送到更广泛的受众层级，除非该内容已获得原受众团队的明确批准或升级许可。 - **防范交叉综合泄密（Cross Synthesis Risks）**：当系统将散落于不同工件、面向不同受众的零散事实综合推导成一段全局叙事时，极易无意间向外部暴露单项原本受限的信息。“You collect all these two or three facts together and then form a narrative on top of it. When this happens, who should it be visible? If it's made public, then probably there is a risk that this has exposed the information that should not be visible to few people.” 因此，交叉推理生成的复合洞察必须经过最严格的受众交集校验。 ### [11:50](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=710s) 隐私边界与提取授权机制 在组织内部的多源情报融合中，数据的权限和边界管理至关重要。Subramanian 阐述了系统在处理多源叙事时所恪守的核心安全原则：生成的新叙事必须仅对所有相关信源受众的“交集”可见（intersection of audiences）。这意味着如果两份材料分别有不同的受众权限，融合生成的洞察绝不能突破原有的安全隔离，向权限不足的人群泄露另一部分信息。与此同时，短暂性的沟通必须保持其短暂属性（“Ephemeral stays ephemeral”）。无论是随意的单次即时聊天还是临时对话，都应仅维持原样，绝对不能被固化成组织的核心事实。 此外，系统在信息抽取上严格遵循“提取前先授权”（consent before extraction）的原则。在 Rhythm 提取任何项目工件（artifacts）中的数据之前，必须先获得相应项目负责人（project owner）的明确许可。Rhythm 具备主动探索能力，能够向项目团队推荐可能相关的其他潜在信息源；但即便系统发现了这些具有高价值的参考资料，它也绝不会越权自行抓取。系统只提供来源建议，最终决定权完全在项目负责人手中，只有当负责人手动、明确地将这些源头添加进来后，它们才会被纳入组织的整体智能库中。 - **受众交集可见**：生成的新叙事与见解只对所涉及工件权限的交集群体开放。 - **短暂对话隔离**：临时即时通讯内容不沉淀为核心事实，避免上下文断裂产生误导。 - **前置授权抽取**：所有工件接入前必须经过负责人授权，杜绝无感越权采集。 - **推荐但不擅采**：系统可自主识别并建议相关信源，但必须由负责人主动确认添加。 --- ### [12:43](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=763s) 针对管理层的“感知查看者洞察”架构 过于严苛的权限限制往往会带来新的业务痛点：拥有最高机密权限的高层管理人员（Leadership）如果只能看到脱敏或被裁剪后的普通公共洞察，这些分析对他们校准全局战略而言将变得毫无用处。为了解决这一两难境地，Rhythm 提出了“感知查看者洞察”（Viewer-Aware Insights）架构。该架构先运行一条公共信息流水线，收集公开工件并建立共享状态，生成并固化一套基础的公共洞察版本（public insights version）。 ```mermaid flowchart TD accTitle: Viewer-Aware Insights Architecture accDescr: Pipeline showing how public and restricted artifacts are merged dynamically based on viewer identity. subgraph BasePipeline["公共流水线阶段"] A["公开工件<br>Public Artifacts"] --> B["证据收集与共享状态<br>Evidence & Shared State"] B --> C["基础公共洞察<br>Public Insights"] end subgraph DynamicPipeline["动态权限修饰阶段"] D["保密工件<br>Restricted Artifacts"] --> E{"验证查看者权限<br>Viewer Access?"} E -->|"有权限"| F["同构隐私流水线抽取<br>Extract via Privacy Pipeline"] C --> G["最终定制洞察视图<br>Decorated View"] F --> G end ``` 当特定用户打开系统查看洞察时，动态匹配机制随即启动。在早期的证据收集阶段，系统已经将部分工件打上了“受限”（restricted）标签。系统会实时核验当前查看者的访问权限：如果发现该用户拥有某些保密工件的阅读权限，就会将这部分高密工件送入同一套符合隐私规范的流水线中，并将提取出的高级洞察实时“修饰”（decorate）叠加到原有的公共洞察之上。 Subramanian 举了一个具体案例来解释这种机制的运作方式：某项目在推进过程中遇到了一个关键里程碑延期，该延期可能会对整个产品线产生严重后果；在公共视图中，项目正面临重大风险。然而，负责统筹该板块的副总裁（VP）手中掌握着一项公司正在进行的保密收购谈判，被收购工具的功能恰好可以覆盖该延期项目的诉求。由于其他团队成员无权知晓收购事宜，他们看到的是局限于自身业务范围的项目风险；而当副总裁登录系统时，系统自动叠加了保密收购的进展洞察。副总裁据此可以评估出全局风险可控，从而避免在不必要的补救措施上过度投入资金或调整整个产品组合。 --- ### [15:09](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=909s) 信任构建、来源溯源与项目管理革新 组织对系统的信任不能建立在“因噎废食”的假设之上。如果简单地告诫员工“不要把高密信息放进工具里”，工具将失去其在现代协作中的意义。Rhythm 允许团队将各种工件放入系统，所有安全控制交由底层细粒度权限逻辑托管。系统中的所有追踪项（tracked items）均以结构化模型存储，且与对应的工件隐私规则紧密绑定。同时，Rhythm 生成的每一条洞察都必须配备明确的引文与溯源出处（citations and provenance）。Subramanian 强调，将每一段洞察与原始信源严密关联，不仅赋予了结论极高的可信度，也正是这种机制使得系统能够准确地将对应的数据隐私策略绑定至每一条具体事实。 在应用成效方面，Rhythm 在过去几个月内彻底颠覆并取代了传统的周报状态同步流程。过去每逢周五下午，项目经理（PM）和协调员必须逐一联系工程师与各跨职能团队，打听项目进度、梳理矛盾信息，整理出冗长的报告并在周一早晨向领导层汇报。而在引入 Rhythm 之后，这一切繁琐的线下跨团队沟通完全被自动化流水线取代，系统在周一早晨直接输出整理完毕的高质量报告，人工仅需花费约 5 分钟进行微调；而用户的微调修改又会反哺给系统进行自主学习（self-learning）。 - **工时大幅缩减**：以往耗费数小时的人工跨部门沟通与对齐，被彻底压缩至 5 分钟的高效校对。 - **规模化落地**：已正式服务 Meta 内部 200 多个活跃工程项目。 - **上层衍生生态**：基于 Rhythm 沉淀的权威规范信息（canonical information），团队搭建了更上层的战略辅助工具，能自动识别当前在研项目与既定战略之间的空白缺口（gaps），并主动给出新项目与新提案的填补建议。 --- ### [17:48](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=1068s) 走向工程丰沛时代 在演讲的最后，Subramanian 展望了由这类智能系统引领的组织未来变革。他指出，整个行业正在迎来一个“工程丰沛”（engineering abundance）的伟大时代。AI 工具的发展不再仅仅是充当简单的查询机器人或代码补全插件，而是逐步演进为能够感知全局组织脉动的深度协同伙伴。 > “Rhythm 旨在成为具备高能动性的团队成员，通过在合适的时间浮现出合适的信息来积极保护组织，同时主动采取一些唾手可得的补救措施，协助你在工程流水线中创造出更多成果。”（“Rhythm aims to be the high agency teammate that actively protects the organization by surfacing the right information at the right time and also go ahead and proactively take some low-hanging remediation actions and help you achieve a lot more in the engineering pipeline.” —— Sundararajan Subramanian） 这种具备主动性（high agency）的系统，其核心使命是在后台默默吸收组织杂音，主动识别潜在延误和信息隔阂，在风险酿成事故前就将其暴露给有权处理的人。通过将工程师和项目管理人员从无休止的状态对齐与低效汇报中解放出来，组织能够将生产力真正聚焦于高价值的战略执行与产品创新中。 ## Notable quotes > [01:42](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=102s) So what you need is an AI that does not track an org. It has to protect its vision. — Sundararajan Subramanian > [09:52](https://www.youtube.com/watch?v=vnEWk2PQY4E&t=592s) The AI must read everything but it must never surface the information to someone who should not see it. — Sundararajan Subramanian ## Mentioned Sundararajan Subramanian, Meta, Rhythm AI, GB300 GPU, Google Sheets, Excel [Source](https://www.youtube.com/watch?v=vnEWk2PQY4E)

Source quotes

[1:42] Sundararajan Subramanian: So what you need is an AI that does not track an org. It has to protect its vision. [9:52] Sundararajan Subramanian: The AI must read everything but it must never surface the information to someone who should not see it.

Source quotes

[9:52] Sundararajan Subramanian: The AI must read everything but it must never surface the information to someone who should not see it. [1:42] Sundararajan Subramanian: So what you need is an AI that does not track an org. It has to protect its vision.

Original transcript

[0:05] Hello everybody. [0:07] What if your orgi [0:09] told you what's broken even before you [0:12] asked? [0:14] The industry has always had some sort of [0:19] organizational intelligence. Now at the [0:22] start it was the humans keeping the [0:24] ledgers, deriving reports out of it, [0:27] sending paper mails out of it, status [0:30] reports to the leadership and across [0:32] different layers of an organization. [0:34] Then it evolved into web- based tooling [0:38] where we have different project [0:41] management software, dashboards, reports [0:45] and communication platform that helped [0:48] provide this organizational [0:49] intelligence. [0:50] Then the real breakthrough came with the [0:53] LLMs. With the LLMs, you can digest all [0:57] these information into a simple [1:00] intelligence layer and then you can have [1:03] different people query this and get the [1:06] right information that they need. This [1:08] is what the LLM has unveiled to us. [1:11] And this is something we call as [1:13] reactive where it's on you to ask the [1:16] right questions. Now where can we take [1:18] this? The proactive or intelligence is [1:21] where you don't even have to ask the [1:23] right questions because it's only good [1:26] when as much as the questions that are [1:28] being asked. But with the proactive [1:30] organ intelligence, it knows the right [1:32] things that needs to be surfaced. And we [1:34] are going to talk about this proactive [1:37] organ intelligence [1:39] in the next few minutes. [1:42] Even before talking about ta let's talk [1:45] about what are the common failure modes [1:47] of the modern organization. The first [1:50] thing is a strategy is being made at the [1:53] top and then as it percolates down the [1:56] layers into where the work is being [1:58] performed it starts mutating and so [2:02] there is a gap and misalignment between [2:04] the strategy and the actual work. Then [2:08] the reverse happens. [2:10] The work is being performed. There are [2:13] some risk. There are some decisions that [2:15] needs to be made. There are some [2:17] guidances that the team wants to hear [2:20] and all these ask it travels up to the [2:23] right altitude and before it reaches the [2:26] right person a time has passed by and [2:29] this has caused a decision latency as [2:31] such. It's too late at that point of [2:34] time. And the third thing is where every [2:38] organization has multiple teams working [2:40] towards a common objective. Now where a [2:44] schedule slippage or a risk or an issue [2:46] in one of the teams could knowingly or [2:50] unknowingly adversely affect the [2:52] outcomes of another teams in the same [2:54] organization. Now who will capture how [2:57] the risks cascade between teams and how [3:00] it affects the overall vision? [3:04] Now these are all the common problems [3:06] that happens in smart organization. So [3:09] what you need is an AI that does not [3:12] track an org. It has to protect its [3:14] vision. So here comes Rythm AI. We are [3:20] going to talk about how Rythm AI offers [3:23] anticipatory intelligence and how does [3:25] it solve some of the hard problems and [3:28] what are the results and what is the [3:30] vision forward. [3:32] What's anticipatory intelligence? [3:34] Anticipator intelligence is an AI system [3:37] with complete organizational awareness [3:40] and then it has to go ahead and then [3:44] identify all the information that's [3:46] required. It does not wait for you to [3:49] ask the right question and then it tells [3:51] you what you need to know right when [3:53] that is needed. Let's take an example. [3:56] There's a team that's coming up with an [3:58] architecture for their new project. The [4:01] architecture demands a training [4:03] requirement and it needs GB300s for as [4:08] training GPUs in the next 1 month. [4:10] Architecture takes dependency on this [4:13] when your architecture document is being [4:16] written. [4:18] Little did they know there are many [4:20] other teams that had the same dependency [4:23] and were extremely delayed because of [4:25] this nonavailability of this particular [4:28] training hardware. Now the architecture [4:32] evolves and by the time they go ahead [4:34] and then start the execution at that [4:36] point of time they hit the ceiling and [4:39] that's when they realize okay this is [4:41] not moving forward. Now this is where we [4:44] should go ahead and when that [4:45] architecture document is being formed. [4:48] If your RGI derives that information and [4:52] highlights that risk right then and [4:54] there and then say hey many teams who [4:58] are actually dependent on this [5:00] particular training hardware have been [5:02] delayed or been blocked. Do you want to [5:04] consider something else or even better [5:07] these are all the some suggestions that [5:08] the way teams have solved these [5:10] problems. Do you want to choose one of [5:12] these other options? This is what the [5:15] anticipatory intelligence look like. [5:18] Having said this, let's look at what the [5:22] difference between reactive AI and [5:24] anticipatory intelligence. In reactive [5:26] AI here in this particular example, when [5:29] the architecture document is being made, [5:31] the person who is making it should go [5:34] and ask, I am taking a dependency on [5:37] this particular training hardware. Do I [5:39] have it? Is there any risks and all that [5:41] stuff if they don't ask it they will [5:43] never go ahead and they get the right [5:44] information. In the anticipatory [5:47] intelligence the risk is surfaced by the [5:51] time the decision or by the time the [5:53] plans are being made and that way it [5:56] proactively monitors and then it [5:58] proactively goes ahead and then gives [6:00] that right information. What makes it [6:02] possible? It has this deep [6:04] organizational context cross- linked [6:06] with other teams and be able to provide [6:08] the right information there and it [6:11] serves the entire earth. It's not only [6:12] your team. It could be from any other [6:15] team. The risk that you face go ahead [6:18] and get surfaced as possible risks for [6:21] the other teams. [6:23] Now how do we achieve this rhythm [6:27] providing the anticipatory intelligence? [6:30] Now the first thing is for any [6:33] intelligence platform you need the [6:35] sources you need the artifacts. What are [6:38] those? They are documents, they are [6:40] dashboards, they are sheets, Google [6:42] sheets, they are excel sheets and many [6:44] many different forms of these resources. [6:47] All these resources has to be reconed by [6:51] something called evidence engine. It [6:53] goes ahead and takes these sources and [6:56] then orders them in priority and then [6:58] goes ahead and then figures out like [7:00] which one takes precedence because some [7:02] of these documents could be old. The [7:04] initial architecture document could be [7:06] old but then the later decision meeting [7:08] notes could be the newest one which will [7:10] take precedence on some of these [7:12] decisions. Then once it sorts and [7:15] organizes these evidences, it goes ahead [7:18] and feeds it into something called [7:19] context engine. The context engine takes [7:22] all these evidences and then creates a [7:25] context out of it in terms of in your [7:28] right priority order. And then this [7:31] context is provided to the state engine [7:33] which goes ahead and divides this [7:36] context into different states like it [7:40] goes ahead and finds out the different [7:42] facts from these artifacts and then [7:44] segregates into what are all the [7:46] objectives what are all the milestones [7:48] what are all the risks what are all the [7:50] things to know and different other [7:52] things and categorize them separately. [7:54] Now once you have this categorized list [7:57] of states or tracked items now this is [8:01] fed into insights generator which [8:04] aggregates all this tracked items and [8:08] provides a human readable insights human [8:12] readable summary for different [8:14] audiences. Right? So now at the team [8:16] level you have these project insights [8:19] that are generated but then for a VP [8:22] they have multiple projects a rollup of [8:25] all these insights happen and then a [8:27] portfolio rollup presence and then it [8:29] would could even go ahead and give you [8:31] an example say that hey across your [8:33] portfolio many people many projects are [8:36] blocked on this training hardware can [8:38] you do something about it then [8:42] what happens is every time the insights [8:44] generated a notifications are sent and [8:47] your project owner goes ahead and then [8:49] looks at these insights there could be [8:52] some corrections that could be made. Now [8:54] when the corrections that are made [8:57] immediately the self-arning loop begins. [9:00] The self-arning loop goes ahead and [9:02] classes these edits as cosmetic or [9:04] language edits versus critical factual [9:07] edits. Now when for those critical [9:09] factual edits it goes ahead and figures [9:11] out why was this edit made? Why was this [9:14] not captured by the state engine [9:17] earlier? And then it goes ahead and then [9:20] records that as a new fact as a new [9:23] trapped item in the state engine so that [9:26] the further insights can make use of [9:29] this and consider this as precedence [9:31] unless it's overridden by a new fact [9:34] from some other artifact. [9:36] This forms the entire core of your [9:39] rhythm engine. Now on top of it, it has [9:42] integration points through CLA [9:44] integrations which other tools can plug [9:46] into and then provide functionality on [9:50] top of it. [9:52] Now the second thing, okay, this all [9:55] seems cool. What's the big problem at [9:58] the real problem is the trust at scale. [10:01] How do you make sure that this is [10:04] trusted? How do you make sure this [10:07] caters to the needs of different layers [10:10] of an organization to be useful? The AI [10:13] must read everything [10:15] but it must never surface the [10:19] information to someone who should not [10:21] see it. If it has all the documents and [10:25] if it surfaces all the information then [10:28] no one will trust it. No one will give [10:29] all the right documents to it. [10:33] We for this we will go ahead and then [10:35] see what are the core principles that it [10:39] should adhere to the here we have this [10:42] five principles for a information access [10:45] respect the original ACL we have [10:49] specific ACL on this facts and each [10:51] artifact is connected to the state so [10:55] when your state is being suffaced in the [10:57] inage you have to make sure that it's [10:59] visible to only those people who the [11:02] original artifact actually had in its [11:05] visibility access. Then no audience esc [11:10] whatever is in a small group of people [11:12] should stay within that unless and until [11:14] explicitly promoted or approved by them. [11:17] Then the cross synthesis this is [11:19] something that's very very easy to miss [11:21] but what happens is you have this [11:23] disperate facts [11:25] actually happening across different of [11:27] these artifacts with different set of [11:30] audiences. You collect all these two or [11:33] three facts together and then form a [11:35] narrative on top of it. When this [11:37] happens, who should it be visible? If [11:40] it's made public, then probably there is [11:42] a risk that this has not [11:46] this has exposed the information that [11:48] should not be visible to few people. [11:50] Then here what we need to make sure is [11:53] that new narrative should be visible [11:56] only to the intersection of the [11:58] audiences. Ephemeral stays ephemeral. So [12:02] whatever is a one-off chat, one-off [12:04] conversation should stay that way and [12:06] you should not become part of the core [12:09] fact. Then consent before extraction for [12:12] everything you have these artifacts [12:14] before extraction. A specific project [12:16] owner has given permission for you to [12:18] extract the information out of it. Not [12:20] only that, Rutam can go ahead and then [12:22] find okay, these are all the possible [12:24] other suggested sources for you, right? [12:27] It should not go ahead and collect [12:28] information from the sources that were [12:30] not explicitly added but can suggest [12:32] some of the sources then the project [12:34] owner should go ahead and explicitly add [12:37] those sources to be part of the or [12:40] intelligence here. [12:43] Having said that, okay, this looks too [12:46] restrictive. How would how would it [12:50] cater to the people who actually are at [12:53] the leadership level who have access to [12:55] confidential information and when the [12:59] rhythm insights do not factor for those [13:01] how would it be useful to them that's [13:03] where we have something called viewer [13:06] aware insights what this viewer aware [13:08] insights does is it runs the first [13:10] public information pipeline as I [13:13] explained earlier it has this public [13:15] information it goes ahead and does this [13:18] evidence collection and then creates [13:20] this shared state generates a public [13:23] insights and saved it as a public [13:26] insights version. Now when a viewer [13:29] comes in to save the insights what it [13:32] does is during the evidence collection [13:34] it also has marled some of these [13:36] artifacts as restricted and only visible [13:39] to few people. So when the viewer came [13:41] in, it goes ahead and looks into those [13:43] artifacts and then say how many of those [13:46] artifacts does this particular viewer [13:48] have access to and then takes those [13:50] artifacts [13:52] that this viewer has access to and then [13:54] generates insights by passing it through [13:57] the same pipeline in a privacy aware and [14:00] decorates the existing public insights [14:03] with those new insights. Now this is how [14:06] it's made useful to everyone. Let's take [14:08] an example here. Let's say you have a [14:10] project being executed. There are [14:12] milestones that are coming up. There are [14:14] milestone that is considered. There is a [14:16] delay and this delay could cause some [14:19] serious consequences for the project. [14:22] But then there is also some acquisition [14:25] talks going on on your tool that should [14:28] satisfy similar requirements as this [14:30] product. But the VP who is overseeing it [14:32] has the information to the acquisition [14:34] not anyone else. So when these insights [14:37] are presented to the VP [14:40] VP to calibrate the risk of the entire [14:44] vision, he goes ahead and then says okay [14:47] I also have this additional insights [14:50] which is based on the acquisition and [14:52] how far it is progressing. So if it goes [14:56] well then I should not risk my entire [14:58] portfolio. So he has this additional [15:00] layer of information presented but none [15:02] of the other team members even a scope [15:04] to the information that is relevant to [15:07] them. [15:09] Here we build the trust into the system. [15:12] There are two ways to deal with it. You [15:13] can go ahead and say do not put any high [15:16] confidential information into the tool [15:18] but that does not serve the purpose [15:20] here. You can throw all the artifacts [15:22] into the system and the system will take [15:25] care of going ahead and providing the [15:27] right access control on each of these [15:30] information. Everything is owner [15:33] controlled. So the owners have to [15:35] explicitly say [15:37] some information or some facts that can [15:39] be made visible outside. Then all the [15:43] artifacts were digested and then created [15:45] as tracked items and a very structured [15:49] model. And each of these items are tied [15:51] to their privacy and the artifacts. Then [15:55] you have this citations or provenence [15:57] where whenever an insight is generated, [16:01] each part of this is tied to the exact [16:05] source where it is taken from which [16:08] provides credibility to that [16:10] information. On top of it, it also helps [16:13] attach the right privacy to each of this [16:15] information. [16:17] So what did we achieve with this? Over [16:21] the past few months, this has replaced [16:24] the status updates. What does it mean? [16:27] Every Friday afternoon, project managers [16:30] and coordinators have to go and reach [16:32] out to different engineers, different [16:34] crossunctional teams to identify where [16:37] everything stand, prepare a report, [16:40] reconcile conflicting information and [16:43] then generate a report that can be [16:46] digested and post it on Monday morning [16:48] and update the leadership on it. Now [16:52] with rhythm [16:54] all you need is you don't have to do [16:56] that coordination on Monday morning you [16:59] have everything compiled and ready for [17:02] you. You can go ahead and make small [17:04] edits that if you find is required which [17:08] will become part of the self-arning and [17:10] it's there. What used to take a lot of [17:13] time a few hours has been crunched to 5 [17:17] minutes of work. This has started [17:19] serving 200 plus projects. On top of it [17:22] with all the rich information that [17:24] exists in rhythm, there are tools that [17:27] are being built on top of this for the [17:29] canonical information. Like there is a [17:32] tool that we have which goes ahead and [17:34] then finds out what are all the gaps in [17:38] achieving our strategy based on the [17:40] existing projects and starts filling in [17:43] the gaps with new ideas, new initiatives [17:46] and stuff. [17:48] That's great. Now, where does it all [17:50] lead to? It all leads to [17:54] engineering abundance. [17:56] This is the time for a great engineering [17:59] abundance. Rytham aims to be the high [18:03] agency teammate that actively protects [18:05] the artition by surfacing the right [18:09] information at the right time and also [18:11] go ahead and proactively take some [18:14] lowhanging remediation actions and help [18:17] you achieve a lot more in the [18:19] engineering pipeline. [18:22] Thank you.