Meta Overhauls Data Architecture as AI Agents Strain Infrastructure

Keynote at AI & Data | Barak Yagour from Meta — @Scale

@Scale

Autonomous AI agents are breaking conventional capacity, identity, and velocity limits, forcing Meta to replace legacy batch storage with real-time streaming architectures capable of sustaining 500 million queries per second.

AI-assisted summary of Keynote at AI & Data | Barak Yagour from Meta — original source. Check the original for full context.

Source published:

Key ideas

  • Automated internet traffic reached 51% last year, expanding eight times faster than human traffic as autonomous agents transition into first-class infrastructure citizens. (4:10)
  • Modern foundation models consume 20 times more tokens per parameter than Chinchilla-optimal thresholds, driving the need for architectures sustaining 500 million QPS and 1 petabyte per second of training read throughput. (8:41)
  • Meta is shifting recommender systems from two decades of statistical pattern matching to LLM-driven intent reasoning, responding to data showing 42% of Instagram users want to directly alter the algorithm. (10:48)
  • Internal deployments such as Clowntown and Malbook demonstrate multi-week autonomous workflows and agent-to-agent negotiation operating without humans in the loop. (13:58)

Detailed notes

## Background Published on YouTube by @Scale on June 19, 2026, this keynote features Barak Yagour, VP of Engineering at Meta, addressing the future of AI. The presentation explores how exploding AI workloads and autonomous agents are reshaping engineering infrastructure across Meta's applications.

## TL;DR Barak Yagour outlines how autonomous software agents and deep reasoning models are breaking legacy infrastructure assumptions around capacity, identity, and deployment velocity. To support these workloads, Meta is overhauling its storage and data pipelines, shifting away from opaque blobs and batch processing toward intelligent, real-time streaming architectures capable of 500 million queries per second and 1 petabyte per second throughput. Recommender systems are simultaneously evolving from passive statistical pattern matching into conversational, LLM-powered intent reasoning that actively interprets user goals. These components unite into a self-reinforcing AI product flywheel, where internal multi-agent orchestration frameworks like Clowntown and Malbook transition humans from being 'in the loop' to governing 'over the loop.'

## Atmosphere & style The presentation carries an authoritative, high-energy engineering tone, blending sober technical realism with the awe of hyperscale computing. Yagour delivers his points with sharp clarity and momentum, making the staggering numbers and paradigm shifts feel immediate and inevitable. Watching the keynote feels like peering inside an operational command center at an inflection point, moving from the friction of legacy systems breaking down to a striking vision of autonomous agent societies managing production software.

## Bulletin - [04:10](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=250s) Automated internet traffic reached 51% last year, expanding eight times faster than human traffic as autonomous agents transition into first-class infrastructure citizens. - [08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) Modern foundation models consume 20 times more tokens per parameter than Chinchilla-optimal thresholds, driving the need for architectures sustaining 500 million QPS and 1 petabyte per second of training read throughput. - [10:48](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=648s) Meta is shifting recommender systems from two decades of statistical pattern matching to LLM-driven intent reasoning, responding to data showing 42% of Instagram users want to directly alter the algorithm. - [13:58](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=838s) Internal deployments such as Clowntown and Malbook demonstrate multi-week autonomous workflows and agent-to-agent negotiation operating without humans in the loop.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=5s) Meta committed an initial $115 million to establish America's Workforce Academy to meet soaring demands for specialized AI and infrastructure talent. - [00:05](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=5s) Coding agents and autonomous operators are moving humans from being 'in the loop' to managing systems 'over the loop'. - [04:10](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=250s) Agent workloads break traditional infrastructure across capacity, identity, and velocity because systems were architected around predictable human sessions. - [04:10](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=250s) A single engineer can deploy hundreds of agents that trigger infrastructure load equivalent to 100,000 human users. - [06:03](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=363s) Adoption of agentic data apps drove 63% of Meta's internal dashboards within three months, triggering a 30x increase in agent queries and requiring new governance frameworks. - [08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) Deep reasoning models require full cross-surface chronological histories rather than legacy sparse feature vectors, making 24-hour batch ETL obsolete. - [10:22](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=622s) Infrastructure storage is moving away from opaque blobs to intelligent layers that dynamically retrieve specific columns and temporal windows to eliminate GPU starvation. - [10:48](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=648s) Recommender systems are evolving toward conversational interfaces powered by chain-of-thought reasoning that differentiate identical queries by user context. - [12:47](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=767s) Autonomous agents, intelligent real-time data storage, and reasoning models form a self-reinforcing product flywheel powering Meta's ecosystem for 3.5 billion users. - [13:58](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=838s) Meta's Clowntown framework allows multi-week semi-autonomous task execution, keeping engineering projects moving without active human presence. - [13:58](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=838s) Experimental environments like Malbook explore synthetic social dynamics where agents interact and negotiate directly with other agents. - [13:58](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=838s) The next frontier for infrastructure scale shifts from supporting billions of humans to managing billions of humans alongside millions of concurrent superintelligent agents.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=5s) The Exploding Scale of AI and Meta's Product Loop

Every year at the @Scale Conference, attendees expect massive scale numbers, but this year represents an unprecedented inflection point where the pace of change enables results that were unimaginable just twelve months prior. The relentless appetite for compute has sparked massive data center construction across the industry at an unseen scale. To meet the soaring demand for specialized talent, Meta recently made an initial $115 million investment to build America's Workforce Academy. Across software engineering, the proportion of AI-generated code is breaking records, and agentic token consumption is multiplying by orders of magnitude quarter-over-quarter.

```mermaid flowchart LR accTitle: Meta Core Product Loop with AI Models accDescr: End-to-end loop showing user interactions generating data that trains models to power recommendation systems and product UI.

subgraph ConsumerSpace["Consumer Interaction"] User["Instagram / App User"] Surface["Product Surfaces / UI"] end

subgraph DataAndAI["Data & Model Processing"] DataStream["User Engagement Data"] Models["Large Language, Vision & Speech Models"] RecSys["Dynamic Recommender Systems"] end

User -->|"interacts with"| Surface Surface -->|"generates"| DataStream DataStream -->|"trains & feeds"| Models Models -->|"powers"| RecSys RecSys -->|"updates & personalizes"| Surface ```

Every day, 3.5 billion people interact with Meta's family of apps, supported by complex infrastructure across machines, energy, data centers, hardware, software, privacy, and integrity. In the traditional product loop, an Instagram user interacts with surface UI, generating massive streams of data. That data feeds dynamic recommender systems powered by large language, vision, and speech models, which in turn shape what the user experiences next. While this entire lifecycle was once exclusively designed, deployed, and monitored by human engineers, coding agents can now one-shot large codebases, manage data pipelines, evaluate analytical reports, execute machine learning experiments, and even construct other models.

- **Talent Investment**: Meta committed an initial $115 million toward building America's Workforce Academy. - **Daily Meta User Base**: 3.5 billion people engage daily across Meta's family of applications. - **Agent Evolution**: Moving from task-specific coding assistants to autonomous operators that can run with proper guardrails for hours, days, or weeks while humans transition from being "in the loop" to "over the loop."

---

[04:10](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=250s) Infrastructure Citizens: When Capacity, Identity, and Velocity Break

A massive structural inversion is occurring in infrastructure. Last year, automated internet traffic officially surpassed human traffic, reaching 51% and expanding eight times faster than human traffic. Inside internal systems, agents are no longer just software features; they have become first-class infrastructure consumers. For twenty years, infrastructure architectures were optimized around predictable, session-based human habits bounded by biological constraints—humans sleep, eat, and get distracted, whereas autonomous agents operate continuously without natural pauses.

```mermaid flowchart TD accTitle: The Three Core Infrastructure Failure Modes Caused by AI Agents accDescr: Diagram illustrating how agents cause capacity, identity, and velocity to break simultaneously in systems designed for humans.

Agents["First-Class Infrastructure Agents"]

subgraph FailureModes["Simultaneous System Bottlenecks"] Cap["1. Capacity Breaks<br/>(1 engineer = 100s of agents = 100k user load)"] Ident["2. Identity Breaks<br/>(Not a human badge, service, or static batch job)"] Vel["3. Velocity Breaks<br/>(Code written in seconds vs. slow legacy CI/CD pipelines)"] end

Agents --> Cap Agents --> Ident Agents --> Vel ```

This sudden shift breaks three fundamental systems simultaneously:

1. **Capacity Breaks**: Historically, one engineer equated to one unit of load. Today, a single engineer can spin up dozens or hundreds of agents that spawn sub-agents, enabling an organization of 1,000 engineers to generate the infrastructure load of 100,000 users overnight. 2. **Identity Breaks**: An agent does not hold an employee badge, it was not provisioned as a static service by a production engineer, and it cannot be categorized as a dumb batch job because it makes active, dynamic decisions. 3. **Velocity Breaks**: GitHub Copilot already writes 46% of code for its users. While agents author code in seconds, legacy CI/CD pipelines, build systems, test suites, and deployment monitors remain tied to human cycle times.

- **Automated Web Traffic**: Automated bots/agents account for 51% of all internet traffic and are growing 8x faster than human traffic. - **GitHub Copilot Adoption**: Writes 46% of code for developers using the platform.

---

[06:03](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=363s) Democratizing Data Apps and Establishing Agent Governance

Data sits at the core of all business decisions, recommendation engines, and next-generation foundation models. In February, Meta introduced the ability to build agentic data applications. Within just three months, 63% of all dashboards published across the company were built through data apps, driving a 30x explosion in agentic queries. However, eliminating traditional human data analysts—who historically served as domain gatekeepers understanding regulatory, privacy, and contextual nuances—introduces severe enterprise risks when questions are posed by agents or non-experts lacking context.

To address these vulnerabilities, Meta and industry partners are restructuring data interfaces, compliance guardrails, and observability frameworks:

- **Dinkar's Session**: Covers moving away from rigid data interfaces toward end-to-end data apps that let non-experts and agents assemble custom data products independently. - **Komal's Session**: Focuses on agent governance, safety, and compliance as mandatory prerequisites rather than optional add-ons: "autonomy without governance is chaos." - **Alex from Snorkel's Session**: Highlights agent observability, evaluations, benchmarks, and data quality standards when agents transition from advisory recommendation systems to autonomous decision-makers.

---

[08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) The Scale of Modern Reasoning and Real-Time Data Infrastructure

The volume and complexity of data required for model training and inference is expanding at an extraordinary rate. Since 2022, the ratio of training data to active parameters in open-weight LLMs has surged by 3.1x each year. State-of-the-art models are trained on 20 times more data per parameter than the optimal ratios originally proposed by the 2022 Chinchilla scaling laws. Beyond language models, core business systems like recommendation and ranking engines (e.g., Lexis) require ever-increasing volumes of rich data.

```mermaid flowchart LR accTitle: Evolution from Pattern Matching to Deep Reasoning Data Requirements accDescr: Comparison of sparse feature vectors used in pattern matching versus full cross-surface behavioral history required for reasoning models.

subgraph Legacy["Pattern Matching"] Sparse["Sparse Feature Vectors<br/>(Summarized Signals)"] BatchETL["24-Hour Batch ETL"] end

subgraph Modern["Deep Reasoning Models"] FullHistory["Full Cross-Surface Behavioral History<br/>(Complete Action Stream)"] StreamPipeline["Real-Time Streaming Pipelines"] end

Sparse -.->|"evolves to"| FullHistory BatchETL -.->|"replaced by"| StreamPipeline ```

The underlying architectural demand is driven by the fact that "reasoning is hungry." Unlike legacy pattern matching that operates on sparse, summarized feature vectors, advanced reasoning requires a full cross-surface chronological view of user behavior over time. These models are already driving substantial performance gains across Meta's ad-ranking and organic recommendation pipelines. Supporting these workloads demands a complete overhaul of storage and processing architectures:

- **Target Throughput**: Meta is engineering systems capable of sustaining 500 million queries per second (QPS) and 1 petabyte per second of training data read throughput. - **Chinchilla Over-Training**: Modern foundation models consume 20x more tokens per parameter than the 2022 Chinchilla scaling optimal threshold. - **Real-Time Streaming**: Traditional 24-hour batch ETL pipelines are being replaced with low-latency streaming architectures because reasoning over real-time intent cannot tolerate batch delays. - **Intelligent Storage**: Data storage layers must evolve from passive repositories into active, intelligent components capable of supporting real-time streaming and high-bandwidth retrieval.

[10:22](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=622s) Rethinking the Storage Layer: From Opaque Blobs to Intelligent Data Retrieval

The historic approach to managing user data relied on storing information as opaque blobs with zero internal awareness of what was contained inside. This lack of visibility triggered massive over-fetching and severe GPU starvation, as compute pipelines were forced to pull unnecessary data across the network. To resolve this bottleneck, Meta is fundamentally restructuring the storage tier so that the storage layer itself understands the data it serves. Rather than relying on rigid, fixed schemas, the new infrastructure pulls only the exact columns and time ranges required, delivering data in flexible formats designed to feed modern architectures efficiently.

- **Opaque Blobs vs. Intelligent Storage**: Replaced blind blob retrieval with a storage layer that inspects and interprets data natively to end GPU starvation. - **Selective Retrieval**: The system dynamically fetches targeted columns and specific temporal windows instead of entire user histories. - **Deep-Dive Session**: The engineering mechanics of this architecture are explored further in the conference's dedicated sequence storage session.

[10:48](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=648s) The Next Era of Recommender Systems: From Pattern Matching to Intent Reasoning

Recommender systems represent the heartbeat of Meta's product ecosystem, connecting, entertaining, and serving billions of people daily while driving meaningful double-digit revenue growth year over year. Yet for nearly two decades, the core philosophy behind recommender systems has remained virtually unchanged: observe user behavior, extract statistical patterns, and predict the next action. While these systems grew increasingly sophisticated, they remained confined to pattern matching. Large language models shatter that paradigm. For the first time, recommender systems can move beyond correlating behavioral signals to actively reasoning about human intent—understanding who a person is rather than merely what they clicked.

This architectural evolution directly addresses a profound consumer demand: 42% of Instagram users report that they want to change the algorithm, not merely tweak a minor setting, but fundamentally alter what surfaces in their feeds. Meeting this expectation requires abandoning the passive model where the algorithm is "something that happened to you" and replacing it with an experience directed by the user. The North Star is fully conversational recommendations powered by chain-of-thought reasoning, balancing an individual's historical preferences against their real-time context. Under this paradigm, identical queries yield vastly different results tailored to intent: a query for "soccer" returns entertaining highlights for a casual fan, but serves structured training drills to a competitive club athlete.

```mermaid flowchart TD accTitle: Evolution of Recommendation Systems accDescr: Diagram contrasting traditional pattern-matching recommendation systems with LLM-based intent reasoning systems.

subgraph Legacy["Legacy Paradigm (Past 2 Decades)"] A["Observed Behavior & Clicks"] --> B["Statistical Pattern Matching"] B --> C["Passive Ranking (Algorithm happens to you)"] end

subgraph Modern["Modern LLM Paradigm"] D["User Directives & Natural Language"] --> E["Intent Reasoning & Chain-of-Thought"] F["Real-Time Context & History"] --> E E --> G["Conversational Recommendation (User directs algorithm)"] end ```

[12:47](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=767s) Closing the Loop: The AI-Driven Product Flywheel

Bringing these architectural threads together reveals that storage optimization, agentic workflows, and reasoning models are not isolated developments—they represent a total transformation of Meta's core product loop (Consumer, Product, Data, and Recommendations) serving 3.5 billion people daily. Software agents are no longer external tools used merely to build the loop; they are becoming active "citizens" inside it, analyzing data, running discoveries, and operating alongside human engineers.

Concurrently, data infrastructure is transitioning from slow batch pipelines and opaque blobs into intelligent, real-time storage layers capable of satisfying the extreme freshness, scale, and flexibility demands imposed by reasoning models. This feedback cycle operates as an accelerating flywheel rather than linear progress: autonomous agents streamline data accessibility, enriched data enables deeper model reasoning, and the demands of reasoning push agent capabilities and infrastructure forward once again.

```mermaid flowchart LR accTitle: The AI Product Flywheel accDescr: The self-reinforcing flywheel linking Agents, Data Infrastructure, and Reasoning Models.

Agents["Autonomous Agents<br/>(Citizens in the loop)"] -->|"Increases accessibility"| Data["Intelligent Data Layer<br/>(Real-time, flexible)"] Data -->|"Powers deep context"| Reasoning["Reasoning Models<br/>(Intent-driven recommendations)"] Reasoning -->|"Drives new demands & scale"| Agents ```

[13:58](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=838s) Internal Agent Deployments and the Autonomous Future at Scale

Within Meta, autonomous agents are already operating in production across several experimental and enterprise frameworks:

- **Domain Task Agents**: Specialized agents embedded across engineering organizations to execute code generation, operational troubleshooting, and deep data analysis. - **Clowntown**: A multi-agent orchestration framework designed for long-horizon autonomy. Engineers deploy multi-week agents that make semi-autonomous system decisions without requiring a human in the loop, allowing projects to progress even while team members are away on vacation. - **Malbook**: An experimental agent-to-agent social network where autonomous systems interact, communicate, and negotiate directly with one another while human engineers act as "flies on the wall" observing synthetic social dynamics. - **The Frontier Loop**: An executive-level synthesis engine that aggregates complex, high-context documents, project updates, and organizational communications, converting raw operational noise into synthesized briefs primed for leadership decision-making.

A decade ago, the @Scale conference originated around the challenge of architecting systems capable of supporting a billion people. The mandate for the coming era centers on scaling infrastructure to support billions of humans alongside millions of superintelligent agents operating concurrently. In this emerging reality, machine learning models will no longer depend strictly on manual human training cycles; instead, autonomous agents will train, continuously tune, and adapt models in real time. This co-creation between humans and agents marks an inflection point where engineering focus shifts away from incremental 10% gains toward inventing entire categories of software that do not yet exist.

## Notable quotes > [00:05](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=5s) humans transition from being 'in the loop' to 'over the loop.' — Barak Yagour

> [04:10](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=250s) Historically, one engineer equated to one unit of load. Today, a single engineer can spin up dozens or hundreds of agents that spawn sub-agents, enabling an organization of 1,000 engineers to generate the infrastructure load of 100,000 users overnight. — Barak Yagour

> [04:10](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=250s) An agent does not hold an employee badge, it was not provisioned as a static service by a production engineer, and it cannot be categorized as a dumb batch job because it makes active, dynamic decisions. — Barak Yagour

> [06:03](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=363s) autonomy without governance is chaos — Komal

> [08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) reasoning is hungry — Barak Yagour

> [10:48](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=648s) 42% of Instagram users report that they want to change the algorithm, not merely tweak a minor setting, but fundamentally alter what surfaces in their feeds. — Barak Yagour

> [10:48](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=648s) something that happened to you — Barak Yagour

> [10:48](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=648s) The North Star is fully conversational recommendations powered by chain-of-thought reasoning, balancing an individual's historical preferences against their real-time context. — Barak Yagour

> [13:58](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=838s) flies on the wall — Barak Yagour

> [13:58](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=838s) The mandate for the coming era centers on scaling infrastructure to support billions of humans alongside millions of superintelligent agents operating concurrently. — Barak Yagour

## Mentioned Barak Yagour, Meta, @Scale Conference, America's Workforce Academy, Instagram, GitHub Copilot, Dinkar, Komal, Alex, Snorkel, Chinchilla, Lexis, Clowntown, Malbook, The Frontier Loop

## Video description Hear from Barak Yagour, VP of Engineering at Meta, about the future of AI.

Want to learn more about the @Scale Conference? Visit our website here: atscaleconference.com

[Source](https://www.youtube.com/watch?v=L0BSZb7ujC8)

中文简报

AI智能体令基础设施承压，Meta全面重构数据架构 自主AI智能体正在突破传统的容量、身份和速度极限，迫使Meta用能够承受每秒5亿次查询的实时流式架构替代传统的批处理存储。 Barak Yagour概述了自主软件智能体与深度推理模型如何打破传统基础设施关于容量、身份和部署速度的假设。为了支持这些工作负载，Meta正在全面重构其存储和数据流水线，从不透明的大型二进制对象（blob）和批处理转向智能、实时的流式架构，该架构能够支持每秒5亿次查询以及每秒1 PB的吞吐量。与此同时，推荐系统正从被动的统计模式匹配演变为由大语言模型驱动的对话式意图推理，从而主动理解用户目标。这些组件融合为一个自我强化的AI产品飞轮，其中Clowntown和Malbook等内部多智能体编排框架，正推动人类从“人在回路”（in the loop）转向“人在回路之上”（over the loop）进行治理。 去年自动化互联网流量占比达到51%，扩张速度是人类流量的八倍，自主智能体正在转变为基础设施中的一等公民。 (4:10) 现代基础模型每个参数消耗的token数量是Chinchilla最优阈值的20倍，催生了对支撑5亿QPS以及每秒1 PB训练读取吞吐量架构的需求。 (8:41) Meta正将其推荐系统从历经二十年的统计模式匹配转向LLM驱动的意图推理，以应对42%的Instagram用户希望直接改变算法的调查数据。 (10:48) Clowntown和Malbook等内部部署展示了无需人工介入即可运行的长达数周的自主工作流以及智能体之间的协商。 (13:58)

Reading summary

At Meta's @Scale conference, Barak Yagour argued that scaling AI requires restructuring foundational infrastructure because autonomous agents are shifting from development assistants to primary consumers that challenge legacy capacity, identity, and velocity models. Agents as Infrastructure Citizens Breaking Human-Centric Constraints (4:10) Software infrastructure has historically assumed human operating constraints, characterized by discrete, session-based traffic. With automated internet traffic reaching 51% and growing eight times faster than human traffic, internal autonomous agents are becoming first-class consumers that disrupt capacity, identity, and velocity. A single engineer can now direct dozens or hundreds of sub-agents, allowing an organization of 1,000 engineers to place demands equivalent to 100,000 users onto internal systems. This continuous operation breaks existing systems across three fronts. Identity models fail because agents lack human employee badges, static service provisioning, or static batch patterns. Concurrently, software velocity creates severe bottlenecks: while tools like GitHub Copilot generate 46% of code for its users in seconds, legacy CI/CD pipelines, build systems, and deployment verifications remain anchored to human review cycles. Democratizing Data Apps and the Challenge of Agent Governance (6:03) Following Meta's release of agentic data application tooling in February, 63% of internal dashboards were created through data applications within three months, driving a 30-fold increase in agentic queries. While this democratization lets non-specialists and autonomous agents construct tailored reporting pipelines quickly, it bypasses human data analysts who previously served as domain gatekeepers for contextual, legal, and privacy constraints. Meta speakers emphasized that scaling autonomous agents requires strict governance and observability frameworks. Sessions highlighted transitioning from rigid data interfaces to integrated data apps, instituting safety guardrails under the principle that autonomy without governance results in chaos, and establishing rigorous benchmarks to monitor agents as they evolve from advisory roles into direct execution. Re-engineering Data and Storage Layers for Intent-Based Reasoning (8:41) Modern reasoning architectures are replacing pattern matching based on sparse feature vectors with full chronological user histories across surfaces. Foundation models are being trained on 20 times more data per parameter than the thresholds suggested by the 2022 Chinchilla scaling laws, with the training data-to-parameter ratio expanding 3.1-fold annually since 2022. To avoid real-time delays, Meta is replacing 24-hour batch ETL with streaming architectures designed to sustain targets of 500 million queries per second and 1 petabyte per second of training data read throughput. Supporting this throughput requires overhauling storage architectures away from opaque blobs, which historically caused network overhead and GPU starvation. The redesigned storage tier inspects data natively, dynamically retrieving only targeted columns and specific temporal windows instead of fetching unbounded user histories. This selective retrieval feeds intent-reasoning recommender systems, aligning with data indicating that 42% of Instagram users want to actively direct and modify their recommendation algorithms. Internal Agent Deployments and Long-Horizon Autonomy (13:58) Meta has deployed autonomous agent frameworks across several production and research environments. These implementations include domain-specific operational agents, Clowntown—a multi-agent framework managing multi-week autonomous workflows without humans in the loop—and Malbook, an experimental network for tracking synthetic agent-to-agent negotiations. At the executive level, the Frontier Loop engine parses and synthesizes high-context enterprise documentation into operational briefs for decision-makers. Meta projects that future engineering infrastructure must scale to accommodate billions of human users alongside millions of autonomous agents. A key open operational challenge remains shifting engineering workflows from incremental human-driven coding cycles toward systems where agents autonomously train, tune, and iterate models within rigorous governance boundaries.

Detailed notes — zh

## Background 该视频由 YouTube 频道 @Scale 于 2026 年 6 月 19 日发布，内容为 Meta 工程副总裁 Barak Yagour 在 @Scale 大会上的主题演讲。根据页面介绍，演讲聚焦于人工智能的未来，深入探讨了支撑下一代系统所需的技术演进与架构变革。 ## TL;DR Barak Yagour 在演讲中指出，AI 正在从人类主导的编码工具演进为具备高度自主性的 Agent 革命，迫使工程师从“在闭环内”转向“在闭环上”的全局监督。与此同时，不断激增的机器流量与数据规模对现有基础设施的容量、身份认证和交付速度带来了巨大冲击。为了应对推理模型对高质量数据的极度饥渴，Meta 正在重构其底层存储与实时流架构，并将推荐系统的核心逻辑从被动统计模式匹配升级为理解用户意图、支持对话式交互的自主推理体系。此外，Meta 内部已在前沿测试多 Agent 编排框架及自主决策系统，全面构筑服务数十亿人类与数百万超智能体的下一代基础设施。 ## Atmosphere & style 整场演讲充满前瞻性与技术宏大感，语调沉着而富有紧迫感，展现出顶尖科技公司身处技术剧变风暴眼时的清醒洞察。讲者没有停留在空泛的概念炒作，而是以真实、硬核的基础设施断裂点与工程瓶颈为抓手，引发观众对系统可承受极限的深度思索。整场分享节奏紧凑、视野开阔，既有对现实治理混乱的严峻剖析，又描绘出 Agent 自主协作与对话式推荐的颠覆性未来，带给听众一种见证时代底层范式重构的震撼与兴奋感。 ## Bulletin - [04:07](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=247s) 全网自动化机器流量占比已达到 51% 并成为互联网主体，而在 Meta 内部，Agent 正在以每个季度跨数量级的速度消耗 Token，并引发技术底座在容量、身份鉴权与交付速度三大维度上的系统性断裂。 - [06:03](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=363s) Meta 内部 63% 的新仪表板已由 Agent 数据应用自动生成，直接推动查询量暴增 30 倍，促使基础设施朝着 5 亿 QPS 及 1 PB/s 训练数据读取吞吐量的实时流与语义存储架构演进。 - [11:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=701s) 大语言模型的推理能力正颠覆推荐系统过去二十年的模式匹配逻辑，Meta 确立了以思维链（CoT）为核心的全对话式推荐愿景，让用户从被动接受算法投喂转向主动掌控算法。 - [13:59](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=839s) Meta 内部正在试验包括多智能体自主编排框架 Clowntown、智能体社交网络 Malbook 以及高管决策支持系统 Frontier Loop 在内的前沿 Agent 体系，探索无人类在环的自主协同运作。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=5s) 算力饥渴完全没有放缓迹象，Meta 投资 1.15 亿美元设立美国劳动力学院以填补 AI 基础设施人才缺口，支撑日活 35 亿用户的超大规模产品矩阵。 - [02:01](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=121s) 工程师角色发生根本性范式转移，从具体编码与模型调试的“在闭环内”（in the loop），升维至全局监督与战略引导的“在闭环上”（over the loop）。 - [04:07](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=247s) 网络自动化流量（51%）超越人类流量，一个千人工程师团队通过多层裂变 Agent 即可在内网制造相当于十万用户的并发负载，冲击传统容量假设。 - [04:07](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=247s) 现有身份鉴权体系难以定义兼具动态决策权的 Agent，它既非自然人也非标准微服务或离线批处理任务，带来了全新的系统治理难题。 - [04:07](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=247s) 开发瓶颈从代码编写转移至交付流水线，即使 Agent 几秒内即可生成代码，后续的编译、测试与 CI/CD 验证仍受物理限制。 - [06:03](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=363s) 数据民主化伴随可信危机，非专业人员与 Agent 自主构建数据应用需匹配严格的治理、可观测性与溯源机制，避免陷入无序混乱。 - [08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) 前沿模型的训练数据量与参数比值已达 Chinchilla 缩放法则推荐值的 20 倍，且模型推理需要长时间维度的全量交互历史而非稀疏摘要特征。 - [08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) 为支撑高达 5 亿 QPS 和 1 PB/s 的吞吐需求，传统离线批处理 ETL 正在被高新鲜度的实时流数据骨干所取代。 - [10:22](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=622s) 底层存储正在进行语义智能化重构，摒弃不透明的二进制 Blob 存储以消除数据过度拉取和 GPU 饥饿，支持按需读取列与灵活 Schema 传输。 - [10:22](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=622s) 推荐系统从过去二十年被动的统计模式匹配，演进为借助 LLM 真正理解用户意图与用户画像的全新交互形态。 - [11:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=701s) Meta 设立对话式推荐的北极星愿景，利用思维链（CoT）推理实时动态调整分发，响应 42% Instagram 用户想要主动掌控算法的意愿。 - [12:47](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=767s) 智能体深度参与、数据实时语义化与推荐系统推理化三者相互加速，重塑服务 35 亿人的核心产品循环飞轮。 - [13:59](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=839s) Meta 内部积极探索自主 Agent 框架（如 Clowntown 和 Malbook），基础设施的核心命题已演变为同时承载数十亿人类与数百万个超智能体的协同并发运转。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=5s) 算力饥渴与 Agent 革命的开端 在过去十年的 @Scale 大会上，系统构建的讨论焦点始终围绕着“行星级规模”（planet scale）展开。然而今年的氛围截然不同，技术变革的步伐呈现出爆发式增长，在短短一年内将曾经不可想象的事情变为日常现实。AI 引发的对计算资源的极端饥渴完全没有放缓的迹象，全行业正在以史无前例的规模兴建数据中心。为了应对激增的人才缺口，Meta 刚刚完成了一笔 1.15 亿美元的初始投资，用于建设“美国劳动力学院”（America's Workforce Academy），专为这一领域培养未来的专业人才。 在整个软件行业，AI 生成代码的比例屡创新高，Agent（智能体）消耗的 Token 数量更是呈爆发式增长，环比每个季度都出现数量级的跃升。但在编码之外，技术架构底层的根本逻辑也在被重构。每天有 35 亿人使用 Meta 旗下的各类产品，支撑如此庞大应用家族的复杂度是极其惊人的——从底层机器、数据中心、能源调度、硬件架构，到上层软件、产品生态、数据完整性（integrity）与隐私合规，背后凝结着无数工程师的心血。而当前正在席卷而来的 Agent 革命，正以前所未有的姿态渗透进这一体系，在加速创新步伐的同时，也颠覆了传统的运作规则。 - Meta 向“美国劳动力学院”注资 1.15 亿美元，旨在解决 AI 基础设施领域的人才短缺瓶颈。 - Meta 旗下产品矩阵目前每日触达的活跃用户规模达到 35 亿人。 - Agentic Token 的消耗量正在以季度为单位实现跨数量级的指数级增长。 --- ### [02:01](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=121s) 核心产品循环的演进：从 In The Loop 到 Over The Loop 以一位普通的 Instagram 用户为例，产品开发世界中存在着一个经典的核心循环：用户在各个前端交互界面产生大量行为数据，这些数据经由大语言模型、视觉模型与语音模型支撑的推荐系统进行处理，并与业务决策共同作用，动态调整该用户看到的内容流，从而形成完整的业务闭环。在 Agent 出现之前，这套闭环中的所有环节——从设计、构建到运维和迭代优化——全部完全依赖于专业的人类工程师。 这一格局在今年发生了剧烈演变。编程 Agent 不仅能够胜任常规编码，甚至已经进化到能够真正“一次性生成（one-shot）大段核心代码”的水平；与此同时，Agent 开始展现出管理数据、分析业务报表、构思机器学习实验甚至自主训练新模型的能力。接下来技术探索的前沿，在于人类愿意赋予 Agent 多大程度的自主权。在适当的安全护栏（guardrails）保护下，Agent 将能够连续数小时、数天甚至数周完全自主工作，宛如一位“极其聪明、执行力极强且在获得上下文后能独立做决策的员工”。在这种全新范式下，人类工程师的角色正在发生质的转变，从原本深陷在具体执行层面的“在闭环内”（in the loop），彻底升维为全局监督与战略引导的“在闭环上”（over the loop）。 ```mermaid flowchart TD accTitle: 核心产品循环向自主 Agent 的架构演进 accDescr: 展示用户行为数据在推荐系统中的流转，以及人类角色从在环操作转变为在环监督的过程。 subgraph UserInteraction["用户与产品交互层"] A["终端用户 (如 Instagram)"] -->|"产生交互行为与数据"| B["产品核心界面与业务逻辑"] end subgraph ModelEngine["模型与推荐引擎"] B --> C["数据处理与模型训练\n(LLM / 视觉 / 语音)"] C --> D["动态推荐系统与业务决策"] D -->|"个性化反馈内容"| A end subgraph AgentShift["研发与运维范式转移"] E["自主 Agent\n(写代码 / 跑实验 / 构建模型)"] -->|"自主执行数小时至数周"| C F["人类工程师与专家"] -->|"从 In the loop 转向 Over the loop\n(设定上下文 / 监督引导)"| E end ``` --- ### [04:07](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=247s) 基础设施三大断裂：容量、身份与交付速度 互联网在去年迎来了一个历史性拐点：全网自动化流量（机器流量）正式超越了人类流量，占比达到 51%，并且其扩张速度是人类流量的 8 倍。而在 Meta 内部基础设施中，这种流量倒挂同样正在发生——Agent 已经不再是单纯交付给用户的一项功能，而是变成了基础设施的第一类（First-class）消费者。 过去二十年里，整套技术基础设施的底层假设都是针对人类行为模式量身定制的：会话具有规律性、流量相对可预测，且存在天然的速率限制（人类需要睡眠、进餐、容易分心）。然而 Agent 不知疲倦，当它们接管系统交互时，现有的技术底座瞬间在三个关键维度上遭遇崩塌： 1. **容量（Capacity）的断裂**：以往一名工程师在系统中往往只代表一个单位的负载；而在今天，一名工程师可以随意启动数十个甚至上百个 Agent，而每个 Agent 又能进一步裂变衍生出子 Agent（sub-agents）。一个千人规模的工程团队，一夜之间就能在内网制造出相当于十万名真实用户并发的恐怖负载。 2. **身份（Identity）的断裂**：Agent 既不是拥有员工工牌的自然人，也不是生产工程师按流程部署的标准线上服务，更不是传统的离线批处理任务（batch job）——它拥有动态的决策权。面对这种模糊的中间态实体，现存的身份鉴权系统难以适配，全行业仍在艰难探索治理方案。 3. **交付速度（Velocity）的断裂**：例如 GitHub Copilot 已经为用户编写了 46% 的代码。虽然 Agent 可以在数秒内吐出代码，但后续的编译、单元测试、CI/CD 持续集成、灰度发布与线上监控流水线，并不会仅仅因为“作者是一台机器”而自动加速。瓶颈从代码编写转移到了系统集成与验证上。 基础设施的核心拷问由此诞生：“当 Agent 全面接管工作时，现有的技术架构到底能不能扛得住？” --- ### [06:03](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=363s) 数据民主化背后的治理风险与可信危机 数据是支撑商业决策、核心产品、推荐系统以及下一代大模型训练的真正命脉。Meta 于今年二月正式推出了构建“Agent 数据应用”（Agentic data apps）的能力。这一工具上线仅三个月，公司内部新发布的仪表板（Dashboards）中就有 63% 是完全通过数据应用生成的，直接推动了 Agent 数据查询量实现高达 30 倍的迅猛增长。 这种将数据获取门槛平民化、让全员能够基于数据敏捷决策的趋势表面上看极具价值，但其背后潜藏着极其严重的治理隐患。过去，从庞杂数据中提炼商业洞察高度依赖人类专家，这些专家不仅负责数据清洗与指标建模，还深刻理解数据背后的业务逻辑、隐私红线与合规风险，本身就是一道严密的“质量防线”。当这道防线被移除，海量完全没有上下文认知、也不懂法律法规的 Agent 或非技术人员开始随意发起查询，一系列尖锐的问责难题便随之浮现： - 它们给出的数据和分析洞察真的值得信赖吗？ - 它们是否有权限调阅并给出这些答案？ - 它们是否遵守了人类员工被要求遵循的合规指南？ - 当下游业务因此做出错误决策时，系统是否有能力进行精细化的溯源与排错（backtrack）？ 正如后续演讲嘉宾所阐释的核心观点：非专业人员与 Agent 自主构建定制化数据产品虽然打破了传统工具的桎梏，但“没有治理的自主就是混乱”（autonomy without governance is chaos）。治理、安全、隐私与合规不再是锦上添花的附属品，而是支撑这一整套架构在超大规模下平稳运行的前提；同时，针对做出具体决策而非仅仅提供建议的 Agent，必须引入全新的可观测性（Observability）、评估基准（Evals）与严苛的数据质量校验机制。 --- ### [08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) 推理时代的算力与数据渴求：重构存储与实时流架构 伴随着治理困境的，是数据总量与数据复杂度的几何级暴增。自 2022 年以来，开源权重（open weight）大语言模型中“训练数据量与激活参数量之比”正以每年 3.1 倍的速度飙升。近期部分先进模型的训练数据量与参数比值，甚至达到了 2022 年 Chinchilla 缩放定律（Scaling Laws）所给出的最优推荐值的 20 倍。 更重要的是，大模型对数据的依赖正在从单纯的规模扩张转向形态的本质改变——“推理是饥渴的”（Reasoning is hungry）。传统的模式匹配算法往往仅依赖稀疏特征向量（sparse feature vectors）和高度浓缩的行为摘要信号；但如果要让模型具备真正的推理能力，它需要获取用户跨不同产品界面、贯穿长时间维度的完整交互历史，而不是过滤后的片段摘要。这种高维度数据正在被全面用于训练 Meta 的下一代生成式模型，并显著提升了广告投放和自然内容分发的核心表现。 为了支撑这一极其繁重的数据吞吐需求，Meta 正在全力构建一套能够承载“每秒 5 亿次查询”（500 million QPS）以及“每秒 1 PB 训练数据读取吞吐量”（1 petabyte per second）的全新数据基础设施。这直接催生了底层架构的范式转变： - **数据新鲜度（Freshness）成为第一优先级**：当推荐与排序算法需要实时推理用户当前的意图时，系统根本无法忍受运行周期长达 24 小时的传统离线 ETL 批处理流水线。实时流数据（real-time streaming data）正迅速取而代之，成为排序推荐系统的全新骨干支柱。 - **存储架构智能化**：仅靠提升带宽和降低延迟已无法解决问题，底层的存储系统本身必须被赋予理解数据与智能调度的能力。 - 自 2022 年起，开源 LLM 的训练数据与激活参数比值每年以 3.1 倍的速度增长。 - 近期前沿模型的参数平均数据配比达到了 Chinchilla 缩放法则最优值的 20 倍。 - Meta 正在打造的数据底层吞吐目标高达 5 亿 QPS，训练数据读取带宽达 1 PB/s。 ### [10:22](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=622s) 存储层革新与推荐系统的范式转移 在传统的数据架构中，用户数据通常以不透明的二进制大对象（opaque blobs）形式存储，底层存储系统对其中的具体内容毫无感知。这种机制导致了极严重的数据过度获取（over-fetching）和 GPU 饥饿（GPU starvation）现象。Meta 目前正推动存储层自身的智能化改造，让底层存储能够直接理解其所服务的数据语义，从而仅按需拉取相关的列和时间范围，并采用具灵活性而非固定 Schema 的格式进行高效传输。 推荐系统不仅是 Meta 产品矩阵的关键组成部分，更是整个商业变现的基石。推荐系统每天连接、娱乐和服务全球数十亿用户，并持续推动着年对年两位数（double-digit）的显著营收增长。然而，在过去近二十年里，推荐系统的核心逻辑始终未变，即“观察行为、寻找模式、预测下一步动作”——哪怕算法变得空前精细，其本质依然只是更高阶的模式匹配（pattern matching）。 大语言模型（LLM）彻底打破了这一旧范式。推荐系统第一次不再仅仅停留在行为信号的关联上，而是能够真正理解用户的意图（intent）。系统不仅知道用户“点击了什么”，更能推断“用户是谁”，这从根本上颠覆了算法能为用户提供的内容，也对底层数据层提出了全新的要求。 - 传统存储将用户数据作为不透明 Blob 处理，导致过度拉取和 GPU 饥饿问题严重。 - 智能存储层转向理解数据语义，支持灵活 Schema，仅检索特定列与所需时间窗口。 - 推荐系统是 Meta 核心业务，驱动着两位数的年对年营收增长。 - 近二十年的推荐算法局限于“观察-匹配-预测”，而 LLM 实现了向“理解意图与用户本身”的跨越。 --- ### [11:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=701s) 从被动投喂到自主掌控：对话式推荐的北极星愿景 一项调研数据显示，高达 42% 的 Instagram 用户明确表示希望能够改变算法。这并非希望简单调整某项设置，而是渴望从根本上改变呈现在自己眼前的内容。这已经不是普通的“功能需求”（feature request），而是倒逼团队重新构想整个推荐系统。 Meta 设定的“北极星”（North Star）目标，是将算法从“被动降临在你身上的规则”转变为“由你主动指挥的工具”。未来的系统将支持全对话式推荐（conversational recommendations）：用户可以直接告诉系统自己想多看什么、看腻了什么，或者此刻当下的心情如何。推荐系统不再依赖关键词匹配，而是利用思维链（Chain of Thought, CoT）进行深度推理，结合用户的历史习惯与当前上下文环境，实现实时自适应调整。以搜索“足球（soccer）”为例，当系统理解了用户意图后，针对寻找赛事集锦的普通球迷与寻找训练教案的俱乐部运动员，将自动推理并交付截然不同的内容。 ```mermaid flowchart TD accTitle: 对话式推荐系统工作流 accDescr: 展示从用户意图表达到大模型思维链推理，再到针对不同用户分发个性化结果的流程。 subgraph UserIntent ["用户交互层"] A["用户意图输入<br/>(自然语言表达 / 即时心境)"] end subgraph LLMReasoning ["意图推理核心"] B["思维链推理 (Chain of Thought)<br/>权衡历史行为 vs 即时上下文"] end subgraph TailoredResults ["差异化分发"] C["普通球迷<br/>(比赛精彩集锦)"] D["俱乐部运动员<br/>(专业训练教案)"] end A --> B B --> C B --> D ``` --- ### [12:47](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=767s) 飞轮效应：核心产品循环的全面重塑 将上述技术演进整合起来看，它们绝非孤立的创新，而是对 Meta 每日服务 35 亿人口的“消费者-产品-数据-推荐”核心产品循环（product loop）的全面重塑。 1. **智能体（Agents）角色转变**：从仅作为幕后构建循环的工具，转变为深度参与产品循环各阶段的“公民”（citizens），与人类协同进行系统运营、发现洞察、分析数据和创造内容。 2. **数据基础设施重构**：从传统的批处理、不透明存储迈向实时数据流与具备语义理解的智能存储，以支撑推理模型对数据新鲜度、扩展性和灵活性的极高需求。 3. **推荐系统本质演进**：从被动的统计模式匹配，演变为可理解意图、可由用户交互指挥的自主推理。 这三者之间形成了一个相互加速的自增强飞轮（flywheel）：智能体让数据调用更加便捷，高质量数据赋能更强大的推理能力，而推理模型的升级又对智能体和数据基础设施提出更高诉求，推动技术实现非线性的指数级发展。 --- ### [13:59](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=839s) Meta 内部的多智能体探索与未来图景 智能体已经在 Meta 全球业务的各个角落落地生根，除覆盖编码、数据分析、运维等特定领域的任务型智能体（task agents）外，Meta 还在内部积极试验多个前沿项目： - **Clowntown**：一个多智能体编排框架，致力于实现高度的“智能体自主性”（agentic autonomy）。它甚至允许工程师在离岗休假数周的情况下，交由智能体在无人类介入（no human in the loop）的场景下自主进行半自主决策与系统维护。 - **Malbook**：正在探索的“智能体对智能体”（agent-to-agent）社交网络。人类在其中仅作为“墙上的苍蝇”（flies on the wall，旁观者），观察和研究智能体彼此之间是如何自发沟通、协调与博弈谈判的。 - **Frontier Loop**：旨在加速高管层决策的项目。该系统自动汇聚并整合来自全公司的高上下文文档与最新动态，自主提炼出可直接用于最终决策的综合上下文。 回顾十余年前，@Scale 大会聚焦于如何构建能够服务 10 亿级用户的规模化系统。而当下的核心命题转变为：如何构建同时承载数十亿真实人类与数百万个超智能体（superintelligent agents）协同并发运转的基础设施。未来的模型训练甚至不再需要人类干预，而是由智能体实现自我训练、持续微调与实时自适应。这种人机协同带来的将不再是 10% 的边际递增，而是孕育出当下完全无法想象的全新产品形态。 ## Notable quotes > [02:01](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=121s) 从原本深陷在具体执行层面的“在闭环内”（in the loop），彻底升维为全局监督与战略引导的“在闭环上”（over the loop）。 — Barak Yagour > [02:01](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=121s) 宛如一位“极其聪明、执行力极强且在获得上下文后能独立做决策的员工”。 — Barak Yagour > [04:07](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=247s) Agent 已经不再是单纯交付给用户的一项功能，而是变成了基础设施的第一类（First-class）消费者。 — Barak Yagour > [04:07](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=247s) 当 Agent 全面接管工作时，现有的技术架构到底能不能扛得住？ — Barak Yagour > [06:03](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=363s) 没有治理的自主就是混乱（autonomy without governance is chaos）。 — Barak Yagour > [08:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=521s) 推理是饥渴的（Reasoning is hungry）。 — Barak Yagour > [10:22](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=622s) 系统不仅知道用户“点击了什么”，更能推断“用户是谁”，这从根本上颠覆了算法能为用户提供的内容。 — Barak Yagour > [11:41](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=701s) 将算法从“被动降临在你身上的规则”转变为“由你主动指挥的工具”。 — Barak Yagour > [13:59](https://www.youtube.com/watch?v=L0BSZb7ujC8&t=839s) 如何构建同时承载数十亿真实人类与数百万个超智能体（superintelligent agents）协同并发运转的基础设施。 — Barak Yagour ## Mentioned Barak Yagour, Meta, Instagram, GitHub Copilot, @Scale Conference, America's Workforce Academy, Chinchilla, Clowntown, Malbook, Frontier Loop ## Video description Hear from Barak Yagour, VP of Engineering at Meta, about the future of AI. Want to learn more about the @Scale Conference? Visit our website here: atscaleconference.com [Source](https://www.youtube.com/watch?v=L0BSZb7ujC8)

Source quotes

[2:01] Barak Yagour: 从原本深陷在具体执行层面的“在闭环内”（in the loop），彻底升维为全局监督与战略引导的“在闭环上”（over the loop）。 [2:01] Barak Yagour: 宛如一位“极其聪明、执行力极强且在获得上下文后能独立做决策的员工”。 [4:07] Barak Yagour: Agent 已经不再是单纯交付给用户的一项功能，而是变成了基础设施的第一类（First-class）消费者。 [4:07] Barak Yagour: 当 Agent 全面接管工作时，现有的技术架构到底能不能扛得住？ [6:03] Barak Yagour: 没有治理的自主就是混乱（autonomy without governance is chaos）。 [8:41] Barak Yagour: 推理是饥渴的（Reasoning is hungry）。 [10:22] Barak Yagour: 系统不仅知道用户“点击了什么”，更能推断“用户是谁”，这从根本上颠覆了算法能为用户提供的内容。 [11:41] Barak Yagour: 将算法从“被动降临在你身上的规则”转变为“由你主动指挥的工具”。 [13:59] Barak Yagour: 如何构建同时承载数十亿真实人类与数百万个超智能体（superintelligent agents）协同并发运转的基础设施。

Source quotes

[0:05] Barak Yagour: humans transition from being 'in the loop' to 'over the loop.' [4:10] Barak Yagour: Historically, one engineer equated to one unit of load. Today, a single engineer can spin up dozens or hundreds of agents that spawn sub-agents, enabling an organization of 1,000 engineers to generate the infrastructure load of 100,000 users overnight. [4:10] Barak Yagour: An agent does not hold an employee badge, it was not provisioned as a static service by a production engineer, and it cannot be categorized as a dumb batch job because it makes active, dynamic decisions. [6:03] Komal: autonomy without governance is chaos [8:41] Barak Yagour: reasoning is hungry [10:48] Barak Yagour: 42% of Instagram users report that they want to change the algorithm, not merely tweak a minor setting, but fundamentally alter what surfaces in their feeds. [10:48] Barak Yagour: something that happened to you [10:48] Barak Yagour: The North Star is fully conversational recommendations powered by chain-of-thought reasoning, balancing an individual's historical preferences against their real-time context. [13:58] Barak Yagour: flies on the wall [13:58] Barak Yagour: The mandate for the coming era centers on scaling infrastructure to support billions of humans alongside millions of superintelligent agents operating concurrently.

Original transcript

[0:05] You know, when you come to At Scale [0:06] Conference, you expect to see huge scale [0:08] numbers. [0:09] Every year over the past decade, we've [0:11] been talking about building systems at [0:13] planet scale. And we will do that today [0:15] as well. [0:16] But this year feels truly different. The [0:18] exceptional pace of change is showing [0:20] astonishing results, enabling what was [0:22] unimaginable just a year ago. [0:25] I want to take you on a journey to peek [0:27] inside a fast-evolving world. [0:29] So, over the next 20 minutes, I want to [0:31] introduce you into what is becoming our [0:33] day-to-day reality. And hopefully send [0:35] you into the rest of the day with [0:37] thoughts and ideas about the revolution [0:39] brewing in the world of AI and data. [0:41] And how it is going to enable new [0:43] innovation all around us. [0:46] And the place to start is the crazy [0:47] appetite for compute that AI has [0:49] unleashed. [0:51] Which, frankly, doesn't seem to be [0:52] slowing down anytime soon. [0:55] Everyone in the industry is building [0:56] more data centers than ever before. At [0:59] Scale, we have never seen. [1:01] We at Meta just made an initial [1:03] investment of $115 million towards [1:06] building America's Workforce Academy [1:08] in order to train the future talent into [1:10] this space to address the growing talent [1:12] need. [1:14] Across the industry, the percentage of [1:16] code generated by AI is hitting record [1:18] numbers. Agentic token consumption is [1:21] through the roof and growing orders of [1:23] magnitude quarter over a quarter. [1:26] But coding is not the only thing that is [1:27] changing. [1:29] Every day, 3 and 1/2 billion people [1:32] engage with one of Meta's products. [1:34] There is obvious complexity in [1:36] supporting a family of apps at that [1:37] scale. [1:38] In terms of machines, data center, [1:40] energy, hardware, software, product, [1:43] integrity, privacy, and the people [1:46] behind all of this that make this magic [1:47] happen. [1:49] Let's peek inside the curtain and talk [1:51] about how the agentic revolution is [1:52] bringing about unprecedented change. [1:55] Accelerating some things, upending [1:57] others [1:58] and creating the conditions for new [2:00] innovation. [2:01] Let's look at our core product loop in [2:03] the product development world. [2:06] Let's start with a consumer. [2:08] Say an Instagram user. [2:10] As she engages with the product, the [2:11] core loop is working behind the scenes. [2:14] Let's look at the core component of this [2:16] loop. We start with a core product that [2:18] has the various surfaces that the [2:19] consumer directly interacts with. As she [2:22] engages with the product, she [snorts] [2:24] generates a lot of data. Her product [2:26] experience is powered by the recommender [2:28] systems enabled by large language, [2:30] vision, and speech models. [2:33] These models have been trained on vast [2:35] swaths of relevant data. [2:37] Both the business product decisions and [2:39] the dynamic recommender systems come [2:41] together to influence and empower what [2:43] she sees as the product thereby [2:45] completing the loop. [2:48] In a world before agents, all these [2:50] functions were exclusively designed, [2:52] built, operated, and improved by [2:55] talented folks like yourself. [2:58] We know that agents can code now. [3:00] And in fact this year, the coding agents [3:02] have evolved to a level that they can [3:04] truly one-shot big chunks of code. [3:07] But we also now realize that agents can [3:10] be really good at managing data, [3:12] analyzing reports, [3:14] coming up with ML experimentation, even [3:16] building other models. [3:18] And that is sort of where we are today. [3:21] So where do we go from here? [3:23] The next frontier seems to be about [3:25] expanding the level of autonomy we are [3:27] prepared to give our agents [3:29] such that they can, with proper [3:31] guardrails, [3:32] really start doing work autonomously for [3:34] hours, days, perhaps even weeks. [3:38] Think of it as having a really smart and [3:40] driven employee who is able to make [3:41] decisions when provided with context on [3:43] the problem. [3:45] And us as humans are there to guide and [3:48] supervise the work. [3:50] Obviously, this opens up a whole new set [3:52] of problems. And we will get into those [3:54] today. [3:55] But this is an entirely new and exciting [3:57] paradigm. [3:58] Agents are among us, and we are heading [4:01] for a world where humans move from [4:02] spending time in the loop to spending [4:04] time over the loop. [4:07] I will touch on three threads now. [4:10] Starting with the agents among us as the [4:12] infrastructure citizens. [4:15] And let's first talk about the [4:16] inflection point happening now. [4:19] Last year, we saw automated traffic [4:21] surpass human traffic on the internet. [4:24] 51% [4:26] and it's going eight times faster than [4:28] human traffic. [4:29] That's the world outside. When we look [4:31] inside at our own infrastructure, the [4:33] same inversion is happening. [4:35] Agents aren't a feature we shipped. [4:37] They're becoming first-class [4:38] infrastructure consumers. [4:40] And that changes everything. [4:43] We spent 20 years building [4:45] infrastructure optimized for human [4:47] request patterns. Predictable, [4:50] session-based, with natural rate limits. [4:52] Humans sleep, they eat, they get [4:54] distracted. Agents don't. [4:57] Three things break almost [4:59] simultaneously. [5:01] Capacity break. One engineer used to [5:03] mean one unit of load. Now one engineer [5:06] spawns tens of agents, maybe hundreds of [5:08] agents, each spawning sub agents. Your [5:11] thousand-person org can generate the [5:13] load of a hundred thousand users [5:15] overnight. [5:17] Identity breaks. An agent isn't a human [5:19] user. It doesn't have a badge. It's not [5:22] a service. No production engineer [5:24] deployed it. It's not a batch job. It [5:27] actually makes decisions. [5:29] We are seeing the industry start to [5:30] explore approaches on how to deal with [5:32] this, but there's a lot more to figure [5:34] out. [5:35] And then velocity breaks. [5:37] GitHub Copilot now writes 46% of the [5:40] code for its users. [5:41] That code still needs to be built, [5:43] tested, deployed, and monitored. [5:46] The agent writes the code in seconds, [5:48] but your your CI/CD doesn't speed up [5:51] just because the author is a machine. [5:54] So, here's the question this conference [5:56] is really about. [5:57] Can your infrastructure handle what [5:59] happens when agents do the work? [6:03] One of the key areas where we want to [6:05] see agents having a big impact is data. [6:08] Data is at the heart of the business. It [6:11] feeds business decisions, products, [6:13] recommender systems, LLMs, and next-gen [6:16] models. [6:17] Let's see how the data is transforming. [6:21] In February this year, we launched the [6:22] ability to create agentic data apps. [6:25] And in a short 3-month period, 63% of [6:28] dashboards published in the company were [6:30] created using data apps. [6:33] And as mentioned, we saw 30x growth in [6:35] agentic queries. [6:36] This is fueling some of that growth. [6:39] Democratizing the access to data and [6:40] empowering people to do data-driven [6:42] decision-making is obviously a good [6:44] thing. [6:45] But is that the whole story? [6:48] Let's look at how people leverage data. [6:51] Traditionally, [6:52] extracting answers from data involved [6:54] human experts that both curated the data [6:57] and analyzed it. [6:58] They understood the data as well as the [7:00] privacy and regulatory aspects of it, [7:02] and often optimized data for the use [7:04] case. [7:05] Traditional level of expertise played [7:07] both as a quality control and as someone [7:09] who could help navigate the questions [7:11] and answers. [7:12] What happens when that person is no [7:14] longer in the loop? [7:15] And what happens when there are way more [7:16] questions, be it agents or humans with [7:19] less context or no context at all? [7:22] How can we trust the data and insights [7:24] they shared? [7:25] They don't have the context. They don't [7:27] understand regulations. Were they [7:29] supposed to give us that answer? [7:31] Did they follow the guidance and [7:32] guidelines we expect human employees to [7:34] follow? [7:35] How can we backtrack if something is [7:37] going wrong? [7:39] This becomes a a risk. [7:42] Even though the humans are not going [7:43] away, this inherently complex workflow [7:46] of navigating data is evolving to [7:48] accommodate agents and is becoming even [7:51] more complex. [7:52] But there are things we can do and some [7:54] of the talks today will touch exactly on [7:56] those points. [7:57] Dinkar will talk about the [7:58] transformation from traditional data [8:00] interfaces to full-fledged data [8:02] applications [8:03] built by non-experts, [8:05] employees and agents building custom [8:08] data products without being limited to [8:10] the existing tools. [8:12] We want to enable agents to move [8:13] independently and tackle even more [8:15] challenging issues, but autonomy without [8:17] governance is chaos. [8:20] Komal will share how agents governance, [8:22] safety, privacy, compliance isn't a nice [8:24] to have. It's the prerequisite for any [8:27] of these to work at scale. [8:29] And Alex from Snorkel will discuss [8:30] observability in the context of agents, [8:33] Evals, benchmarks, and the role of data [8:35] quality when your agent is not just [8:37] recommending, it is making decisions. [8:41] Trust and governance are not the only [8:43] challenges we are facing. [8:45] The amount of data and its complexity is [8:47] growing rapidly as well. [8:49] Did you know that the ratio of training [8:51] data to active parameters in open weight [8:53] LLM models has grown 3.1 times per year [8:56] since 2022? [8:58] Recent models have been trained with 20 [9:00] times more data per parameter than the [9:02] optimal ratio suggested by the 2022 [9:04] Chinchilla scaling laws. [9:06] And it is not just LLM models who need [9:08] more and more diverse data. [9:10] Evolution of critical business [9:11] solutions, like Lexis, require more data [9:14] to continue improving growth. [9:18] Reasoning is hungry. [9:21] Pattern matching works on sparse feature [9:22] vectors, a few signals summarized, [9:25] whereas reasoning needs full view of [9:27] what the person has done across surfaces [9:29] over time, not just a summary, the full [9:32] behavioral history. [9:34] Our Our generation models learn from [9:36] these and they're already boosting [9:38] Meta's top AI models for both ads and [9:40] organic content. [9:42] And the scale is staggering. We're [9:44] building towards 500 million queries per [9:47] second and a petabyte per second of [9:49] throughput for training data reads. [9:52] As you can imagine, this requires us to [9:53] rethink the entire data infrastructure. [9:56] There are a couple of shifts here that I [9:57] want to highlight. [9:59] Freshness matters. [10:01] If the algorithm reasons about current [10:03] intent, it can't wait for the pipeline [10:05] to run every 24 hours. [10:07] Freshness becomes a first-class [10:09] requirement, and so real-time streaming [10:11] data, not batch ETL, is fast becoming [10:14] the backbone of ranking and [10:15] accommodation pipelines. [10:17] But that alone is not enough. [10:20] Storage needs intelligence, too. [10:22] The old approach stored user data as [10:24] opaque blobs, no awareness of what's [10:26] inside. [10:28] This led to massive over-fetching and [10:30] GPU starvation. [10:32] We're now making the storage layer [10:33] itself understand the data it serves, [10:35] fetching only the columns and time [10:37] ranges that matter, in formats that [10:39] provide flexibility, and not a fixed [10:41] schema. [10:43] Our sequence storage talk this afternoon [10:45] goes deep on exactly how. [10:48] Let's go deeper into recommendation [10:50] systems, a key aspect of our product [10:52] space and in many ways the heart of our [10:54] business. [10:56] Recommender systems allow us to connect, [10:58] entertain, and serve billions of people [11:01] every day. [11:02] They also drive meaningful double-digit [11:04] growth in revenue year over year. [11:07] But for almost two decades, recommender [11:09] systems have had the same essential [11:11] idea. [11:12] Observe behavior, find patterns, predict [11:15] what's next. [11:16] Increasingly more and more sophisticated [11:18] pattern matching, but still pattern [11:20] matching. [11:22] LLMs break that paradigm. For the first [11:25] time, we can build recommender systems [11:26] that don't just correlate signals, they [11:29] reason about intent. They understand who [11:31] you are, not just what you clicked. [11:34] And that changes everything, what the [11:36] algorithm can do for people, and what is [11:38] needed from the data layer underneath. [11:41] Here's a stat. 42% of Instagram users [11:45] said they wanted to change the [11:46] algorithm. [11:47] Not adjust the setting. Fundamentally [11:50] change what gets shown to them. [11:52] That's not a feature request. This [11:53] requires to rethink the whole thing. So, [11:56] imagine, what if the algorithm wasn't [11:58] something that happened to you, but [12:00] something you direct? [12:02] What if you could talk to it, tell it [12:04] what you want more of, what you're tired [12:06] of, what you're in the mood for right [12:08] now, and it reasons about it. Not [12:10] keyboard matching, actual reasoning. [12:13] Chain of thought, weighing your history [12:15] against your current context, adapting [12:17] in real time. [12:18] That's the North Star. Fully [12:20] conversational recommendations, [12:22] where the system doesn't just rank [12:24] content, it understands the request. [12:27] The same query, say soccer, brings forth [12:29] completely different results, depending [12:31] on whether you're a casual fan looking [12:32] for highlights, [12:34] or club athlete looking for training [12:35] drills. [12:36] Because the system reasoned about you [12:39] and your intent, not just the word [12:40] you're looking for. [12:42] Later today, the team will take you deep [12:44] into how we're building this. [12:47] So, let us bring those threads together, [12:49] because they are not three separate [12:51] stories. [12:52] Remember our product loop from the [12:53] beginning? [12:55] Consumer, product, data, [12:56] recommendations. [12:58] The core cycle that serves 3 and 1/2 [13:00] billion people every day. [13:02] What we've just walked through is a [13:04] transformation of that entire loop. [13:07] Agents are no longer just building the [13:08] loop. They're becoming citizens within [13:10] this. [13:11] They operate, discover, analyze, and [13:13] create alongside us at every stage. [13:17] Data is transforming from batch to real [13:19] time, from opaque blobs to intelligent [13:21] storage, because agents and reasoning [13:24] models demand freshness, flexibility, [13:26] and scale that our previous generations [13:28] were never designed for. [13:31] Recommendations are evolving from [13:32] pattern matching to genuine reasoning. [13:34] From an algorithm that happens to you to [13:37] one you direct. [13:39] And here's what gives me genuine [13:40] excitement. Each one accelerates the [13:42] other. Agents make data more accessible. [13:45] Better data makes reasoning possible. [13:47] Reasoning creates new demands that push [13:49] agents and data infrastructure forward [13:51] again. [13:52] It's not linear progress. It's a [13:54] flywheel. [13:55] And there is more than the product loop [13:57] that is evolving. [13:59] Here's a glimpse of what agents are [14:01] already in use at Meta. [14:03] Innovation coming from all across the [14:04] company. [14:05] And you will hear more from Meta [14:07] colleagues in the panel discussion later [14:08] today about some of these projects. [14:11] Right off the bat, no surprises here, we [14:14] have a whole slate of task agents [14:16] focused on domain-specific activities [14:17] like coding, data analysis, operations. [14:22] Clowntown is a multi-agent orchestration [14:24] framework that is enabling agentic [14:27] autonomy. Engineers going on vacation [14:29] and letting multi-week agents make [14:31] semi-autonomous decisions without human [14:33] in the loop. [14:35] Malbook is exploring an agent-to-agent [14:37] social network. [14:39] With humans as flies on the wall [14:40] observing how agents communicate and [14:42] negotiate among themselves. [14:45] And the frontier loop is exploring [14:47] opportunities to accelerate executive [14:49] decision-making. [14:50] Autonomously taking all the high-context [14:53] docs and updates from across the [14:54] organization. And synthesizing context [14:57] ready for decision-making. [14:59] And that leads me to the final prompt I [15:01] would leave you with. [15:02] Over a decade ago, at scale started as a [15:05] conversation about building systems for [15:07] a billion people. [15:09] Today we're here to figure out what [15:10] happens when those systems serve [15:12] billions of people and millions of [15:14] superintelligent agents operating [15:16] simultaneously. [15:18] What happens when models are no longer [15:20] trained by you? But rather an agent that [15:23] trains itself and continuously tunes and [15:25] adapts without waiting for the next [15:27] training cycle. [15:28] When we get this right, when agents and [15:31] humans co-created the scale, we're not [15:33] talking about 10% improvements anymore. [15:35] We're talking about entirely new [15:37] categories of products that don't exist [15:39] yet. [15:40] The talks you'll hear today aren't [15:42] academic. They are glimpse into the [15:44] future being brought into the present. [15:47] The infrastructure you will build this [15:48] year will power things none of us can [15:50] imagine yet. [15:52] So, what will scale infrastructure [15:54] enable you to build? [15:55] Let's dive into it. [15:57] Thank you.