Meta Deploys Distilled LLMs to Give Instagram Users Direct Feed Control

@Scale

Meta's Tune-Your-Algorithm framework uses distilled LLMs and reasoning pipelines to let users explicitly direct their recommendation feeds, cutting compute costs by 100x while replacing black-box collaborative filtering.

AI-assisted summary of Tuning Your Algorithm with MRS Memory System and Think-Then-Recommend (TTR) | Qi Guo from Meta — original source. Check the original for full context.

Source published: 2026-07-08

Detailed notes

## Background Published by @Scale on YouTube on July 7, 2026, this talk features Qi Guo from Meta presenting Tune-Your-Algorithm (TYA), an AI-powered agentic recommendation system on Instagram. The page description highlights how TYA grants users natural-language control over their feed using the MRS Memory System (Biography) and Think-Then-Recommend (TTR) reasoning architecture.

## TL;DR The presentation details Meta's shift from black-box collaborative filtering to transparent, user-steerable recommendation systems via Tune-Your-Algorithm (TYA). TYA leverages 'Biography' to translate raw interaction histories into structured, hierarchical user interest profiles, cutting compute costs by 100x through model distillation. It then pairs this memory with 'Think-Then-Recommend' (TTR), an explicit reasoning step that decomposes complex or ambiguous natural-language user inputs into precise vector search sub-goals. Early MVP results show strong user adoption, paving the way for Meta's vision of proactive, multi-turn conversational feeds.

## Atmosphere & style The presentation carries an optimistic, highly technical, and forward-looking energy typical of an engineering tech keynote. Qi Guo delivers a structured, authoritative narrative that bridges industry pain points with cutting-edge architectural solutions. The tone strikes a balance between practical product engineering and visionary ambition, grounded by clear real-world examples and a pragmatic 'MVP-first' development philosophy.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=y-FiLm__jVY&t=5s) Recommendation systems have evolved into a core economic engine generating ~$1 trillion annually, yet modern Collaborative Filtering still creates a 'black-box echo chamber' that ignores dynamic individual motivations. - [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) To address organic user demand for algorithmic transparency and control, Meta developed Tune Your Algorithm (TYA), an AI-powered agentic recommendation framework. - [04:23](https://www.youtube.com/watch?v=y-FiLm__jVY&t=263s) TYA provides an intuitive 3-step user experience (View, Tune, Sync) using interactive natural-language interest pills, driving strong early product-market fit and user sentiment. - [06:57](https://www.youtube.com/watch?v=y-FiLm__jVY&t=417s) The system operates on a 5-step agentic loop connecting direct human feedback to large-scale retrieval and ranking pipelines. - [07:51](https://www.youtube.com/watch?v=y-FiLm__jVY&t=471s) The MRS Memory System ('Biography') transforms raw interaction logs into a multi-level, hierarchical natural-language representation of user interests. - [10:02](https://www.youtube.com/watch?v=y-FiLm__jVY&t=602s) Through LLM distillation and domain-specific fine-tuning, Meta reduced Biography's compute costs by 100x while outperforming a 100x larger base model. - [10:02](https://www.youtube.com/watch?v=y-FiLm__jVY&t=602s) Biography enables proactive interest discovery by mapping conceptual relationships and expanding user affinity graphs across related entities. - [11:34](https://www.youtube.com/watch?v=y-FiLm__jVY&t=694s) Think-Then-Recommend (TTR) introduces an explicit reasoning step to resolve semantic ambiguity and decompose vague user prompts into personalized sub-goals. - [13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) TTR effectively handles abstract intent realization, connecting high-level life goals (e.g., 'I want to be healthier') with a user's unique context to deliver tailored content. - [14:42](https://www.youtube.com/watch?v=y-FiLm__jVY&t=882s) Meta adopted an MVP-first strategy to validate user demand, using familiar interest pills before expanding toward a fully multi-turn conversational North Star vision. - [16:12](https://www.youtube.com/watch?v=y-FiLm__jVY&t=972s) Next-generation technical focus areas include persistent cross-sessional memory, proactive interest discovery, intent lifecycle management, and multi-step tool orchestration. - [18:27](https://www.youtube.com/watch?v=y-FiLm__jVY&t=1107s) Achieving high ROI and low latency for scaling agentic recommendation systems across billions of users relies on heavy model distillation, fine-tuning, and GPU pipeline automation.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=y-FiLm__jVY&t=5s) The $1 Trillion Industry and the Black-Box Echo Chamber

The talk opens with a high-level overview of the scale and history of recommendation systems (RecSys). Quoting Nvidia CEO Jensen Huang, Qi Guo highlights that recommendation engines are "everywhere"—powering everything from video recommendations and targeted advertising to e-commerce and movie suggestions. By 2025, recommendation systems serve as a core economic backbone, driving approximately $1 trillion in digital platform revenue and growing at a 35% compound annual rate for enterprise adoption across cloud and AI services.

RecSys has evolved dramatically over four decades: from 1990s rule-based logic to 2000s matrix factorization, 2010s deep learning, and finally the 2020s LLM and Transformer era. Despite these advancements, modern recommendation systems still rely heavily on Collaborative Filtering (CF). While CF is extremely powerful at surfacing content based on co-engagement patterns, it fundamentally creates a "black box echo chamber." It limits users to seeing what similar users watch, failing to account for dynamic, highly individual motivations—whether a user comes to a platform for pure entertainment, staying up-to-date on news, or deep-dive learning.

- **Market Scale:** RecSys accounts for ~$1T in digital platform revenue (2025) with a 35% YoY growth rate in enterprise cloud/AI integration over the next decade. - **Evolutionary Eras:** - **1990s:** Rule-based systems - **2000s:** Matrix factorization - **2010s:** Deep learning - **2020s:** LLMs and Transformer-based recommendation models - **Collaborative Filtering Limitation:** Creates an echo chamber where "one size doesn't really fit all" across diverse user motivations.

---

[02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) "The Algo": Latent User Demand and Meta's Vision

User desire to explicitly steer recommendation algorithms is not new, but it has recently erupted into public discourse. Qi Guo points to popular organic user behaviors, such as "The Algo" meme on Threads, where users post natural-language posts explicitly telling the platform, "Hey, I want to see more of this. I want to see less of that." Similar conversations occur on X, where users ask Elon Musk for direct algorithmic control. Even Meta CEO Mark Zuckerberg has noted a vision where users will converse with Meta platforms to shape their feed experience down to pixel-level generation.

To address this demand, Qi Guo and his team across Meta Recommendation Systems (MRS), Instagram, and Facebook built **Tune Your Algorithm (TYA)**. TYA provides an AI-powered agentic recommendation framework consisting of two main innovations: the **MRS Memory System (Biography)** and **Think-Then-Recommend (TTR)**.

- **Latent Demand Manifestations:** "The Algo" text posts on Threads and open user queries on X. - **Zuckerberg's Vision:** Conversational steering moving toward real-time pixel-level generated recommendations. - **Collaborating Teams:** Meta Recommendation Systems (MRS), Instagram, and Facebook engineering groups.

---

[04:23](https://www.youtube.com/watch?v=y-FiLm__jVY&t=263s) TYA User Experience and Early Product-Market Fit

From a user perspective, Tune Your Algorithm (TYA) replaces implicit, opaque feedback loops with transparent, natural-language controls. When users open the TYA interface, they are presented with a concise natural-language summary of what the system thinks they have been up to lately, accompanied by interactive "interest pills" (keywords summarizing detected interests). Users can directly tune their feed by adding or removing these interest pills. Upon editing, the system immediately syncs the updated topics as a "new contract" for future recommendations.

Early results show strong product-market fit, particularly among younger audiences. Users report that modifying their feed with a single tap and seeing immediate results feels "like a dream come true." Press coverage has praised TYA as a "new level of control that feels long overdue," contrasting it with traditional platforms like TikTok, which only offer a static, predefined list of generic topic choices.

- **3-Step User Journey:** 1. **View:** Read natural-language summary of recent activity and topic interest pills. 2. **Tune:** Add, edit, or remove interest pills using natural language. 3. **Sync:** System adopts the updated profile as the active contract for recommendations. - **User Reaction:** Described as an "Aha moment" due to real-time feed updates without tedious implicit scrolling/curation. - **Press Contrast:** Moves past TikTok's rigid, fixed topic lists by offering fully personalized, unconstrained interest customization.

---

[06:57](https://www.youtube.com/watch?v=y-FiLm__jVY&t=417s) Under the Hood: The 5-Step Agentic Recommendation Loop

Behind the user interface, TYA operates as an agentic loop comprising five distinct architectural stages. The workflow connects human feedback directly to large-scale retrieval and ranking pipelines.

1. **User Input:** Natural language prompts or edits to interest pills submitted via the UI. 2. **Biography Memory System:** Merges new user inputs with the system's existing knowledge base of user history. 3. **Think-Then-Recommend (TTR):** Executes reasoning to translate user inputs and memory into personalized sub-goals. 4. **Retrieval & Ranking:** Queries large-scale embedding spaces to fetch and rank contextually relevant items. 5. **Output:** Delivers a fully tailored feed response directly aligned with the user's explicit request.

---

[07:51](https://www.youtube.com/watch?v=y-FiLm__jVY&t=471s) The MRS Memory System: "Biography"

At the center of TYA is **Biography**, an LLM-based memory framework that transforms raw, unstructured user interaction logs into structured text that large language models can analyze. Every interaction—videos watched (e.g., clips of *Suits*, *The Big Bang Theory*, or soccer highlights), photos liked, groups joined, creators followed, and search queries performed—is converted into structured natural language strings detailing watch times and action types.

Biography feeds these textual logs into a specialized LLM to construct a multi-level, hierarchical representation of the user: - **Topics & Subtopics:** High-level categories like Basketball, Soccer, or TV Shows. - **Named Entities & Keywords:** Specific entities like Stephen Curry, NBA, Cristiano Ronaldo, or *Suits*. - **Granular Micro-Interests:** Hyper-specific sub-goals, such as "Curry's famous night-night gestures."

- **Input Signals Translated:** Video watch histories, photo likes, search logs, group memberships, followed creators. - **Hierarchical Output Levels:** Topics/Subtopics → Named Entities/Keywords → Granular Micro-Interests.

---

[10:02](https://www.youtube.com/watch?v=y-FiLm__jVY&t=602s) Model Distillation and Interest Discovery

Running standard LLMs at Meta's scale for every single user interaction is prohibitively expensive. To overcome this, the engineering team spent six months refining the Biography architecture using model distillation and domain-specific fine-tuning. By training a smaller "student" model using the TYA feedback loop and in-domain data, they successfully **reduced computational costs by 100x** while outperforming a general-purpose model 100 times larger.

Beyond summarizing past behaviors, Biography actively powers **interest discovery** by mapping conceptual relationships in an embedding space. For instance, if a user likes Stephen Curry, Biography expands to nearby nodes like the Golden State Warriors and other NBA teams (e.g., Miami Heat). It can then bridge to regional sports like MLS (Inter Miami) and key players like Lionel Messi. Similarly, detecting interest in Super Bowl videos allows the system to surface related pop-culture entities like Taylor Swift and Travis Kelce.

- **Cost Reduction:** 100x compute cost reduction via LLM distillation and domain-specific fine-tuning. - **Performance:** Student model outperforms a 100x larger base LLM on domain-specific user understanding. - **Discovery Graph Examples:** - *Stephen Curry* → Golden State Warriors / NBA → Miami Heat → Inter Miami (MLS) → Lionel Messi. - *Super Bowl* → NFL → Taylor Swift & Travis Kelce.

---

[11:34](https://www.youtube.com/watch?v=y-FiLm__jVY&t=694s) Think-Then-Recommend (TTR): Decomposing Complex Intents

To utilize Biography's rich memory profiles for real-time recommendations, Meta introduced **Think-Then-Recommend (TTR)**. Simple prompts like "I want to see more soccer" carry vast semantic ambiguity based on who is asking. For a Lionel Messi fan, "soccer" should surface Inter Miami, Argentina national team, and World Cup news. For a youth athlete or parent, it should mean training drills and skill tutorials. For a Real Madrid fan, it means Cristiano Ronaldo highlights and Portuguese national team coverage.

TTR inserts a explicit "thinking step" between intent detection and retrieval. It combines the raw user prompt with the Biography memory profile, running reasoning operations to generate nuanced "reasoning factors." These factors map ambiguous prompts into tailored sub-goals, which are then converted into vectors to query billion-scale embedding spaces for item retrieval.

- **Messi Fan Prompt:** Soccer → Inter Miami / Argentina / World Cup news. - **Parent/Athlete Prompt:** Soccer → Fundamental drills and physical training videos. - **Real Madrid Fan Prompt:** Soccer → Cristiano Ronaldo highlights / Portugal national team. - **Scale:** Performs vector space retrieval across billion-scale item catalogs.

---

[13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) Handling Abstract Intent Realization

TTR excels at breaking down broad, non-specific human goals—such as "I want to be healthier" or "I just moved to San Francisco"—into actionable recommendation pipelines.

For example, if a 30-year-old San Francisco resident who likes Stephen Curry and AI types *"I want to be healthier,"* TTR's reasoning step breaks the goal into distinct sub-intents: 1. **Physical Fitness:** Increasing workout frequency. 2. **Health & Diet:** Improving nutrition. 3. **Sleep & Habits:** Optimizing daily routines.

During the recommendation phase, TTR marries these sub-intents with the user's specific context. Instead of serving generic fitness clips, the system prioritizes basketball workout routines (leveraging the Stephen Curry affinity) and smart fitness technology (leveraging the AI/tech interest).

---

[14:42](https://www.youtube.com/watch?v=y-FiLm__jVY&t=882s) Building the MVP vs. The North Star Vision

Meta intentionally took an agile "MVP mindset" when rolling out TYA. Qi Guo uses an engineering analogy: *"The North Star is building a car, but we started by building a scooter, a bike, and a motorbike first, instead of building isolated wheels, doors, and windows."*

By starting with a UI mechanism users already understood—interactive topic/interest pills—the team validated user demand and gathered production feedback without overwhelming users with complex conversational interfaces. Having proven the concept, Meta is now transitioning toward its **North Star Vision**: an agentic, multi-turn conversational recommendation engine that persists across user sessions.

- **MVP Approach:** Familiar UI (interest pills / topic-following paradigm) to test user feedback quickly. - **North Star Vision:** Multi-turn, multi-session conversational interface that maintains long-term memory and proactively manages dynamic user goals.

---

[16:12](https://www.youtube.com/watch?v=y-FiLm__jVY&t=972s) Next-Generation Capabilities and Technical Roadmap

To bridge the gap between the MVP and the North Star vision, Meta is actively developing next-generation core capabilities across both Biography and TTR.

**Key Technical Focus Areas:** - **Persistent Cross-Sessional Memory:** Upgrading Biography to maintain conversation context across multiple sessions over time. - **Proactive Interest Discovery:** Continuous background operation that detects latent intents and surfaces new topics even without direct user prompts. - **Intent Lifecycle Management:** Distinguishing long-term user interests (e.g., lifelong hobbies) from temporary, short-term intents (e.g., planning a weekend trip or moving to a new city). - **Multi-Step Reasoning & Tool Orchestration:** Expanding TTR's ability to break complex goals into structured sub-goals while coordinating diverse platform tools to fulfill user goals holistically. - **Efficiency at Scale:** Maintaining strict latency and infrastructure cost boundaries as increasingly complex LLM reasoning steps are introduced into production.

[18:27](https://www.youtube.com/watch?v=y-FiLm__jVY&t=1107s) Scalability, ROI Optimization, and Future Outlook for Agentic Recommendations

Qi Guo concludes the presentation by highlighting the operational and engineering challenges of transitioning from traditional recommendation architectures to modern, reasoning-augmented AI systems. Moving to LLM-based agentic frameworks requires a fundamental shift in how system efficiency and return on investment (ROI) are approached. He emphasizes that scaling these complex systems to billions of users demands aggressive pipeline automation to optimize compute overhead and resource allocation effectively.

To bring these agentic models into a "high ROI space," Meta relies heavily on techniques such as model distillation, specialized fine-tuning, and advanced GPU pipeline automation. Guo expresses immense excitement about the future trajectory of AI-driven, agentic recommendation systems, inviting researchers and engineers who share a passion for user agency, transparency, and scalable LLM architectures to reach out and collaborate on advancing the North Star vision of Tune-Your-Algorithm (TYA).

- **Core Techniques for Efficiency**: Utilizing model distillation, targeted fine-tuning, and end-to-end GPU automation to lower latency and operational costs compared to standard recommendation systems. - **Strategic Goal**: Achieving high ROI while serving rich, personalized sub-goals generated by the Think-Then-Recommend (TTR) and MRS Memory frameworks at scale. - **Call to Action**: Inviting community collaboration and outreach to tackle ongoing technical challenges in next-generation, user-tunable AI systems.

## Notable quotes > [00:05](https://www.youtube.com/watch?v=y-FiLm__jVY&t=5s) one size doesn't really fit all — Qi Guo

> [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) Hey, I want to see more of this. I want to see less of that.

> [04:23](https://www.youtube.com/watch?v=y-FiLm__jVY&t=263s) like a dream come true

> [04:23](https://www.youtube.com/watch?v=y-FiLm__jVY&t=263s) Aha moment

> [04:23](https://www.youtube.com/watch?v=y-FiLm__jVY&t=263s) a new level of control that feels long overdue

> [13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) I want to be healthier

> [13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) I just moved to San Francisco

> [14:42](https://www.youtube.com/watch?v=y-FiLm__jVY&t=882s) The North Star is building a car, but we started by building a scooter, a bike, and a motorbike first, instead of building isolated wheels, doors, and windows. — Qi Guo

## Mentioned Qi Guo, Meta, Nvidia, Jensen Huang, Threads, X, Elon Musk, Mark Zuckerberg, Instagram, Facebook, MRS Memory System, Biography, Think-Then-Recommend, TTR, Tune Your Algorithm, TYA, TikTok, Stephen Curry, NBA, Golden State Warriors, Miami Heat, Inter Miami, MLS, Lionel Messi, NFL, Taylor Swift, Travis Kelce, Cristiano Ronaldo, Real Madrid, Suits, The Big Bang Theory

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

中文简报

Meta 部署蒸馏 LLM 让 Instagram 用户自主掌控信息流 Meta 的 Tune-Your-Algorithm 框架利用蒸馏 LLM 和推理管线使用户能够明确引导其推荐流，在取代黑箱协同过滤的同时将计算成本降低 100 倍。该演示详细介绍了 Meta 如何通过 Tune-Your-Algorithm (TYA) 从黑箱协同过滤转向透明且可由用户引导的推荐系统。TYA 利用“Biography”将原始交互历史转化为结构化、层级化的用户兴趣画像，并通过模型蒸馏将计算成本降低 100 倍。随后，它将该记忆与“Think-Then-Recommend”(TTR) 相结合，这一显式推理步骤可将复杂或模糊的自然语言用户输入分解为精确的向量检索子目标。早期 MVP 结果显示出强劲的用户采纳度，为 Meta 实现主动式、多轮对话流的愿景铺平了道路。推荐系统已演变为每年创造约 1 万亿美元产值的核心经济引擎，但现代协同过滤仍会构建出忽视个体动态动机的“黑箱回音室”。 (0:05) 为满足用户对算法透明度和控制权的自发需求，Meta 开发了基于 AI 的智能体推荐框架 Tune Your Algorithm (TYA)。 (2:40) TYA 利用交互式自然语言兴趣标签提供直观的三步用户体验（查看、调优、同步），取得了强劲的早期产品市场契合度和积极的用户反馈。 (4:23) 该系统基于一个五步智能体循环运行，将人类直接反馈与大规模检索和排序管线相连接。 (6:57) MRS 记忆系统（“Biography”）将原始交互日志转化为用户兴趣的多层次、层级化自然语言表征。 (7:51)

Reading summary

Meta Recommendation Systems engineering lead Qi Guo outlines "Tune Your Algorithm" (TYA), an agentic recommendation framework combining an LLM memory system ("Biography") with a reasoning engine ("Think-Then-Recommend" or TTR). Designed to move beyond the black-box constraints of collaborative filtering, TYA gives users direct natural-language control to steer their algorithmic feeds. The Limits of Collaborative Filtering (0:05) Recommendation systems underpin roughly $1 trillion in digital platform revenue in 2025 and are growing at a 35% compound annual rate across enterprise cloud and AI services. Despite evolving over four decades—from 1990s rule-based models through matrix factorization and deep learning to modern Transformers—recommendation engines still depend heavily on collaborative filtering (CF). Guo points out that CF inherently produces a black-box echo chamber by serving content based strictly on co-engagement patterns of similar users, failing to account for dynamic individual motivations like learning, news gathering, or entertainment. Latent demand for direct feed control has surfaced publicly through user behaviors like "The Algo" meme on Threads and open requests on X. Meta CEO Mark Zuckerberg has similarly outlined a vision where conversational steering will eventually allow users to direct feeds down to pixel-level generation. In response, Meta Recommendation Systems, Instagram, and Facebook engineering groups developed TYA to replace opaque implicit signals with transparent user agency. User Experience and the Agentic Loop (4:23) TYA presents users with a concise natural-language summary of their detected recent activity alongside interactive "interest pills." Users can add, modify, or delete these topic pills, establishing an explicit "contract" that instantly updates subsequent recommendations without requiring repetitive implicit scrolling. Guo contrasts this flexible natural-language customization with traditional platforms like TikTok, which rely on static, predefined category lists. Behind the interface, TYA functions as an agentic loop across five stages. First, the user submits a natural-language request or interest pill adjustment. Second, the Biography memory system combines this input with historical user data. Third, the Think-Then-Recommend (TTR) engine executes reasoning to break inputs into personalized sub-goals. Fourth, these sub-goals query billion-scale embedding spaces to retrieve and rank relevant items. Finally, the system outputs an updated feed directly aligned with the explicit user request. Biography: Memory Modeling and Distillation (7:51) At the foundation of TYA is Biography, an LLM-based memory framework that transforms raw, unstructured interaction logs—including watch times for video clips, photo likes, search queries, group memberships, and followed accounts—into structured text. Biography then builds a multi-level user profile spanning broad categories, named entities, and granular micro-interests (such as "Stephen Curry's night-night gestures"). Biography also enables interest discovery by mapping conceptual relationships across embedding spaces, such as linking Stephen Curry to the NBA, Inter Miami, and Lionel Messi, or connecting Super Bowl content to Taylor Swift and Travis Kelce. Running large foundational models for every user interaction at Meta's scale was computationally prohibitive. To address this, the engineering team spent six months applying domain-specific fine-tuning and model distillation to produce a smaller student model. Guo notes that this distillation reduced computational costs by 100x while outperforming a general-purpose model 100 times larger on domain-specific user understanding. Think-Then-Recommend Intent Decomposition (11:34) To handle ambiguous user inputs, Meta implemented Think-Then-Recommend (TTR), which inserts an explicit reasoning step between intent detection and item retrieval. A generic prompt such as "I want to see more soccer" carries different meanings depending on context: a Lionel Messi supporter expects Inter Miami and Argentina national team content, a youth athlete requires skill drills, and a Real Madrid fan seeks Cristiano Ronaldo highlights. TTR's reasoning factors map the ambiguous prompt against the user's Biography profile to create tailored sub-goals for vector retrieval. TTR also breaks down broad, abstract objectives into concrete recommendations. For example, if a San Francisco resident with existing interests in Stephen Curry and AI inputs "I want to be healthier," TTR decomposes the goal into sub-intents covering workout frequency, nutrition, and sleep habits. During the recommendation phase, the system contextualizes these sub-intents by prioritizing basketball-specific training drills and smart fitness technology rather than generic fitness videos. Architectural Roadmap and Scalability (14:42) Meta introduced TYA through an agile minimum viable product (MVP) approach—which Guo compares to building a scooter before a car—using familiar interest pills to validate user demand before introducing complex conversational agents. The long-term "North Star" roadmap aims for a multi-turn, multi-session conversational recommendation engine with persistent cross-sessional memory, proactive interest discovery that surfaces topics without explicit prompting, and intent lifecycle management that distinguishes lifelong interests from temporary events like moving cities. Sustaining operational efficiency and high return on investment (ROI) remains the primary technical hurdle in deploying agentic recommendation systems. To maintain strict latency and infrastructure cost limits when serving billions of users, Meta relies heavily on model distillation, targeted fine-tuning, and automated GPU pipelines as it transitions from traditional scoring models to multi-step reasoning frameworks. Meta's deployment of TYA demonstrates that LLM-driven memory and reasoning can shift recommendation feeds from opaque black boxes toward explicit user steering. However, the long-term viability of this agentic model depends on whether automated GPU pipelines and model distillation can continuously hold the computational costs and latencies of multi-step LLM reasoning within sustainable limits at planetary scale.

精炼阅读摘要

Meta推荐系统团队的Qi Guo指出，传统协同过滤驱动着约1万亿美元的数字平台收入，但其黑盒机制已难以满足用户复杂的动态意图；为此Meta推出了Tune Your Algorithm（TYA）框架，结合Biography记忆系统与Think-Then-Recommend（TTR）推理机制，使用户能通过自然语言直接干预推荐算法。传统协同过滤黑盒与算法调控需求 (0:05) Qi Guo援引Nvidia首席执行官Jensen Huang的观点指出，推荐系统无处不在，到2025年已成为驱动约1万亿美元数字平台收入的经济支柱，并在企业云与AI服务中保持35%的复合年增长率。推荐技术虽历经1990年代规则系统、2000年代矩阵分解、2010年代深度学习以及2020年代Transformer和大语言模型等阶段，底层依然重度依赖协同过滤（CF）。然而，协同过滤仅依据群体互动模式推荐内容，容易形成“黑盒回音室”，无法适应用户在娱乐、获取资讯或深度学习等差异化动机间的即时切换。近期用户对主动掌控算法的意愿集中爆发，例如Threads平台上流行的“The Algo”文字梗帖，以及X平台上用户直接向Elon Musk表达算法调控诉求。Meta首席执行官Mark Zuckerberg亦曾描绘过用户通过对话定制信息流、甚至实现像素级内容生成的未来构想。为响应这一显式控制需求，Meta推荐系统团队（MRS）联合Instagram与Facebook工程部门启动了TYA框架的研发。 TYA交互模式与五步智能体闭环 (4:23) 相比TikTok等平台提供的静态、预设话题分类，TYA引入了透明的自然语言与“兴趣药丸”（interest pills）交互模式。用户界面呈现出三步工作流：首先展示系统生成的近期活动自然语言摘要与检测到的兴趣关键词标签；用户随后可直接添加、修改或删除这些标签；保存后，系统立即将更新后的画像作为后续推荐的“新契约”进行同步，早期反馈在年轻受众中表现出显著的产品契合度。在底层架构上，TYA通过一个包含五阶段的智能体循环（Agentic Loop）运作：首先接收前端提交的自然语言或标签编辑；随后交由Biography记忆系统将其与历史知识库融合；接着由TTR执行推理将输入拆解为个性化子目标；然后进入多维向量空间执行检索与精排；最终向用户输出完全对齐显式指令的定制化信息流。 Biography记忆建模与百倍算力蒸馏 (7:51) Biography系统将用户观看视频（如《Suits》、《生活大爆炸》或足球集锦）、点赞、加入小组、关注创作者及搜索词等非结构化交互日志，统一转换为包含操作类型与观看时长的结构化自然语言文本。专用LLM在此基础上提取出三层画像：高层主题（如篮球、电视节目）、命名实体（如Stephen Curry、NBA）以及极其具体的微观兴趣（如“Curry著名的晚安庆祝动作”）。为解决超大规模并发下运行标准大语言模型的昂贵算力开销，Meta团队耗时六个月对Biography展开模型蒸馏与领域微调。训练出的学生模型在专有任务理解上超越了体积大百倍的通用基础模型，同时将计算成本降低了100倍。此外，Biography基于向量空间实现了概念拓展与兴趣发现，例如从Stephen Curry节点延伸至金州勇士与迈阿密热火，并进一步跨界关联至迈阿密国际及Lionel Messi，或由超级碗关联到Taylor Swift与Travis Kelce。 TTR推理机制破解歧义与抽象意图 (11:34) 针对用户输入的高度歧义性，Meta设计了Think-Then-Recommend（TTR）推理机制。以“我想看更多足球”为例，梅西球迷往往期望获取迈阿密国际和阿根廷国家队资讯，年轻球员或家长需要基础训练教程，而皇家马德里球迷则倾向于Cristiano Ronaldo高光集锦。TTR在接收到请求后插入显式“思考阶段”，结合Biography记忆生成细分“推理因子”，将模糊指令转译为具体子目标，再投射至十亿级候选库中检索。面对“我想变得更健康”或“我刚搬到旧金山”等抽象意图，TTR可将其拆解为体育锻炼、健康饮食与作息规律等具体维度。若输入者为偏好Stephen Curry并关注AI技术的30岁旧金山居民，TTR在推荐阶段会结合其背景，优先提供篮球专项训练方案与智能健身硬件相关内容，避免推送泛化且低效的通用健身素材。敏捷演化路径与下一代工程挑战 (14:42) Meta在部署TYA时采取了敏捷的MVP策略，Qi Guo将其比作“造车愿景下先造滑板车、自行车和摩托车，而非孤立地制造车门或车轮”。团队先通过用户熟悉的标签修改界面验证核心价值，随后逐步推进至具备跨会话持久记忆、能够主动探测潜在意图的“北极星”多轮对话智能体架构。技术演化重点涵盖长短期意图生命周期区分、复杂多步推理以及跨应用平台工具编排。在工程规模化方面，Guo指出LLM智能体推荐系统必须应对严苛的投资回报率（ROI）考量。将多步推理模型拓展至数十亿用户，需要极具侵略性的自动化管线优化。Meta正通过模型蒸馏、领域特定微调与全链路GPU计算管线自动化来压低端到端延迟和运行开销，并呼吁学术界与工业界共同攻关下一代可控AI推荐系统的架构难题。 Qi Guo总结强调，由大语言模型与记忆推理驱动的智能体系统代表了推荐架构的必然方向，它打破了传统协同过滤的数据黑盒；但如何在数十亿级用户生产环境中兼顾复杂推理的深度、毫秒级响应延迟与GPU资源的高ROI，仍是未来落地的关键挑战。

Detailed notes — zh

## Background 本视频由 @Scale 于 2026 年 7 月 7 日发布在 YouTube 平台。视频简介指出，Meta 推荐系统团队的 Qi Guo 带来了关于 Tune-Your-Algorithm（TYA）的深度技术分享，介绍了该系统如何借助基于大语言模型的 MRS 记忆系统（Biography）与“思考后推荐”（TTR）框架，让 Instagram 用户不仅能透明看懂推荐算法，还能直接使用自然语言自主调校信息流。 ## TL;DR 演讲直面传统协同过滤构建的“黑盒回音室”局限，系统阐述了 Meta 创新的 TYA 推荐系统如何通过五步 Agentic 闭环赋予用户直接定义内容分发契约的掌控权。底层核心依赖 MRS 记忆系统（Biography）将多模态异构行为转化为结构化语义画像，并通过模型蒸馏在降低百倍算力成本的同时实现高质量跨实体兴趣发现；结合 Think-Then-Recommend（TTR）推理引擎，系统能深度拆解用户的模糊提示词与复杂生活意图，将其映射为精准的个性化子目标并在十亿级索引中召回。最后，团队展示了从经典标签交互 MVP 逐步演进至支持跨会话长期记忆、工具编排与静默主动性的下一代智能体推荐愿景。 ## Atmosphere & style 整场演讲呈现出典型硅谷顶级工业界技术大会的务实风格与前沿视野，主讲人 Qi Guo 的叙事冷静严谨且条理清晰，在直击用户痛点与拆解底层系统工程架构之间切换自如。从引用黄仁勋论断和社交网络模因开场，到深入分享模型蒸馏与意图分解的硬核工程实践，演讲节奏紧凑且充满突破感。整场分享不仅让听众窥见推荐系统从被动黑盒走向主动智能体交互的范式跃迁，更传递出极具感染力的产品务实主义与技术自信。 ## Bulletin - [00:05](https://www.youtube.com/watch?v=y-FiLm__jVY&t=5s) Meta 推出全新智能体推荐系统 Tune-Your-Algorithm（TYA），打破传统协同过滤的黑盒回音室，允许用户通过自然语言直观透视并调校自身推荐流。 - [06:57](https://www.youtube.com/watch?v=y-FiLm__jVY&t=417s) TYA 核心由 MRS 记忆系统（Biography）和 Think-Then-Recommend（TTR）推理机制驱动，将异构用户行为转化为结构化语义，并将抽象诉求分解为个性化子目标。 - [10:02](https://www.youtube.com/watch?v=y-FiLm__jVY&t=602s) 通过知识蒸馏与领域任务微调，团队打造出体积仅为原模型 1% 的学生模型，在将推理成本降低 100 倍的同时超越了基线模型的准确率。 - [16:13](https://www.youtube.com/watch?v=y-FiLm__jVY&t=973s) 未来系统将聚焦跨会话持久记忆、意图生命周期管理和多步推理规划，实现从单次兴趣选择向多轮对话与主动推荐的北极星演进。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=y-FiLm__jVY&t=5s) 推荐引擎支撑着数字平台每年约 1 万亿美元的营收，但现有协同过滤体系容易陷入黑盒回音室，无法理解用户的动态心智与深层诉求。 - [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) 针对用户在社交平台上对算法掌控权的强烈呼声，Meta 研发了 TYA 系统，主打可见化、自主调校及新契约同步三步交互体验。 - [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) TYA 在年轻群体中展现出强劲的自然增长与 PMF，用户无需花费数天被动刷视频驯化算法，点击即可实时重构信息流契约。 - [06:57](https://www.youtube.com/watch?v=y-FiLm__jVY&t=417s) TYA 运行在涵盖用户输入、记忆合成、推理规划、检索排序及定向输出的五步 Agentic 闭环流程中。 - [06:57](https://www.youtube.com/watch?v=y-FiLm__jVY&t=417s) MRS 记忆系统（Biography）负责将用户的停留时长、图文点赞、关注社群等异构行为转化为自然语言日志，并提取跨层级结构化意图。 - [10:02](https://www.youtube.com/watch?v=y-FiLm__jVY&t=602s) 借助知识蒸馏，团队成功研发出体积仅 1% 的学生模型，削减 100 倍推理成本并反超大模型效果，解决了高并发在线部署的算力瓶颈。 - [10:02](https://www.youtube.com/watch?v=y-FiLm__jVY&t=602s) Biography LLM 具备知识图谱扩展能力，能够基于球星实体自主关联到球队、同城跨界联赛及流行文化符号，实现跨域兴趣发现。 - [11:34](https://www.youtube.com/watch?v=y-FiLm__jVY&t=694s) TTR（Think-Then-Recommend）将显式提示词与用户记忆切片融合，解决单一关键词在不同人群背景下语义高度分化的问题。 - [13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) TTR 能够将抽象高阶目标（如“我想更健康”）拆解为体能、饮食、作息等多维度子目标，并与个人兴趣基因交叉生成定制化召回向量。 - [14:42](https://www.youtube.com/watch?v=y-FiLm__jVY&t=882s) 团队践行从滑板车到自行车的 MVP 敏捷迭代哲学，先通过经典的兴趣标签心智验证用户掌控权，再平滑演进至复杂多轮对话。 - [16:13](https://www.youtube.com/watch?v=y-FiLm__jVY&t=973s) 下一代系统的关键支柱包括跨会话持久记忆和意图生命周期管理，可有效避免过期需求（如已结束的旅游计划）长期污染推荐流。 - [16:13](https://www.youtube.com/watch?v=y-FiLm__jVY&t=973s) TTR 将承担智能大脑职能，编排异构检索器和业务工具以实现用户全生命周期的多步推理规划与高 ROI 规模化落地。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=y-FiLm__jVY&t=5s) 推荐系统的演进与“协同过滤黑盒”的局限 Meta 推荐系统（MRS）团队的 Qi Guo 登台分享了名为 Tune-Your-Algorithm（简称 TYA）的全新 AI 推荐系统方案，该系统允许用户直接使用自然语言来“调校”推荐算法。Qi 引用英伟达（NVIDIA）CEO 黄仁勋（Jensen Huang）的论断指出，推荐系统（RecSys）是过去 15 年来最为关键的 AI 落地应用。从短视频推荐、广告精准定向，到电商购物与电影流媒体，推荐算法无处不在。在商业价值上，截至 2025 年，推荐引擎已支撑起各大数字平台约 1 万亿美元的年营收；未来 10 年，企业在云端及 AI 服务中采用推荐引擎的复合年均增长率仍将达到 35%。回顾技术演进，推荐系统经历了从规则到智能的跨越： - **1990 年代**：以规则引擎（Rule-based Systems）为主导。 - **2000 年代**：矩阵分解（Matrix Factorization）技术兴起。 - **2010 年代**：深度学习（Deep Learning）开启了现代 AI 推荐纪元。 - **2020 年代**：大语言模型（LLM）与 Transformer 架构被全面引入推荐领域。尽管架构不断升级，现今绝大多数推荐系统仍严重依赖协同过滤（Collaborative Filtering, CF）。协同过滤依据群体协同互动行为进行内容分发，虽然极具统计威力，却也构建了一个难以穿透的“黑盒回音室”（black-box echo chamber）——用户往往只能被动消费生态中其他用户所消费的内容。这种“千人一面”（one-size-fits-all）的范式无法满足用户多样且动态变化的心智诉求。有的用户刷视频纯粹为了消遣娱乐，有的为了追踪时事新闻，有的则是为了深度学习特定领域的知识；仅凭隐式协同过滤，根本无法解析背后的真实意图。 ### [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) 用户对算法掌控权的呼声与 TYA 产品定位让用户“调校算法”的潜在需求早在十年前就已存在，而近年来这一诉求以更为外显的方式爆发。例如在 Threads 平台上，“Dear Algorithm”（致算法）甚至演变成了一种热门模因（meme）——用户在帖子里直接用自然语言向算法喊话：“我想多看点这个，少看点那个。”在 X（原 Twitter）等平台上，埃隆·马斯克（Elon Musk）也频频公开回应用户希望用自然语言重塑算法流的诉求。Meta CEO 马克·扎克伯格（Mark Zuckerberg）亦曾明确展望：在不久的将来，用户能够直接同系统对话，推荐系统甚至可以在像素级实时生成个性化内容。正是基于这一愿景，Qi 与来自 MRS 以及 Instagram、Facebook 的团队联合研发了 Tune-Your-Algorithm（TYA）。该系统在产品交互上分为三个清晰步骤： 1. **算法可见化（See Your Algorithm）**：用户打开交互界面时，系统会以通俗自然的语言概述其近期的兴趣动态，并附带一系列“兴趣标签药丸”（interest pills），以关键词形态直观呈现系统对其兴趣的理解。 2. **算法自主调校（Tune Your Algorithm）**：用户可以直接添加或删除这些兴趣药丸，对推荐系统的画像进行增删修正。 3. **建立新契约（Sync New Contract）**：推荐系统立即同步用户修改后的结果，以此作为后续内容分发的新契约。尽管尚处初期，TYA 已展现出极高的产品市场契合度（PMF），在年轻群体中更是呈现出强劲的自然增长。多位内测用户直呼这“简直像梦想成真（like a dream come true）”。相较于早期平台死板的话题关注机制，TYA 提供了前所未有的“即时性与精准度”：用户无需再花费数天通过隐式的点赞、浏览来被动调教算法，只需点击按钮，信息流便在眼前即刻重构。科技媒体也给予了高度评价，称这种“对算法拥有绝对控制权的体验早该到来”，并认为这是攻克推荐系统中最关键却最晦涩难懂的一环所迈出的里程碑式一步。 ```mermaid flowchart TD accTitle: TYA 五步 Agentic 闭环流程 accDescr: 展示用户输入、记忆合成、推理分解、检索排序到最终输出的闭环 subgraph UserSide["用户侧交互"] A["步骤 1: 用户自然语言交互 / 增删兴趣药丸"] E["步骤 5: 即时个性化内容流展现"] end subgraph CoreSystem["TYA 核心 Agentic 闭环"] B["步骤 2: MRS 记忆系统 (Biography) 融合历史画像与新输入"] C["步骤 3: 思考后推荐 (TTR) 推理与意图拆解"] D["步骤 4: 召回与重排序 (Retrieval & Ranking)"] end A --> B B --> C C --> D D --> E E -.->|"反馈与持续迭代"| A ``` ### [06:57](https://www.youtube.com/watch?v=y-FiLm__jVY&t=417s) 底层架构：Agentic 闭环与 MRS 记忆系统（Biography）在工程实现上，TYA 运行在一个由五大步骤组成的 Agentic 闭环中： - **步骤 1（用户输入）**：接收用户在 UI 上的自然语言表述或对兴趣药丸的增删操作。 - **步骤 2（记忆合成）**：MRS 记忆系统（内部代号 **Biography**）接管输入，将其与系统中已沉淀的用户画像进行综合融合。 - **步骤 3（推理规划）**：采用“思考后推荐”（Think-Then-Recommend, 简称 TTR）机制，进行深层逻辑推理，打通用户意图与长期记忆。 - **步骤 4（检索排序）**：驱动下游的候选集召回（Retrieval）与精排（Ranking）通道。 - **步骤 5（定向输出）**：输出完全符合用户此番调校契约的定制化推荐内容。整个方案的核心基石正是 MRS 的记忆系统——Biography。该系统的使命是将用户在平台上的所有异构行为全部转化为大语言模型（LLM）能够理解的高维语义。例如，针对一位常看美剧《金装律师》（Suits）、《生活大爆炸》（The Big Bang Theory）同时又热衷篮球和足球锦标赛的用户，Biography 会抓取其观看时长、互动频率、点赞图片、加入的社群、关注的创作者以及搜索记录，统一转换为自然语言文本日志（如“该画像用户多次完整观看了 XYZ 视频，点赞了关于 XYZ 的图文”），随后送入专门训练的 Biography LLM 中进行层级化抽象： - **宏观类目与子领域**：例如篮球、足球、情景喜剧等。 - **具体实体与关键词**：如斯蒂芬·库里（Stephen Curry）、NBA联赛、《金装律师》剧集、克里斯蒂亚诺·罗纳尔多（Ronaldo）等。 - **超细粒度微观意图**：例如“库里标志性的‘Night Night’晚安庆祝动作”这一层级的细分情节与名场面。 ### [10:02](https://www.youtube.com/watch?v=y-FiLm__jVY&t=602s) 模型蒸馏优化与图谱式兴趣发现直接在线部署通用大型 LLM 理解全量用户行为的计算开销极其高昂。为解决算力瓶颈，团队在过去 6 个月内对 Biography 模型进行了高强度的爬坡优化（hill-climbing）。通过知识蒸馏（Knowledge Distillation）与领域内特定任务微调（Fine-tuning），团队将用户在 TYA 中的交互反馈数据作为闭环训练信号引入蒸馏过程。最终，团队成功研发出体积仅有原大模型百分之一的学生模型（Student Model），不仅将**推理成本直接削减了 100 倍**，且在实际推荐理解与准确率指标上**超越了原本大 100 倍的模型**。此外，Biography LLM 还承担着跨模态与跨实体的“兴趣发现”（Interest Discovery）职能，而非局限于机械式的存量总结。当系统识别到某位用户是斯蒂芬·库里的粉丝时，大模型会基于全局知识图谱向外拓宽： - 联想至金州勇士队（Golden State Warriors）及其他 NBA 球队（如迈阿密热火队 Miami Heat）； - 顺着“迈阿密”这一地域实体，横向探索至美国职业足球大联盟（MLS）球队迈阿密国际（Inter Miami CF）； - 进而自然关联到足坛巨星梅西（Lionel Messi），将内容精准拓展至足球跨界领域； - 同样，在识别到“超级碗”（Super Bowl）视频的高频偏好后，模型会自主关联到流行巨星泰勒·斯威夫特（Taylor Swift）及其男友特拉维斯·凯尔西（Travis Kelce）等关联文化符号。 ### [11:34](https://www.youtube.com/watch?v=y-FiLm__jVY&t=694s) Think-Then-Recommend（TTR）：结合记忆系统的个性化意图推理在现代推荐系统中，理解用户的显式指令往往充满挑战。当用户输入一个看似直接的提示词——比如“我想看更多足球（soccer）”——这个词在不同人群的认知中有着完全截然不同的含义。对于梅西（Messi）的死忠粉而言，“足球”意味着美国职业足球大联盟（MLS）、迈阿密国际（Inter Miami）、阿根廷国家队以及世界杯的最新战讯；但如果换作一名运动员或正在指导孩子的家长，“足球”则可能指向专业足球训练教学视频和提升控球技巧的战术动作；而对于皇家马德里（Real Madrid）的球迷，最相关的可能是克里斯蒂亚诺·罗纳尔多（Ronaldo）的高光集锦或葡萄牙国家队的世界杯动态。单纯依靠字面匹配根本无法满足这种因人而异的深层诉求。为了弥合字面提示与实际意图之间的巨大鸿沟，Meta 团队引入了 Think-Then-Recommend（TTR，先思考后推荐）方法，将底层由 MRS（Memory System）生成的 Biography 结构化用户记忆与推理能力进行深度融合。在 TTR 架构中，系统左侧接收用户的自然语言提示词（如“soccer”），右侧挂接来自 Biography 的用户画像记忆切片。系统内部的核心“思考模块（thinking blob）”会展开推理，解析出针对该用户的推理因子（reasoning factors），进而生成独属于该用户意图的向量表征。借助这套向量，系统能够在十亿量级（billion scale）的内容池与嵌入空间中执行近邻内容检索，为拥有不同背景特征的用户输出高度差异化且切中痛点的个性化推荐结果。 ```mermaid flowchart LR accTitle: Think-Then-Recommend (TTR) 推理与检索架构 accDescr: 展示用户提示词与用户记忆如何经过 TTR 推理模块生成个性化向量并在十亿级内容库中完成检索。 subgraph Inputs["输入层"] Prompt["用户提示词 (如: 'Soccer')"] BioMem["MRS Biography 结构化用户记忆"] end subgraph Thinking["TTR 推理引擎"] Reasoning["思考模块 (Thinking Blob) 意图分解与个性化推理"] Embedding["生成多维度 个性化向量表征"] end subgraph Retrieval["大规模检索层"] Corpus[("十亿级内容库 (Billion-scale Index)")] Output["个性化推荐结果 (球迷集锦 / 训练教程)"] end Prompt --> Reasoning BioMem --> Reasoning Reasoning --> Embedding Embedding --> Corpus Corpus --> Output ``` - **个性化理解差异**：单一关键词“足球”可映射为特定球星动态、青训实操教学或特定俱乐部的高光集锦，其上下文完全取决于用户历史画像。 - **十亿级嵌入检索**：TTR 推理生成的个性化因子直接转化为语义表征，支撑海量推荐池中的高召回与精确匹配。 --- ### [13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) 复杂意图拆解与子目标映射：以“我想更健康”为例除了处理具有多义性的简单关键词，TTR 的更强之处在于解析复合型、抽象化的高阶意图，例如“我想变得更健康（I want to be healthier）”或“我刚刚搬到了旧金山”。以一位居住在旧金山、现年 30 岁且平时对斯蒂芬·库里（Stephen Curry）和人工智能（AI）表现出浓厚兴趣的用户为例，当他在 TYA 界面中输入“我想变得更健康”时，传统推荐引擎往往无所适从，只能随机推送泛化的健身短视频。 TTR 则通过两阶段的认知跃迁来攻克这一难题：首先是**意图分解阶段**。系统像人类专家一样进行深度逻辑推演，将“变得更健康”这一宏观目标拆解为具体的行动子目标（sub-goals），涵盖体能锻炼（physical fitness）、健康饮食（health and diet）、改善睡眠质量以及培养规律生活习惯等维度。紧接着进入**个性化映射与推荐阶段**，系统将分解出的子目标与用户存储于 Biography 中的兴趣基因进行交叉：在体能训练维度，系统不会推送枯燥的器械深蹲视频，而是针对他对库里的喜爱定向推荐篮球技巧与球场体能训练内容；在科技与生活习惯维度，系统则顺理成章地推荐结合 AI 与前沿智能硬件的健身科技资讯。这种“先细化拆解、再结合画像具象化”的机制，通过多模态向量召回实现了既贴合意图又深度个性化的推荐落地。 - **多层次意图分解**：将抽象诉求拆解为体能、饮食、睡眠和习惯等多条支线。 - **偏好交叉匹配**：将篮球球星兴趣投射到体能动作，将 AI 极客属性投射到智能穿戴与健康科技产品。 --- ### [14:42](https://www.youtube.com/watch?v=y-FiLm__jVY&t=882s) 从 MVP 到北极星愿景：滑板车思维与多轮对话演变回顾 TYA（Tune-Your-Algorithm）产品的演进路线，团队始终秉持着极度务实的最小可行产品（MVP）迭代哲学。演讲者用了一个经典的类比：“我们的北极星愿景是造出一辆汽车，但我们不会一开始就去分别制造轮子、车门或车窗；相反，我们选择先做滑板车，接着升级成自行车，再到摩托车，最后走向汽车。”这种渐进式思路确保了团队在每一步都能交付端到端的完整用户体验，并在真实交互中快速摸索产品与市场的契合度（PMF）。在 TYA 的 MVP 阶段，团队特意选用了用户最为熟悉的心智模型——主题关注与兴趣池界面。虽然 TikTok 和 X（前 Twitter）等竞品以及 Meta 早期系统都曾尝试过主题选择，但那些传统版本通常受限于僵硬、预设且不可定制的有限标签库，完全不具备个性化。TYA 的 MVP 既保留了用户无门槛操作的经典 UI 结构，又在底层彻底打通了自然语言输入通道。伴随着 MVP 在透明度与用户控制感方面收获的热烈正向反馈，团队明确了下一步的跨越方向：正式走出单次被动的兴趣选择框，迈向允许用户通过多轮对话（multi-turn conversation）自由表达深层生活诉求与细粒度偏好的全新形态。 --- ### [16:13](https://www.youtube.com/watch?v=y-FiLm__jVY&t=973s) 下一代 TYA 系统核心支柱：长期记忆、工具编排与工程效率为了落地下一代 TYA 系统的北极星蓝图，Meta 正在聚焦构建两大核心支柱能力：首先，系统必须具备强大的跨会话多轮对话理解力，能够像资深私人助手一样，在数天或数周的多轮交互中精准维持上下文记忆；其次，系统需要具备极强的“静默主动性（proactivity）”——即便用户没有主动打字交互，系统也能在后台自适应运转，敏锐捕捉隐式与显式行为，主动挖掘用户潜在兴趣并提供个性化的满足方案。围绕这两项核心目标，团队正在对底层技术栈进行全方位的架构升级： - **MRS 记忆系统（Biography）的深化**： - **跨会话持久记忆（Persistent cross-sessional memory）**：支撑长时间跨度的多轮对话，维持稳定的用户理解基座。 - **持续双轨学习**：无缝聚合隐式交互（如滑动停留、完播率）与显式自然语言反馈，加速新兴趣发现。 - **意图生命周期管理（Lifecycle management）**：精准区分不同意图的时间属性，例如将“赴旧金山旅游”识别为短周期意图，并在行程结束后适时淡化，避免无效信息的长久堆积。 - **Think-Then-Recommend（TTR）推理体系的扩展**： - **多步推理规划（Multi-step reasoning）**：针对“刚搬到新城市”等复杂生活转折，自动制定递进式的子目标探索链条。 - **用户全旅程推荐（User journey fulfillment）**：为用户不同生命阶段的需求提供具有温度和连贯性的个性化触达。 - **智能工具编排（Tool orchestration）**：让 TTR 作为核心大脑，协调底层推荐池中各种异构检索器、排序模型与业务工具，以全局视角完成复杂意图的交付。 - **规模化生产与工程效率（Efficiency and Scale）**：大模型与智能体化组件的引入必然伴随巨大的算力成本开销。在十亿级用户体量下，系统无法直接承受未经优化的重型推理链。团队正全力通过模型蒸馏（distillation）、专属微调（fine-tuning）以及底层 GPU 自动化集群调度等工程手段，大幅压缩算力损耗，确保 AI Agent 推荐系统在高投资回报率（high ROI）的健康区间内长期稳定运转。演讲最后，团队也向行业同仁发出热情邀约，共同探索智能体推荐系统的下一代范式。 ## Notable quotes > [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) like a dream come true — 内测用户 > [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) Dear Algorithm > [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) 我想多看点这个，少看点那个。 — 用户 > [02:40](https://www.youtube.com/watch?v=y-FiLm__jVY&t=160s) 对算法拥有绝对控制权的体验早该到来 — 科技媒体 > [06:57](https://www.youtube.com/watch?v=y-FiLm__jVY&t=417s) Night Night > [11:34](https://www.youtube.com/watch?v=y-FiLm__jVY&t=694s) soccer — 用户 > [13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) 我想变得更健康（I want to be healthier） — 用户 > [13:22](https://www.youtube.com/watch?v=y-FiLm__jVY&t=802s) 我刚刚搬到了旧金山 — 用户 > [14:42](https://www.youtube.com/watch?v=y-FiLm__jVY&t=882s) 我们的北极星愿景是造出一辆汽车，但我们不会一开始就去分别制造轮子、车门或车窗；相反，我们选择先做滑板车，接着升级成自行车，再到摩托车，最后走向汽车。 — Qi Guo ## Mentioned Meta, Qi Guo, Tune-Your-Algorithm (TYA), MRS Memory System (Biography), Think-Then-Recommend (TTR), 黄仁勋 (Jensen Huang), 英伟达 (NVIDIA), 马克·扎克伯格 (Mark Zuckerberg), 埃隆·马斯克 (Elon Musk), Instagram, Facebook, Threads, X (原 Twitter), TikTok, 斯蒂芬·库里 (Stephen Curry), 金州勇士队 (Golden State Warriors), 迈阿密热火队 (Miami Heat), 迈阿密国际 (Inter Miami CF), 莱昂内尔·梅西 (Lionel Messi), 克里斯蒂亚诺·罗纳尔多 (Cristiano Ronaldo), 皇家马德里 (Real Madrid), 泰勒·斯威夫特 (Taylor Swift), 特拉维斯·凯尔西 (Travis Kelce), 《金装律师》 (Suits), 《生活大爆炸》 (The Big Bang Theory), NBA, MLS, 超级碗 (Super Bowl) ## Video description Users have long wanted to understand and control the algorithms that shape their recommendations, but enabling meaningful user agency over recommendations has remained challenginging — until now. We present Tune-Your-Algorithm (TYA), an AI-powered agentic recommendation system on Instagram that gives users transparent visibility into the recommendation algorithm and the ability to tune it using natural languages. TYA is built on two key innovations: (1) the MRS Memory System (Biography), an LLM-based framework that summarizes user engagement histories into rich, structured, and interpretable user interest and intent representations at scale; and (2) Think-Then-Recommend (TTR), a reasoning-augmented approach that decomposes user interests and complex user intents into personalized sub-goals for personalized and contextualized recommendations. Early results show strong product-market fit with positive user feedback on transparency and user agency enabled by TYA. We discuss the end-to-end architecture, production learnings, technical challenges we are actively tackling, and the path towards our north star vision. Learn more about the @Scale conferences here: atscaleconference.com [Source](https://www.youtube.com/watch?v=y-FiLm__jVY)

Source quotes

[2:40] 内测用户: like a dream come true [2:40] Dear Algorithm [2:40] 用户: 我想多看点这个，少看点那个。 [2:40] 科技媒体: 对算法拥有绝对控制权的体验早该到来 [6:57] Night Night [11:34] 用户: soccer [13:22] 用户: 我想变得更健康（I want to be healthier） [13:22] 用户: 我刚刚搬到了旧金山 [14:42] Qi Guo: 我们的北极星愿景是造出一辆汽车，但我们不会一开始就去分别制造轮子、车门或车窗；相反，我们选择先做滑板车，接着升级成自行车，再到摩托车，最后走向汽车。

Source quotes

[0:05] Qi Guo: one size doesn't really fit all [2:40] Hey, I want to see more of this. I want to see less of that. [4:23] like a dream come true [4:23] Aha moment [4:23] a new level of control that feels long overdue [13:22] I want to be healthier [13:22] I just moved to San Francisco [14:42] Qi Guo: The North Star is building a car, but we started by building a scooter, a bike, and a motorbike first, instead of building isolated wheels, doors, and windows.

Original transcript

[0:05] Welcome in. Today I'm super excited to [0:07] share with you something called Tune [0:08] Your Algorithm, [0:10] which is an AI-powered recommendation [0:12] system that allow users to tune the [0:14] algorithm using natural languages. [0:18] Rexis, or recommendation system, has [0:20] been the most important AI application [0:22] for last 15 years, as Jensen Huang, [0:25] Nvidia CEO, called out. It's everywhere. [0:28] Right? If you think about it, all the [0:29] video recommendations as targeting, [0:32] shopping, or movie recommendation, it's [0:34] all based on recommendation systems. [0:37] It's also served as economic backbone. [0:40] As of 2025, it accounts for about $1 [0:42] trillion in revenue across major digital [0:45] platforms. [0:47] Yet, it's still growing very fast. [0:49] In the next 10 years, [0:51] there's 35% growth year over year for [0:55] enterprises to use recommendation engine [0:58] as cloud and AI services. [1:01] Over years, [1:03] there's a lot of evolution for [1:04] recommendation system from rules to [1:07] intelligence. [1:08] In the '90s, we have rule-based systems. [1:12] In the 2000s, it was matrix [1:14] factorization. [1:15] In 2010s, we start entering the AI era [1:19] with deep learnings. [1:21] In 2020s, [1:23] as LLMs, large language models, become [1:26] popular, [1:28] people start exploring LLMs [1:31] for recommendation system. [1:32] And, you know, transformers [1:35] that inspired by LLMs has been, [1:38] you know, tremendously powerful for [1:40] recommendation systems. [1:43] Yet, recommendation system today still [1:46] heavily relies on something called [1:48] collaborative filtering, or CF, [1:50] uh which recommend content based on [1:52] co-engagement from others, which is very [1:54] powerful, but it forms a black box echo [1:57] chamber. [1:59] So, users can only see what others [2:01] people see in the ecosystem, which is [2:03] insufficient for, you know, satisfying [2:06] users' dynamic user needs. [2:08] Right? Users like a video for different [2:10] reasons. And they come to a platform [2:13] with different motivations. For example, [2:15] users come here maybe for entertaining. [2:17] Some users might come here to stay up to [2:19] date for news. Some others might come [2:23] here to learn about certain topics. [2:25] So, one size doesn't really fit all. [2:28] If you use collaborative filtering, [2:31] then we're going to have a one-size [2:32] system for recommendation that wouldn't [2:35] satisfy the user needs that are very [2:37] diverse and dynamic. [2:40] So, why do we want to tune our [2:41] algorithm? As mentioned, this is a need [2:44] latent demand [2:46] even from a decade ago. And more [2:48] recently, this has become very sort of [2:51] verbal from some of the users. [2:53] The algo, for example, is a popular meme [2:55] on Threads, where users tells the [2:59] recommendation system by posting letters [3:02] in natural languages to say, "Hey, I [3:04] want to see more of this. I want to see [3:06] less of that." Right? And this is not an [3:08] isolated behavior. On a platform like X, [3:11] Elon Musk is also respond to some of the [3:13] users [3:14] who like to change the algorithms using [3:17] natural languages. [3:19] As a matter of fact, [3:20] Meta CEO Mark Zuckerberg [3:23] called out in the near future [3:25] users can talk to Meta [3:27] and recommendation system can be even [3:29] generated from pixel level. [3:32] I'm Chi [3:33] from Meta recommendation system. [3:36] This is a project I've been working on [3:38] across [3:39] like with my colleagues across MRS, Meta [3:41] recommendation system, along with our [3:44] colleagues from Instagram [3:46] and Facebook. [3:48] And this is such a challenging problem. [3:51] So, we're really excited about our [3:52] progress here. [3:53] So, we start with [3:55] introducing our proposal on Tune Your [3:57] Algorithm. We're going to talk about the [3:59] key components [4:00] including MRS memory system [4:03] which we call Biography [4:05] and a system that recommend paradigm [4:07] which use the memory system to do [4:09] personalized recommendation [4:12] in response to user steering. [4:14] I'm going to also discuss our [4:16] thoughts on the future [4:19] for this agentic recommendation system. [4:22] To begin with, let's talk about Tune [4:23] Your Algorithm. [4:26] Here is our solution [4:28] Tune Your Algorithm or TYA powered by [4:31] the MRS memory system called Biography. [4:34] As users come to the UI, they can start [4:36] with seeing your algorithms which are [4:39] represented in natural language as [4:41] summaries of what you're up to lately [4:44] along with interest pills that's [4:46] summarizing the system's understanding [4:48] of user's interest in keywords. [4:50] The users can tune your algorithms [4:53] by add or remove their interest pills. [4:56] And here comes third step number three. [4:59] The recommendation system will sync your [5:01] algorithm after user edits as the new [5:04] contract for the recommendation. [5:08] As of now, TYA shows pretty strong [5:10] product market fit. [5:12] It's still in a very early stage. [5:14] We receive overwhelming positive [5:16] feedback from users and we see strong [5:18] growth over time especially among young [5:21] audience. [5:23] So, here are some example quotes from [5:25] our users. [5:27] Users are loving having better control [5:30] from TYA. [5:31] Some of the users says, "It's like a [5:33] dream come true." And some users call [5:35] out, you know, the old experience of [5:37] following topics such as other platforms [5:39] like X were quite limited and this [5:43] innovation removes that particular [5:44] constraint. And people have a moments by [5:47] this immediacy and accuracy from the TYA [5:50] experience. [5:51] They really like they can control the [5:54] feed experience [5:55] by clicking a button and everything [5:57] change immediately. [5:59] And they found it to be really cool. And [6:01] it's not like the old experiences used [6:03] to have to scroll and curate their feed [6:06] using a specific implicit feedback like, [6:08] you know, click and views. [6:11] Media and press also give a lot of, you [6:14] know, positive feedback for this [6:15] experience. [6:17] Some of them called out this is a new [6:18] level control that feel long overdue. [6:22] Right? And then some of them also called [6:24] out understanding algorithm is probably [6:27] most important [6:28] but least understood part of the [6:30] recommendation system. So, this is a [6:31] very big step towards that. Some of the [6:33] press also called out the leadership [6:35] from TYA [6:37] paradigm, right? Unlike TikTok, only a [6:40] short list of fixed topics are presented [6:43] to user, with TYA user kit express [6:46] their interest [6:48] in a unconstrained way [6:51] with a very personalized list of [6:53] interest that they can add it to the TYA [6:55] experience. [6:57] Under the hood, this is how the TYA [6:59] system look like. It's a agentic loop of [7:02] five steps. [7:04] The step number one, user input [7:06] is, you know, what we showed earlier in [7:07] the slides of how you should interact [7:09] with the TYA system. [7:11] Once the input from user is received by [7:14] the TYA system [7:15] the biography memory system will take [7:17] that user input [7:19] and synthesize it with what we know [7:22] about user in the current system. [7:24] And then after that we go to step number [7:26] three [7:27] the sting that sting that recommend or [7:29] TDR to do the additional reasoning [7:32] to connect user input and the memory [7:35] system [7:36] to provide personalized recommendation. [7:39] And that goes to step number four where [7:41] we have retrieval ranking [7:44] that leads to the output that's [7:46] personalized to this particular request. [7:51] And now let's talk a little bit more [7:52] about the MRS memory system, what we [7:55] call biography. [7:57] So this is really a centerpiece of [8:00] the 20 algorithm as your algorithm. [8:03] What we've done here is turn everything [8:05] that we know about users, every [8:07] interaction the user has in a system, [8:10] into something an LLM or large language [8:12] model can understand. [8:14] Right here's the example of engagement [8:16] history for the users. [8:17] We can see that this user is interested [8:19] in, you know, looking at TV shows like [8:21] Suits and Big Bang Theory, and [8:23] interested in basketball and soccer [8:25] videos. And then we can take this [8:27] engagement and turn it to natural [8:29] languages [8:31] as illustrated at the bottom. [8:33] We can say, "Hey, this is, you know, a [8:35] user of this particular profile, and [8:38] they watched this video for about XYZ [8:41] for, you know, several of the time. [8:44] They liked the other photos [8:46] about XYZ, [8:48] you know, and and so on and so forth." [8:50] With this sort of natural languages [8:52] representation of the user's history, [8:55] then we can run LLM to summarize and [8:58] have an in-depth understanding of the [8:59] user. [9:01] On the right-hand side, we have a [9:04] flowchart to illustrate the biography [9:06] system. [9:07] On the left, we have different data [9:09] sources such as the video you used to [9:12] watch, the groups they joined, the [9:14] creator they follow, the searches they [9:16] perform, [9:18] everything about a user, and we turn [9:19] them into text as we listed on the [9:21] left-hand side, [9:22] and goes through the biography LLM [9:24] model, [9:25] and that generates a hierarchical [9:27] representation of the user. Ranges from [9:30] topic subtopic such as basketball, [9:33] soccer, TV shows, to keywords, named [9:37] entities such as Stephen Curry, NBA, [9:41] suits, the TV shows, or Ronaldo for [9:45] soccer. [9:46] And then we have more granular [9:47] representation of users' interest, such [9:50] as, you know, Curry's famous 99 [9:53] gestures. These are particular instance [9:56] users will be interested at granular [9:58] interest level. [10:02] As you can see, LLM can have very [10:05] in-depth understanding of a user. [10:07] However, it's very expensive if you [10:09] don't do additional optimization for [10:11] LLM. [10:12] To this end, we spend a lot of time [10:15] improving biography modeling through [10:17] distillation and fine-tuning. Here's a [10:19] graph that we can show the hill-climbing [10:21] of biography model on 6-month period [10:23] time. What's worth noting is that with [10:26] this approach, we were able to reduce [10:29] the cost by 100 times with distillation [10:32] and with the TYA feedback loop, [10:35] training within domain data, [10:37] we were able to outperform a model 100 [10:39] times bigger [10:41] with a smaller student model from [10:42] biography. [10:45] Biography LLM can also facilitate [10:48] interest discovery in addition to [10:51] interest summary. [10:52] In early example, whenever user is a [10:54] Stephen Curry fan, from there we can [10:57] find nearby interests such as the Golden [10:59] State Warriors, NBA, other teams of NBA [11:03] such as Miami Heat for recommendations. [11:05] And going from Miami Heat, we might also [11:07] recommend, you know, other sports for [11:10] the area such as MLS teams, [11:14] Inter Miami for example, and then [11:16] recommending Messi, which users might be [11:18] interested as well. [11:20] Similarly, you know, for a Super Bowl, [11:23] you know, video that we see users [11:24] interested in, [11:26] we might also recommend Taylor Swift and [11:28] Travis. [11:29] Those are related entity for the Super [11:31] Bowl interest. [11:34] Next, let's talk about think and [11:35] recommend, which is a critical step to [11:38] make use of the memory system from [11:40] biography for three hours of [11:42] recommendation. [11:44] Think about this particular example [11:45] here. If users say I want to see more [11:47] soccer, this could actually mean [11:49] different things for different users. [11:52] For a Messi fan, this might lead to MLS, [11:56] Inter Miami, Argentina teams, and the [11:59] World Cup news. [12:01] But for an athlete and parent, soccer [12:04] might mean soccer training videos, [12:06] different drills that can help them [12:08] improve their skills. [12:10] And for Real Madrid fan, for example, [12:13] highlights about Ronaldo [12:15] is probably more relevant and the [12:16] Portuguese national team [12:19] for the World Cup is probably more [12:20] relevant to them. [12:22] So, this information was captured [12:23] through biography. [12:25] And with biography, [12:27] then combined with [12:29] TTR approach, we can generate more [12:32] personalized recommendation. So, here's [12:33] the illustration of TTR. On the left [12:35] hand side, we have the user prompt, in [12:37] this case, soccer. [12:38] And then on the right hand side, we have [12:39] all the different user profile generated [12:42] by biography memory system. With the TTR [12:45] and reasoning, we can then [12:48] generate these reasoning factors to [12:49] better represent [12:52] the soccer intent and retrieve nearby [12:55] content [12:57] in the same embedding space [12:59] at billion scale for recommendation. [13:02] This is a lot of illustration of [13:04] flowchart where, you know, starting from [13:05] the user prompt, soccer, adding the user [13:08] memory from biography, we have this [13:10] thinking blob that generate the [13:12] embedding for different users. Like you [13:14] can see, for the different user profile [13:16] in different colors, we can generate [13:18] corresponding recommendations after TTR [13:21] steps. [13:22] TTR can also support more complex needs [13:25] and intents, such like I want to be [13:27] healthier or I want to learn about [13:29] soccer. [13:31] I, you know, just moved to San [13:32] Francisco, right? Take I want to be [13:34] healthier, for example. [13:36] Let's say this is a user of 30 years old [13:38] living in San Francisco and they are [13:39] interested in Stephen Curry and AI. They [13:42] now come to UI system and say something [13:44] like I want to be healthier. [13:46] What are the things that we can [13:47] recommend to this user? [13:49] With a thinking step, we can first of [13:51] all break down [13:52] I want to be healthier into sub intents [13:55] or sub goals. Like I want to be [13:57] exercising more, you know, physical [13:58] fitness. I want to be eating better, [14:01] health and diet. And the user might also [14:03] want to sleep better and just have [14:06] better habits and so on and so forth. [14:09] Now, in the recommendation phase, we can [14:11] take users personal preferences into [14:13] account. For example, recommending [14:15] workouts for users physical fitness [14:18] instead of recommending arbitrary videos [14:21] of workouts, we can recommend basketball [14:24] videos because because user's interested [14:26] in Stephen Curry. Similarly, we can also [14:27] recommend smart finish technology given [14:30] user's interest in AI and tech. And as [14:33] we sort of have this thinking step of [14:35] revisement of the interest, then we can [14:38] recommend the videos using the embedding [14:40] approach. [14:42] That's a quick summary of the TY system [14:44] and the two key components. As we're [14:46] moving towards the future of this [14:48] agentic recommendation system, [14:51] we're thinking about evolving from our [14:53] MVP, the minimum viable product, [14:56] towards the North Star vision. [14:58] We made a very dedicated decision to [15:01] start simple. [15:02] You know, as hindsight shows, this is [15:04] the MVP mindset. The North Star is to [15:06] building a car, but we'll start building [15:08] with a scooter, a bike, a motorbike [15:11] instead of just building wheels and [15:13] doors and windows. [15:15] So, this way we can really learn about [15:17] the experiences and sort of [15:19] understanding where things would need to [15:20] improve so on and so forth. So, in the [15:23] case of MVP for the TYA experience, we [15:25] start with something people are more [15:27] familiar with, which is interest PO, [15:29] even though we're still opening it up [15:31] for people to express their interest in [15:35] with natural languages, [15:36] but the UI is familiar, like topic [15:38] following, you know, as we mentioned [15:40] earlier, which is something many other [15:43] platforms like TikTok and X already [15:44] provide. Meta in some older versions of [15:47] the system also provide that, but it's [15:49] very limited. It's only a fixed set of [15:51] topics and it's not personalized. [15:53] Now that we've proven this is something [15:55] user resonate, [15:57] we are more marching towards this North [15:59] Star vision, where we allow multi-turn [16:02] conversation [16:04] for the users [16:05] that can express their interest [16:07] and preferences beyond interest. [16:12] So, as next steps, [16:13] we are building towards this next-gen [16:15] TYA system. [16:17] These are a few areas that we'll be [16:20] focusing on. For the North Star, we're [16:22] thinking about two key capabilities. [16:24] Number one, we want to make sure the TYA [16:27] system can receive multi-turn [16:28] conversational inputs from the user, [16:31] which is something pretty standard these [16:33] days for an AI system, and this needs to [16:35] be persistent across sessions, so the [16:39] memory system from biography is going to [16:40] be critical. [16:42] Number two, we want to make sure, even [16:45] without user talking, the system [16:47] operating in the background continuously [16:49] to help user discover their interest [16:51] and detect the underlying intents that [16:53] they have and provide personalized [16:55] fulfillments. [16:57] So, given these two sort of [16:59] goals, [17:00] we need to build out some core [17:01] capabilities. [17:03] And here's a breakdown of that. For MI's [17:06] memory system biography, we need to have [17:08] a persistent cross-sessional memory [17:10] that can be useful to enable this [17:12] multi-turn conversation TYA. [17:15] And we also need to continuously learn [17:16] from users [17:18] implicit and explicit behaviors to [17:20] facilitate the interest discovery. [17:23] And we also need to go beyond interest [17:26] to detect users' interests [17:28] and with life cycle management. As some [17:30] of the intents are long-lasting, some of [17:32] them may be short-term. For example, if [17:33] users are traveling. [17:35] And in terms of think then recommend [17:37] TTR, we also need to make sure we have [17:39] the phone capabilities such as [17:41] multi-step reasoning to stop to support [17:44] the complex information needs [17:46] just mentioned earlier like I just moved [17:49] to San Francisco, that can break it down [17:51] into subgoals [17:53] that are personalized for fulfillment. [17:55] And we also like to make sure we can [17:57] build out a user journey for [17:58] recommendation that can provide personal [18:01] touches of fulfillment for these [18:03] different intents. [18:05] Last but not least, we also need to make [18:06] sure [18:08] the TTR system can really coordinate and [18:10] orchestrate all the tools for the [18:13] recommendation system so that we can [18:14] fulfill users' goals in the most [18:16] holistic way. [18:18] Efficiency and scale they still critical [18:21] as we move on to build up different AI [18:23] and components. [18:25] The cost is going to be higher, you [18:27] know, compared to a traditional system. [18:29] And how to go through automation such as [18:32] distillation, fine-tuning, and GPU [18:35] automation to put everything into a high [18:37] ROI space, this is another very critical [18:40] area for next generation AI gently [18:43] recommendation system. This is something [18:45] we are really excited about. If you are [18:47] also interested, please reach out.