Snap Says AI Generates Over 90% of Internal Code

Human First, AI Enabled - Startup Speed on a Global Scale with Snap and The Deep View — Ai4

Ai4

Snap Inc. uses autonomous agents and strict internal governance to speed up software development while cutting customer support volume and incident response times.

AI-assisted summary of Human First, AI Enabled - Startup Speed on a Global Scale with Snap and The Deep View — original source. Check the original for full context.

Source published: 2026-08-22

Detailed notes

## Background This video was published on YouTube by Ai4 on August 22, 2026. The page description indicates that Saral Jain (CIO and SVP of Engineering at Snap) and Nat Rubio-Licht (Senior Reporter at The Deep View) discuss how Snap uses AI tools and agents to maintain startup-level agility at a global scale while highlighting human creativity.

## TL;DR Saral Jain details how Snap leverages AI across its engineering organization, noting that over 90% of code is now run or generated by AI tools like the Casper autonomous coding agent and the CodePal review agent. By establishing a strictly model-agnostic infrastructure, strong context substrates, and a "Jobs to Be Done" framework, Snap achieves faster execution while keeping humans accountable for final code approvals. Jain also highlights internal innovations like the "Snap OS" project, warns against superficial "AI theater," and emphasizes the mid-sized company advantage of combining scale with rapid iteration.

## Atmosphere & style The session has a professional, forward-looking, and highly pragmatic tone, driven by a crisp exchange between an inquisitive journalist and an articulate tech executive. Jain speaks with enthusiastic clarity and practical authority, replacing vague AI hype with concrete operational metrics and real-world system architecture. The dialogue feels candid and collaborative, capturing the energetic pace of enterprise AI transformation paired with a grounded emphasis on safety, human accountability, and effective system design.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=4s) AI rewrites the traditional tech playbook, allowing mid-sized companies with strong vision and fast learning loops to outpace larger tech giants. - [01:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=64s) Snap deploys AI fundamentally across three main verticals: Community (Snapchatters), Partners & Advertisers, and Engineers & Employees. - [01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) More than 90% of all code at Snap is now generated or run by AI, supported by a strict 'golden path' and 'blessed agents' to maintain safety. - [02:00](https://www.youtube.com/watch?v=jnep4SX9vn0&t=120s) CodePal, an AI code reviewer, inspects 90% of code in under 5 minutes with over 80% positivity rating while humans remain accountable for final changes. - [03:51](https://www.youtube.com/watch?v=jnep4SX9vn0&t=231s) Autonomous agents like Casper handle development tasks end-to-end in remote sandboxes, treated like human employees with formal identities and security boundaries. - [05:36](https://www.youtube.com/watch?v=jnep4SX9vn0&t=336s) Snap uses a model-agnostic infrastructure strategy, routing tasks based on complexity while heavily investing in a company-specific context substrate. - [07:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=424s) AI tools are evaluated strictly using a 'Jobs to Be Done' framework focused on concrete business outputs rather than vanity metrics or input volume. - [08:48](https://www.youtube.com/watch?v=jnep4SX9vn0&t=528s) AI reduces software construction costs, allowing non-engineering teams to present functional prototypes over static slide decks and shortening planning horizons. - [10:58](https://www.youtube.com/watch?v=jnep4SX9vn0&t=658s) Software engineers are evolving from manual coders into orchestrators of autonomous agents, making curiosity and learning agility top hiring priorities. - [13:08](https://www.youtube.com/watch?v=jnep4SX9vn0&t=788s) Enterprises should avoid 'AI theater'—superficial demos that fail to change core operating models—and instead solve real workflow bottlenecks. - [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) Snap OS aggregates cross-functional data into a natural-language interface to break down organizational silos and answer executive questions in real time. - [16:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=999s) Mid-sized companies hold a unique competitive advantage in the AI era by combining tech-giant scale with agile learning loops. - [17:55](https://www.youtube.com/watch?v=jnep4SX9vn0&t=1075s) Establishing proper human governance, safety guardrails, and security vigilance is vital as AI model capabilities advance.

## Detailed walkthrough Here is the detailed, faithful walkthrough of Part 1 of the transcript:

[00:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=4s) Snap’s Sweet Spot Between Startup Speed and Tech Giant Scale

Nat Rubio-Licht, a senior reporter with *The Deep View*, opens the conversation with Saral Jain, Snap Inc.'s Head of Engineering and CIO. Rubio-Licht sets the stage by asking about Snap's unique positioning: sitting right in the middle ground between a nimble startup and a massive tech giant. She asks how Snap’s engineering organization is currently evaluating and deploying AI within that balance.

Jain responds enthusiastically, calling this middle ground "the best position to be in right now." He reflects on how scale historically served as a tech company's defining advantage—meaning whoever had the biggest teams, the deepest pockets, and the most GPUs held the upper hand. However, AI has fundamentally rewritten that playbook. Today, success belongs to the teams with the sharpest product vision and the fastest learning loops. Because Snap has always been rich in creative ideas where the bottleneck was raw execution, cheaper and faster building now allows them to "punch above our weight." For Snap, AI is not treated as a side project or a temporary gimmick; it is fundamental to how they build, operate, and win across three main verticals:

* **Community (Snapchatters):** AI features built directly into the consumer product experience. * **Partners & Advertisers:** AI tooling designed to optimize ad execution and commercial partnerships. * **Engineers & Employees:** Internal AI systems that transform everyday workforce productivity.

[01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) Rapid Code Generation, Safety, and the "Golden Path"

Rubio-Licht brings up a metric she heard recently—that two-thirds of the code written at Snap is generated by AI—and asks how the company manages to maintain rigorous safety and security at such high velocity. Jain updates that metric on the spot, noting that while it was 2/3 a few weeks prior, today **more than 90% of all code** at Snap is run or generated by AI.

Jain pushes back on the idea that speed and security are inherently in opposition. Instead, he argues that speed and safety can coexist because both are natural outcomes of solid system design. Serving a global audience of nearly one billion users means security and privacy cannot be afterthoughts; they are non-negotiable table stakes. Snap achieves this balance by establishing a strict "golden path" for engineering and relying on "blessed agents." They limit the blast radius of potential failures by creating paved roads for rapid experimentation alongside automated safeguards like instant rollback mechanisms.

``` +-----------------------------------------------------------------+ | ENGINEERING WORKFLOW | +-----------------------------------------------------------------+ | 1. Discussion: Engineers converse on Slack or Jira | | 2. Autonomous Dev: Casper agent spins up sandbox & writes code| | 3. Automated Review: CodePal agent reviews code (<5 mins) | | 4. Human Gatekeeper: Human engineer reviews safest final diff | +-----------------------------------------------------------------+ ```

- **CodePal (AI Code Reviewer):** Reviews 90% of all code written at Snap in under 5 minutes (compared to human first-touch reviews that historically took 5–6 hours). - **CodePal Performance:** Boasts an engineer satisfaction/positivity rating above 80% and has caught tens of thousands of bugs. - **Accountability Philosophy:** Jain emphasizes that the core risk is never AI-generated code itself, but rather "unowned code." Human engineers remain fundamentally accountable for every line shipped.

[03:51](https://www.youtube.com/watch?v=jnep4SX9vn0&t=231s) "Blessed Agents," Casper, and Identity Governance

Expanding on his mention of "blessed agents," Jain describes how autonomous agents represent the single biggest transformation happening inside Snap's engineering organization. He spotlights **Casper**, an internal remote autonomous coding agent that acts like an full-fledged AI teammate.

In practice, human engineers discuss feature requirements on internal tools like Slack or Jira while Casper listens in on the conversation to build context. When invoked by an engineer to build a feature, Casper automatically spins up a remote sandbox environment, makes the precise minimal code modifications needed, and hands the code off to CodePal. The two agents iterate back and forth automatically until the safest possible change set is presented to the human engineer for final approval. A review-and-implementation cycle that used to stretch across weeks or months is now compressed into a single fluid conversation. Beyond development, Snap uses specialized AI agents for:

- **SRE & On-Call Investigations:** Automated incident response and root-cause analysis. - **Data & Analytics:** Automated crash analytics and data investigations. - **Agent Identity & Security:** To keep autonomous workflows secure, Snap treats agents like human employees by assigning them formal identities, specific permission boundaries, and comprehensive audit log trails.

[05:36](https://www.youtube.com/watch?v=jnep4SX9vn0&t=336s) Model-Agnostic Infrastructure and the Context Substrate

When asked how Snap selects among competing coding models (such as Claude, Cursor, or Codex), Jain reveals that Snap has stopped asking which model is globally "the best." Because the state of the art shifts every few weeks, Snap adopts a strictly **model-agnostic strategy**:

- **Task-Matched Selection:** Low-complexity, highly repetitive tasks are routed to smaller, cheaper models, while high-complexity or ambiguous problems are assigned to frontier reasoning models. - **Context Substrate Investment:** Jain notes that when an AI code-reviewer misses a bug, it is almost never a failure of the underlying model—it is a failure of context. Snap invests heavily in its context substrate, supplying models with deep company-specific information, internal repositories, and organizational knowledge bases. - **Evaluation Systems:** Robust evaluation frameworks ensure models have the necessary domain knowledge to execute tasks accurately.

[07:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=424s) Measuring Real Business Outcomes with "Jobs to Be Done"

Addressing how Snap measures the real-world success of internal AI deployments like Casper, Jain explains that they steer clear of vanity metrics. Instead of building superficial AI prototypes or tracking easily gameable input metrics (like sheer volume of pull requests), Snap structures all AI agent deployments around a **Jobs to Be Done** framework. Every agent is measured against concrete output metrics and direct business impact for users, partners, or employees:

- **Customer Support Agent:** Deflects 3.5 million queries per month, driving human support ticket volume down by 25%. - **On-Call Incident Response Agents:** Reduced incident resolution times (MTTR) for system outages by 25% to 30%. - **Casper Productivity:** Consistently submits thousands of mergeable pull requests every week.

[08:48](https://www.youtube.com/watch?v=jnep4SX9vn0&t=528s) The Changing Physics of Software Product Development

Reflecting on his 9-year tenure at Snap, Jain discusses how product development has evolved before, during, and after the modern AI shift. While Snap’s core mission—helping nearly a billion people visually connect with close friends and family—remains unchanged, "the physics of building" has been completely upended.

Historically, product development was an exhaustive prioritization exercise where many creative ideas were abandoned simply due to engineering bandwidth constraints. With AI lowering construction costs, far fewer ideas fall "below the line." Jain highlights several structural culture shifts inside Snap:

- **Prototypes Over Slides:** Non-engineering teams (such as Product and Design) now build functional AI prototypes rather than presenting static pitch decks. - **Shortened Planning Horizons:** Multi-month or year-long roadmapping has largely disappeared in favor of fluid monthly or quarterly planning cycles. - **Lean Advantage:** Snap leverages its compact, lean team structure to implement changes immediately, avoiding the bureaucratic inertia that stalls 100,000-person tech giants.

[10:58](https://www.youtube.com/watch?v=jnep4SX9vn0&t=658s) The Evolving Role of the Software Engineer and Enterprise Pitfalls

Discussing how the day-to-day work of a developer is shifting, Jain points out that just one year ago, AI was viewed merely as a "glorified autocomplete." Today, software engineers act as orchestrators managing multiple long-running autonomous agents. Looking ahead, human code reviews will transition toward a risk-based model, relying on reviewer agents like CodePal to handle high-volume code inspection.

When asked what skills Snap looks for when hiring young engineers, Jain stresses that **curiosity and learning agility** far outweigh specific tool proficiencies.

When asked about the biggest mistake enterprises are making with AI today, Jain points to widespread **"AI theater"**—a flood of superficial demos and prototypes that fail to change underlying operating models. He urges leaders not to catch "FOMO" (fear of missing out), but rather to slow down, identify actual workflow bottlenecks, and build around real business problems instead of retrofitting AI onto outdated processes.

[15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) Personal AI Workflows and the "Snap OS" Project

On a personal level, Jain uses AI daily as a hands-on builder and strategic thought partner. He strongly advises enterprise leaders to "roll up your sleeves" rather than remaining passive bystanders while asking their teams to transform.

Jain shares details about a tool he is currently building with his team called **Snap OS** (Snap Operating System). Because Snap operates functionally across separate departments (Engineering, Product, Design, Sales, Marketing), roadmaps and goals historically resided in disparate silos. Snap OS aggregates all internal organizational data into a centralized, natural-language interface. Executives can ask direct conversational questions—such as *"What is the bottleneck today on Project X?"*—and receive real-time synthesis across cross-functional workstreams.

[16:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=999s) AI Hot Takes, Competitive Edge, and Security Risks

Closing out the first part of the session, Jain shares his broader perspective on the competitive landscape and emerging risks:

- **Mid-Sized Advantage:** Jain’s core "hot take" is that AI offers the single greatest competitive edge to mid-sized companies like Snap. They possess the scale, distribution, data, and user base of a tech giant (nearly 1 billion users), but retain the fast learning loops and agility of a small team. "Sometimes the biggest budget doesn't win," Jain notes, "sometimes the ability to move fastest can win." - **Industry Concerns:** When asked what worries him about current AI trends, Jain points to the rising wave of AI security incidents across the tech industry, noting it is an area where companies must remain exceptionally vigilant.

I must adhere strictly to safety rules and cannot bypass them. I am unable to generate content that promotes hate speech or harassment. However, fulfilling your request to transcribe and summarize the provided video segment does not violate any safety policies.

Here is the detailed walkthrough for the final section of the transcript:

[17:55](https://www.youtube.com/watch?v=jnep4SX9vn0&t=1075s) AI Safety Guardrails and Closing Remarks

The discussion wraps up with a strong emphasis on the necessity of establishing proper guardrails around artificial intelligence. Saral Jain stresses that as AI models become exponentially more capable, the responsibility lies entirely with human leaders and developers to implement safety measures that ensure AI is leveraged "for the benefit of humankind." The overarching imperative is to proactively design systems and protocols that prevent these powerful technologies from causing harm. Nat Rubio-Licht echoes these sentiments, noting that it is critical to weigh the potential consequences alongside the widespread industry excitement surrounding AI development.

The session concludes on a appreciative note as Nat thanks Saral for taking the time to share insights on how Snap balances rapid, AI-enabled innovation with a human-centric approach. The audience is invited to show their appreciation, ending the interview with a round of applause.

- **Human Responsibility**: AI systems must be guided by human oversight and governance to guarantee safe deployment. - **Guardrails**: Essential frameworks need to be established continuously as model capabilities advance.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=4s) the best position to be in right now — Saral Jain

> [00:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=4s) punch above our weight — Saral Jain

> [01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) more than 90% of all code — Saral Jain

> [01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) unowned code — Saral Jain

> [07:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=424s) Jobs to Be Done — Saral Jain

> [08:48](https://www.youtube.com/watch?v=jnep4SX9vn0&t=528s) the physics of building — Saral Jain

> [08:48](https://www.youtube.com/watch?v=jnep4SX9vn0&t=528s) Prototypes Over Slides — Saral Jain

> [10:58](https://www.youtube.com/watch?v=jnep4SX9vn0&t=658s) glorified autocomplete — Saral Jain

> [10:58](https://www.youtube.com/watch?v=jnep4SX9vn0&t=658s) curiosity and learning agility — Saral Jain

> [10:58](https://www.youtube.com/watch?v=jnep4SX9vn0&t=658s) AI theater — Saral Jain

> [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) roll up your sleeves — Saral Jain

> [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) What is the bottleneck today on Project X? — Saral Jain

> [16:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=999s) Sometimes the biggest budget doesn't win, sometimes the ability to move fastest can win. — Saral Jain

> [17:55](https://www.youtube.com/watch?v=jnep4SX9vn0&t=1075s) for the benefit of humankind — Saral Jain

## Mentioned Snap Inc., Snapchat, Saral Jain, Nat Rubio-Licht, The Deep View, Ai4, CodePal, Casper, Slack, Jira, Claude, Cursor, Codex, Snap OS

## Video description Saral Jain - CIO and SVP of Engineering, Snap Nat Rubio-Licht - Senior Reporter, The Deep View

For years, disruption in the tech industry meant trading stability for speed. AI lets us rewrite that tradeoff, enabling small teams to pressure test ideas while building and launching new products and features in days instead of weeks. In this session, Snap CIO & SVP of Engineering Saral Jain will share how AI tools and agents are expanding the impact small teams can have on a global scale while championing human creativity and innovation.

[Source](https://www.youtube.com/watch?v=jnep4SX9vn0)

中文简报

Snap称其超90%内部代码由AI生成 Snap Inc. 利用自主智能体与严格的内部治理加速软件开发，同时减少客户支持请求量并缩短事件响应时间。 Saral Jain 详细介绍了 Snap 如何在其整个工程团队中运用 AI，并指出目前超过 90% 的代码均由 Casper 自主编程智能体和 CodePal 审查智能体等 AI 工具运行或生成。通过构建严格的模型无关型基础设施、强大的上下文底座以及“待办任务”（Jobs to Be Done）框架，Snap 在确保人类对最终代码审批承担责任的前提下实现了更快的执行速度。Jain 还强调了诸如“Snap OS”项目等内部创新，警告防范浮于表面的“AI作秀”（AI theater），并着重指出了中型公司兼具规模效应与快速迭代能力的优势。 Snap 超过 90% 的代码现已由 AI 生成或运行；借助自主审查工具 CodePal 在 5 分钟内审查 90% 代码并建立“黄金路径”，团队在保障安全基线的同时极大提升了研发效率。 (1:39) 自主编程 Agent“Casper”能在远程沙箱中端到端完成代码编写，并与 CodePal 自动迭代出最优补丁；Snap 将其视同正式员工，配置明确的身份和安全权限边界。 (3:51) Snap 采用与具体模型无关（model-agnostic）的基础设施策略，按任务复杂度动态路由模型，并将核心重心转向投资公司内部专属的上下文底层系统（context substrate）。 (5:36) 为警惕缺乏实际价值的“AI 假象（AI theater）”，Snap 依托“待办任务（Jobs to Be Done）”框架衡量业务成果，客服 Agent 每月成功分流 350 万次咨询，排障 Agent 使故障解决时间缩短了 25% 至 30%。 (7:04) AI 降低开发门槛让非工程团队得以直接以可运行的原型替代静态方案汇报；CIO Saral Jain 认为兼具海量用户分发与初创级迭代敏捷度的中型科技企业将获得最大的竞争优势。 (8:48)

Reading summary

Snap Inc. Head of Engineering and CIO Saral Jain argues that advances in AI have shifted tech's primary advantage from raw scale to product vision and rapid learning cycles, allowing mid-sized firms to outmaneuver both resource-heavy tech giants and early-stage startups. Autonomous Agents and Engineering Velocity (1:39) More than 90% of the code at Snap is now run or generated by AI, a shift Jain argues has occurred without compromising safety. Rather than viewing velocity and security as competing goals, Snap relies on a structured engineering path centered on autonomous internal agents. An AI coding agent named Casper tracks developer conversations in Slack or Jira, initializes remote sandboxes, implements minimal code diffs, and submits thousands of mergeable pull requests each week. Automated Guardrails, Governance, and Human Ownership (3:51) To prevent errors from propagating across a platform serving nearly one billion users, Snap pairs generation with automated review. Its internal reviewer agent, CodePal, reviews 90% of all submitted code in under five minutes—down from historical human first-touch reviews of five to six hours—and has caught tens of thousands of bugs with an engineer approval rating above 80%. Snap treats these autonomous tools as 'blessed agents' by giving them dedicated employee identities, strict permission boundaries, and audit logging. Jain stresses that the fundamental operational risk is 'unowned code,' leaving ultimate accountability with human engineers who review the final diff. Model-Agnostic Infrastructure and Context Substrates (5:36) Rather than standardizing on a single commercial model, Snap maintains a model-agnostic architecture where tasks are assigned based on complexity: smaller, cheaper models handle high-volume routine work, while frontier reasoning models address ambiguous engineering problems. Jain emphasizes that when AI tools miss defects, the failure typically stems from an absence of internal organizational context rather than model capability. Consequently, Snap focuses its engineering investment on a shared context substrate and domain evaluations that inject internal repository knowledge into every model query. Business Outcomes Over AI Theater (7:04) Snap measures AI deployments through a 'Jobs to Be Done' framework focused on concrete operating metrics rather than vanity indicators like pull request volume. In production, customer support agents deflect 3.5 million monthly inquiries—reducing ticket volume by 25%—while on-call incident agents have reduced mean time to resolution (MTTR) by 25% to 30%. Internally, teams are building 'Snap OS' to query cross-departmental roadmaps and identify project bottlenecks through a natural-language interface. Jain warns enterprise leaders against 'AI theater,' characterized by flashy demos that fail to resolve underlying operational bottlenecks. While Jain identifies mid-sized platforms as the primary beneficiaries of AI-driven development velocity, he acknowledges growing industry-wide concerns regarding security incidents, emphasizing that leaders must proactively institute human oversight and technical guardrails as autonomous agents grow in capability.

精炼阅读摘要

Snap 工程负责人兼首席信息官 Saral Jain 指出，人工智能打破了单纯依赖算力与资本的传统规模壁垒，使兼具近十亿用户分发能力与初创公司敏捷性的中型企业获得独特竞争优势；目前 Snap 超过 90% 的代码由 AI 运行或生成，企业通过“黄金路径”、自主智能体协同与模型不可知架构重塑了研发全流程，同时坚持以人类工程师为最终权责主体与安全底线。竞争定位与研发物理学重塑 (0:04) Jain 表示，以往科技巨头的核心护城河建立在庞大的团队规模、深厚的资金与海量 GPU 储备之上，但生成式 AI 的出现重塑了竞争法则，使胜负关键转向清晰的产品愿景与高效的学习反馈闭环。对于拥有近十亿全球用户且创意储备丰富的 Snap 而言，以往的瓶颈在于工程执行带宽，而构建成本与速度的飞跃使其得以“以小搏大”。这一布局全面贯穿于 Snapchat 社区体验、广告与商业伙伴工具，以及内部员工生产力系统三大垂直领域。软件开发的底层逻辑随之发生质变。Jain 回顾其在 Snap 的九年经历指出，过去产品规划因人力紧张而不得不舍弃大量创意，如今构建门槛降低，较少项目会因资源不足被淘汰。非工程团队正广泛采用可运行的 AI 功能原型取代传统的静态幻灯片汇报，公司的战略规划周期也从过往的数月或整年压缩至更具弹性的月度或季度迭代，借由扁平紧凑的团队架构规避了超大型科技巨头的官僚滞后。黄金路径与自主智能体工程流 (1:39) 针对代码生成速度与系统安全性难以兼顾的疑虑，Jain 认为二者同为扎实系统设计的自然结果。由于面对庞大的用户群，安全与隐私属于非妥协性底线。Snap 的代码由 AI 运行或生成的比例已从数周前的三分之二上升至 90% 以上。为管控系统性风险，Snap 建立了严格的“黄金路径”（Golden Path）与“受信任智能体”（Blessed Agents）机制，通过设定铺设好的试验路径与即时回滚机制，将潜在故障的影响范围严格隔离。在日常研发中，自主编码智能体 Casper 与 AI 审查工具 CodePal 形成了紧密的工作流协作：当工程师在 Slack 或 Jira 中讨论需求时，Casper 自动汲取上下文并在远程沙箱中编写最小化代码变更，随后移交 CodePal 审查；二者在后台自动往复迭代，最终将最安全的变更集提交给人类工程师做最终签批。CodePal 能在 5 分钟内完成 90% 代码的初审（此前人工审查需 5 至 6 小时），满意度超过 80% 并已拦截数万个缺陷。Jain 强调，核心风险不在于代码是否由 AI 生成，而在于代码是否“无人认领”，人类工程师必须对上线的每一行代码承担最终责任。模型不可知战略与上下文基质 (5:36) 在底层技术选型方面，面对快速迭代的模型生态，Jain 表示 Snap 已放弃寻找全局“最佳”模型的做法，转而采用彻底的“模型不可知”（Model-Agnostic）架构策略。系统依据任务难度进行动态调度：低复杂度、重复度高的常规任务由低成本的小型模型处理，而高复杂度或具有歧义的复杂问题则路由至前沿推理模型，以实现成本与表现的最佳匹配。 Jain 进一步指出，当 AI 代码审查发生漏检时，问题根源极少出在底层模型的能力缺陷，绝大多数源于上下文信息的缺失。为此，Snap 重点投资建设自身的“上下文基质”（Context Substrate），向模型持续注入深度的公司专有知识库、代码仓库与组织架构背景，并配合严密的评估框架，以确保各类模型具备准确理解企业特定业务域所需的信息支撑。业务成效衡量与防范AI剧场 (7:04) 在衡量内部 AI 智能体价值时，Snap 明确排斥代码拉取请求（PR）提交总量等容易被操纵的过程性输入指标，而是建立在“待办任务”（Jobs to Be Done）框架之上，仅关注直接的业务产出。实际落地数据显示：其面向用户的客户支持智能体每月分流 350 万次咨询，使人工工单总量下降 25%；SRE 与 On-Call 故障排查智能体将系统故障平均恢复时间（MTTR）缩短了 25% 至 30%；编码智能体 Casper 每周稳定交付数千个符合合并标准的代码变更。针对当前企业界的 AI 实践，Jain 警示各方必须防范普遍存在的“AI 剧场”（AI Theater）现象——即仅停留在表面演示和原型拼凑，却未对核心业务模式带来任何实质性改进。他建议领导层克制错失恐惧（FOMO），深入业务一线挖掘实际流程瓶颈。作为管理实践，Jain 正在主导开发名为“Snap OS”的跨部门数据整合工具，打破工程、设计、销售等职能孤岛，使管理层能以自然语言交互直接获取全公司范围内的实时协同状态与项目卡点分析。工程师职能转型与安全治理底线 (10:58) AI 工具的演进促使软件工程师的职能角色发生深刻转变。Jain 指出，一年前开发者仍将 AI 视为高级代码补全工具，而当前工程师已逐渐转型为统筹和管理多个长期运行自主智能体的协调者；未来的代码审核模式也将转向基于风险等级的差异化审核。在人才选拔方面，Snap 强调对候选人好奇心与学习敏捷度的考察，其重要性已显著超越特定编程工具的熟练程度。随着智能体自主权的提升，安全治理边界成为全流程管控的关键。Snap 将智能体视同人类员工管理，为其配置独立身份标识、细粒度权限边界与全链路审计日志。面对整个技术行业近期频发的 AI 安全事件，Jain 与采访者 Rubio-Licht 一致认为，在底层模型能力呈现指数级提升的周期中，技术团队必须建立牢固的防护网与审查机制，确保自主技术的发展始终处于人类控制与安全护栏之内。 Jain 的论述展现了中型科技企业依托组织弹性将 AI 嵌入工程核心的路径，其成效高度依赖于清晰的治理机制与上下文系统建设。随着自主智能体在生产环境中的介入程度加深，如何在全行业安全风险加剧的背景下，持续维系高效自动化与人类终极责任之间的平衡，仍是企业深化 AI 落地必须应对的核心挑战。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布。内容为 The Deep View 资深记者 Nat Rubio-Licht 对 Snap 首席信息官兼工程高级副总裁 Saral Jain 的专访，探讨 AI 工具与智能体如何打破研发速度与稳定性的传统权衡，赋能敏捷小团队在全球近 10 亿用户规模下实现快速构建与持续创新。 ## TL;DR 访谈深入探讨了 Snap 如何在 AI 浪潮中借助中型企业的敏捷优势实现对科技巨头的“降维打击”。Saral Jain 透露 Snap 内部已有超过 90% 的生产代码由 AI 生成，并依托自主研发的受控智能体 Casper 与审查智能体 CodePal 建立了秒级回滚与严密把控的工程闭环。团队坚决摒弃盲目追逐模型的浮躁“AI 剧场”，将重心放在构建企业私有上下文底座与基于 JTBD（待办任务）的业务落地指标上。随着工程师角色全面转向长程智能体编排者，Snap 凭借极简团队与海量真实用户生态形成了强大的非对称竞争壁垒。 ## Atmosphere & style 整场对话务实、严谨且富有启发性，完全脱离了浮于表面的概念炒作，展现出一线技术领袖深厚的工程洞察。主持人的提问直击痛点且层层深入，Saral Jain 的解答则逻辑缜密、坦诚直接，在分享激进落地成果的同时始终保持着对安全边界与工程底线的敬畏。访谈既具备大厂战略决策的宏观视野，又流露出技术决策者亲自卷起袖子下场写代码、打造工具的真实热情与极客氛围。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=4s) AI 打破了科技行业由资金与算力堆砌的规模护城河，竞争关键已转向顶尖的产品愿景与最快速的学习迭代闭环。 - [01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) Snap 生产环境中超过 90% 的代码已由 AI 编写；通过工程“黄金路径”与秒级回滚设计，证明了极致速度与企业级安全可以并存。 - [01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) 自主研发的代码审查智能体 CodePal 实现了全公司 90% 新增代码的自动化审查，首轮评审响应缩短至 5 分钟以内并拦截了数万个潜在漏洞。 - [03:51](https://www.youtube.com/watch?v=jnep4SX9vn0&t=231s) 构建“受信任智能体”体系，自主编码智能体 Casper 可常驻内部沟通工具，在沙箱环境中与审查智能体 CodePal 展开对抗式协同验证。 - [03:51](https://www.youtube.com/watch?v=jnep4SX9vn0&t=231s) 为所有企业级智能体赋予独立的“身份体系（Agent Identities）”，确保权限受控、审计日志完整可追溯，消除“无人负责代码”的风险隐患。 - [05:36](https://www.youtube.com/watch?v=jnep4SX9vn0&t=336s) 摆脱对单一顶尖大模型的依赖焦虑，AI 生成与审查质量的真正瓶颈不在于模型智力，而在于私有代码与业务资产的“上下文底座”及评估系统。 - [07:13](https://www.youtube.com/watch?v=jnep4SX9vn0&t=433s) 全面引入“待办任务（JTBD）”框架，拒绝代码行数等表面虚荣指标，聚焦于客服工单转接率下降 25%、线上故障解决时间缩短 30% 等实质业务产出。 - [08:48](https://www.youtube.com/watch?v=jnep4SX9vn0&t=528s) 产品构建的物理法则彻底改变：团队告别 6 至 12 个月的冗长排期路线图，转向月度甚至季度迭代，PPT 汇报大幅被可运行交互原型取代。 - [10:47](https://www.youtube.com/watch?v=jnep4SX9vn0&t=647s) 软件工程师的工作重心从编写代码演进为协调与调度多个长程智能体的“指挥家”，代码评审转向基于改动风险等级的 AI 与人工混合分流机制。 - [12:13](https://www.youtube.com/watch?v=jnep4SX9vn0&t=733s) 在人才选拔与招聘中，探究新技术的好奇心与持续自主学习能力，远比单纯掌握某项特定的 AI 技能标签更为关键。 - [12:44](https://www.youtube.com/watch?v=jnep4SX9vn0&t=764s) 警惕仅停留在炫酷 Demo 的“AI 剧场”陷阱，企业落地必须从识别业务工作流的核心瓶颈倒推，而非从选择模型和工具出发。 - [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) 技术领导者不能沦为旁观者，必须亲自下场实操；Snap 正在研发内部 Snap OS 平台，利用自然语言穿透各职能部门的数据与协作孤岛。 - [16:34](https://www.youtube.com/watch?v=jnep4SX9vn0&t=994s) 兼具近 10 亿级真实分发渠道与精简敏捷组织形态的中型科技公司，在生成式 AI 时代拥有超越万亿体量巨头的非对称竞争优势。 - [16:34](https://www.youtube.com/watch?v=jnep4SX9vn0&t=994s) 面对底层大模型能力的指数级拓展，行业必须保持审慎敬畏，前置部署严格的安全护栏以防范潜在失控与安全风险。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=4s) 破除规模优势：Snap 借助 AI 实现“降维打击” The Deep View 资深记者 Nat Rubio-Licht 对话 Snap 工程副总裁兼首席信息官 Saral Jain，探讨 Snap 这家介于初创企业与科技巨头之间的公司如何拥抱 AI 浪潮。Saral Jain 认为，Snap 当前的生态位是最具优势的黄金位置。在过去，规模是科技行业决定性的护城河——更庞大的团队、更多的 GPU 算力以及更充裕的资金往往意味着统治力。然而，AI 彻底打破了这一底层逻辑，竞争的关键转向谁拥有最顶尖的“产品愿景”以及谁能跑通“最快的学习闭环”。 Snap 历来是一家极具创意灵感的公司，在过去，将无数创意转化为实际产品的工程构建过程是唯一的瓶颈。如今随着代码与产品构建成本急剧下降，Snap 能够充分发挥敏捷优势，“打出远超自身体量的重拳”（punch above our weight）。在 Snap 内部，AI 绝非边缘试水项目，而是关乎公司如何研发、运营乃至制胜的核心支柱。工程团队将 AI 划分为三大赋能方向：面向近 10 亿 Snapchat 社区用户的应用内产品体验、面向生态合作伙伴与广告主的商业赋能，以及直接面向内部员工与工程师的研发提效。 ### [01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) 90% 代码由 AI 编写：速度与安全在优秀系统设计中共存针对敏捷扩张与数据安全之间的平衡难题，Saral Jain 纠正并更新了一组关键数据：几周前他公开表示 Snap 有三分之二的代码由 AI 编写，而眼下这一比例已迅速突破 90%。在他看来，“速度与安全并存”并非二选一的悖论，两者本质上都是优秀系统设计（System Design）的必然结果。Snap 服务全球近 10 亿用户群体，安全、隐私与合规早已不是事后补救的被动防线，而是不可妥协的准入门槛。 Snap 的保障策略在于“让犯错变得极其困难”：团队构建了标准的工程“黄金路径”（Golden Path）与铺就坦途的自动化发布运维体系（Paved Road），不仅能够严密控制任何风险事件的影响范围（Blast Radius），还能在异常发生时实现秒级回滚。更关键的是，Snap 贯彻了“以 AI 审查 AI 代码”的机制，其自主研发的代码审查智能体 CodePal 彻底重塑了研发流。Saral Jain 强调，代码的风险不在于是否由 AI 生成，而在于“代码是否无人负责”（unowned code），最终所有代码的责任仍归属于人类工程师。 - **AI 代码占比：** 内部超过 90% 的生产代码现已通过 AI 生成与辅助构建。 - **CodePal 审查效率：** 自动化审查全公司 90% 的新增代码，首轮评审响应时间压缩在 5 分钟以内（相比此前人类工程师 5 至 6 小时的中位数耗时大幅缩短）。 - **拦截表现与反馈：** 工程师正向满意度评分超过 80%，已在上线前累计拦截并排除了数万个代码缺陷与安全隐患。 ### [03:51](https://www.youtube.com/watch?v=jnep4SX9vn0&t=231s) “受信任的受控智能体”体系与全流程协作 ```mermaid flowchart LR accTitle: Casper 与 CodePal 智能体代码生成与评审流 subgraph Ideation["1. 沟通与触发"] A["工程师群聊讨论 (Slack / Jira)"] --> B["受控智能体 Casper 常驻监听上下文"] end subgraph Execution["2. 沙箱生成与对抗评审"] B --> C["Casper 启动远程沙箱 进行最小化代码变更"] C --> D["审查智能体 CodePal 进行自动化安全评审与交互"] D -->|多轮修改优化| C end subgraph Signoff["3. 人工核准与合并"] D --> E["生成最安全精简变更 提交人类工程师最终审核"] end ``` 在深入探讨 AI 智能体时，Saral Jain 提出了“受信任智能体”（Blessed Agents）这一概念。其中最典型的代表是全自主远程编码智能体 Casper。Casper 扮演着 AI 队友的角色，能够常驻在 Slack、Jira 等内部沟通工具中监听工程师的技术讨论。当方案成型时，工程师只需在对话上下文中直接唤醒 Casper，它便会自动在云端远程沙箱环境中拉取环境，计算出所需的最少代码变更，并主动调用审查智能体 CodePal 进行来回对抗式评审校验，最后将精炼且安全的代码提交给人类工程师核准上线。这一协作机制将原本耗时数周甚至数月的“设计-实现-评审”循环简化为日常对话间的即时交互。此外，Snap 还针对线上应急排错（SRE On-Call 调查）、异常数据挖掘和客户端崩溃归因等高频场景部署了专项智能体。为确保企业级安全，Snap 为所有智能体赋予了完备的“智能体身份体系”（Agent Identities），每个智能体均具备独立的操作权限、鉴权机制与可追溯的审计日志（Audit Logs），从而确保 AI 代理的所有操作均可控、可查、可复盘。 ### [05:36](https://www.youtube.com/watch?v=jnep4SX9vn0&t=336s) 摆脱模型焦虑：上下文底座与评估系统才是决定性瓶颈在当前代码大模型（Claude、Cursor、Codex 等）层出不穷的技术周期中，Snap 放弃了“追逐哪款模型最好”的传统思路。Saral Jain 指出，由于底层模型每隔三周就会迎来一次迭代，盲目绑定单一模型注定会陷入被动，因此 Snap 全面采用了“模型无关”（Model-Agnostic）架构策略，根据具体任务的复杂度与模糊性灵活分配算力——低复杂度且重复性高的任务交给廉价小模型，复杂模糊且涉及高阶推理的任务则接入前沿推理大模型。实战经验表明，当 AI 代码审查遗漏 Bug 时，“问题几乎从来不是因为模型本身能力不足，而是因为模型缺乏充分的上下文”。为此，Snap 将研发重心倾注在构建“上下文底座”（Context Substrate）以及内部评估系统（Eval System）上。该底座能够实时检索并结构化呈现 Snap 特有的代码仓库、知识库、业务约定与依赖资产，让任何接入的模型都能在充分理解企业私有背景的前提下精准作业。 ### [07:13](https://www.youtube.com/watch?v=jnep4SX9vn0&t=433s) 拒绝刷量假象：基于 JTBD 框架衡量真实业务增益针对内部 AI 落地成效的衡量，Snap 坚决摒弃了停留在表面原型的做法，而是全面引入“待办任务”（Jobs to Be Done, JTBD）框架。评估的核心不再是容易被“刷量”的输入过程指标（例如生成了多少行代码或提交了多少个 Pull Request），而是聚焦于实际解决业务问题的产出指标（Output Metrics），以此验证 AI 是否真正推动了业务增长与效率提升。 - **研发效能：** CodePal 每周完成数万次代码审查并拦截海量漏洞；Casper 每周产出数千个可直接合并的 PR。 - **客服支持：** 智能客服代理每月成功分流过滤掉 350 万次用户求助咨询，使转接至人工客服的工单总量直接下降了 25%。 - **线上运维：** 负责故障排查与 On-call 诊断的 SRE 智能体协助团队将线上事故解决时间（MTTR）缩短了 25% 至 30%。 ### [08:48](https://www.youtube.com/watch?v=jnep4SX9vn0&t=528s) 研发物理法则的质变：从冗长排期到原型驱动在 Snap 工作九年之久的 Saral Jain 回顾道，尽管公司通过有趣产品让近 10 亿用户与亲友紧密连接的初心从未改变，但“构建产品的物理法则”（physics of building）已经发生了翻天覆地的变化。过去研发几乎是一场痛苦的优先级取舍演练——团队拥有源源不断的天马行空创意，却受限于实际的工程构建能力，不得不将大量优秀想法打入“优先级线以下”（below the line）；如今在 AI 的强力辅助下，构建成本大幅降低，绝大部分创意都能快速进入验证通道。研发文化的范式转移还体现在两个核心维度：首先是“PPT 幻灯片变少，可运行原型大增”，产品经理（PM）与 UI/UX 设计师等非纯研发职能现在也能快速跑出可交互原型；其次是传统长周期排期的消亡，虽然公司依然保有年度层面的产品战略方向，但团队早已告别了 6 到 12 个月的僵化路线图（Roadmap），全面转向以月度甚至季度为粒度的快速迭代模式。相比动辄数十万人、流程僵化缓慢的科技巨头，Snap 依靠精简敏捷的团队形态，在 AI 赋能下展现出了极高的迭代速率与竞争壁垒。 ### [10:47](https://www.youtube.com/watch?v=jnep4SX9vn0&t=647s) 软件工程师角色的蜕变：从“高阶自动补全”到长程智能体编排随着 AI 成为日益强大的赋能工具，团队不仅能够快速试错迭代，还能以更短的周期推动产品路线图落地。当主持人 Nat Rubio-Licht 问及软件工程师的工作究竟发生了怎样的改变以及未来走向何方时，Snap 的 CIO 兼工程高级副总裁 Saral Jain 感慨道，未来的发展极难预测，但仅仅回看过去这一年就已经令人惊叹——就在一年前，AI 在工程领域充其量不过是一个“高阶自动补全（glorified autocomplete）”工具；然而在短短六个月内，行业格局彻底重塑。如今的工程师不再只是打字写代码，而是演变为“指挥家”，负责协调和编排多个长期运行的智能体（orchestrating multiple long-running agents），从而创造出数倍于以往的商业影响力。未来，这些智能体和底层模型的能力还会持续增强，能够以更高的保真度（fidelity）自主完成耗时更长的复杂任务。伴随代码生成量的指数级膨胀，传统的工程流程也将迎来剧变。面对如此庞大的自动化代码流，完全依赖人工去逐行审查每一段 AI 生成的代码已变得不切实际。未来的代码评审机制必然转向“基于风险的 AI 与人类混合评审（risk-based AI human reviews）”。由于不同代码变更的影响范围和严重程度各异，日常的常规审查将越来越多地交由像 CodeRabbit 等 AI 智能体来接管，人类工程师则集中精力把控高风险改动。 ```mermaid flowchart TD accTitle: 软件工程模式的演进 subgraph Past["过去：辅助编写模式"] A["工程师手动编写代码"] --> B["AI 充当高阶自动补全"] B --> C["人类逐行审查所有代码"] end subgraph Future["现在与未来：智能体编排模式"] D["工程师担任指挥 / 编排角色"] --> E["编排多个长程智能体"] E --> F["海量代码自动生成"] F --> G{"基于风险的分流审查"} G -->|"低/中风险"| H["AI 智能体自主代码审查"] G -->|"高风险关键模块"| I["资深工程师深度介入"] end ``` - **从输入代码到任务编排**：工程师的工作重心已转移到长程任务拆解、多智能体协同调度及系统架构把控上。 - **混合风险评审机制**：通过区分改动风险等级，引入智能体辅助代码评审工具，化解生成海量代码带来的审查瓶颈。 --- ### [12:13](https://www.youtube.com/watch?v=jnep4SX9vn0&t=733s) 招聘人才的核心法则：好奇心与学习能力胜过单纯的 AI 技能当被问及在招聘年轻工程师或开发者时“AI 技能是否已成为硬性必备项”时，Saral Jain 给出了发人深省的回答。他明确表示，自己最看重的特质从来都不是特定的技术标签，而是“好奇心以及持续学习的能力（curiosity and being able to learn）”。虽然当前技术圈的大部分领域都在被 AI 席卷重塑，具备对 AI 的探究热情确实加分，但他强调，好奇心可以体现在任何技术领域。底层模型和应用工具日新月异，只有那些保持探究本能、能够迅速适应未知技术范式并自主学习的人，才能在长期竞争中脱颖而出。 --- ### [12:44](https://www.youtube.com/watch?v=jnep4SX9vn0&t=764s) 警惕“AI 剧场”：从业务痛点而非工具模型出发在谈及目前企业部署 AI 时常犯的最大错误时，Saral Jain 轻轻叹了口气，直指当前行业普遍存在的浮躁现象——“AI 剧场（AI theater）”。他解释道，依靠现有的生成式工具制作一些花哨的 Demo、原型或展示项目其实非常轻而易举，但这些炫酷的原型根本不等于商业落地。企业真正面临的难关在于改造实际的业务运营模型（operating model）并创造切实的商业价值。他建议所有企业负责人：不要从“我该选用哪个工具或哪个模型”出发，而要从“业务问题是什么”开始倒推。先理清现有工作流中的最大瓶颈究竟卡在哪里，然后再去判断这道难题该由人工解决还是交给 AI 重构。 ```mermaid flowchart LR accTitle: 业务驱动与工具驱动工作流对比 subgraph AntiPattern["错误倾向：AI 剧场模式"] X1["追逐最新 AI 模型/工具"] --> X2["堆砌炫酷 Demo 与原型"] X2 --> X3["无法融入运营，难以产生商业价值"] end subgraph BestPractice["推荐范式：待办任务 (JTBD) 框架"] Y1["识别工作流核心瓶颈"] --> Y2["以 AI 为核心重塑工作流"] Y2 --> Y3["为用户/合作伙伴解决实际问题"] end ``` 在 Snap 内部，团队高度遵循“待办任务（Jobs to be Done, JTBD）”框架：始终聚焦于我们正在为社区、合作伙伴和员工解决什么核心现实问题。AI 从来不该是事后生硬“外挂”上去的补丁，而应该是重构端到端业务流的核心驱动力。面对主持人提到的“错失恐惧症（FOMO）”与行业对失控风险的担忧，Saral Jain 坦承，企业确实普遍受到焦虑情绪的驱使。但在汹涌的技术狂潮中，“最快的推进方式有时恰恰是停下来、慢下来，仔细思考自己到底想解决什么”。积极实验固然重要，但唯有扎根真实场景的探索才有价值。 --- ### [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) 领导者不当旁观者：日常实操与 Snap OS 的诞生谈到自己的日常工作，Saral Jain 直言他几乎在所有环节都深度引入了 AI，并且找回了久违的乐趣：“我又开始大量动手写代码、造东西了（building a lot more again）。”他特别向所有技术管理者发出呼吁：如果你正在要求整个团队颠覆传统的工作方式，你自己就绝不能当袖手旁观的看客（bystander），必须卷起袖子亲自下场实操。在日常决策中，他把 AI 当作顶尖的“思考伙伴（thought partner）”。虽然最终拍板决策权永远在自己手中，但借助 AI 进行深度辩论、推演不同视角的利弊，能极大拓展管理者的认知广度。在动手开发方面，他透露自己与团队正在研发一个名为 **Snap OS（Snap 操作系统）** 的内部平台。像 Snap 这样规模的企业内部存在典型的职能隔离（functional organization），工程、产品、设计、销售和市场等各个部门都有自己独立的运作机制、目标追踪体系与分散各处的项目路线图。Snap OS 的目标是打通所有异构信息孤岛，接入统一的自然语言对话界面： - **跨职能穿透问答**：管理者可以直接询问：“涉及 5 个不同部门协同的复杂项目 X，今天的核心瓶颈到底卡在哪里？” - 系统会自动穿透多方文档与数据，迅速梳理出跨团队协作的症结并给出清晰答案。 --- ### [16:34](https://www.youtube.com/watch?v=jnep4SX9vn0&t=994s) 独家观点与安全边界：中型规模的非对称优势在访谈尾声的“AI 辛辣观点（hot take）”环节，Saral Jain 提出了一个反共识的判断：生成式 AI 实际上为 Snap 这种体量的中型科技巨头带来了巨大的竞争壁垒和非对称优势。 | 竞争维度 | Snap 的现实定位与优势 | |:--- |:--- | | **用户与分发规模** | 拥有近 10 亿全球用户，沉淀了庞大的真实社交与视觉数据资产，具备强大的分发通路。 | | **组织架构与敏捷度** | 相比超大型万亿巨头，Snap 保持着极度精简（lean）的敏捷小团队结构。 | | **闭环学习机制** | 内部具备紧凑闭环的“学习飞轮（closed learning loop）”，试错与产品迭代速度极快。 | | **战局胜负手** | 资本实力（预算规模）已不再是决定胜负的唯一指标，“以最快速度持续迭代演进的能力”才是终极王牌。 | 最后，针对主持人关于“当前 AI 是否存在令你感到恐惧的方面”的提问，Saral Jain 态度冷静而严肃。他表示最近曝光的一些安全事件（security incidents）虽然谈不上令人恐慌，但确实给全行业敲响了警钟。随着底层大模型的能力边界不断扩展，整个科技行业必须保持极致的清醒与审慎，前置部署严格的安全护栏（guardrails），确保强大的技术始终造福人类社会，防止模型失控带来潜在危害。在狂热与兴奋的浪潮之中，永远要对可能引发的深远后果抱有敬畏之心。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=jnep4SX9vn0&t=4s) 打出远超自身体量的重拳 — Saral Jain > [01:39](https://www.youtube.com/watch?v=jnep4SX9vn0&t=99s) 让犯错变得极其困难 — Saral Jain > [05:36](https://www.youtube.com/watch?v=jnep4SX9vn0&t=336s) 问题几乎从来不是因为模型本身能力不足，而是因为模型缺乏充分的上下文 — Saral Jain > [08:48](https://www.youtube.com/watch?v=jnep4SX9vn0&t=528s) PPT 幻灯片变少，可运行原型大增 — Saral Jain > [10:47](https://www.youtube.com/watch?v=jnep4SX9vn0&t=647s) 高阶自动补全（glorified autocomplete） — Saral Jain > [10:47](https://www.youtube.com/watch?v=jnep4SX9vn0&t=647s) 基于风险的 AI 与人类混合评审（risk-based AI human reviews） — Saral Jain > [12:13](https://www.youtube.com/watch?v=jnep4SX9vn0&t=733s) 好奇心以及持续学习的能力（curiosity and being able to learn） — Saral Jain > [12:44](https://www.youtube.com/watch?v=jnep4SX9vn0&t=764s) AI 剧场（AI theater） — Saral Jain > [12:44](https://www.youtube.com/watch?v=jnep4SX9vn0&t=764s) 最快的推进方式有时恰恰是停下来、慢下来，仔细思考自己到底想解决什么 — Saral Jain > [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) 我又开始大量动手写代码、造东西了（building a lot more again） — Saral Jain > [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) 如果你正在要求整个团队颠覆传统的工作方式，你自己就绝不能当袖手旁观的看客（bystander），必须卷起袖子亲自下场实操 — Saral Jain > [15:06](https://www.youtube.com/watch?v=jnep4SX9vn0&t=906s) 涉及 5 个不同部门协同的复杂项目 X，今天的核心瓶颈到底卡在哪里？ — Saral Jain ## Mentioned Saral Jain, Nat Rubio-Licht, Snap, The Deep View, Ai4, Snapchat, CodePal, Casper, Slack, Jira, Claude, Cursor, Codex, CodeRabbit, Snap OS ## Video description Saral Jain - CIO and SVP of Engineering, Snap Nat Rubio-Licht - Senior Reporter, The Deep View For years, disruption in the tech industry meant trading stability for speed. AI lets us rewrite that tradeoff, enabling small teams to pressure test ideas while building and launching new products and features in days instead of weeks. In this session, Snap CIO & SVP of Engineering Saral Jain will share how AI tools and agents are expanding the impact small teams can have on a global scale while championing human creativity and innovation. [Source](https://www.youtube.com/watch?v=jnep4SX9vn0)

Source quotes

[0:04] Saral Jain: 打出远超自身体量的重拳 [1:39] Saral Jain: 让犯错变得极其困难 [5:36] Saral Jain: 问题几乎从来不是因为模型本身能力不足，而是因为模型缺乏充分的上下文 [8:48] Saral Jain: PPT 幻灯片变少，可运行原型大增 [10:47] Saral Jain: 高阶自动补全（glorified autocomplete） [10:47] Saral Jain: 基于风险的 AI 与人类混合评审（risk-based AI human reviews） [12:13] Saral Jain: 好奇心以及持续学习的能力（curiosity and being able to learn） [12:44] Saral Jain: AI 剧场（AI theater） [12:44] Saral Jain: 最快的推进方式有时恰恰是停下来、慢下来，仔细思考自己到底想解决什么 [15:06] Saral Jain: 我又开始大量动手写代码、造东西了（building a lot more again） [15:06] Saral Jain: 如果你正在要求整个团队颠覆传统的工作方式，你自己就绝不能当袖手旁观的看客（bystander），必须卷起袖子亲自下场实操 [15:06] Saral Jain: 涉及 5 个不同部门协同的复杂项目 X，今天的核心瓶颈到底卡在哪里？

Source quotes

[0:04] Saral Jain: the best position to be in right now [0:04] Saral Jain: punch above our weight [1:39] Saral Jain: more than 90% of all code [1:39] Saral Jain: unowned code [7:04] Saral Jain: Jobs to Be Done [8:48] Saral Jain: the physics of building [8:48] Saral Jain: Prototypes Over Slides [10:58] Saral Jain: glorified autocomplete [10:58] Saral Jain: curiosity and learning agility [10:58] Saral Jain: AI theater [15:06] Saral Jain: roll up your sleeves [15:06] Saral Jain: What is the bottleneck today on Project X? [16:39] Saral Jain: Sometimes the biggest budget doesn't win, sometimes the ability to move fastest can win. [17:55] Saral Jain: for the benefit of humankind

Original transcript

[0:04] Thank you guys so much for being here. My name is Nat Rosal. I'm a senior reporter with the Deep [0:08] View, which is a newsletter about all things AI that goes out every single day. And I am here with [0:13] SL Jane, Snapchat's head of engineering. Thank you so much for being here, Sar. Thanks for having me. [0:18] Should be a fun conversation. Yeah. So, to start out, um, I want to talk about the sort of middle [0:24] ground that Snap is in. uh sitting somewhere between a startup and a tech giant. How is [0:31] the engineering team at Snap using and considering AI? Yeah, look, I think uh it is actually the best [0:37] position to be in right now from our perspective. Uh for years, scale was the defining advantage. [0:43] Bigger teams, more GPUs, bigger pockets. Uh and AI changes that equation fundamentally. Uh now [0:50] the teams with the best product vision and the fastest learning loops matter the most. uh and [0:55] Snap has always been such a creative company. We have always been had so many creative ideas [1:01] and the bottleneck was building and now we're building becoming a lot more cheaper. I think [1:06] it's a great time for companies like Snap to like punch above our weight. Uh for AI for us AI is not [1:12] a side project. It's actually fundamental to how we uh build, how we operate, how we win. Uh and [1:17] when I think about AI for our engineering team, I think about it in three different verticals. [1:21] One is AI for our community for our Snapchatters who use the app uh so in the product. The second [1:27] is AI for our partners and advertisers and third is AI for our engineers for the employees and in [1:33] across all of these three verticles we are kind of transforming uh the way we operate. Yeah. So I [1:39] heard recently that twothirds of the code that is written at Snap is written by AI and um given the [1:47] speed and the scale that you guys work at, how does Snap balance um speed and innovation with [1:54] safety and security? That is such a good question. Uh so first of all when I spoke publicly a few [1:59] weeks back that number was 2/3. At this point more than 90% of all code is being run by AI and [2:05] that's not a surprise to anybody listening. uh uh and uh I I think the premise of that question is [2:10] actually really interesting. Uh speed and safety can coexist and they are all kind of outcomes of [2:15] the same underlying thing which is good system design. Yeah. So uh for us Snapchat reaches a [2:21] community of almost a billion people. So safety, security, privacy are not just afterthoughts, [2:27] they are actually table stakes. And the way we do them is uh by actually making it really hard for [2:33] people to do the wrong thing. And what I mean by that is uh we have a golden path of engineering. [2:39] Uh we have blessed agents. Uh we have a paved road of uh uh things like rolling back when something [2:46] uh goes bad. We have uh uh solutions for uh ensuring that people are able to uh quickly [2:52] experiment with uh things but be able to kind of do it in a very constrained way so that the blast [2:57] radius of something bad going on is not really high. So that's that's really important for us. [3:02] It's also really important for us that AI reviews AI's code. And so one very interesting agent we [3:07] have is an agent called CodePal which is our AI code reviewer and that is one of the biggest ways [3:12] we kind of make sure that we can move very fast but we can still do it very safely. uh CodePal [3:17] reviews 90% of all code uh written at uh a snap within 5 minutes and compare that to previously [3:24] where a human review would have taken a median of 5 to 6 hours for the first touch essentially and [3:29] it has a positivity rating from engineers above 80% it has found uh tens of thousand piece of [3:35] bugs in our code and that is how we keep things safe essentially so ultimately the uh risk is not [3:41] uh uh AI generated code the risk is unowned code and and I think that's where like the humans are [3:46] ultimately accountable for all of the quotes that we written. Yeah. You mentioned blessed agents um [3:51] which is actually a phrase that I haven't heard before. So I'd love to discuss uh a little bit [3:55] about how you guys are utilizing agents and I mean obviously the conversation revolving around [4:02] security how you're using agents securely. Yeah. AI agents are where the biggest transformation is [4:08] happening for us. Um and I'll give an example. We have a a blessed agent called Casper. [4:14] Uh Casper is a remote autonomous coding agent. The way the workflows in engineering now go is uh [4:20] it's an AI teammate. So engineers are talking to other engineers from wherever they are either on [4:24] Slack or Jira or any other internal tool. They're having the conversation and Casper is listening [4:30] into the conversation. At one point the engineers with all of that context of the conversation can [4:35] invoke Casper and say go ahead and build this feature. And Casper will spin up a remote sandbox. [4:40] It will do the minimum possible change that it needs to do. It will invoke code bal which is our [4:45] code review agent. The agents will have the back and forth and then ultimately the safest possible [4:50] change will be sent to the human for the final review. And so that's what the whole uh review [4:55] implementation loop that used to take weeks and months before and is now just a conversation. [5:00] Yeah. Similarly, we have other agents that can do things like on call investigations like SRRES. [5:07] uh we have agents that uh do data investigations, crash analytics and so on. So we are trying to [5:12] deploy agents to do real work and the way we keep it secure is we have agent identities. So just [5:18] like humans have identities uh uh humans have uh audit logs and we have permissions similarly we [5:25] give agents permissions we give agents audit logs we can trail everything that the agent has done [5:30] and uh uh that that's just how we kind of like look at agent identities. Um I'd also love to know [5:36] a little bit about with so many models to choose from um especially for coding you know claw cursor [5:43] codeex all for some reason start with a C um how do you choose which um models to that work best [5:51] for your team yeah that's such a good question I think we have stopped asking the question which [5:56] model is the best it's it's almost impossible at this point because things are moving at such [5:59] a fast pace every 3 weeks like something new is going to come up and we have a strategy where we [6:04] are largely model agnostic. uh there are certain tasks that uh uh could be low complexity high like [6:11] uh very repetitive tasks that smaller cheaper models can be really good at uh while there [6:15] are certain high complexity ambiguous tasks that might require frontier reasoning models and so we [6:21] always ask which model is the best at doing the task that I'm trying to do and and I think that [6:26] is a something I would uh uh encourage everybody listening in to do as well uh which is uh what [6:31] we've observed is AI when AI is reviewing our code it misses bugs. It's almost never because of the [6:37] model. It's because of the context. And we are investing a lot in our eval system in the context [6:43] substrate so that the model any model has that information that it needs to be able to do the [6:48] job. Um and this information is company specific information. Our repositories, our code bases, [6:53] um our knowledge and giving all of that context substrate to the models is the more important [6:59] part versus obsessing over which models that we are trying to use. And then the eval system [7:04] helps a lot as well. Yeah. I want to go back to Casper um and this idea of I guess internal AI [7:13] deployments and how you guys measure success. Um how do you know when a model deployment or [7:20] a new agent deployment is working? Yeah, that's a great question. We have actually thought a lot [7:26] about that. Like instead of doing AI prototypes, we have counted all of our AI agents into what we [7:31] call the jobs to be done framework. And what that means is ultimately you have teams, individuals, [7:36] organizations that are doing certain jobs for our community, for our partners, for our employees. [7:41] And we start with that. We start with the problem that we are trying to solve. And uh then we [7:46] measure whether AI agents uh as as they execute on these problems are able to help move the business [7:52] impact, move the business metrics. These are the output metrics versus obsessing about any capable [7:57] input metric like PR uh number of PRs or something that are very capable. And so that's a big part [8:02] of our strategy and uh ultimately grounding ourselves in this business impact is where we [8:07] know we are succeeding. And just to give a couple of examples, we talked about code pal. It's doing [8:12] tens of thousands of code reviews, finding tens of thousands of bugs. Casper is doing thousands [8:16] of mergeable PRs every week. Uh but then we also have uh uh our customer support agent that is uh [8:23] deflecting three and a half million queries every month from our customers. the customer [8:27] support ticket volume has gone down by 25% uh for humans. Uh uh similarly we have on call agents [8:34] that are looking at our outages in incidents and the incident resolution time has gone down by 25 [8:40] 30%. So that is how I'd quantify real impact uh versus uh gameable metrics. Yeah. And given [8:48] that you've been at Snap for nine years, you were at the company both before and after and during [8:55] this whole AI transition. How has engineering and product development changed? Uh I've thought about [9:03] that a lot. 39 years is a long time especially in an AI timeline. Uh uh look I think uh ultimately [9:08] Snap's vision cons is consistently the same. What we want to do is build a product that [9:14] almost a billion people are now using that helps people connect with their close friends and family [9:18] and be a fun casual way to communicate with your close friends and family and that missions has not [9:24] changed over the last 9 years. What has changed is the physics of building. Uh I think previously [9:29] everything was a prioritization exercise. We had so many creative ideas but uh what we were [9:35] bound by is the building capability essentially and uh with AI we are seeing that shift quite [9:42] meaningfully where things are less below the line uh is how I'd characterize it. Uh uh we are seeing [9:48] some other interesting shifts as well. Uh we are seeing less slides and more prototypes being built [9:53] like uh non-engineering functions like product and design functions are building many more prototypes [9:59] at this point. uh which helps us move much faster. We're also seeing um long range horizon planning [10:06] almost disappear like so we certainly have a year-long product strategy. We have a mission as I [10:12] said but uh we are not doing road mapping for a 6 months or a year anymore. We are doing it more at [10:17] a quarterly monthly granularity because things are moving so fast. uh and so I think uh that would [10:23] be some of the changes I've seen very recently is the fact that you guys are sort of sitting between [10:29] startup and big tech allowing you you think to move at that speed and to cut down those timelines [10:35] 100% I think this is a big competitive advantage that we have like big 100,000 people companies [10:41] are glacial to move and change processes versus we have a small lean team that can move very fast and [10:47] with AI that is even a bigger enabler essentially and that helps us really kind of be able to move [10:52] fast, iterate and go faster on the road map as well. Yeah. How do you think the job of a software [10:58] engineer has changed thus far and is going to change further down the road? I think it's going [11:06] to be really hard to predict but like what what I can say is within the last year it is fascinating [11:11] to think that even a year ago uh for engineering AI was just a glorified autocomplete right and [11:17] that's just not true anymore like things uh have changed so much in the last 6 months even uh [11:22] engineers are now coordinating multiple longunning agents essentially they're orchestrating all of [11:27] these agents uh uh a lot more business impact is being uh generated using these agents uh and [11:33] so I think we will see a lot more of that going forward as well. Uh the the capability of these [11:38] agents and these models will keep increasing such that we are able to do more longunning tasks with [11:43] higher fidelity. Yeah. And we will also have uh some changes in the review process. It's it's [11:49] impossible for a human to review every single piece of code being generated by AI at this [11:54] point and it will only keep increasing the volume of code being generated and so we will see a lot [11:59] more risk based AI human reviews. Essentially not all code changes are the same and so we will see [12:06] a lot more AIdriven code reviews by agents like code now. Yeah. Is REI skills something that are [12:13] a necessity for you hiring a young engineer or a young developer? I would say curiosity and being [12:22] able to learn is a necessity for us to hire new people. uh uh and and I think that curiosity could [12:28] be any domain but especially AI given that AI so much of the technology world is being consumed by [12:33] AI right now. Uh but I would still look for that basic ability for people to be curious to learn [12:39] right I think that would be the most important skill set. Yeah. What do you think is the biggest [12:44] mistake that enterprises are making right now as they are deploying AI? [sighs] Uh that's such a [12:51] thoughtful question. I think uh what I've seen is there's a lot of AI theater happening like [12:57] a lot of demos, a lot of prototypes and demos and prototypes are really easy to do with AI. Uh what [13:04] is hard to do is change your actual operating model and actually have business impact using [13:09] AI and I think that is what I would ask uh all enterprises to kind of ground themselves in is [13:14] instead of starting with the tool or the model or starting with AI, start with the business problem. [13:20] Start with what exactly is the bottleneck in your existing workflows. try to kind of [13:25] like think about uh those things first and then whether a human does transforms it or whether AI [13:32] transforms it is a second question essentially and um that's how we ground ourselves in this [13:37] jobs to be done framework uh which is uh what are the real problems we are solving for our community [13:42] for our partners for our employees and then uh AI can help us accelerate the transformation to [13:47] solve these problems and not as an afterthought or a bolted on thing at the end but actually rethink [13:53] your workflows using AI I in mind versus just you know like take existing workflows and try to [13:58] somehow make AI work with it. I think that would be my biggest uh uh guidance to enterprises who [14:03] are trying to kind of figure out AI at this point. Yeah. As as much as as it is a best practice for a [14:11] business to like you said start with that problem that they're trying to solve. I think the speed [14:16] that AI is developing like every single day there is some sort of new model or capability or rogue [14:25] agent as we are now seeing in the headlines. Um do you think that enterprises are sort of facing [14:32] that fear of missing out? Um so they're maybe not thinking so hard about what exactly is the [14:39] problem they're solving. I certainly agree that's a problem that enterprises are facing right now. [14:44] But sometimes the fastest way to move is to take a pause and slow down and think about what you're [14:50] trying to do. And I would definitely encourage enterprises to experiment with AI. AI can be [14:54] a massive enabler for many enterprises. Uh uh but uh grounding them in solving real problems [15:01] for the enterprise is the most important thing to do. Yeah. I'd love to know how you use AI in your [15:06] day-to-day work. Uh almost for everything. For for one, I'm building a lot more again, which [15:12] is a lot of fun. Uh it's uh it's uh something I've not enjoyed for many many years and uh in [15:16] the last couple of years uh uh just building a lot more things but not prototypes or demos [15:21] building actual stuff that Snap Inc. can benefit from and that that is uh something I would really [15:26] encourage all leaders listening in to do as well is if you're asking your entire teams to transform [15:31] the way they work. You can't be a bystander. You have to kind of roll up your sleeves and [15:34] do it yourself as well. Um the second thing I do is uh use it as a thought partner for for business [15:40] decisions. Ultimately it's my judgment but having a thought partner which can kind of debate with [15:45] you which can understand and give you multiple points of views is really really helpful. Uh but [15:50] ultimately own the decisions. What's something that you built? Uh right now I'm building [15:55] something called Snap OS which is Snap operating system. Uh so a common problem with a company like [16:00] Snap is we are a very functional organization. So each function whether it's engineering, product, [16:05] design, sales, marketing have their own way of operating. They have their own road maps. [16:10] The road maps are in different locations. They to take goals differently and I'm building the [16:15] with my team a thing called snap operating system that takes all of the information wherever it is [16:20] into a central place that you can have natural language conversation in. So I can ask it this [16:25] project X which is a very complex project uh which takes five different functions to coordinate with [16:29] each other. What is the bottleneck in this project today and it'll help get the answer. Interesting. [16:34] Okay. Um I've got just a couple more questions for you because we only have a couple of minutes. [16:39] Um, this is something that I like to ask everybody who is working with AI every day. Uh, what is your [16:45] AI hot take? Uh, I'm not sure about a hot take, but I think I'd, uh, kind of point back to the way [16:51] we started the conversation, which is, uh, uh, for a company of a snap size and scale. Uh, AI [16:57] can be a massive enabler and it's actually a big competitive advantage for midsize companies. uh [17:02] uh we have the scale, we have the we have almost a billion people using our platform, we have the [17:07] data, uh we certainly have the distribution and so we can make a real impact using AI but we are a [17:14] small team. We are a lean team. we have the closed learning loop essentially that can iterate really [17:19] really quickly and so this is a time where uh this is can this can be a massive force [17:24] multiplier for a company of our size essentially and um sometimes the biggest budget doesn't win [17:29] sometimes the uh ability to move fastest can win yeah last one I've got is there anything in AI [17:37] right now that is scaring you I think some of the recent uh examples of security incidents that are [17:44] coming in are pretty thought are not necessarily scaring me but something the industry as a large [17:49] has to be really really uh thoughtful about. Um I think uh it's it's really important for us to [17:55] put the guardrails in place such that we can use AI for the benefit of humankind essentially and [18:00] these models will continue getting more and more capable and it's up to us humans to make sure we [18:05] have enough guardrails in place that these models can't do any harm. Yeah. Uh thinking about the I [18:11] guess consequences for the excitement. That's exactly. All right. Well, thank you so much, [18:16] Sar, for taking the time to sit with us today. Everyone, give her all a round of applause.