What It Takes to Build an AI That Actually Knows You

发布时间    来源
Episode 设置


登录已过期或未登录,无法修改。请先登录后再试。

以下是根据提供的英文文本翻译的简体中文摘要: 以下是完全基于所提供的文字记录总结的播客节目内容: **嘉宾与公司概览** * 肯是第一人称实验室(First Person Labs)的创始人,这是他在播客中首次谈论该公司的工作。 * 第一人称实验室专注于创建一个人的模型(理解他们的记忆、优先级、偏好、目标、价值观、承诺和意图),并把控制权交还给这个人。 **核心产品:“第二”(Second)** * 该公司创造了一个名为“第二”的个人代理/模型。 * “第二”充当个人管家或代理(类似于“蝙蝠侠的阿尔弗雷德”),安全地保存个人的生活背景,并作为一个值得信赖的保险库。 * 第一人称实验室作为一个独立的实体存在,而不是像 Grokbot 或 Codex 那样本身就是一个代理。其他专门的代理(例如 Grokbot、Codex 或短暂代理)与用户的“第二”进行对接,以了解用户的需求。 * 主持人设想,拥有许多个人代理只是过渡性的,“第二”旨在成为所有个人数据的唯一受信任代理/保险库,与所有其他软件和硬件代理进行交互。 * “Second”在用户使用电脑(应用切换、网页停留、打字、Wi-Fi连接)时,全天候24小时持续分析无限的事件流。它以层级方式构建对用户试图做的事情的理解,并与连接器(如 slack、signal、whatsapp 和原始计算机流)相配对,以理解用户。 * 它能捕捉事物并提供主动帮助——例如实时更新活动天气预报或就正在进行的对话向用户发送提醒——范围从被动生产力(简单的电子邮件/表情包通知)到反思性生产力(跨越漫长的时间范围和优先级将各个点连接起来)。 * 肯指出,“第二”现在就可以使用(在 `two dot computer`),尽管他们正在扩大推理和容量规模,并以负责任的方式推出它。 **安全、架构与隐私** * 构建一个能够看到用户所做一切的系统会立即引发信任问题,因此第一人称实验室优先考虑隐私和安全架构。 * 数据和推理是人工智能的两个组成部分。用户可以选择他们带来的推理。如果他们不信任第一人称实验室,他们可以带上自己的令牌(tokens)在笔记本电脑上进行本地数据分析。 * 或者,“第二”提供了一个使用受信任执行(机密计算/机密推理)的推理提供商,其中模型在带有特殊开关的 GPU 内部运行。这提供了硬件保证(由英伟达和第三方提供商支持),没有人(包括提供商)能够窥视内部或查看数据。 * 这不同于“零数据保留”,后者仅仅是基于提供商承诺的政策承诺。机密推理使用硬件签名使隐私在技术上得到证明。 * 存储选项包括“Second Local”(数据保留在本地,类似于密码保险库)或云托管版本,其中数据首先在本地加密(类似于 iCloud 或密码管理器),并且仅在内存中或完全在用户可以验证的安全机密容器内部进行短暂解密。 * 防止大语言模型(LLMs)泄漏信息依赖于确定性保证(在应用程序级别实现)和基于推理的保证(随模型能力扩展并使用上下文来指导关于共享什么数据的决策)。 **成本、模型与性能** * 24小时全天候运行分析计算机流的后台代理消耗大量的令牌。肯提供了一个数学示例:以 Opus 定价,24小时全天候运行10个代理,速度为每秒60个令牌,每天的成本将达到1,000美元,而使用 DeepSeek Flash 则将成本降至每天60美元。自该公司成立以来,令牌价格已经下降了50%以上。 * 该系统旨在具有便携性并在任何模型上运行,利用工具包(harness)设计来确定何时使用哪些模型以实现成本效率(例如,使用昂贵的智能模型来处理诸如理解用户每周目标的高级目标,以及使用像杰夫这样的模型进行检索)。 **用例与未来愿景** * 令人惊奇的开发用例:“第二”充当了编程的超级记忆。用户不再需要向编程代理解释任何事情;他们只需截取任何屏幕截图,说“修复”,代理就知道该修复什么。它还实现了“上下文连续性”,使用户能够无缝地从笔记本电脑上的浏览过渡到在手机上编写电子邮件,并具有最新的上下文。 * 产品界面:目前,该应用在后台运行,请求权限并生成 API(例如用户10分钟前做了什么、下周的优先级、提醒、维基以及用户生活中的人)。用户可以通过文本和 API 进行交互,或者使用音频模式(例如在散步时通过音频呼叫 Grokbot)与他们的“Second”进行交互。 * 未来演变:该公司计划最终将界面向环境计算、具身计算和泛在计算(如贾维斯或科幻文化系列人工智能同伴)演变,每个人都有一个“第二”,所有设备都充当该核心模型的输入和输出。

Here is a summary of the podcast episode based exclusively on the provided transcript: **Guest and Company Overview** * Ken is a founder at First Person Labs, talking about the company's work for the very first time on the podcast. * First Person Labs focuses on creating a model of a person (understanding their memories, priorities, preferences, goals, values, commitments, and intent) and putting that person in control. **The Core Product: "Second"** * The company creates a personal agent/model called a "Second." * A Second acts as a personal butler or agent (akin to an "Alfred to Batman"), securely holding a person's life context and serving as a trusted vault. * First Person Labs acts as its own thing, rather than being an agent itself like Grokbot or Codex. Other specialized agents (such as Grokbot, Codex, or ephemeral agents) interface with the user's Second to understand what the user wants. * The host envisions that having many personal agents is transitory, and the Second is intended to be the one trusted agent/vault for all personal data that interacts with all other software and hardware agents. * The Second continuously analyzes the infinite stream of events as a user uses their computer (app switches, pauses on websites, typing, wi-fi connections) 24/7. It hierarchically builds up an understanding of what the user is trying to do, which is paired with connectors (like slack, signal, whatsapp, and raw computer streams) to understand the user. * It catches things and provides proactive help—such as updating event weather forecasts in real-time or pinging users regarding ongoing conversations—ranging from reactive productivity (simple email/meme notifications) to reflective productivity (connecting dots across long time horizons and priorities). * Ken notes that Second is available today (at `two dot computer`), though they are scaling up inference and capacity and rolling it out responsibly. **Security, Architecture, and Privacy** * Building a system that sees everything a user does raises immediate trust concerns, so First Person Labs prioritizes privacy and security architecture. * Data and inference are the two components of AI. Users can choose what inference they bring. If they do not trust First Person Labs, they can bring their own tokens to analyze data locally on their laptop. * Alternatively, Second provides an inference provider using trusted execution (confidential computing/confidential inference), where a model runs inside GPUs with a special toggle. This provides a hardware guarantee (backed by NVIDIA and third-party providers) where no one—including the provider—can see inside or view the data. * This differs from "zero data retention," which is merely a policy promise based on a provider's word. Confidential inference uses hardware signatures to make privacy technically provable. * Storage options include "Second Local" (data stays local, similar to a password vault) or a cloud-hosted version where data is encrypted locally first (similar to iCloud or password managers) and only decrypted ephemerally in memory or entirely inside a secure confidential container that the user can attest. * Preventing LLMs from leaking information relies on deterministic guarantees (implemented at the app level) and reasoning-based guarantees (scaling with model capability and using context to guide decisions on what data to share). **Cost, Models, and Performance** * Running background agents analyzing a computer stream 24/7 consumes a significant number of tokens. Ken provides a math example: running 10 agents 24/7 at 60 tokens per second using Opus pricing would cost $1,000 per day, whereas using DeepSeek Flash drops the cost to $60 per day. Token prices have already dropped by more than 50% since the company started. * The system is built to be portable and run on any models, utilizing a harness design to determine which models get used when for cost efficiency (e.g., using expensive smart models for high-level goals like understanding a user's weekly goals, and using models like Jeff for retrieval). **Use Cases and Future Vision** * Surprising development use cases: Second acts as a super memory for coding. Users no longer need to explain anything to coding agents; they can just screenshot anything, say "fix," and the agent knows what to fix. It also enables "context continuity," allowing a user to seamlessly move from browsing on a laptop to writing an email on a phone with up-to-the-minute context. * Product interface: Currently, the app runs in the background, asks for permissions, and produces APIs (such as what the user did 10 minutes ago, priorities for next week, reminders, wiki, and people in the user's life). Users can interact via text and APIs, or use audio modes (like calling Grokbot on audio while walking) to interface with their Second. * Future evolution: The company plans to eventually evolve the interface toward ambient, embodied, and ubiquitous computing (like Jarvis or sci-fi culture series AI companions), where every person has a Second and all devices act as inputs and outputs to that same core model.

摘要

AI agents are getting smarter, but every new agent still has to learn who you are. What if your preferences, priorities, relationships, and digital context could move with you instead? First Person Labs co-founder Ken Liu joins a16z’s Martin Casado and Joel de la Garza to introduce Second, a personal context layer designed to understand what you’re doing across your digital life and make that context available to the agents you choose. They discuss how Second can proactively connect information across conversations and applications, why protecting such personal data requires a different security architecture, and the longer-term vision of AI that doesn’t just respond to prompts, but understands enough context to know when and how to help. Timestamps: 0:00 Intro 0:50 What Is First Person Labs? 2:43 The Alfred Model: Your Agent's Butler 3:56 What Your Computer Already Knows About You 5:59 Proactive AI: Reactive vs. Reflective 8:16 Why You Only Need One Second 10:27 Can You Trust Something That Sees Everything? 15:41 Can an LLM Keep Your Secrets? 17:40 Cost, Models, and the Target User 20:20 Use Cases and the Model Mix 23:34 How You Use Second (APIs and Apps) 26:20 Toward Jarvis: AI Everywhere 30:43 Fort Knox Meets Delightful UX Resources: Try Second: https://2.computer Learn more about First Person Labs: https://fpl.company/ Follow Ken Liu: ​​​​https://www.linkedin.com/in/kenziyuliu/ Follow Martin Casado on X: https://x.com/martin_casado Follow Joel de la Garza: https://x.com/YODELAGARZA Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

GPT-4正在为你翻译摘要中......

中英文字稿