Blaxel 加入 Baseten,共建智能体基础设施的未来

Baseten Blog ·

一份描述 Blaxel 面向自主智能体的技术基础设施的来源——涵盖微型虚拟机沙盒性能、生产环境网络设计、推理共置要求以及基础性的基础设施设计原则——基于归属于 Blaxel 的陈述。 阅读 4 条观点,查看支持证据与原始来源。

Amir Haghighat, Tuhin Srivastava, Paul Sinai

理解这篇

4 个要点

综合解读

  1. 具备 25 毫秒挂起/恢复能力的微型虚拟机沙盒

    Blaxel 为智能体构建了隔离的微型虚拟机沙盒,实现安全的代码执行和状态持久化。这些沙盒可在 25 毫秒内挂起和恢复——比其他沙盒产品快多达 5 倍——并且能够以接近零成本保持空闲数月,同时保留状态。

    支持这项说法 1

    每个智能体都拥有自己的微型虚拟机。我们让它们变得极快:在 25 毫秒内挂起和恢复,比其他沙盒产品快多达 5 倍。一个沙盒可以以接近零的成本空闲数月,并在模型完成下一句话之前恢复运行,因此状态得以持久保存,而无需为你并未使用的计算资源付费。

    Amir Haghighat, Tuhin Srivastava, Paul Sinai · 段落 6

    原始摘录
    Every agent gets its own microVM. We made them extremely fast: suspend and resume in 25 milliseconds, up to 5x faster than other sandbox products. A sandbox can sit idle for months at close to zero cost and come back before the model finishes its next sentence, so state persists without paying for compute you aren't using.
    上下文

    沙盒是首要任务,因为没有它们其他一切都无法运作。智能体会编写并运行代码、调用工具,并在长时间任务中携带状态,而在共享进程中无法安全地做到这一点。

    原始上下文

    Sandboxes came first because nothing else works without them. Agents write and run code, call tools, and carry state across long tasks, and you can't do that safely in a shared process.

    回到原文语境 →
  2. 重建的网络层,具备隔离和访问控制能力

    Blaxel 重建了其网络层,用于将智能体与工具、MCP 服务器、API 以及其他智能体相连接——其设计专门满足生产团队对隔离和访问控制的要求。

    支持这项说法 1

    代理需要与各类对象交互:工具、MCP服务器、API,以及其他代理。我们重建了网络层,将所有这些组件连接起来,并提供了生产团队实际能批准的隔离机制和访问控制。

    Amir Haghighat, Tuhin Srivastava, Paul Sinai · 段落 9

    原始摘录
    And agents have to talk to things. Tools, MCP servers, APIs, other agents. We rebuilt a networking layer that connects all of these pieces with the isolation and access controls a production team will actually sign off on.
    回到原文语境 →
  3. 推理必须紧邻计算、存储和网络

    Blaxel 团队观察到,面向自主智能体的推理不再是一项远程服务调用:它必须与计算、存储和网络共置,以避免在每次智能体循环中累积延迟。

    支持这项说法 1

    我们未能自主掌控的原语是推理。我们已经看到,开放权重模型与定制模型正成为自主智能体在生产环境中部署方式的核心。推理不再是一项可以从另一个数据中心调用的服务。它必须紧邻计算、存储和网络;否则延迟会在每一次循环中不断累积。

    Amir Haghighat, Tuhin Srivastava, Paul Sinai · 段落 10

    原始摘录
    The primitive we didn't own was inference. We've seen open-weight & custom models become core to how autonomous agents get deployed in production. Inference is no longer a service you call from another data center. It has to sit next to compute, storage, and network; or latency compounds on every loop.
    回到原文语境 →
  4. 将性能、可靠性、安全性和所有权作为首要考量

    Blaxel 和 Baseten 都从零开始构建其基础设施——优先考虑性能、可靠性、安全性、可扩展性、灵活性、成本效益,以及客户的可见性、控制权和所有权——而非选择最快推向市场的路径。

    支持这项说法 1

    两家公司都是从零开始构建的,目的是真正将工作负载扩展到满足客户需求的程度,而不是选择最快推向市场的路径。我们都已决定,性能、可靠性、可扩展性、安全性、灵活性和成本效益必须作为首要考量。而且我们都认为,客户需要最大程度的可见性、控制权和所有权。

    Amir Haghighat, Tuhin Srivastava, Paul Sinai · 段落 16

    原始摘录
    Both companies built from the ground up for what it would take to truly scale the workload to our customers’ requirements, rather than taking the fastest path to market. We had both decided that performance, reliability, scalability, security, flexibility, and cost-efficiency needed to be first-class considerations. And we each believed customers need maximum visibility, control, and ownership.
    回到原文语境 →

关键段落4

带明确归属与语境的原文片段。打开原始文本核查出处。

基础设施设计理念

将性能、可靠性、安全性和所有权作为首要考量

两家公司都是从零开始构建的,目的是真正将工作负载扩展到满足客户需求的程度,而不是选择最快推向市场的路径。我们都已决定,性能、可靠性、可扩展性、安全性、灵活性和成本效益必须作为首要考量。而且我们都认为,客户需要最大程度的可见性、控制权和所有权。

原始摘录
Both companies built from the ground up for what it would take to truly scale the workload to our customers’ requirements, rather than taking the fastest path to market. We had both decided that performance, reliability, scalability, security, flexibility, and cost-efficiency needed to be first-class considerations. And we each believed customers need maximum visibility, control, and ownership.
智能体沙盒性能

具备 25 毫秒挂起/恢复能力的微型虚拟机沙盒

每个智能体都拥有自己的微型虚拟机。我们让它们变得极快:在 25 毫秒内挂起和恢复,比其他沙盒产品快多达 5 倍。一个沙盒可以以接近零的成本空闲数月,并在模型完成下一句话之前恢复运行,因此状态得以持久保存,而无需为你并未使用的计算资源付费。

原始摘录
Every agent gets its own microVM. We made them extremely fast: suspend and resume in 25 milliseconds, up to 5x faster than other sandbox products. A sandbox can sit idle for months at close to zero cost and come back before the model finishes its next sentence, so state persists without paying for compute you aren't using.
上下文

沙盒是首要任务,因为没有它们其他一切都无法运作。智能体会编写并运行代码、调用工具,并在长时间任务中携带状态,而在共享进程中无法安全地做到这一点。

原始上下文

Sandboxes came first because nothing else works without them. Agents write and run code, call tools, and carry state across long tasks, and you can't do that safely in a shared process.

面向智能体的推理共置

推理必须紧邻计算、存储和网络

我们未能自主掌控的原语是推理。我们已经看到,开放权重模型与定制模型正成为自主智能体在生产环境中部署方式的核心。推理不再是一项可以从另一个数据中心调用的服务。它必须紧邻计算、存储和网络;否则延迟会在每一次循环中不断累积。

原始摘录
The primitive we didn't own was inference. We've seen open-weight & custom models become core to how autonomous agents get deployed in production. Inference is no longer a service you call from another data center. It has to sit next to compute, storage, and network; or latency compounds on every loop.
面向智能体的生产环境网络

重建的网络层,具备隔离和访问控制能力

代理需要与各类对象交互:工具、MCP服务器、API,以及其他代理。我们重建了网络层,将所有这些组件连接起来,并提供了生产团队实际能批准的隔离机制和访问控制。

原始摘录
And agents have to talk to things. Tools, MCP servers, APIs, other agents. We rebuilt a networking layer that connects all of these pieces with the isolation and access controls a production team will actually sign off on.

来源与研究方法

这些观点均关联原始来源。转述已明确标注,不作为逐字原话展示。

打开转录或来源材料 (在新标签页中打开)报告问题