在这期节目中，主持人采访了 Grepper 首席执行官、曾在斯坦福读研时的同窗杰德（Jed），深入探讨了系统思维、工程学，以及人工智能如何改变软件开发和运维。 对话从轻松回顾他们在导师尼克·麦基恩（Nick McKeown）手下读研的时光开始。杰德分享了自己丰富的背景，从早期在硬件领域从事 SHA-3 的工作，到在 Nysera 的经历、在麻省理工学院（MIT）跟随尼古拉·泽尔多维奇（Nikolai Zeldovich）做博士后、在 App Dynamics 被思科（Cisco）收购前的任职经历，以及在 Imply 担任的领导职务。由于观察到工程团队在可观测性上大量投入，以此作为可靠性的替代指标，杰德受到启发创立了 Grepper，旨在帮助企业大幅降低遥测成本，并将这些信号输入到人工智能驱动的故障排查系统中。 谈到人工智能对软件开发的更广泛影响，杰德指出，早期的人工智能工具仅限于基本的自动补全，但它们已经有了显著的演变。对于懂得如何组织复杂任务的顶级工程师来说，人工智能极大地加速了产出。杰德描述了他使用 Windows Linux 子系统（WSL）、Claude Code 以及 Opus 5.5 等模型的个人工作流，这使他能够在大约 10 天内（而不是过去的 6 个月）交付庞大的、达到生产级别的项目，例如自动升级的数据湖架构。不过，他强调，优秀工程师与糟糕工程师之间的根本差距反而拉大了；糟糕的工程师并没有变成优秀的工程师，但优秀的工程师变得极其高效。 随后，讨论转向了将人工智能应用于分布式系统和基础设施运维的复杂性。虽然编程主要是一个语法驱动的领域，非常适合映射到海量的训练数据集，但运维则涉及厚尾异常、不可预测的运行时语义以及现实世界杂乱的复杂性。尽管一些怀疑论者（如主持人）认为运维过于混乱且状态空间过于庞大，不适合人工智能，但杰德则更为乐观。他解释说，Grepper 处理运维人工智能的方法不是从指标飙升进行滞后的回溯，而是利用预测模型来标记异常和新出现的日志模式。然后，人工智能代理（agent）可以通过结构化的工作流、假设和验证步骤来隔离更改并解决问题。 尽管取得了这些进展，但杰德承认，通过人工智能端到端地运行大型分布式集群仍然在未来。代码生成享有即时的经济采用率，因为它直接针对工程生产力，而人工智能驱动的 SRE 和安全工具仍在成熟中，面临着价值认知以及对抗性或罕见生产状态的纯粹不可预测性的障碍。 最后，两人探讨了人工智能的更广泛发展轨迹，将数字领域（其解空间与训练分布紧密契合）与现实世界（经济均衡、系统约束和仿真限制决定了可行性）进行了对比。杰德提出，随着人工智能工具的扩展，人类可能会越来越多地塑造物理和数字环境，以更好地适应人工智能系统高效运行所需的要求。

In this episode, the host interviews Jed, CEO of Grepper and a former Stanford grad school peer, to dive deep into systems thinking, engineering, and how artificial intelligence is transforming software development and operations. The conversation begins with a lighthearted look back at their time in grad school under advisor Nick McKeown. Jed shares his extensive background, from his early work on SHA-3 in hardware to his time at Nysera, a postdoc at MIT with Nikolai Zeldovich, his stint at AppDynamics before its Cisco acquisition, and his leadership role at Imply. Motivated by observing engineering teams heavily spending on observability as a proxy for reliability, Jed started Grepper to help companies massively reduce telemetry costs and feed those signals into AI-driven troubleshooting systems. Shifting to the broader impact of AI on software development, Jed notes that while early AI tools were limited to basic autocomplete, they have evolved significantly. For top-tier engineers who understand how to structure complex tasks, AI has drastically accelerated output. Jed describes his personal workflow using Windows Subsystem for Linux (WSL), Claude Code, and models like Opus 5.5, enabling him to ship massive, production-grade projects—such as auto-promotion data lake architectures—in about 10 days instead of the previous six months. However, he emphasizes that the fundamental disparity between good and bad engineers has only widened; bad engineers have not become good ones, but good engineers have become vastly more productive. The discussion then pivots to the complexities of applying AI to distributed systems and infrastructure operations. While coding is primarily a syntax-driven domain that maps well to a massive training dataset, operations involve heavy-tailed exceptions, unpredictable runtime semantics, and the natural world's messy complexities. Although some skeptics, like the host, view operations as too chaotic and state-space-heavy for AI, Jed is more optimistic. He explains that Grepper approaches operational AI not by lagging backtracking from spiked metrics, but by utilizing predictive models to flag anomalies and newly emerged log patterns. An AI agent can then run through structured workflows, hypotheses, and verification steps to isolate changes and resolve issues. Despite these advancements, Jed acknowledges that running a large distributed cluster end-to-end via AI is still on the horizon. Code generation enjoys immediate economic adoption because it directly targets engineering productivity, whereas AI-driven SRE and security tooling are still maturing, facing hurdles in value perception and the sheer unpredictability of adversarial or rare production states. Finally, the two explore the broader trajectory of AI, comparing the digital domain—where the solution space aligns closely with the training distribution—with the real world, where economic equilibrium, system constraints, and simulation limits dictate feasibility. Jed suggests that as AI tools expand, humanity may increasingly shape physical and digital environments to better fit what AI systems require to operate efficiently.