嘿，是西格诺兹的伊丽莎白！
这篇通讯是一次诚实的尝试来谈论所有的事情——可观察性，OpenTeleometry，开源和工程介于两者之间！
&
这幅画花了6天，5小时才被煮熟，希望我们服务。
现在有两种流行的预言浮在科技圈里
第一个称 SRE是所有软件工程的未来，随着AI写出越来越多的代码，留下的人类将会是那些保持系统生命力的人。 第二个称AI会活吞掉所有技术工作，包括SREs. 如果你是SRE，试图弄清楚你的星期二在2027年会是什么样子
让我们通过观察已经发生的事情，当AI处理你95%的事件反应，你得到更糟糕处理5% 这实际上很重要？
我们大部分人已经用AI做日常工作(我们的大脑有点被炸了！), SREs也一样。 今天的讨论议程不是AI是否取代 SREs，而是AI是否悄悄地使 SREs 能力降低，整个辩论的基础框架来自1983年的研究论文
自动化的讽刺：第一部分
作者注：以下是自动化史的简短前奏
1983年，一位名为Lisanne Bainbridge的认知心理学家发表了一篇名为"自动化讽刺"的论文。 它成为人类因素研究中引用最多的论文之一，其核心论点在今天几乎不适相关。
班布里奇研究了工厂和工业系统使蓝领运营商的工作自动化时发生的事情。 这些发现是自相矛盾的，揭示了一个过程越自动化，人类操作员在难得的瞬间，自动化失败，和他们得到的实践越少，和他们变得越糟糕正是在这些干预。 自动化本来就是为了把人类从循环中除去，却让他们有最差的工作，即被他们越来越没有准备的罕见的、高度的危机所吸引的长时间被动监测。
有印象吗？
基本上，我在AI革命和工业自动化之间划出一个平行点。 工业自动化重塑了蓝领工作，AI通过接管日常认知任务，模式匹配，分解，已知问题解析，让人类面临罕见的，复杂的，模棱两可的问题，对知识工作者做同样的事情。
真正的问题需要深层的专业知识，准确的专业知识奖杯当你停止练习。
现在，我们用AI agents来重现这个模式，软件系统的利害关系只是不断增长。
SRE中的大赦国际现状
让我们来总结一下今天在遗址可靠性工程世界中的情况。
何者已经自动或大量人工智能协助？
提醒降噪和智能组合、针对已知问题的运行图执行、日志模式检测和异常标记、以及来自历史事件数据的基本根源建议、以及诸如重启坠机舱或扩大运行热度服务等众所周知的故障模式的自动补救，如今都是相当自动化的。
未来前景如何？
一些直接目标包括跨度量衡的多信号关联，日志，和痕迹，自主的根因子分析部分理解的故障，在用户受到影响之前的预测性事件检测，AI驱动的改变风险评估和自动回滚。
PagerDuty 帧这个作为阶梯模型。
- 第1级事件：已知问题和已知的解决方案完全自动化。
- 第二级事件：部分理解的问题得到大赦国际的建议，并经过人类验证。
- 3级事件：小说、复杂、连锁故障仍由人主导，AI提供辅助环境。
但是
如果人类SREs(好的，现在我们必须像人类一样使用形容词)只与3级事件接触，即小说，从不出现前的停机，他们在哪里建立直觉来处理？知觉通常是从多年的亲身操作事件反应，通过重复构建的模式识别，以及那种对系统只有从痛苦的醒来在奇点时才发展到解决导致系统下沉的bug的直觉层次的理解。
死亡危机
照片开始模糊让我们看看AI引发的跨越多个领域的脱杀研究
在医学方面，最近一项研究发现，使用人工智能协助进行多肽检测的内科医生发现，在使用人工智能一段时间后，他们的无辅助检测率从28%下降到22%。 他们越来越糟糕了他们应该成为专家并不是因为他们忘记了理论而是因为他们不再锻炼技能
在航空方面，研究表明，严重依赖自动驾驶系统的长途飞行员在情况意识和人工飞行能力方面经历了可衡量的退化。 问题变得非常严重，联邦航空局现在规定更多的人工飞行时间来抵消这种影响。
在过去一年的某个时候，AI不再成为我偶尔找到的工具，而成为我总想得到的第一件事。 我的本能是尽可能卸载，问题是，这些时刻正在成为我脑中唯一的锻炼，我能感觉到生锈。
我们可以在这里画一个图案。 你越是让系统处理，你自己处理事情就越糟糕，而这里才是真正危险的部分，你并不觉得会发生这种情况。 它被掩盖为超生产力。 认知研究表明，由于AI工具使任务感觉更加容易，提高了可见性能，用户往往无法准确判断自己技能的真实状况。 你感到自己有能力，仪表板看起来是绿色的，然后在星期三，一个小事件达到，它与AI所看到的模式不符，你意识到肌肉萎缩了。
对 SREs来说，这表现在具体的方式上，比如我们停止读取原始的日志流，因为AI总结了它们，在事件发生时我们停止形成假设，因为AI暗示了根本原因，我们停止构建系统架构的心理模型，因为AI为我们绘制了依赖性图，每个个体看起来都像生产力胜利。 集体地说，当事情以无人预料的方式发生时，它们就把使SRE有效的专门技能掏空了。
研究者开始称它为永不杀人。 杀死意味着你曾经有能力但后来就失去了能力永远不杀意味着你从来没有开发过它首先。 对于初级SRE在AI处理大部分事件响应工作流程的环境中进入今天的球场，通过亲身实践建立基础直觉和肌肉的机会正在消失。
训练管道本身断裂，不能自愈。
SRE意识到他们的技能在有辱人格，更依赖AI来补偿，这进一步降低了他们的技能，造成了一种恶性循环，从中逃生是困难的。
我们能做些什么呢？
我们绝对不会拒绝人工智能工具；我们正在采用它，并将它整合起来，这比以往任何时候都更加有力，因为这是唯一的出路。
值得考虑的几种办法：
故意无效率。 正如联邦航空局授权人工飞行时间，即使自动驾驶完全有能力，SRE小组也可以指定某些事件，即使是AI能够处理的事件，作为人类实践的机会。 这可被视为一项长期投资，以保持技能新鲜，尽管这样做可能以超快解决方案为代价。
建为人行于上，非人行于上。 人类同意AI建议的制度与人类与AI一起积极处理问题的制度之间有着有意义的区别。 前者将人类置于班布里奇(大约40年前写过研究论文的女士)所显示的监督角色中，导致警惕衰减，后者则让他们保持认知接触。
让zoom出局
更大的图片
我们在这里讨论的每件事，自动化的讽刺，消除杀机的风险，永不杀机的问题，集体应用远远超出了SRE. 软件工程作为一个整体，正在导航同样的张力。 随着AI写更多的代码，审查更多的PR，处理更多的调试，同样的问题也适用。
我们谈论SRE，是因为那是我们生活在西格诺兹的世界。 我们建造了一个开源可观察性平台，这种工具可以让SRE获得他们需要的测量，痕迹和日志来深入了解他们的系统。 对我们来说，这个解毒问题并不是一个空谈问题；它直接决定了我们如何将AI建设成我们的产品。
我们的方法是从AI助手开始，帮助SREs利用LLMs的力量，同时使人类牢牢控制。 最终，我们将让更多的自主权
我们在这个空间的一个优势是，作为一个可观测平台，我们坐在数据本身，测量，痕迹，以及SRE所依赖的日志上。 大多数AI SRE产品如今都通过API与可观察性工具融合，因为我们拥有数据层，我们可以建立更深，更了解背景的AI能力，来理解你的系统就像经验丰富的SRE那样。
我们的目标不是AI取代SRE，是AI给SRE充电。 与目前的神话不同，我们认为，人类对于最重要的决定，特别是那些影响生产基础设施的决定，仍然至关重要。
SRE的未来是人类的，AI有意设计让人类保持锋利，参与，为真正重要的5%做好准备。
我们的创始人几天前张贴的LinkedIn文章激励我写这个。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
💌 Hey there, it’s Elizabeth from SigNoz!
This newsletter is an honest attempt to talk about all things - observability, OpenTelemetry, open-source and the engineering in between!
&
This piece took 6 days, 5 hours to be cooked, hope we served. 🌚
There are two popular prophecies floating around tech circles these days.
The first says SRE is the future of all software engineering, that as AI writes more and more code, the humans who remain will be the ones keeping systems alive. The second says AI will devour every tech job alive, SREs included. Neither is particularly useful if you’re an SRE trying to figure out what your Tuesday will look like in 2027.
Let’s ask a more grounded question by looking at what’s already happening: When AI handles 95% of your incident response, do you get worse at handling the 5% that actually matters?
Most of us already use AI for our daily work (our brains a little fried!), and so do SREs. Today’s discussion agenda is not whether AI replaces SREs, but whether AI is quietly making SREs less capable and whether anyone will notice anything before the next novel outage hits. The foundational framework for this entire debate comes from a 1983 research paper that’s eerily prescient.
The Ironies of Automation: Part-I
Note from author: Below is a brief precursor to History of Automation, which you might enjoy if you are into History and Cultures (like me 😉).
In 1983, a cognitive psychologist named Lisanne Bainbridge published a paper called Ironies of Automation. It became one of the most cited papers in human factors research, and its core argument is almost uncomfortably relevant today.
Bainbridge studied what happened when factories and industrial systems automated the work of blue-collar operators. The findings were paradoxical and revealed that the more you automate a process, the more critical the human operator becomes during the rare moments automation fails and the less practice they get, and the worse they become at exactly those interventions. Automation, which was inherently designed to remove humans from the loop, left them with the worst possible job, i.e., long stretches of passive monitoring punctuated by rare, high-stakes crises they were increasingly unprepared for.
Ring any bells yet? 🙂
Basically, I’m drawing a parallel between the AI revolution and industrial automation. Industrial automation reshaped blue-collar work by taking over routine physical tasks, and the workers who remained had to handle exceptions they’d lost the muscle memory for. AI is doing the same thing to knowledge workers by taking over the routine cognitive tasks, the pattern matching, the triage, the known-issue resolution and leaving humans with the rare, complex, ambiguous problems.
The exact problems that require deep expertise, the exact expertise that atrophies when you stop practising.
Now we’re replaying this pattern with AI agents, and the stakes in software systems are only growing.
Current State of AI in SRE
Let’s take stock of where things stand today in the world of site reliability engineering.
What’s already automated or heavily AI-assisted?
Alert noise reduction and intelligent grouping, runbook execution for known issues, log pattern detection and anomaly flagging, and basic root-cause suggestions from historical incident data, and auto-remediation for well-understood failure modes like restarting a crashed pod or scaling up a service that’s running hot, are all fairly automated today.
What’s on the horizon?
Some immediate targets include multi-signal correlation across metrics, logs, and traces, autonomous root-cause analysis for partially understood failures, predictive incident detection before users are affected, AI-driven change risk assessment and automated rollbacks.
PagerDuty frames this as a tiered model.
- Tier 1 incidents: Known issues with known fixes get fully automated.
- Tier 2 incidents: Partially understood problems receive AI recommendations with human validation.
- Tier 3 incidents: Novel, complex, cascading failures stay human-led, with AI providing supporting context.
But here’s the catch.
If human SREs (okay, now we have to use adjectives like human ) only engage with Tier 3 incidents, i.e. the novel, never-before-seen outages, where do they build the intuition to handle them? Intuition is usually developed from years of hands-on incident response, pattern recognition built through repetition, and the kind of gut-level understanding of a system that only develops from painfully waking up at the odd hour to solve the bug that brought the system down.
The Crisis of Deskilling
This is where the picture starts getting blurry. Let’s look at some emerging research on AI-induced deskilling across multiple fields, which paints a consistent and concerning picture.
In medicine, a recent study found that endoscopists who used AI assistance for polyp detection saw their unassisted detection rates drop from 28% to 22% after a period of AI use. They got worse at the thing they were supposed to be experts in, not because they forgot the theory, but because they stopped exercising the skill.
In aviation, research has shown that long-haul pilots who rely heavily on autopilot systems experience measurable degradation in situational awareness and manual flying ability. The problem got serious enough that the FAA now mandates more manual flying time to counteract the effect.
Somewhere over the past year, AI stopped being a tool I occasionally reached for and became the first thing I reach for, always. My instinct now is to offload as much as possible and apply my own thinking only where it’s absolutely unavoidable. The problem is that these moments are becoming the only exercises my brain gets, and I can feel the rust.
We can draw a pattern here. The more you let the system handle, the worse you get at handling things yourself and here’s the truly dangerous part, you don’t feel it happening. It gets masked as hyper-productivity. Cognitive research suggests that because AI tools make tasks feel easier and enhance visible performance, users are often unable to accurately judge the true status of their own skills. You feel competent, dashboards look green, and then on a Wednesday, a novel incident hits that doesn’t match any pattern the AI has seen, and you realise the muscle has atrophied.
For SREs, this manifests in specific ways, like we stop reading raw log streams because the AI summarises them, we stop forming hypotheses during incidents because the AI suggests root causes, we stop building mental models of system architecture because the AI maps dependencies for us, and each of these individually looks like a productivity win. Collectively, they hollow out the very expertise that makes an SRE effective when things go sideways in ways nobody anticipated.
But there’s something even more concerning than deskilling, and researchers have started calling it never-skilling. Deskilling means you once had a capability but have since lost it. Never-skilling means you never developed it in the first place. For junior SREs entering the field today in an environment where AI handles most of the incident response workflow, the opportunities to build foundational intuition and muscle through hands-on practice are vanishing.
The training pipeline itself is broken and not self-healing.
SREs realise their skills are degrading and lean more on AI to compensate, which further degrades their skills, creating a vicious cycle from which escape is difficult.
What Can We Do About It?
We are definitely not rejecting AI tooling; we are adopting it and integrating it stronger than ever before, because that’s the only way forward.
A few approaches worth considering:
Deliberate inefficiency. Just as the FAA mandates manual flying time even when the autopilot is perfectly capable, SRE teams can designate certain incidents, even the ones the AI could handle, as human-practice opportunities. This can be considered as a long-term investment to keep skills fresh, although it might come at the cost of a super-fast solution
Build for human-in-the-loop, not human-on-the-side. There’s a meaningful difference between a system where a human approves an AI’s recommendation and one where a human actively engages with the problem alongside AI. The former keeps humans in a supervisory role that Bainbridge (the lady who wrote that research paper about 40 years ago) showed leads to vigilance decay, and the latter keeps them cognitively engaged.
Let’s zoom out and take a look at the bigger picture.
The Bigger Picture
Everything we’ve discussed here, the ironies of automation, the deskilling risk, the never-skilling problem, collectively applies well beyond SRE. Software engineering as a whole is navigating the same tension. As AI writes more code, reviews more PRs, and handles more debugging, the same questions apply.
We’re talking about SREs specifically because that’s the world we live in at SigNoz. We build an open-source observability platform, the kind of tool that gives SREs the metrics, traces, and logs they need to understand their systems deeply. For us, this deskilling question is not a rhetorical fad; it directly shapes how we’re building AI into our product.
Our approach is to start with an AI assistant that helps SREs leverage the power of LLMs while keeping humans firmly in control. Eventually, we’ll enable more autonomy but within clear guardrails, and only as trust is earned.
One advantage we have in this space is that, as an observability platform, we sit on the data itself, the metrics, traces, and logs that SREs rely on. Most AI SRE products today integrate with observability tools through APIs, which means they’re working with a limited, second-hand view of your systems. Because we own the data layer, we can build much deeper, more context-aware AI capabilities that understand your system the way an experienced SRE would.
And to answer the burning question in your head, our goal isn’t AI that replaces SREs. It’s AI that supercharges SREs. Unlike ongoing lore, we believe humans will remain essential for the decisions that matter most, especially those that impact production infrastructure.
The future of SRE is human with AI intentionally designed to keep humans sharp, engaged, and ready for the 5% that really counts.
Here’s the LinkedIn post our founder posted a few days ago, which inspired me to write this.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.