• Human Infrastructure
  • Posts
  • Human Infrastructure 432: AI Cookbook Review, Latency Monitoring, Systems Thinking for Automation, and More

Human Infrastructure 432: AI Cookbook Review, Latency Monitoring, Systems Thinking for Automation, and More

AuthorAuthor

Drew Conry-Murray & Ethan Banks
January 22, 2026

THIS WEEK’S MUST-READ BLOGS 🤓

Pat Allen is killing it with his blog series on the hurdles you face as you move from a technical role into management. The latest installment is about resisting the impulse to immediately begin making network changes. He offers practical advice on documenting and understanding the network, asking the right questions to help you understand why particular choices were made, and why things that look “wrong” to you might be there for good reasons. He addresses potential political pitfalls, and how to avoid getting stuck in them. He also shares some of his mistakes that he’s made in hopes that you don’t have to make them, too. - Drew 

AJ Murray reviews the “AI Networking Cookbook,” the latest project from the indefatigable Eric Chou. AI is, of course, everywhere. If you’re feeling overwhelmed by hype, overly confident that AI is all BS, or worried about falling behind, this book should be in your hands. AJ’s review gets into why, but here’s the kicker: “If you’re a network engineer wondering how AI actually fits into your job—not the hype, but the practical reality—AI Networking Cookbook is one of the best places to start.” - Drew

AI certainly invokes strong feelings. On the one hand, there’s Eric’s new AI Networking Cookbook mentioned above. That is, let’s make use of the tool, because at the end of the day, AI is just another tool. Let’s stop demonizing it and make it part of our network operations workflows. Here’s a book to get you going.

On the other hand, folks like Andrew Alston argue vociferously against AI, at least for some use cases such as an LLM generating code. Andrew’s got his reasons why, and they’re well thought out. The piece is a great read you should take the time to consume.

On yet another hand, there’s the question of the business perspective. If the business cares about productivity, and you can get working code delivered more quickly if an LLM vomits the code forth for you, what technique is going to win?

In the short term, the business will probably lean into AI-generated code—a cheaper, faster result that, to an outsider, is indistinguishable from what a human would generate. In the long term? Well, if the AI code is badly documented hot garbage, today’s shortcut becomes tomorrow’s obstacle—technical debt the business can’t move beyond. I’m not sure we’ve had enough time with AI generated code to see what the ultimate impact will be. But businesses often don’t make decisions for the long term anyway. Gotta get that stock ticker up today!

So, think about the issue from your own point of view. AI generated code makes you ignorant of what’s going on, because it’s a code abstraction layer. If you didn’t know how to code in the first place, you don’t care. You simply used AI to Harry Potter a result. Magic! But if you do know how to code, you’re making yourself dumber on the altar of productivity. The more you use AI, the more your programming skills will wane.

That said, I feel this is a nuanced discussion. Let’s say you know how to use OSPF. Cool. But can you explain how shortest path actually works in detail? With graph theory and math and Dijkstra's algorithm and so on? Do you need to know? Some would argue you do. Others would point out that if you can configure OSPF to do the thing…that’s good enough. Is it good enough? Of course, it depends. It depends on your career goals, your job responsibilities, your role within a team, whether or not you want to write an OSPF daemon in your spare time, etc.

My point is that, like an LLM, OSPF is a sort of abstraction layer. You don’t have to know how it works to invoke its powers. If you do know, that helps a lot..but you don’t have to know.

There’s more to consider here, but I feel I’ve overstayed my welcome in this thought exercise. If you read all the way through, thank you for indulging me! - Ethan

Chris Grundemann says organizations need to engage in systems-level thinking to be successful in network automation. And this thinking needs to take people and how they work into account, not just technology and tools. It also requires organizations to get comfortable with not knowing, and being willing to conduct experiments that might not work. He writes “...you need “a general understanding of critical patterns” and then you collaborate with actors in the ecosystem to test simple ideas.” His post also offers four principles to help guide a network automation project. - Drew 

MORE BLOGS

Most networks are assembled. Meter networks are built – from the ground up.

Meter delivers vertically integrated networks designed for outcomes, not a list of parts. Hardware, software, and operations work together to deliver predictable performance, security, and reliability.

With Meter, teams avoid stitching together vendors, licenses, and support contracts. Instead, they get one system that’s designed, deployed, and operated end to end.

That means:

  • Enterprise hardware designed in-house

  • Unified software platform across WAN, LAN, and Wi-Fi

  • 24/7 operations with one owner accountable for performance

  • Faster deployments and fewer failure points

The result is networking that scales cleanly and behaves consistently across every site.

Learn how Meter is building enterprise networking differently at meter.com.

TECH NEWS 📣

The title above is from one of the sections in Doug Dawson’s Technology Short Takes piece.

Doug says, “Lasers can create a cooling effect, and this has been used in the past to chill antimatter and to study quantum phenomena. How does this work? Lasers tuned to a specific frequency and targeted at a small area on the surface of a certain element can cool it instead of heating it. Small means an area in the order of hundreds of microns.” He continues, explaining the specifics.

Why is this interesting? Because this method of cooling could save significantly over water circulation used today. Plus, the chips themselves might run longer. - Ethan

TL;DR. Flash drives have service lives. If you re-use an SSD you were using in one drive array in another vendor’s drive array to save money, Dell argues that’s a risky act. The big idea is you’ve already burnt through a bunch of the SSD’s life when it was in the other array, so re-using it is a bad idea.

There are market forces at play here. Dell, not a flash-only storage provider, is throwing shade at companies including VAST Data and PureStorage who are. Dell doesn’t want you to change storage providers. VAST, Pure, et al. want you change to them, but not have to spend a mountain of cash to populate their arrays with new SSDs…if you can even get them. - Ethan

Sergiu Gatlan reports, “The new policy will apply to devices where the Microsoft 365 Copilot and Microsoft Copilot are both installed, the Microsoft Copilot app was not installed by the user, and the Microsoft Copilot app was not launched in the last 28 days.

"Admins can now uninstall Microsoft Copilot for a user in a targeted way by enabling a new policy titled RemoveMicrosoftCopilotApp," the Windows Insider team said.

"If this policy is enabled, the Microsoft Copilot app will be uninstalled, once. Users can still re-install if they choose to. This policy is available on Enterprise, Pro, and EDU SKUs.” - Ethan

MORE NEWS

FOR THE LULZ 🤣

RESEARCH & RESOURCES 📒

Mircea has written an open-source tool called latency-monitor to measure latency and one-way delay on TCP and UDP networks. The blog post covers why he wrote this tool and how it differs from existing options such as ping and Smokeping. He shares configuration examples and talks about how to extract metrics using Clickhouse. It’s still early days for the project, so Mircea writes “Issues, ideas, and pull requests are all welcome.” - Drew

Ivan Pepelnjak reports significant changes behind the scenes in netlab. He covers some of that in detail in How Moving Away from Ansible Made netlab Faster.

But here’s the TL;DR. “I completely rewrote netlab’s device configuration file generation during the New Year break. netlab Release 26.01 no longer uses Ansible Jinja2 functionality and works with Ansible releases 12/13, which are used solely for configuration deployment. I had to break a few eggs to get there; if you encounter any problems, please open an issue.”

More changes that found their way into netlab 26.01 are documented for you when you click through. - Ethan

Where in the US are the best opportunities for people in STEM fields? WalletHub put together a ratings system that accounted for factors including job openings in STEM fields, median wages, the quality of local universities, and quality of life. Using these metrics, they ranked 100 US cities. The results are interesting, particularly if you’re contemplating a change. - Drew  

MORE RESOURCES

  1. Packet Pushers Salary Survey

  2. ping6.it - IPv4 vs IPv6, side by side

  3. Tier 1 Analysis - 53bits 

INDUSTRY BLOGS & VENDOR ANNOUNCEMENTS 💬 

TIL—BGP Vortex. Felix Stöger explains.

“A fundamental problem of BGP is so-called “route oscillations,” where ASes continuously switch between routes for the same IP prefixes. This strains a router’s CPU and can result in loss of connectivity for parts of the network. Previous research has shown that oscillations can be prevented if ASes follow specific routing rules.

However, our research shows how indefinite route oscillations can be created (BGP Vortex) by exploiting two widely used BGP policies:

  • lowering preference of certain routes (“local preference lowering”), and

  • blocking ASes from receiving certain routes (“selective NOPEER”).”

Felix goes on to discuss how a BGP Vortex can be used as a DoS attack and three mitigation strategies. He references an academic paper that dives deeper into this research which can be found here. - Ethan

My take is that the Ubiquiti networking folks are aiming for broader enterprise & MSP acceptance with this “Fabrics” announcement. To be clear, this isn’t an Ethernet fabric. It’s about simplifying and scaling how you manage Ubiquiti networks. - Ethan

The inaugural Toronto Network Operators Group (TORNOG) gathering is scheduled for April 13th, 2026. The organization’s mission is to “bring together network engineers, architects, operators, and enthusiasts to share knowledge, discuss best practices, and build professional relationships.” Registration is open and there’s an open call for talks. If you’re in Toronto, this is a great opportunity to connect with your networking community. - Drew 

Opengear has announced two new products for Out-of-Band management. The CM8000 series is a serial-first platform targeted at “core branches, small aggregation sites, and network closets. The series supports 4 or 8 serial ports, integrates with PDUs for remote power cycling, offers flexible Ethernet uplinks via RJ-45 and SFP connectors for copper or fiber connectivity.” The OM1300 offers serial access and switching in the same device.  - Drew  

ServiceNow has released a new report looking at the risk landscape created by technologies including digitization, cloud computing, and AI. The company interviewed 1,000 risk and security executives to find out “where risk is intensifying, how preparedness varies by industry and region, and what leaders are doing to close the resilience gap.” The report is free to download with no need to exchange contact info. - Drew   

I’m never sure what to make of patent announcements by tech companies. In general, I feel like the US patent system is a bit of a mess, and that the whole process is less about encouraging innovation and more akin to nuclear deterrence (i.e., my company won’t launch an infringement lawsuit against you if you won’t launch one against us). In any case, the AI network ops startup Selector now has 8 patents in its arsenal. In a hotly contested market such as AI, that’s a worthwhile achievement. - Drew 

MORE INDUSTRY NOISES

DYSTOPIA IRL 🐙

  1. How AI Destroys Institutions - Boston University School of Law via Social Science Research Network | Elsevier

TOO MANY LINKS WOULD NEVER BE ENOUGH 🐳

LAST LAUGH 😆

Shared on the Packet Pushers Community Slack by Matthew. Come join us there!