• Human Infrastructure
  • Posts
  • Human Infrastructure 435: Network Complexity, AI Intensity, Attention Deficiency, and More

Human Infrastructure 435: Network Complexity, AI Intensity, Attention Deficiency, and More

AuthorAuthor

Ethan Banks & Drew Conry-Murray
February 12, 2026

THE NETWORKING COMPLEXITY PROBLEM

The topic of networking complexity has popped up in the last few weeks. Ryan Hamel and I connected at NANOG 96 to record a Heavy Networking podcast episode with the premise boring is good. That’ll publish in the next few weeks. Sometime after Ryan and I recorded, I noticed this viral gem posted by Russ White on LinkedIn.

I don’t want to bother positing, “Are our networks too complex?” because the answer is, in most cases, yes. Enterprises, service providers, and probably everyone but the hyperscalers are building networks bordering on incomprehensible to the all but senior people who’ve been around to witness networking’s steady growth of acronyms over the last 30+ years.

My work on the N Is For Networking podcast series has reinforced for me that networking rarely throws things away. Almost everything I had to learn as a CCNA/CCNP/CCIE in the 2000-2008 time frame is still around. In production. We don’t replace as much as we stack on top of.

There’s so much more that we’ve added since my certification days. As a sampling, SPB, LISP, DNSSEC, SRv6, EVPN/VXLAN, YANG, microsegmentation, automation (around which an entire community & conference series has been built), SD-WAN, SASE, and multi-cloud networking are all new on some level or other. Plus we’ve bolted on new features to sundry protocols & transports that have been around forever. I give you Ultra Ethernet, for instance.

Is all of this new stuff necessary? Depends on who you ask and what problem you’re trying to solve, as well as what’s meant by necessary. No matter your take, I think we can agree that network design has become too complex and overcomplicated, both broadly and specifically, in a great many cases.

Notable curmudgeon Randy Bush summed it up in 2005 in his short critique of the IETF when he griped, “Instead of hard thought, rigorous, and simple designs, every possible feature gets added and many competing proposals are approved. This last is like throwing spaghetti at the wall to see what sticks, an amusing tactic to everyone but the wall.

The operators are the wall. And they pay capital cost and operational expense to deploy complex features which vendors market as needs to the users. And then everyone wonders why the margins went down and the prices stayed up.”

I am merely making an observation. I don’t have a concrete answer to the complexity problem…aside from this. Network architects should be revising their designs by taking away as much as possible.

  • “As simple as possible, but no simpler.” - a reductive version of an Einstein quote

  • “Simplify, then add lightness.” - Colin Chapman

  • “Keep it simple, stupid.” - Kelly Johnson

Here’s a talk by David Meyer going back to NANOG 58. Watch Macro Trends, Complexity, and Software Defined Networking (2013) for more thoughts on the implications of complexity. - Ethan

PACKET PUSHERS AUDIENCE SURVEY

Hey all, we’ve rolled out our annual audience survey. We’d deeply appreciate if you could take ten minutes to respond. The information we collect on listening habits and audience demographics helps us lure sponsors into our web…er, I mean, explicate the value of a sponsored engagement.

As always, we don’t share any contact info or individual details with anyone, including sponsors. We only report aggregate data via our media kit. If you’ve got a little time to spare, please hit the link. Thank you so much! - Drew

THIS WEEK’S MUST-READ BLOGS 🤓

Michael Becze walks through a simple lab exercise configuring two Cisco Catalyst switches to stretch a VLAN over VXLAN using MP-BGP EVPN as the control plane. The point here isn’t to reflect a full production network design, but to show how the VXLAN & EVPN related components fit together to accomplish the L2 stretch. You get to learn VLAN to VNI mappings and VTEPs, as well as how to deal with tunnel overhead without fragmentation. Plus MAC learning over BGP, of course! - Ethan

The author makes a distinction between vibe coding and a term he’s coined called automatic programming.

“If vibe coding is the process of producing software without much understanding of what is going on, ... automatic programming is the process of producing software that attempts to be high quality and strictly following the producer's vision of the software, … with the help of AI assistance.”

I think this is a key distinction. The more I use LLMs, the more this mirrors my experience. I’m evolving my initial reaction to LLMs from “fancy autocomplete for lazy people” to a more nuanced view of what they do & what they are good for.

I have discovered that the more intelligently, specifically, and qualified I ask an LLM a question or give a direction, the more relevant the result tends to be. That means I have to know what I’m talking about to start with. If I don’t know what I’m talking about, I can’t supply a sufficiently developed prompt to guide the LLM into producing something useful.

Put another way, LLMs save me time by abstracting the fussiness of syntax across a variety of knowledge domains. My need for domain specific knowledge and systems level thinking doesn’t go away. But I don’t have to spend as much time linting YAML or composing an IOS stanza. Or in the case of my car maintenance hobby, figuring out what Subaru called a specific part. - Ethan

This post previews ongoing research into AI use at a 200-person company. So far, the researchers have discovered that voluntary AI use by employees has increased the number of projects employee take on (including projects outside their specialties), added to the cognitive load of developers as they check the work of their vibe-coding colleagues, and amped up multi-tasking (which creates the feeling of being more productive, but also comes with the costs of continual attention-switching, a longer to-do list, and increased pressure). 

OK, but if this was all voluntary, what’s the big deal? The researchers say “What looks like higher productivity in the short run can mask silent workload creep and growing cognitive strain as employees juggle multiple AI-enabled workflows. Over time, overwork can impair judgment, increase the likelihood of errors, and make it harder for organizations to distinguish genuine productivity gains from unsustainable intensity.”

I think they might be on to something. I’ve yet to run across an information technology that creates less work. It’s an interesting post, and worth your time to read and consider. - Drew 

In a similar vein to the post above, Tom Hollingsworth writes about how much of our attention is being constantly pulled at by app alerts, be it LinkedIn or Facebook or a dental app (yes, that’s a real thing). And every alert means a potential interruption that pulls our attention away from what we were doing to look at this new thing. 

Tom writes “We suck at multitasking. No, don’t fight me on this. We really do. We think we are good at it because we can jump back and forth between things but in reality we are way more efficient when we work on things in uninterrupted blocks of time.”

He offers a few suggestions for how to get some of your precious attention back, but I think you probably already have a pretty good idea of the temptations that exist within your own electronic ecosystem. Oh, new Slack notification! Be right back… - Drew 

MORE BLOGS

TECH NEWS 📣

TL;DR. Nvidia and OpenAI aren’t fast friends because reasons, so Nvidia isn’t gonna invest $100B in OpenAI to buy Nvidia GPUs. They’ll probably invest something, although maybe nothing. Who knows? It’s hard to parse the weasel words. Meanwhile, Nvidia is working with Groq and OpenAI with Cerebras. And while those deals are still in the barely comprehensible billions, no one’s spending $100 billion. It’s as if numbers are real all of a sudden. Huh. - Ethan

TL;DR. AI is generating a lot of sloppy pull requests, and project maintainers are having to wade hip-deep through it all. Ain’t nobody got time for PR shenanigans, so something’s got to be done. What? No one’s sure yet. But conversations are happening. - Ethan

A heads up for the home automation crowd. - Drew 

Not at all concerning. - Drew  

MORE NEWS

FOR THE LULZ 🤣

RESEARCH & RESOURCES 📒

Alex Krentsel presented at NANOG 96 about simplifying software defined networks. His premise was to do away with the centralized, hierarchical controller model and out-of-band control plane network, moving to a distributed controller architecture with in-band signaling. He calls this model distributed software defined networking or dSDN.

The driver for this research is the number of failures still experienced in centralized controller SDN (cSDN) networks, the root cause of which is often controller-related issues or control plane network problems. Alex’s theory is that simplifying SDN networks would lead to more stability without losing SDN’s novel forwarding benefits or even having to make tradeoffs such as poor convergence time.

Enabling the dSDN architecture is modern router multi-core CPUs & network operating systems that can run containers, along with on-box telemetry and interfaces such as gRIBI. In tests with production-level code, dSDN performed well at scale, many details of which Alex & his team document in this highly readable paper.

Look for a Heavy Networking podcast episode coming soon where I and guest co-host Scott Robohn talk to Alex about dSDN. I think there’s something here worthy of consideration & further discussion. - Ethan

Echo is an iOS and iPadOS SSH client based on Ghostty. Replay describes Echo as, “a fast, modern SSH client for iOS and iPadOS, built for the new era of rich terminal-based tools and AI coding agents.” Also, it’s theme-able.

Not free, but really cheap at $2.99 in the app store and no subscription. - Ethan

Ivan Pepelnjak reports, “The videos from the Network Observability webinar with Dinesh Dutt are now available without a valid ipSpace.net account. Enjoy!” - Ethan

Chris Greer has a free DNS course on YouTube. Here’s what’s in lesson 2: “In this video, part 2 of the How DNS Works series, we break down when DNS uses UDP, when it switches to TCP even on the client end, and why it matters.” - Drew 

Google is keeping an eye on how malicious actors are using AI. They are particularly worried about nation-state actors using techniques such as “knowledge distillation” to clone proprietary (i.e. Google’s) models. That’s not a problem for the rest of us, but Google is also watching how threat actors are using LLMs to write more effective phishing messages and tiptoeing into agentic AI to help them develop malware and attack tools. Those are problems for the rest of us. - Drew  

MORE RESOURCES

INDUSTRY BLOGS & VENDOR ANNOUNCEMENTS 💬 

The NAF team is building out the content track for AutoCon5. Got something to contribute to the network automation community? This is your time to shine. Click through for the details on what they’re looking for.

I’ve been to every AutoCon event, and generally speaking, the nerdier the content, the better. Go deep with the tools, techniques, workflows and platforms you’ve figured out from your own experience. Don’t feel limited to just open source tools. Solutions built with commercial tools are viable submissions, too. Talks about the business & money side of things are also useful & appreciated.

AutoCon5 will be held in Munich, Germany the week of 8 - 12 June 2026. The submission deadline is 9 March. Click the links above for details on how to submit a workshop or talk, or to learn how to attend the event in person. The Packet Pushers team will be there covering the conference, and we hope to see you there! - Ethan

In this essay, Rachel Thomas compares the vibe coding reward system to the sort of positive feelings we get when gambling.

“With vibe coding, people often report not realizing until hours, weeks, or even months later whether the code produced is any good. They find new bugs or they can’t make simple modifications; the program crashes in unexpected ways. Moreover, the signs of how hard the AI coding agent is working and the quantities of code produced often seem like short-term indicators of productivity. These can trigger the same feelings as the celebratory noises from the multiline slot machine.”

That is, a small win in the form of code that seems to work can mask a long-term loss in the form of code that’s poor quality and difficult to maintain. But LLMs are designed to keep us prompting them, as Rachel explains.

“Both slot machines and LLMs are explicitly engineered to maximize your psychological reaction. For slot machines, the makers want to maximize how long you play and how much you gamble. LLMs are fine-tuned to give answers that humans like, encouraging sycophancy and that they will keep coming back.”

I’m not saying we shouldn’t use LLMs to generate code. I am saying we have to know how to use them in the context of skills we already have to not create a new form of loathsome technical debt. - Ethan

I have no words. - Ethan

I had words. Johna and I talked about this on a recent Network Break. - Drew

Orange Business is offering WAN and SD-WAN services that support post-quantum cryptography (PQC) via Cisco 8000 series routers. Orange Business says PQC can be “implemented as a software feature” for organizations that want to protect against “harvest now, decrypt later” attacks. PQC-enabled WAN services are available now; SD-WAN with PQC will be available in the third quarter of this year. - Drew 

BlueCat has announced a SaaS offering for DDI and other services. Called Horizon, BlueCat says this is a “...shared control plane for network services, policy, identity, telemetry, analytics, automation, and AI-assisted intelligence. This unified approach platform lets teams apply consistent governance, correlate signals, surface prioritized insights, and take coordinated action across DNS, DHCP, IPAM, security, and network performance, enabling networks to automatically adapt and remediate issues as conditions change, without forcing infrastructure replacement or disruptive migrations.” - Drew 

Portnox makes a zero-trust, passwordless access product for Web and enterprise applications. In its latest release, it’s extending that capability to console-based applications. From the press release: “Organizations can now eliminate passwords and credentials from administrative access via Remote Desktop Protocol (RDP), Secure Shell (SSH), Virtual Network Computing (VNC), and Teletype Network (Telnet).” I know none of you reading this have Telnet enabled anywhere, right? Right? - Drew 

Cisco rolls out the G300, its highest-ever throughput Silicon One ASIC. Aimed at AI data centers, you can have it in a whitebox running SONiC or in purpose-built Cisco switches, including a liquid-cooled model. It also supports the P4 programming language if you want to tweak packet processing functions, though Cisco assumes most customers will come to Cisco for P4 programs. - Drew 

MORE INDUSTRY NOISES

  1. From the stupid DNS tricks department: ipasn.net - Geoff Huston via the APNIC Blog

DYSTOPIA IRL 🐙

TOO MANY LINKS WOULD NEVER BE ENOUGH 🐳

LAST LAUGH 😆