• Human Infrastructure
  • Posts
  • Human Infrastructure 426: Teaching GitHub Copilot, Why You Should Do Hard Things, and More

Human Infrastructure 426: Teaching GitHub Copilot, Why You Should Do Hard Things, and More

AuthorAuthor

Drew Conry-Murray & Ethan Banks
November 20, 2025

THIS WEEK’S MUST-READ BLOGS 🤓

Angie Jones is a contributor to the open source project Goose. She says she and other reviewers were having trouble keeping up with contributions to the project, so she turned on a Copilot Code Review agent. Other maintainers said reviews were “noisy” and most of its comments were low quality, so they wanted to turn it off. 

Her response? “Here’s what I know from helping engineers work with AI: you don’t throw in the towel. You don’t disable. You tune. You teach the model how you want to work, not just hope for the best.” And that’s what she did. This post has detailed information on how she instructed Copilot to focus on some principles and priorities that would produce more useful code reviews. She shares instructions she used and key takeaways from her efforts. As more of our lives will be spent interacting with agents and models, tuning and teaching are going to become useful skill sets. - Drew 

Andy Lapteff is speaking at AutoCon 4 about his painful journey into network automation. As a self-proclaimed “CLI Lifer” who struggled with computer programming in college, he got into networking because he wanted to work in tech but didn’t want to be a developer.

But here comes network automation, which requires some facility with programmatic frameworks, Python, Git, and other coding elements. The post linked above isn’t specifically about network automation, but the principles he lays out in the post apply, including “growth is a direct result of discomfort” and “consistency is where transformation begins.” Whether the hard thing you’re going to do is learn network automation or something else, there’s good, hard-won advice here. - Drew 

Terrestrial fiber optic networks can be impacted by backhoes and boat anchors. Satellite networks don’t have to worry about orbital fiber cuts, but they do have to worry about space weather. Solar storms can cause geomagnetic disturbances in the atmosphere; these disturbances can cause drag on satellites in low earth orbit. This drag causes a loss of altitude, which can lead to packet loss and latency. This blog post summarizes the results of a study of the impact of four different solar storms on a Starlink network. It also links to the original research. - Drew 

Cisco’s Unified Edge is a converged infrastructure platform aimed at edge locations that are generating data that can be fed into AI. Instead of shunting all that data to the cloud or an off-site data center, you can run inferencing AI workloads close to where the data is being created. Phil gets into some of the specs in the box, talks about Cisco’s rationale for the product, and does some competitive analysis with other options. - Drew  

Jason Gintert doesn’t think that AI is going to replace us, largely because AI doesn’t actually think. AI doesn’t know the context of your request—your intent. But can it help you do your job? Absolutely.

Jason sees, “a future where AI handles the mundane pattern matching and correlation analysis then engineers focus on architecture, automation and complex troubleshooting. We're not being replaced as much as we're being promoted to work on problems that actually require contextual thinking which IMHO, is a good thing!” - Ethan

MORE BLOGS

Your Foundation for Enterprise AI Is Here
Enterprise AI demands more than cloud alone. Equinix Distributed AI delivers the secure, globally consistent infrastructure you need to train where compute is abundant, process sensitive data where it's stored and run inference at the edge, where data is generated.

AI infrastructure without trade-offs: Deploy confidently in liquid-cooled data centers purpose built for high-density AI, delivering power, performance and reliability.

Low-latency edge-to-cloud connectivity: Connect to 225+ cloud on-ramps and thousands of partners via private edge links for seamless data flow.

Predictable costs, neutral ecosystem: Optimize AI spend with transparent pricing and no hidden fees. Use a vendor-neutral ecosystem to avoid lock-in.

Explore how Distributed AI can simplify your architecture, surface hidden efficiencies and ensure a scalable roadmap.

Explore Distributed AI Solutions → https://www.equinix.com/lp/equinix-distributed-ai

Curious about Distributed AI? We're collecting questions from the Packet Pushers community for an upcoming follow-up episode. Send us what you want to know about Equinix and distributed AI infrastructure: [email protected]

TECH NEWS 📣

A fun bit of history in this story, including Gelsinger employing some techno-babble to keep from getting yelled at for etching his initials in silicon. - Drew 

From the article: “Azure was hit by the "largest-ever" cloud-based distributed denial of service (DDoS) attack, originating from the Aisuru botnet and measuring 15.72 terabits per second (Tbps), according to Microsoft.” - Drew  

I guess “pyramid” is the wrong shape for this scheme in which Company A invests billions in Company B while Company B buys billions in services from Company A. Maybe a better description of this financial instrument is “human centipede.” - Drew  

This is a weird one. Extropic is making a chip that uses probabilistic bits.

“Extropic calls its processors thermodynamic sampling units, or TSUs, as opposed to central processing units (CPUs) or graphics processing units (GPUs). TSUs use silicon components to harness thermodynamic electron fluctuations, shaping them to model probabilities of various complex systems, such as the weather, or AI models capable of generating images, text, or videos.”

Why? Energy efficiency. 1000’s of times more energy efficient. If Extropic has success with AI applications, this tech could address the power problem holding back AI data center builds now. That’s a very optimistic best case scenario, but it’s still interesting. - Ethan

TL;DR. The German government wants to give everything in the upper 6GHz range to mobile operators, leaving nothing in that spectrum for license-free WiFi use. Fiber optic network operators want to see that upper 6GHz spectrum available so that their capacity can be utilized. - Ethan

MORE NEWS

FOR THE LULZ 🤣

A few weeks ago it was AWS. This week it’s Cloudflare. Who got next? Thanks to Danilo for sharing in the Packet Pushers Community Slack.

RESEARCH & RESOURCES 📒

Dan Ryan has released a side project into the wild! On LinkedIn, Dan says, “If you’ve ever spent time digging through multiple vendor datasheets trying to find (or compare) AP specs, you know how painful that process can be. I deal with that constantly, so I built something to make the job easier, not just for me, but for anyone working in Wi-Fi.

Introducing the Enterprise AP Comparison Database: A reference that lets you quickly browse, filter, and compare access points. Everything is structured, sortable, and easy to navigate, and new APs can be added or updated in minutes.”

Thanks, Dan! - Ethan

Hey, Netherlands folks. Here’s a meetup for you my friend Bart Dorlandt let me know about. February isn’t that far away, so register soon! - Ethan

All the talks from NANOG 95 have been posted! I was at this event and posted many talk summaries to my LinkedIn. The topic that people were most interested in covered Energy Efficient Routing.

EER is in development in the IETF. In part, EER is about using a dynamic routing protocol—ISIS thus far—to share metadata via TLVs to a routing infrastructure that can facilitate the powering down of unused links to save energy. Speaker Colby Barth talked in detail about how this works and shared data about how early testing has gone. - Ethan

MORE RESOURCES

INDUSTRY BLOGS & VENDOR ANNOUNCEMENTS 💬 

That headline is admittedly hard to parse. Here’s my take on it after a briefing with the Gluware team, walking through their explainer PDF, and watching three demo videos (scroll through this page) showing off various use cases.

Gluware has added AI to their network automation & orchestration platform to help you make the most of the platform’s capabilities. Without AI, you need a lot of time, training, and experience to really get your head around everything that the Gluware platform has to offer. With AI, you can create models or customize models, interact with multiple sources of data including data that lives outside of Gluware (such as NetBox and ServiceNow), and perform other complex, fussy, tedious automated tasks a whole lot faster than if you’d had to learn all of the Gluware platform nuances artisanally.

The big question then becomes…how do you know that what the AI is spitting out in response to your plain English prompt is worth anything? How do you know you’ve got a valid model or regex, etc.? Gluware told me that the AI interacts with their validation engine to prove without any doubt that the output will work in your network.

Put another way, you aren’t #yolo vibecoding with the Gluware approach. You’re asking an AI in plain language for something that might be hard to do by hand and getting back a validated result. I think the extra special magic here is in that validation. That matters. - Ethan

I’m intrigued by Itential’s newly announced FlowAI. It’s trying to do some hard things. While AI agents backed by LLMs can analyze network data and suggest changes using natural language, those changes need to be translated into commands and configs using the correct syntax. And workflows need to be in place to ensure that the correct changes are applied to the necessary devices in the proper order. 

For those and other reasons (such as hallucinations), network engineers are hesitant to let agents run willy-nilly across the production environment. FlowAI uses a tiered approach to try to leverage the best of both worlds: the analytic capabilities of agents with the deterministic execution required for changes to be safely made to network infrastructure.

There’s more details about how FlowAI works in this documentation from Itential. - Drew

I found it hard to be more concise than Cloudflare themselves, so here you go.

“The issue was not caused, directly or indirectly, by a cyber attack or malicious activity of any kind. Instead, it was triggered by a change to one of our database systems' permissions which caused the database to output multiple entries into a “feature file” used by our Bot Management system. That feature file, in turn, doubled in size. The larger-than-expected feature file was then propagated to all the machines that make up our network.

The software running on these machines to route traffic across our network reads this feature file to keep our Bot Management system up to date with ever changing threats. The software had a limit on the size of the feature file that was below its doubled size. That caused the software to fail.”

Lots of 500 errors resulted for sites (such as our very own and much beloved PacketPushers.net) that went on for a while, especially as pent-up demand for various sites surged through Cloudflare post-fix.

A workaround for those of us impacted would have been updating the DNS record hosted at Cloudflare so that the site would be served natively instead of front-ended by CF. But…during the outage, accessing the Cloudflare dashboard was impossible. But there’s a way around that catch-22 for the future, if you do the homework outlined in this post: How to disable CloudFlare proxying when you can't access the dashboard. - Ethan

MORE INDUSTRY NOISES

  1. Announcing NetBox Observability: Infrastructure Monitoring That Understands Your Design - NetBox Labs

  2. NetBox Integration with Cisco Meraki Now Available in Customer Preview - NetBox Labs

  3. Kentik Launches the First Agentic AI Solution That Deeply Understands the Network - BusinessWire   

  4. Windows CLAT Enters Private Preview: A Milestone for IPv6 Adoption - Microsoft Networking Blog

  5. Introducing AWS Site-to-Site VPN 5 Gbps Tunnels to support high throughput workloads - AWS Networking & Content Delivery Blog

  6. Introducing QUIC Protocol Support for Network Load Balancer: Accelerating Mobile-First Applications - AWS Networking & Content Delivery Blog

  7. Beyond Deterministic Automation: Why AI Reasoning is the Future of Infrastructure Orchestration - Itential Blog

  8. Nokia deepens commitment to AI data center and edge innovation with Open Compute Project Platinum membership - Nokia 

DYSTOPIA IRL 🐙

TOO MANY LINKS WOULD NEVER BE ENOUGH 🐳

LAST LAUGH 😆

A deep nerd reference from Kaj. Shared on the Packet Pushers Community Slack.