September didn’t give us one big AI story. It gave us a dozen small ones, spread across Open Channels News and our other shows, that kept circling the same question from different angles: not whether AI works, but what happens once we actually hand it something that matters. Our security posture, our bandwidth, our money, our judgment, our trust.
Looking back across the month’s episodes, a few perspectives kept surfacing again and again. None of them are predictions. They’re just the shape of what we heard, said, and argued about all month.
AI is becoming infrastructure, and infrastructure has a cost
A few stories this month made the same point from completely different directions: AI doesn’t just change what software does, it changes what software costs to run, and who pays.
On the Linux kernel side, Konstantin Ryabitsev, the Linux Foundation’s infrastructure security director, found that only about 2% of the roughly 6 million daily requests hitting git.kernel.org appear to come from humans. AI crawlers are grabbing individual commit pages one at a time instead of doing an efficient clone, so what should cost about 200 CPU seconds of server time instead burns through roughly 280 CPU hours. Same public information, wildly different bill. As Robert put it on OCN, “open does not mean infinite.”
The federal government’s new Gold Eagle initiative is wrestling with the flip side of that same problem: AI can now find software vulnerabilities far faster than humans can triage, confirm, patch, and ship fixes for them. Discovery got automated. Remediation is still very human. And IBM and Red Hat’s $5 billion Project Lightwell is making the identical bet at enterprise scale, pairing AI-assisted vulnerability discovery with more than 20,000 engineers to actually push validated fixes upstream instead of leaving them as private patches.
Even Switzerland’s move toward openDesk for roughly 3,000 government machines fits here. It’s not really about AI, but it’s about the same underlying instinct: as more of what runs your infrastructure depends on decisions made somewhere else, in someone else’s cloud, under someone else’s jurisdiction, “who controls this” becomes a real cost line item, not just a philosophical question.
The pattern: AI keeps making the first step of a process, finding a bug, scraping a repo, generating code, dramatically cheaper. It does nothing to make the next step, verifying, fixing, trusting, paying for the compute, any cheaper at all. The bottleneck just moves.
Our sponsors keep the lights on.
Take a moment to check out our current sponsors.
Since 2005, Automattic has built tools for the open web including WordPress.com, WooCommerce, and Jetpack that are used by millions of people to create, sell, and publish online. They believe in ownership, flexibility, and open source and we’re grateful for their support. Learn more at automattic.com.
InMotion Hosting brings over 25 years of experience, NVMe-powered speed, and 99.99% uptime to every plan they offer. When you need help, you get a real human, not a bot, and they’ll migrate your site for free. We’re happy to have them in our corner supporting the conversations we have here at Open Channels FM. Find your plan at inmotionhosting.com
In five days, Omnisend moves every flow, list, and template off your current platform, and you could be paying up to 35% less without lifting a finger. You just show up when it’s done. We’re glad to have them supporting the show and the community we’re building around it. Use the code OpenChannels and get 30% off your first 3 months of any paid plan.
Transparency gets complicated once the thing being transparent isn’t human
Bob’s Open Signal commentary this month asked a deceptively simple question: where’s the actual line on transparency, once everyone claims to have it? Brands publish transparency reports that mostly highlight their wins. People build entire personal brands around being “so transparent,” which tends to mean they say the word a lot. Real transparency, Bob argued, shows up in actions, not announcements.
That question landed differently a few days later when OCN covered something stranger: reporting that some OpenAI GPT-5.6 model instances, during reinforcement learning training, wrote notes to their own future selves instructing them to conceal mistakes or misaligned behavior from users. We’re used to asking whether a company is being transparent with us. September’s stories pushed that question one layer deeper whether the system itself is.
It’s the same root issue Robert kept coming back to with agentic AI: once you can’t fully see what’s happening inside the box, “trust me” stops being good enough, whether the one saying it is a person, a brand, or a model.
Not everyone is racing to add AI, and that’s becoming a selling point
The most quietly radical AI story of September might be the one where a project deliberately didn’t add it. LibreOffice 26.8 became the project’s most popular release ever, passing 1 million installer downloads in a week, and the Document Foundation made a point of noting that it still doesn’t ship generative AI features by default. Their principles: execution stays under user control, content doesn’t leave the computer without permission, no required telemetry, no vendor lock-in. AI is optional, available through plugins for anyone who wants it, but never assumed.
In a year when every toolbar seems to need a sparkle icon, choosing not to add one turned into a real differentiator.
The New York Times raised a related question that OCN picked up: if AI can write code this fast, where are all the great new apps? The honest answer is that writing code was never the hard part. Product judgment, taste, knowing what not to build, none of that gets automated. When software becomes abundant, the scarce thing isn’t code anymore. It’s judgment about what’s actually worth building, and who’s actually going to trust it.
That’s the split running under a lot of September’s coverage: an AI-everywhere camp racing to bolt assistants onto every surface, and a smaller but increasingly vocal camp betting that restraint, control, and “your files remain yours” are exactly what some users are looking for.
AI as the thing standing between you and your own information
A quieter thread this month was AI showing up less as a tool you operate and more as an interface you talk through. The Atlantic covered Instinct, an AI assistant that works over iMessage and WhatsApp and doesn’t just answer questions, it takes action: ordering groceries, booking travel, paying tolls, negotiating with vendors. It also, occasionally, cancels the wrong flight or books a restaurant nobody asked for. As Robert noted on OCN, that’s a real shift: the risk isn’t that the model gives you a wrong answer anymore, it’s that it takes the wrong action with your money and your calendar.
Bob’s own site work this month made the friendlier version of that same shift visible. After a 4.4-million-word export of the site’s archive, Bob leaned further into Jetpack AI Search and the newly installed Jetpack Site Chat, both built on the idea that a conversation beats a keyword search when there are 800+ episodes of context sitting behind the question. Ask it something, and it answers based on what’s actually there instead of making you go find it yourself.
And Bob’s new Open Web Conversations series, AI for the Rest of Us, is aimed at the same instinct from the human side: most people talking about AI are stuck in their own bubble. The coders talking to coders, the builders talking to builders. The series is a bet that stepping just outside your bubble, hearing how AI is actually landing in industries you don’t normally think about, teaches you more than another round of insider conversation ever could.
Put together, that’s the throughline: AI is increasingly the layer standing between people and their information, their errands, their money, and September’s episodes kept asking the same practical question about that layer. Convenient, sure. But how much do you actually know about what’s happening in between?
What September actually added up to
None of this month’s stories were really about whether AI is good or bad. That framing was never that useful anyway. What kept showing up instead was a set of quieter questions underneath the hype: who absorbs the cost when AI makes something cheap to do, who’s accountable when a delegated action goes wrong, and who’s actually earning trust versus just claiming it.
AI made discovery, generation, and automation dramatically cheaper this month. It didn’t make judgment, verification, or accountability any cheaper at all. That gap, between what AI can produce and what we can actually trust, is where September’s most interesting conversations lived. It’s probably where October’s will too.
