Pattern Recognized

Every time you open an AI chat, it knows nothing about you, so you retype your whole situation from scratch, slightly differently every time, and wonder why the answers wander. In this episode, I hand you the fix, which is a folder with three shelves: the facts about your organization, samples of what you actually sound like, and the decisions you have already made and do not want to relitigate. You do not write any of it from scratch, because most of it is already sitting in your annual report, your grant proposals, and your sent folder, waiting to be harvested in an afternoon. I also get specific about the failure mode nobody plans for, which is a folder that quietly goes stale and starts confidently describing an organization that no longer exists. And I draw a hard line around what never goes in it, because the whole premise here is handing your material to a machine and that deserves a boundary you set in daylight.

Security researchers documented what they assess to be the first ransomware attack run end to end by an AI agent, with no human at the keyboard, and the way it got in should make every small organization leader look twice at their own setup. In this episode I walk through what actually happened in plain language, including the moment the agent hit a wall and fixed its own mistake in thirty-one seconds. Then I separate what the research documented from what the headlines turned it into, because reading an AI claim carefully is a skill that outlasts any single news story. The real shift here is not that attackers got smarter, it is that the cost of running an attack collapsed, and being too small to bother with is no longer the protection it used to be. I close with four unglamorous things you can do this week without spending a dollar, plus the one question to ask whoever runs your technology.

Nearly every organization has at least one process that lives in exactly one person's head, and nearly every leader has been meaning to write it down for years. The reason it never happens is not that writing is hard, it is that when you sit down to document work you do every day, you cannot find the beginning. In this episode I hand over the thing that technical writers and project managers spend careers absorbing by accident: work has visible endings and invisible beginnings, so you start at the end, point at the finished thing, and walk backwards until you hit somebody else's job. From there I walk through the whole build, including why you should talk instead of write, the one move most people skip, which is letting the AI interview you to surface the steps you would never think to mention, and how to bring in the person who does the work as a co-author rather than making them feel audited. And as always we cover when not to use AI here, including what happens when you point at the wrong finished thing and document a broken process backwards, faster and more confidently than before.

Science fiction spent a century imagining how to build an artificial mind, and it imagined almost every recipe except the one we actually used. In this episode I walk through the five recipes for making an AI, from scraping the internet and watching a mind emerge, to hand-built knowledge, curated specialists, systems that learn by doing, and the mostly untraveled road of copying or growing the brain itself. Each recipe gets its pop culture pairing, including Ex Machina, HAL, Data, the Weyland-Yutani synthetics, Skynet, and the Tyrell Corporation, along with the real technology behind it. We dig into why an AI trained on humanity behaves like humanity, flaws included, and why the same strange recipe accidentally produced something like intuition for computers. And we close with the oldest answer in the library, the novel that banned the whole project, for a reason that has nothing to do with robot uprisings and everything to do with you.

Most organizations decide whether AI will work for them before they ever open a tool, and they do it with a single word. In this episode I break down automation versus augmentation, two frames for the exact same technology that produce wildly different results. I bring the receipts from my own two hundred application job search, where full automation produced volume and staying the author produced every interview that mattered. You will leave with a four-question test that picks the right frame for any task and then proves the pick in the business's own numbers, not invented AI metrics. The tool does not decide your outcome; the frame does.
 

I have submitted over two hundred job applications this year, and sixty-three of them I never even looked at. In this episode, I break down the accidental experiment that those applications became: fully automated, fully manual, and a hybrid system built with AI, and only one of the three lanes produced real interviews. I walk through the exact system that worked, a career context library with your master resume, your talked-out project stories, a ranking step before you ever write, and a human review on everything that goes out under your name. Along the way: what two hundred applications taught me about ATS parsing and file formats, why tailoring is telling the truth well while fabricating is lying efficiently, and what hiring managers should learn from seeing their own pipeline from the candidate's side. Whether you are job hunting or doing the hiring, this one shows you where the machines belong and where your judgment does.

The most capable AI model ever released to the public disappeared overnight by order of the United States government, and three weeks later it came back. This episode is the story of what those three weeks were actually like for the people already building on it, because I was one of them. I walk through the Fable 5 shutdown and return as a case study in improvised governance, which is not the same thing as functioning oversight, however relieved we all were when it ended. I also put a name to the experience of watching a brand new capability expand your ambitions and then vanish out from under them: capability whiplash. If your plans depend on a tool someone else can switch off, this one is for you.

The most valuable knowledge in your organization is probably locked in someone's head and has never been written down. In this episode, Josh introduces the Interview Flip: a simple technique that turns the tables on the usual AI interaction by letting the AI ask the questions instead of answering them. The result is a crowbar for tacit knowledge, the expertise so embedded in how you work that you have stopped recognizing it as expertise at all. Plus the two ways this goes wrong, and why a bigger library only helps if what is in it is actually true.

You are probably using about a fifth of what AI can actually do, and it is not your fault. Google spent twenty-five years training you to bring keywords to someone else's library. Grounded AI flips that completely: you bring your own library and just talk to it. In this episode, Josh shows you how to step outside the chat window and give AI the kind of context it was actually built for, your real documents, your projects, your drives, and the live systems behind them. Plus the one thing that turns this whole move into a disaster if you skip it. Join us for another Pattern Recognized!

There is a stage of AI use where you stop doing the task with AI and start building the thing that does the task. Most people do not know it exists. The ones who do mostly assume it is locked behind a computer science degree. In this episode, I explain how anyone can get to Fly-level AI use, creating their own AI tools and Agents.