<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"
	xmlns:content="http://purl.org/rss/1.0/modules/content/"
	xmlns:wfw="http://wellformedweb.org/CommentAPI/"
	xmlns:dc="http://purl.org/dc/elements/1.1/"
	xmlns:atom="http://www.w3.org/2005/Atom"
	xmlns:sy="http://purl.org/rss/1.0/modules/syndication/"
	xmlns:slash="http://purl.org/rss/1.0/modules/slash/"
	>

<channel>
	<title>Ian O&#8217;Byrne</title>
	<atom:link href="https://wiobyrne.com/feed/" rel="self" type="application/rss+xml" />
	<link>https://wiobyrne.com</link>
	<description>Educator, researcher, and writer</description>
	<lastBuildDate>Sat, 25 Apr 2026 15:10:46 +0000</lastBuildDate>
	<language>en</language>
	<sy:updatePeriod>
	hourly	</sy:updatePeriod>
	<sy:updateFrequency>
	1	</sy:updateFrequency>
	<generator>https://wordpress.org/?v=7.0</generator>

<image>
	<url>https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/03/ian-obyrne-avatar.webp?fit=32%2C32&#038;ssl=1</url>
	<title>Ian O&#8217;Byrne</title>
	<link>https://wiobyrne.com</link>
	<width>32</width>
	<height>32</height>
</image> 
<site xmlns="com-wordpress:feed-additions:1">48138478</site>	<item>
		<title>Karpathy Found the Pattern. Educators Have Been Teaching It for Years.</title>
		<link>https://wiobyrne.com/karpathy-found-the-pattern/</link>
					<comments>https://wiobyrne.com/karpathy-found-the-pattern/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Sat, 25 Apr 2026 15:10:44 +0000</pubDate>
				<category><![CDATA[Doing]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7293</guid>

					<description><![CDATA[<p>Andrej Karpathy is a well-known AI researcher and educator. He helped build early AI systems at OpenAI and led AI work at Tesla. For me, what makes him stand out isn’t just the tech, it&#8217;s how clearly he explains it. He’s one of the people helping the rest of us understand what this new wave... <span class="more"><a class="more-link" href="https://wiobyrne.com/karpathy-found-the-pattern/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/karpathy-found-the-pattern/">&lt;span class=&#039;p-name&#039;&gt;Karpathy Found the Pattern. Educators Have Been Teaching It for Years.&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph"><a href="https://en.wikipedia.org/wiki/Andrej_Karpathy" rel="nofollow noopener" target="_blank">Andrej Karpathy</a> is a well-known AI researcher and educator. He helped build early AI systems at OpenAI and led AI work at Tesla.</p>



<p class="wp-block-paragraph">For me, what makes him stand out isn’t just the tech, it&#8217;s how clearly he explains it. He’s one of the people helping the rest of us understand what this new wave of AI actually means.</p>



<p class="wp-block-paragraph">Andrej Karpathy went viral with his idea of <a href="https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f" rel="nofollow noopener" target="_blank">an LLM wiki</a>. This is a way to make AI compound knowledge over time rather than start from scratch. </p>



<p class="wp-block-paragraph">Karpathy identified a major problem in how we use generative AI for queries and learning. Karpathy’s idea of an LLM Wiki is surprisingly simple. Instead of using AI like a chatbot you talk to once, you can use it like a living notebook that thinks with you over time.</p>



<p class="wp-block-paragraph">Think of it like this:</p>



<ul class="wp-block-list">
<li><strong>Normal AI habits (what most people use). </strong>You ask a question. It gives an answer. The conversation disappears into the void. It’s like talking to a very smart stranger every time.</li>



<li><strong>Karpathy’s LLM Wiki idea</strong>. You feed the AI your notes, ideas, and documents. It organizes, connects, and rewrites them over time. It becomes a growing knowledge system, not just a tool</li>
</ul>



<p class="wp-block-paragraph">It’s like building a second brain that helps you think.</p>



<p class="wp-block-paragraph">I recognized this model immediately. It&#8217;s what I have been trying to build in my vault with Obsidian. I&#8217;ve also been trying to teach my students how to build this in their own systems. </p>



<p class="wp-block-paragraph">Let&#8217;s break it down and show how I modified it.</p>



<h2 class="wp-block-heading">From Answers to Systems</h2>



<p class="wp-block-paragraph">Karpathy’s insight is deceptively simple. Instead of asking AI to search your documents every time you have a question, you ask AI to read your documents and build a structured, interlinked knowledge base from them.</p>



<p class="wp-block-paragraph">That system grows and compounds over time. It starts to help you think, not just retrieve answers.</p>



<p class="wp-block-paragraph">Here is my understanding of how Karpathy&#8217;s three-part system works:</p>



<ul class="wp-block-list">
<li><strong>Instructions (Specification)</strong> — what the system is supposed to do</li>



<li><strong>Context (Editable surface)</strong> — the knowledge it reads, writes, and updates</li>



<li><strong>Feedback (Evaluator)</strong> — how do you know if it’s actually improving</li>
</ul>



<p class="wp-block-paragraph">If you’ve spent any time in instructional design, curriculum development, or educational research, this structure should feel familiar. It mirrors how we already design learning systems:</p>



<ul class="wp-block-list">
<li><strong>Learning objectives</strong> → the specification (what success looks like). Objectives define the target.</li>



<li><strong>Student work</strong> → the editable surface (where thinking develops). Student work is the evolving surface.</li>



<li><strong>Assessment rubric</strong> → the evaluator (how we judge progress and quality). Assessment tells us if this is actually working. </li>
</ul>



<p class="wp-block-paragraph">This isn’t a coincidence. It’s the same pattern that shows up anywhere a system needs to learn over time. This is true whether you’re teaching a student, training a model, or building an AI agent inside your knowledge garden.</p>



<h2 class="wp-block-heading">Revising My Systems </h2>



<p class="wp-block-paragraph">I&#8217;ve been building toward this for years. This has shown up in explorations of Evernote, bookmarking (Pinboard, Diigo, Raindrop), blogging, my newsletter, and in my teaching. All of this focused on setting up a second brain and offloading some of the cognitive tasks as I work with a firehose of information. </p>



<p class="wp-block-paragraph">Recently, I&#8217;ve been using Obsidian to create a vault of my ideas, turn them into a digital garden, and create a knowledge system. This includes a focus on raw intake, structured organization, and a layer where ideas are actively developed and connected. It’s less a repository and more a space where thinking compounds.</p>



<p class="wp-block-paragraph">The original setup was built around a <em>maturity model </em>(Seeds, Plants, Evergreens, Groves, Forests). A note started as a rough idea and evolved into a polished, reusable asset. It was a good model for one person moving ideas forward intentionally.</p>



<p class="wp-block-paragraph">But two things broke it.</p>



<p class="wp-block-paragraph">First, the volume. I take tons of notes daily. These include web clippings, podcast notes, videos I&#8217;m watching, fleeting notes, etc. I was using AI tools to help clean these up and add to the system, but it quickly became a mess. The <em>maturity model </em>assumed I had time to tend each note. I didn&#8217;t. </p>



<p class="wp-block-paragraph">The second tension was the blending of my ideas and those from AI. I care a lot about keeping my voice. The research I&#8217;ve done on AI use in education, particularly on agency and friction, has made it clear to me that low-agency AI use is defined by accepting the default. Taking the first draft. Letting the tool structure the thinking. I had started doing that in my own notes without realizing.</p>



<p class="wp-block-paragraph">I needed a system that could handle volume <em>and</em> protect the human layer.</p>



<h2 class="wp-block-heading">Consume, Curate, Create</h2>



<p class="wp-block-paragraph">The rebuild is organized around a simple idea. Different layers of the knowledge system serve different purposes, and AI should be present in some of them and absent in others.</p>



<p class="wp-block-paragraph">Longtime readers of this blog recognize this model of consumption, curation, and creation. I&#8217;ve held onto this model for some time as I think and write about being web literate. </p>


<div class="wp-block-image">
<figure class="aligncenter size-full"><a href="https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image-5.png?ssl=1"><img data-recalc-dims="1" fetchpriority="high" decoding="async" width="680" height="440" src="https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image-5.png?resize=680%2C440&#038;ssl=1" alt="" class="wp-image-7300" srcset="https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image-5.png?w=680&amp;ssl=1 680w, https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image-5.png?resize=300%2C194&amp;ssl=1 300w, https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image-5.png?resize=600%2C388&amp;ssl=1 600w" sizes="(max-width: 680px) 100vw, 680px" /></a></figure>
</div>


<p class="wp-block-paragraph"><strong>01 Consume</strong> is the intake layer. Everything flows in here. Web clippings, podcast notes, paper summaries, videos, chat logs, and quotes. It’s a source layer. Material arrives with minimal processing, and the rule is simple: <em>sources stay sources</em>. Two AI agents (the Inbox Sorter and the Consume Custodian) clean and organize what comes in, ensuring notes are structured and in the right place. They don’t rewrite, merge, or synthesize. Provenance stays intact.</p>



<p class="wp-block-paragraph"><strong>02 Curate</strong> is the wiki layer. This is where the AI does its primary work. A single agent (The Librarian) routes material from 01 Consume, maintains metadata, and builds a structured, evolving knowledge wiki. It operates from explicit rules I’ve defined, reviewing new inputs and updating the system as it grows. I’m largely hands-off here.</p>



<p class="wp-block-paragraph"><strong>03 Create</strong> is where I step in again. This is where I draft blog posts, publish newsletter issues, evergreen notes, and build content. All of it is produced in this layer. AI may assist with drafting or editing, but the initiating voice, the argument, and the choices about what matters are mine. I pick up the pieces the curator has organized and synthesize them into something worth publishing.</p>



<h2 class="wp-block-heading">The Firewall</h2>



<p class="wp-block-paragraph">Here is the thing I keep coming back to when I explain this to people. The layers are not just organizational. They serve as a firewall between AI-generated and human-authored content.</p>



<p class="wp-block-paragraph">In Curate, the agents work. They route, classify, synthesize, and link. But they operate inside hard rules. The Librarian cannot touch the Create layer. The Inbox Sorter cannot invent content. It can only route and flag. Every run is evaluated against a rubric before the outputs are accepted. In Create, this is where I spend most of my time. </p>



<p class="wp-block-paragraph">The blog post you are reading right now is not AI-drafted and Ian-approved. It is Ian-initiated, Ian-structured, Ian-written. To help me understand Karpathy&#8217;s model, I reviewed countless YouTube videos, reviewed a dozen or so blog posts, social media feeds, and GitHub repos. I saved all of this to Consume and let the AI agents clean it up and prepare a couple of wiki entries in Curate. As I added new content, these entries were modified over time. The related notes about Obsidian, my knowledge management system, and other content were also updated in the AI-generated wiki. </p>



<p class="wp-block-paragraph">As I drafted this post in Obsidian, I asked an AI assistant to explain a part of Karpathy&#8217;s model to a layperson, while also ensuring I was correct and not conflating concepts. This is the same way I would use a search tool or my notes as I work. The thinking, the argument, and the voice all stay mine.</p>



<p class="wp-block-paragraph">This matters because it means the system has explicit design friction built into it. In my <a href="https://wiobyrne.com/agency-looks-like-friction/">research on student AI use</a>, I found that high-agency users were defined by their willingness to <em>refuse</em> the AI&#8217;s first output, to push back, to re-prompt with constraints. I am not exempt from that dynamic in my own work. The firewall is how I engineer refusal into the architecture rather than relying on my willpower in the moment.</p>



<h2 class="wp-block-heading">What This Has to Do With You</h2>



<p class="wp-block-paragraph">I share this publicly because the work isn’t just the idea. As always, it’s the system behind it. </p>



<p class="wp-block-paragraph">Learning systems (whether human or AI) depend on three things. What to do, where the work happens, and how you know it’s working. Systems learn when you get these three things right. Get that wrong, and you don’t get knowledge. You get noise that compounds.</p>



<p class="wp-block-paragraph">This is what I mean by AI literacy. Not using tools. Building systems and tools that work for you and your needs. I share this publicly because the idea only matters if it works in practice. This currently works for me.</p>



<p class="wp-block-paragraph">In the next post, I’ll walk through how it actually works. How material flows, what each agent does, and how the system evaluates itself over time.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph"><em>If this resonated, the best next step is to <a href="https://digitallyliterate.net/newsletter/" rel="noreferrer noopener nofollow" target="_blank">subscribe to my newsletter</a>.</em></p>



<p class="wp-block-paragraph"><em>Questions, pushback, or just want to connect at <a href="mailto:hello@wiobyrne.com">hello@wiobyrne.com</a>. Find me at </em><a href="https://wiobyrne.com"><em>wiobyrne.com</em></a><em>, <a href="https://www.linkedin.com/in/wiobyrne/" target="_blank" rel="noreferrer noopener nofollow">LinkedIn</a>, <a href="https://bsky.app/profile/wiobyrne.com">Bluesky</a>, or </em><a href="http://@Wiobyrne@mastodon.social" target="_blank" rel="noreferrer noopener nofollow">Mastodon</a>.</p>



<p class="wp-block-paragraph">Photo by <a href="https://unsplash.com/@lazycreekimages?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Michael Dziedzic</a> on <a href="https://unsplash.com/photos/blue-and-white-water-wave-nbW-kaz2BlE?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/karpathy-found-the-pattern/">&lt;span class=&#039;p-name&#039;&gt;Karpathy Found the Pattern. Educators Have Been Teaching It for Years.&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/karpathy-found-the-pattern/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7293</post-id>	</item>
		<item>
		<title>Writing Is a Process. So Is Losing Your Voice.</title>
		<link>https://wiobyrne.com/writing-is-a-process-so-is-losing-your-voice/</link>
					<comments>https://wiobyrne.com/writing-is-a-process-so-is-losing-your-voice/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Fri, 24 Apr 2026 14:46:06 +0000</pubDate>
				<category><![CDATA[Thinking]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7287</guid>

					<description><![CDATA[<p>I&#8217;ve been getting a lot of great feedback on my last post about building an AI model of my voice. One in particular surfaces a tension I&#8217;ve been having while working with generative AI. Why would I want to create a model that knows how I write? I&#8217;m still confused as to the &#8216;why&#8217; —... <span class="more"><a class="more-link" href="https://wiobyrne.com/writing-is-a-process-so-is-losing-your-voice/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/writing-is-a-process-so-is-losing-your-voice/">&lt;span class=&#039;p-name&#039;&gt;Writing Is a Process. So Is Losing Your Voice.&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph">I&#8217;ve been getting a lot of great feedback on <a href="https://wiobyrne.com/i-trained-a-local-ai-to-write-like-me/">my last post about building an AI model</a> of my voice. One in particular surfaces a tension I&#8217;ve been having while working with generative AI. </p>



<p class="wp-block-paragraph"><em>Why would I want to create a model that knows how I write? I&#8217;m still confused as to the &#8216;why&#8217; — drafting is an important part of the thinking process, working out how thoughts/ideas connect. I don&#8217;t want a machine doing that for me, because I might never come to those connections without going through this process.</em></p>



<p class="wp-block-paragraph">This is a serious argument about what writing actually does. And it&#8217;s absolutely correct.</p>



<p class="wp-block-paragraph">For me, writing is thinking. I think (and learn) through my writing. My writing (and thinking) are always changing. It changed drastically when I started blogging. It changed when I started writing with others. It changed when I started learning (and then unlearning) how to write for academic spaces. And now it&#8217;s constantly in a tug-of-war as I try to figure out how to make my work more approachable and accessible to others. </p>



<p class="wp-block-paragraph">If you hand your drafting to a machine, you risk skipping the cognitive work that generates the thought in the first place. The struggle to find the sentence isn&#8217;t inefficiency. It&#8217;s often where the idea gets made. </p>



<p class="wp-block-paragraph">I&#8217;ve taught writing long enough to know that. More to the point, I&#8217;ve also taught web design, coding, webmaking, and video production long enough to learn a powerful lesson. When you go into the process of making, writing, building, or creating, you never know what will come out the other side. There are certain things (<a href="https://en.wikipedia.org/wiki/Affordance" rel="nofollow noopener" target="_blank">affordances</a>) that help or hinder you as you work and create. For example, when you&#8217;re writing, doing this is different in a Google Doc than it is on a typewriter, with pen/paper, or in Markdown. You&#8217;ll have different results and a different relationship with your own thoughts.</p>



<p class="wp-block-paragraph">A typewriter demands a certain level of commitment; every keystroke is a permanent decision, forcing a linear, deliberate pace. Pen and paper offer a tactile, spatial freedom where arrows, margin doodles, and crossed-out lines map the messy geography of your brain. Google Docs prioritizes the social and the frictionless, where the ghost of a collaborator’s cursor or the safety net of &#8220;version history&#8221; encourages a fast, iterative flow. Even Markdown shifts the focus to structure and logic, stripping away the visual distractions of formatting to keep you locked in the &#8220;code&#8221; of your prose.</p>



<p class="wp-block-paragraph">Ultimately, the tool isn&#8217;t just a passive vessel for your ideas — it is a silent collaborator. Each medium exerts its own gravity, pulling your creative process in directions you didn&#8217;t see coming when you first sat down to work. I&#8217;ve watched this happen with students for years. Give a kid the wrong tool at the wrong moment and the idea doesn&#8217;t just come out badly — it doesn&#8217;t come out at all. The gap between knowing something and rendering it into prose is where a lot of thinking happens. Close that gap too quickly, with the wrong collaborator, and the thinking doesn&#8217;t happen at all.</p>



<p class="wp-block-paragraph">So yes. The commenter is right.</p>



<p class="wp-block-paragraph">But here&#8217;s the thing: that&#8217;s not what I built <em>ian-writer</em> to do.</p>



<h2 class="wp-block-heading">Stop erasing my ideas</h2>



<p class="wp-block-paragraph">I didn&#8217;t build it because I wanted to stop drafting. I built it because something kept erasing my drafts.</p>



<p class="wp-block-paragraph">I&#8217;ve been publishing the Digitally Literate newsletter for over a decade. I wanted to get my drafts out of the older systems and tools that previously held them and archive them on my own site. As you can imagine, when you bring ten years of weekly writing into one space, you&#8217;ll get some drift.</p>



<p class="wp-block-paragraph">I started using AI tools to figure out how to convert the files from HTML and other formats, strip out the unnecessary garbage, and archive them in Markdown. I wanted to identify a throughline that would include my ideas and voice, while also connecting themes and patterns across the individual issues. </p>



<p class="wp-block-paragraph">As I engaged in this process, I kept running into the same problem. After every couple of passes through the issues, the AI model would sand something down. My voice would come back simplified, neutralized, and written in a way I would never actually write. Not wrong, exactly. Just not mine.</p>



<p class="wp-block-paragraph">After a year of pushing back against that cycle of rewriting, correcting, and re-correcting, I started thinking about the problem differently. I wasn&#8217;t losing my voice in the drafting. I was losing it in the translation. Those are different failure modes, and they call for different responses. This was more than a problem of prompt engineering. This was a space of friction in my work, writing, and thinking. I was trying to identify and fight for my voice as a writer. </p>



<p class="wp-block-paragraph">I think the commenter was worried about the first kind of friction. The productive kind, where the struggle to find words is also the struggle to find the thought. That friction is worth protecting. I agree. I have thoughts about that for an upcoming post. </p>



<p class="wp-block-paragraph">What I was trying to fix was the second kind. Translation friction, where you already have the thought and lose it on the way to the page. Not because the thinking was incomplete, but because the tool that was shaping it didn&#8217;t know what my voice was supposed to sound like.</p>



<p class="wp-block-paragraph">Imagine you&#8217;re painting a picture and standing in front of the canvas. You sketch out an idea and start painting. After some time, you step back to look at your work, only to realize that every time you turned your head, an invisible hand reached out and softened your brushstrokes. It blended your intentionally sharp edges into gradients, muted your boldest colors into &#8220;safe&#8221; neutrals, and corrected your perspective into something mathematically perfect — but empty. You didn&#8217;t lose the vision in your mind. You lost the evidence of your hand on the canvas.</p>



<p class="wp-block-paragraph">I didn&#8217;t build a custom system because I was lazy or looking for a shortcut. I built it because I was tired of fighting an &#8220;assistant&#8221; that insisted on tidying up my messy, human truths. I needed a tool that functioned less like an editor and more like a mirror. One that reflected my specific jagged edges instead of buffing them away.</p>



<p class="wp-block-paragraph">This distinction is vital for anyone worried that AI will replace the soul of writing. My work with <em>ian-writer</em> only ever touches the second kind of friction. The blank page still belongs to me.</p>



<h2 class="wp-block-heading">Writing is a process</h2>



<p class="wp-block-paragraph">I want to be precise about what that means in practice, because I think most people imagine a different workflow than the one I actually use.</p>



<p class="wp-block-paragraph">I don&#8217;t open a blank document and start writing. I also don&#8217;t open a blank document and ask the model to write something. In my work, writing is a process, and I&#8217;ve built up a system over the years.</p>



<p class="wp-block-paragraph">My system usually includes a series of text notes I take in passing, or a voice recording while in the car. It includes the podcast I was listening to or the blog posts I came across online. And yes, it includes output from AI queries on a topic. All of this goes into Obsidian as a chaotic dump of half-formed ideas. I start teasing these apart and ask AI models to help me find patterns and threads in my thinking. What am I actually trying to understand in this work? How might I explain it to others?</p>



<p class="wp-block-paragraph">When the AI model responds, I usually engage in a back-and-forth to better understand and push the thinking. Injecting friction into the process. If it responds the way we&#8217;ve all come to expect, a long laundry list of bullet points, hyperlinks, and ideas that have nothing to do with my actual question, I tune out and redirect or delete the chat. It&#8217;s frustrating. When it sends something back in a register that sounds like me, it&#8217;s not generating my thinking. It&#8217;s reflecting it back in a form I can work from.</p>



<p class="wp-block-paragraph">The model never sees a blank page. It sees notes. What it produces isn&#8217;t a draft. It&#8217;s a mirror to help me clarify my thinking.</p>



<p class="wp-block-paragraph">There&#8217;s one more honest piece of the workflow worth naming. Once I have a draft I&#8217;m satisfied with, I&#8217;ll sometimes run it through <em>ian-writer</em> as a final check. Not to correct my grammar, but to catch the places where I&#8217;ve drifted. Where I&#8217;ve over-qualified a claim into meaninglessness, conflated two ideas that needed to stay separate, or slipped back into the kind of academic register that loses everyday readers before they reach the second paragraph. Generic proofreading tools correct toward a standard that isn&#8217;t mine. This is a check run against my own established patterns. Not &#8220;is this correct?&#8221; but &#8220;is this still me, and will it land for someone who isn&#8217;t already living inside my head?&#8221;</p>



<p class="wp-block-paragraph">That distinction doesn&#8217;t resolve every concern. The commenter&#8217;s worry about intellectual atrophy is real and worth holding onto. The risk of offloading too much, of letting the model close the loop before the loop is actually closed, is a failure mode I think about. The original post named this as <em>the tics without the thinking</em>. Close enough to be recognizable, wrong enough to be uncomfortable, and useful precisely because of that discomfort.</p>



<p class="wp-block-paragraph">But building a tool to stop a specific kind of loss is different from building one to skip a necessary kind of struggle.</p>



<p class="wp-block-paragraph">I was protecting my voice. Not replacing it.</p>



<p class="wp-block-paragraph">Photo by <a href="https://unsplash.com/@susan_wilkinson?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Susan Wilkinson</a> on <a href="https://unsplash.com/photos/background-pattern-fMXyMuOxuYU?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>



<p class="wp-block-paragraph"></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/writing-is-a-process-so-is-losing-your-voice/">&lt;span class=&#039;p-name&#039;&gt;Writing Is a Process. So Is Losing Your Voice.&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/writing-is-a-process-so-is-losing-your-voice/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7287</post-id>	</item>
		<item>
		<title>I Trained a Local AI to Write Like Me</title>
		<link>https://wiobyrne.com/i-trained-a-local-ai-to-write-like-me/</link>
					<comments>https://wiobyrne.com/i-trained-a-local-ai-to-write-like-me/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Thu, 16 Apr 2026 16:49:32 +0000</pubDate>
				<category><![CDATA[Doing]]></category>
		<category><![CDATA[AI]]></category>
		<category><![CDATA[writing]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7277</guid>

					<description><![CDATA[<p>A few weeks ago, I pulled up a list of models running on my home server and noticed one I&#8217;d half-forgotten. It was called ian-writer, it was nine gigabytes, and its entire purpose was to sound like me. That&#8217;s a strange thing to sit with. I&#8217;ve been publishing the Digitally Literate newsletter for over a... <span class="more"><a class="more-link" href="https://wiobyrne.com/i-trained-a-local-ai-to-write-like-me/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/i-trained-a-local-ai-to-write-like-me/">&lt;span class=&#039;p-name&#039;&gt;I Trained a Local AI to Write Like Me&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph">A few weeks ago, I pulled up a list of models running on my home server and noticed one I&#8217;d half-forgotten. It was called <em>ian-writer</em>, it was nine gigabytes, and its entire purpose was to sound like me.</p>



<p class="wp-block-paragraph">That&#8217;s a strange thing to sit with. </p>



<p class="wp-block-paragraph">I&#8217;ve been publishing the <a href="https://digitallyliterate.net/newsletter/" rel="nofollow noopener" target="_blank"><em>Digitally Literate</em> newsletter</a> for over a decade. When I started archiving all 400+ issues into Markdown, I leaned on AI tools to help format and clean the files. What I kept noticing was that every revision sanded something down. My voice would come out the other side simplified, neutralized, written in a way I would never actually write. It was honestly infuriating. After more than a year of fighting it, I tried flipping the problem. Instead of asking AI to edit my writing, I asked it to learn from it.</p>



<p class="wp-block-paragraph">So I did it. But what I ended up building wasn&#8217;t just a custom model. It was a writing stack.</p>



<h2 class="wp-block-heading"><strong>Three layers, not one</strong></h2>



<p class="wp-block-paragraph">I currently have three parts of this stack.</p>



<p class="wp-block-paragraph">The first is <em><a href="https://wiobyrne.com/ian-md/">IAN.md</a></em>, a plain-text file I drop into any AI session. It describes who I am, how I work, what I focus on, and how I prefer to collaborate. It tells the model I write a weekly newsletter, that I&#8217;m an educator and researcher, and that I&#8217;d rather hear &#8220;I&#8217;m not sure&#8221; than a confident wrong answer. It&#8217;s a portable identity document. It gives the model context, but it doesn&#8217;t teach it how I sound.</p>



<p class="wp-block-paragraph">The second layer is <em>ian-writer</em>, a local model built with Ollama, software that runs AI on your own hardware without sending anything to a cloud server. I gave it four extended writing samples and a detailed system prompt. No em dashes, no bullet points used as a substitute for actually developing an idea, parenthetical definitions when a technical term appears for the first time, physical space analogies over military or sci-fi framings. The model doesn’t just know about my voice. It starts from those instructions every time. Every session starts with knowing how I should sound.</p>



<p class="wp-block-paragraph">The third layer is editorial judgment. My own. That&#8217;s the part that can&#8217;t be encoded.</p>



<p class="wp-block-paragraph">The two tools fit different moments. <em>IAN.md</em> plus a capable cloud model is where I think and learn. The reasoning is better, and the context travels. <em>ian-writer</em> running locally is where I draft. It&#8217;s slower and less capable, but it doesn&#8217;t pull my sentences toward the center. It&#8217;s a production tool, not a thinking tool.</p>



<h2 class="wp-block-heading"><strong>What the model gets wrong</strong></h2>



<p class="wp-block-paragraph">This is where the story gets more honest.</p>



<p class="wp-block-paragraph"><em>ian-writer </em>is good at surface patterns. Sentence length, the parenthetical habit, the tendency to close a technical piece by connecting it to something human. It is not good at argument. It doesn&#8217;t know when I&#8217;m hedging too much, when a paragraph needs to be cut rather than polished, or when I&#8217;m hiding behind clarity instead of saying something hard. It can reproduce the sound of my writing without the weight of it.</p>



<p class="wp-block-paragraph">Its main failure mode is being&nbsp;<em>too</em>&nbsp;consistent. It can replicate my stylistic tics, but it can’t generate a surprising metaphor or a genuinely new insight. It produces text that sounds like me, but it doesn’t&nbsp;<em>think</em>&nbsp;like me. It’s a&nbsp;voice prosthesis, not a ghostwriter.</p>



<p class="wp-block-paragraph">There&#8217;s also a version of the model that sounds too much like me. The tics without the thinking. Reading a draft from <em>ian-writer</em> that I haven&#8217;t touched yet is a little like hearing a recording of your own voice. Recognizable, slightly off, more instructive than comfortable.</p>



<h2 class="wp-block-heading"><strong>The real product</strong></h2>



<p class="wp-block-paragraph">I didn&#8217;t build a machine that replaces my voice. I built a system that helps me keep it.</p>



<p class="wp-block-paragraph">The paradox of this exercise is that you can only teach an AI to preserve your voice if you first make that voice legible enough to encode.</p>



<p class="wp-block-paragraph">What surprised me most wasn&#8217;t the output. It was the process of building the system prompt. Sitting down to describe my own voice precisely enough to teach it to a model forced me to name patterns I&#8217;d been following intuitively. The parenthetical definitions. The habit of closing on something human. The resistance to bullet points as a substitute for developed thought. I don&#8217;t know that I would have named those things if I hadn&#8217;t had to teach them.</p>



<p class="wp-block-paragraph">That observation matters beyond the tooling. AI systems can help preserve a voice only if you&#8217;ve first made that voice legible enough to encode. That’s less a technical insight than an editorial one. It came from trying to explain my writing to something that didn’t already understand it.</p>



<p class="wp-block-paragraph">If you want to try this, Ollama is a good place to start. Free, open source, runs on most modern hardware with enough memory. The <em>ian-writer </em>model isn&#8217;t public, but the approach is. I&#8217;ll write up a post showing how to build it if you want to give it a try.</p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@mooo3721?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">R Mo</a> on <a href="https://unsplash.com/photos/silver-click-pen-on-white-paper-fPbDggoRaQA?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>



<p class="wp-block-paragraph"></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/i-trained-a-local-ai-to-write-like-me/">&lt;span class=&#039;p-name&#039;&gt;I Trained a Local AI to Write Like Me&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/i-trained-a-local-ai-to-write-like-me/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7277</post-id>	</item>
		<item>
		<title>What Are You Actually Running? How to Audit Your Local AI Setup</title>
		<link>https://wiobyrne.com/audit-your-local-ai-setup/</link>
					<comments>https://wiobyrne.com/audit-your-local-ai-setup/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Wed, 15 Apr 2026 13:36:46 +0000</pubDate>
				<category><![CDATA[Doing]]></category>
		<category><![CDATA[local ai]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7264</guid>

					<description><![CDATA[<p>A methodology for getting honest about what&#8217;s on your machine — before it gets complicated. Most people who run local AI tools have, at some point, lost track of what they&#8217;ve actually built. Not because they weren&#8217;t paying attention. Because the thing grows. I&#8217;d been building and accumulating files for months. An IAN.MD, I use... <span class="more"><a class="more-link" href="https://wiobyrne.com/audit-your-local-ai-setup/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/audit-your-local-ai-setup/">&lt;span class=&#039;p-name&#039;&gt;What Are You Actually Running? How to Audit Your Local AI Setup&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph"><em>A methodology for getting honest about what&#8217;s on your machine — before it gets complicated.</em></p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph">Most people who run local AI tools have, at some point, lost track of what they&#8217;ve actually built.</p>



<p class="wp-block-paragraph">Not because they weren&#8217;t paying attention. Because the thing grows. </p>



<p class="wp-block-paragraph">I&#8217;d been building and accumulating files for months. An IAN.MD, I use to calibrate AI sessions, a local Ollama model called <em>ian-writer</em> which I built to draft in my voice. In addition, I had memory files that persist context across Claude Code sessions, as well as agent rules for my notes vault. Each piece made sense when I added it. But I&#8217;d lost track of how they related to each other.</p>



<p class="wp-block-paragraph">More importantly, I wasn&#8217;t sure what was actually <em>studying</em> me versus what was just <em>describing</em> me. I wanted to know the difference between a file that tells a model who I am, which is doing something different from a model configured to sound like me, and a memory system tracking what I&#8217;ve been working on.</p>



<p class="wp-block-paragraph">I wanted to know what I actually had, what each piece was doing, and where my voice and identity were actually stored, versus where I just assumed they were stored. So I ran an audit.</p>



<p class="wp-block-paragraph">The results were clarifying. Some things I thought were sophisticated turned out to be simple. Some things I&#8217;d dismissed as &#8220;just a text file&#8221; turned out to be doing more work than I realized. And a few things I&#8217;d completely forgotten about were still running, or still being read, or still shaping behavior in ways I wasn&#8217;t accounting for.</p>



<p class="wp-block-paragraph">This post is about the methodology. The audit prompt I used, why I structured it the way I did, and what kinds of questions it&#8217;s designed to answer. The next post shares the actual findings. What I had, what it really was, and the cleanup steps that came out of it.</p>



<h2 class="wp-block-heading">Why Audit at All</h2>



<p class="wp-block-paragraph">The reason to audit a local AI setup is the same reason you&#8217;d audit anything. Because the gap between what you think is true and what is actually true has costs.</p>



<p class="wp-block-paragraph">In this case, the costs are mostly confusion and wasted effort. If you think you&#8217;ve built a custom model when you&#8217;ve actually built a prompt wrapper, you&#8217;ll make decisions (about when to update it, how much to trust it, what it can do) based on a wrong assumption. </p>



<p class="wp-block-paragraph">If you have three different files that all try to define who you are and how you work, and they&#8217;re not synchronized, then whatever AI tool you&#8217;re using is working from a blurry, partially contradictory picture of you. If you have memory files that haven&#8217;t been updated in six weeks, you&#8217;re injecting stale context into every session and wondering why the responses feel slightly off.</p>



<p class="wp-block-paragraph">None of this breaks anything catastrophically. But it adds friction. It makes the system harder to maintain. And it means the &#8220;intelligence&#8221; you&#8217;re getting from these tools is partly shaped by things you&#8217;ve forgotten you put there.</p>



<p class="wp-block-paragraph">There&#8217;s a practical dimension to this, too. Local models don&#8217;t run for free. They consume electricity, processing cycles, and storage. Knowing what&#8217;s actually running, and whether it&#8217;s still earning its keep, is part of responsible homelab maintenance. The audit tells you not just what exists, but what&#8217;s worth keeping.</p>



<p class="wp-block-paragraph">An audit makes the implicit explicit. That&#8217;s most of what it does.</p>



<h2 class="wp-block-heading">What the Audit Covers</h2>



<p class="wp-block-paragraph">Local AI setups tend to fall into the same rough categories, even if they&#8217;re organized differently.</p>



<ul class="wp-block-list">
<li><strong>Models</strong> — the actual AI weights running locally, usually via Ollama or a similar tool. These might be stock models you downloaded, or models you&#8217;ve wrapped with a custom configuration.</li>



<li><strong>Identity and context files</strong> — documents that describe who you are, how you work, and what you care about. These are loaded into sessions (manually or automatically) to give models context without having to re-explain every time.</li>



<li><strong>Agent or instruction files</strong> — files that tell a model how to behave in a specific context. Rules for an inbox-sorting agent. Behavioral guidelines for a writing assistant. These are different from identity files: identity files say who you are; instruction files say <em>what to do</em>.</li>



<li><strong>Memory</strong> — persistent context that carries across sessions. This might be a structured folder of markdown files, a database, or something more exotic. The key question is: what actually gets read, by whom, and when?</li>



<li><strong>Prompts</strong> — templates you reuse. Stage prompts for different kinds of writing. Batch processing prompts. Query templates. These are instructions, not identity.</li>



<li><strong>Scripts and automation</strong> — actual executable code. Bots, runners, pipeline scripts, StreamDeck buttons. These do things; they don&#8217;t just describe things.</li>



<li><strong>Supporting documentation</strong> — setup guides, decision logs, run logs, state files. Useful context, not functional components.</li>
</ul>



<p class="wp-block-paragraph">The audit prompt I wrote asks about all of these. Not just to list them, but to classify them accurately and explain what role each one actually plays.</p>



<h2 class="wp-block-heading">The Prompt</h2>



<p class="wp-block-paragraph">I asked Claude Code to audit my own setup. To go through everything (models, prompts, config files, scripts, identity documents, memory stores) and tell me, as concretely as possible, what I had actually built. Not what I thought I&#8217;d built. Not what the folder names suggested. What was actually there, and what it was actually doing.</p>



<p class="wp-block-paragraph">Here&#8217;s the full audit prompt I used. You can adapt it to your own setup:</p>



<p class="wp-block-paragraph"><em>I want you to audit my local AI setup and explain exactly what I have built.</em></p>



<p class="wp-block-paragraph"><em>Please inspect my workspace, configs, local model files, prompts, markdown files, scripts, and any Ollama-related files. I am specifically trying to understand the difference between:</em></p>



<ul class="wp-block-list">
<li><em>identity files like <code>ian.md</code></em></li>



<li><em>agent or instruction files</em></li>



<li><em>local memory files or stores</em></li>



<li><em>skills/tools/scripts</em></li>



<li><em>custom Ollama models</em></li>



<li><em>any other supporting system components</em></li>
</ul>



<p class="wp-block-paragraph"><em>Your job is to determine, as concretely as possible:</em></p>



<p class="wp-block-paragraph"><strong>1. What artifacts exist</strong>? <em>Find and list all relevant files, folders, configs, and model definitions. Don&#8217;t just list file names. Open and inspect them.</em></p>



<p class="wp-block-paragraph"><strong>2. What each artifact actually is</strong>. <em>For each item, explain whether it is primarily a model, prompt, identity file, memory, knowledge file, agent instruction file, tool, wrapper, script, or something else.</em></p>



<p class="wp-block-paragraph"><strong>3. What role each artifact plays</strong>. <em>Explain what is static vs. dynamic, what is loaded into context vs. baked into a model, what changes over time, what is executable, and what is just text.</em></p>



<p class="wp-block-paragraph"><strong>4. Whether my custom model is a true model or just a configured wrapper</strong>. <em>Determine if it is fine-tuned. Is it a base model plus a system prompt? What base model does it depend on? Is its &#8220;custom-ness&#8221; in the weights or in the context?</em></p>



<p class="wp-block-paragraph"><strong>5. How identity is distributed across the system</strong>. <em>Where does &#8220;me&#8221; currently live — in a markdown file, in a model config, in prompts, in memory, in agent files? Is identity centralized or fragmented?</em></p>



<p class="wp-block-paragraph"><strong>6. How the system is structured overall</strong>. <em>Map the architecture in plain language: model layer, identity layer, instruction layer, memory layer, knowledge layer, tools layer, automation layer.</em></p>



<p class="wp-block-paragraph"><strong>7. Risks, redundancy, and confusion</strong>. Identify<em> duplicated definitions, overlapping files, potentially conflicting memory and identity files, wrappers that look like models, stale or unused components, anything that makes the system harder to maintain.</em></p>



<p class="wp-block-paragraph"><strong>8. Final verdict.</strong> <em>Give me: a clean inventory, a plain explanation of what I&#8217;ve actually built, whether I have a &#8220;custom model&#8221; or a &#8220;configured system&#8221; or a hybrid, and the 3–5 most important cleanup steps.</em></p>



<p class="wp-block-paragraph"><em>Important: Be skeptical of labels. Don&#8217;t just describe file names — open the files and determine what they actually are. If something is ambiguous, say so clearly. Prefer direct evidence over guesses.</em></p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<h2 class="wp-block-heading">Why It&#8217;s Structured This Way</h2>



<p class="wp-block-paragraph">The eight-section structure isn&#8217;t arbitrary. Each section aims to close a different gap between assumptions and reality.</p>



<p class="wp-block-paragraph">Sections 1–3 (inventory, classification, role) are about honest accounting. Take time to review and understand what you have installed and running. Most setups have more stuff than the owner remembers. Getting a complete list is the foundation.</p>



<p class="wp-block-paragraph">Section 4 (true model vs. wrapper) is, for me, the most important question for anyone running Ollama or a similar tool. As detailed in my intro, I wanted to understand the difference between a model fine-tuned on your writing and a base model with a system prompt that tries to sound like you. They feel similar in use. They are architecturally different. They have different implications for how you maintain and update them.</p>



<p class="wp-block-paragraph">Section 5 (identity distribution) surfaces a specific failure mode. For my purposes, I have been identifying ways to use local AI models to think and operate like me. This can be problematic if you have &#8220;me&#8221; defined in multiple places that aren&#8217;t synchronized with each other. One file says you&#8217;re focused on one project; another file, written three months earlier, says something else. The model reads both and gets a contradictory picture.</p>



<p class="wp-block-paragraph">Section 6 (architecture) forces the system into layers. This is useful because it makes visible which tools can communicate with which others. Your Ollama model doesn&#8217;t have access to your Claude Code memory files. Your Telegram bot doesn&#8217;t read your vault agent rules. Understanding the layers tells you why things work the way they do, and why certain things you want aren&#8217;t happening.</p>



<p class="wp-block-paragraph">Sections 7–8 (redundancies and verdict) are the payoff. What needs to be cleaned up, consolidated, or just acknowledged as the intentional complexity it is.</p>



<h2 class="wp-block-heading">What to Do With the Results</h2>



<p class="wp-block-paragraph">The audit gives you a picture. What you do with it depends on what you find.</p>



<p class="wp-block-paragraph">If you find stale files, update them or archive them. If a file says &#8220;current focus&#8221; but reflects work from two months ago, it&#8217;s injecting false context and noise into every session that reads it.</p>



<p class="wp-block-paragraph">If you find identity fragmentation, decide on a canonical source and note where the others differ. Will the real you please stand up? You don&#8217;t need to merge everything. Different tools serve different purposes. But you do need to know which file is authoritative for which purpose.</p>



<p class="wp-block-paragraph">If you find a &#8220;model&#8221; that&#8217;s actually a prompt wrapper, that&#8217;s not a problem; it&#8217;s just a fact. A prompt wrapper is a legitimate and useful tool. But it has different maintenance requirements. To update its behavior, you edit the system prompt and rebuild, not retrain.</p>



<p class="wp-block-paragraph">If you find items that are stored in only one place, back them up. System prompts that exist only inside Ollama&#8217;s blob storage, for instance, are one <em>ollama rm</em> away from being gone.</p>



<p class="wp-block-paragraph">And if you find things you&#8217;ve completely forgotten about, decide whether to keep, update, or remove them. Abandoned components aren&#8217;t neutral. They sit there in directories and get loaded into context or silently influence behavior in ways you&#8217;re no longer tracking.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph">In my next series of posts, I&#8217;ll share what I actually found when I ran this audit on my own setup. I&#8217;ll also detail an honest reckoning about what my &#8220;Ian writer&#8221; model really is, where my identity is actually stored, and the five cleanup steps I&#8217;m working through now.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph"><em>If this resonated, the best next step is to <a href="https://digitallyliterate.net/newsletter/" rel="noreferrer noopener nofollow" target="_blank">subscribe to my newsletter</a>.</em></p>



<p class="wp-block-paragraph"><em>Questions, pushback, or just want to connect — find me at </em><a href="https://wiobyrne.com"><em>wiobyrne.com</em></a><em>, <a href="https://www.linkedin.com/in/wiobyrne/" rel="noreferrer noopener nofollow" target="_blank">LinkedIn</a>, or Bluesky/</em><a href="http://@Wiobyrne@mastodon.social" rel="noreferrer noopener nofollow" target="_blank">Mastodon</a>.</p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@mihaiteslariu0?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Teslariu Mihai</a> on <a href="https://unsplash.com/photos/person-holding-magnifying-glass-with-black-frame-nquLTbqTPLc?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/audit-your-local-ai-setup/">&lt;span class=&#039;p-name&#039;&gt;What Are You Actually Running? How to Audit Your Local AI Setup&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/audit-your-local-ai-setup/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7264</post-id>	</item>
		<item>
		<title>Building Flynn: A Telegram Bot for Obsidian Using Local AI</title>
		<link>https://wiobyrne.com/building-flynn/</link>
					<comments>https://wiobyrne.com/building-flynn/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Tue, 14 Apr 2026 12:57:05 +0000</pubDate>
				<category><![CDATA[Doing]]></category>
		<category><![CDATA[AI]]></category>
		<category><![CDATA[virtual assistants]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7231</guid>

					<description><![CDATA[<p>This is the technical companion to I Built an AI Chief of Staff to Manage My Life Across Five Domains. That post covers the why. This one covers the how. Flynn is a Telegram bot that captures notes from your phone and routes them into your Obsidian vault using a local AI model. It runs... <span class="more"><a class="more-link" href="https://wiobyrne.com/building-flynn/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/building-flynn/">&lt;span class=&#039;p-name&#039;&gt;Building Flynn: A Telegram Bot for Obsidian Using Local AI&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<style>
body.postid-7231 pre.wp-block-preformatted {
  background: #f6f8fb;
  border: 1px solid #d9e2ec;
  border-radius: 8px;
  box-shadow: none;
  color: #1f2933;
  font-size: 15px;
  line-height: 1.55;
  margin: 1.25em 0 1.75em;
  overflow-x: auto;
  padding: 16px 18px;
  white-space: pre;
}
body.postid-7231 :not(pre) > code {
  background: #f6f8fb;
  border: 1px solid #d9e2ec;
  border-radius: 4px;
  box-shadow: none;
  color: #344054;
  display: inline;
  font-size: 0.92em;
  margin: 0;
  padding: 0.08em 0.35em;
}
</style>



<p class="wp-block-paragraph"><em>This is the technical companion to <a href="https://wiobyrne.com/i-built-an-ai-chief-of-staff/">I Built an AI Chief of Staff to Manage My Life Across Five Domains</a>. That post covers the why. This one covers the how.</em></p>



<p class="wp-block-paragraph">Flynn is a Telegram bot that captures notes from your phone and routes them into your Obsidian vault using a local AI model. It runs on your own hardware, costs nothing to operate, and survives reboots as a systemd service.</p>



<p class="wp-block-paragraph">Here&#8217;s how to build it.</p>



<h2 class="wp-block-heading">What you need</h2>



<ul class="wp-block-list">
<li>Linux machine (always-on, or a homelab VM)</li>



<li>Python 3.11+</li>



<li><a href="https://ollama.ai" rel="nofollow noopener" target="_blank">Ollama</a> with at least one model</li>



<li>An Obsidian vault</li>



<li>Telegram account</li>
</ul>



<p class="wp-block-paragraph">The bot runs fine on modest hardware. I run it on an Ubuntu PC with an RTX 4070 for Ollama, but it would work on a Raspberry Pi or a cheap VPS using smaller models.</p>



<h2 class="wp-block-heading">The architecture</h2>



<pre class="wp-block-preformatted">Telegram (your phone)
    ↓
assistant.py (Python, systemd service)
    ↓ detects intent (task vs reflection)
    ↓ parses date reference (today / tomorrow / day name)
    ↓ classifies via
Ollama (local) → Claude API (fallback) → keywords (fallback)
    ↓ writes to
Bot Inbox.md (tagged tasks) + daily note (Tasks Quick Add or Notes)
    ↓ queried by
Domain notes (Tasks plugin) + Flynn.base (Obsidian Bases)
</pre>



<p class="wp-block-paragraph">The key design decision: the AI only does <em>classification</em>. It reads a short note and returns a single domain name. That&#8217;s it. No long-context reasoning, no summarization. A small local model handles this easily and quickly.</p>



<h2 class="wp-block-heading">Step 1: Set up Ollama</h2>



<p class="wp-block-paragraph">Install Ollama and pull a model:</p>



<pre class="wp-block-preformatted">curl -fsSL https://ollama.ai/install.sh | sh
ollama pull llama3.1:8b
</pre>



<p class="wp-block-paragraph">Test it:</p>



<pre class="wp-block-preformatted">ollama run llama3.1:8b "Which is better, cats or dogs? Answer in one word."
</pre>



<p class="wp-block-paragraph">If you&#8217;re on a machine without a GPU, <em>llama3.1:8b</em> still runs on CPU. It&#8217;s slower but fine for occasional classification tasks.</p>



<h2 class="wp-block-heading">Step 2: Create your Telegram bot</h2>



<ol class="wp-block-list">
<li>Open Telegram and search for <code>@BotFather</code></li>



<li>Send <code>/newbot</code> and follow the prompts</li>



<li>Copy the token (looks like <code>123456789:ABCdef...</code>)</li>



<li>Send any message to your new bot</li>



<li>Visit <code>https://api.telegram.org/bot&lt;TOKEN&gt;/getUpdates</code></li>



<li>Find <code>"chat": {"id": 123456789}</code> — save that number</li>
</ol>



<p class="wp-block-paragraph">Keep your token private. Don&#8217;t paste it in public chats, don&#8217;t commit it to git.</p>



<h2 class="wp-block-heading">Step 3: Set up the Obsidian vault structure</h2>



<p class="wp-block-paragraph">Flynn writes tasks to 01 CONSUME/<img src="https://s.w.org/images/core/emoji/17.0.2/72x72/1f4e5.png" alt="📥" class="wp-smiley" style="height: 1em; max-height: 1em;" /> Inbox/Bot Inbox.md in your vault, tagged with #domain/self, #domain/build, etc.</p>



<p class="wp-block-paragraph">Create a <em>00 DOMAINS/</em> folder with one note per domain. Each note needs this frontmatter:</p>



<pre class="wp-block-preformatted">---
title: Self
domain: self
focus: active
energy: medium
next_action: ""
last_updated: 2026-04-02
---
</pre>



<p class="wp-block-paragraph">Add a Tasks plugin query to each note to show open tasks for that domain:</p>



<pre class="wp-block-preformatted">## Open Tasks

```tasks
not done
tags include #domain/self
limit 5
```
</pre>



<p class="wp-block-paragraph">Optionally add a Dataview status callout at the top:</p>



<pre class="wp-block-preformatted">&gt; [!info] Status
&gt; **Focus:** `= this.focus` | **Energy:** `= this.energy`
&gt; **Next:** `= this.next_action`
</pre>



<p class="wp-block-paragraph">This pulls live from frontmatter — when Flynn updates <em>next_action</em> via <em>/focus</em>, the callout updates automatically.</p>



<h2 class="wp-block-heading">Step 4: Install and configure Flynn</h2>



<pre class="wp-block-preformatted">git clone https://github.com/wiobyrne/flynn
cd flynn
python3 -m venv venv
venv/bin/pip install -r requirements.txt
cp .env.example .env
cp config.yaml my-config.yaml
</pre>



<p class="wp-block-paragraph">Edit <em>.env</em>:</p>



<pre class="wp-block-preformatted">TELEGRAM_BOT_TOKEN=your_token_here
TELEGRAM_CHAT_ID=your_chat_id_here
ANTHROPIC_API_KEY=   # optional
</pre>



<p class="wp-block-paragraph">Edit <em>config.yaml</em>:</p>



<pre class="wp-block-preformatted">vault_path: "~/Documents/your-obsidian-vault"

ollama:
  model: "llama3.1:8b"  # match whatever you pulled

domains:
  - id: self
    label: "Self"
    emoji: "<img src="https://s.w.org/images/core/emoji/17.0.2/72x72/1f3c3.png" alt="🏃" class="wp-smiley" style="height: 1em; max-height: 1em;" />"
    description: "health, mental wellness, physical fitness..."
    keywords: [health, gym, workout, sleep, ...]
</pre>



<p class="wp-block-paragraph">The <em>description</em> field is what the AI receives for classification. Make it specific to how you actually think about that area of your life. The <em>keywords</em> field is the fallback if the AI is unavailable.</p>



<h2 class="wp-block-heading">Step 5: Test it</h2>



<pre class="wp-block-preformatted">venv/bin/python assistant.py
</pre>



<p class="wp-block-paragraph">Send <em>/start</em> to your Telegram bot. You should get the welcome message. Send a test message — &#8220;remind me to call the doctor&#8221; — and check your Bot Inbox.md in Obsidian.</p>



<h2 class="wp-block-heading">Step 6: Run as a systemd service</h2>



<p class="wp-block-paragraph">Create <code>/etc/systemd/system/flynn.service</code>:</p>



<pre class="wp-block-preformatted">[Unit]
Description=Flynn AI Assistant Bot
After=network.target

[Service]
ExecStart=/home/youruser/flynn/venv/bin/python /home/youruser/flynn/assistant.py
WorkingDirectory=/home/youruser/flynn
Restart=on-failure
User=youruser

[Install]
WantedBy=multi-user.target
</pre>



<p class="wp-block-paragraph">Note: if your path has spaces (e.g. the bot lives inside an Obsidian vault folder), create a wrapper script:</p>



<pre class="wp-block-preformatted"># /home/youruser/flynn-start.sh
#!/bin/bash
exec "/path with spaces/venv/bin/python" "/path with spaces/assistant.py"
</pre>



<p class="wp-block-paragraph">Then point <em>ExecStart</em> at the wrapper script.</p>



<p class="wp-block-paragraph">Enable and start:</p>



<pre class="wp-block-preformatted">sudo systemctl enable flynn
sudo systemctl start flynn
sudo systemctl status flynn
</pre>



<h2 class="wp-block-heading">How the AI routing works</h2>



<p class="wp-block-paragraph">The classification prompt is simple:</p>



<pre class="wp-block-preformatted">Classify the note below into exactly one domain. Reply with ONLY the domain id.

Domains:
- self: health, mental wellness, physical fitness...
- family: marriage, relationship, home life...
[...]

Note: schedule a workout for later today

Domain:
</pre>



<p class="wp-block-paragraph">The model returns self. That&#8217;s the whole interaction. Short, fast, cheap.</p>



<p class="wp-block-paragraph">The three-tier fallback:<br>1. <strong>Ollama</strong> — local, instant, free<br>2. <strong>Claude Haiku</strong> — if <em>ANTHROPIC_API_KEY</em> is set and Ollama fails<br>3. <strong>Keywords</strong> — always available, no AI required</p>



<p class="wp-block-paragraph">In practice, Ollama handles everything. The fallbacks are there for when your machine is restarting or Ollama isn&#8217;t running yet.</p>



<h2 class="wp-block-heading">The Obsidian dashboard</h2>



<p class="wp-block-paragraph">With the <a href="https://obsidian.md/bases" rel="nofollow noopener" target="_blank">Obsidian Bases</a> core plugin enabled, create <em>Flynn.base</em> to show all five domain notes in a table:</p>



<pre class="wp-block-preformatted">filters:
  and:
    - file.folder == "00 DOMAINS"
properties:
  note.focus:
    displayName: Focus
  note.energy:
    displayName: Energy
  note.next_action:
    displayName: Next Action
views:
  - type: table
    name: Overview
    order:
      - title
      - focus
      - energy
      - next_action
</pre>



<p class="wp-block-paragraph">The Bases table is editable — click any cell to update the frontmatter directly. No need to open the individual domain note.</p>



<h2 class="wp-block-heading">Commands reference</h2>



<figure class="wp-block-table"><table class="has-fixed-layout"><thead><tr><th>Command</th><th>Does</th></tr></thead><tbody><tr><td>/toda<code>y</code></td><td>Morning briefing — domain status and next actions</td></tr><tr><td>/status</td><td>Bar chart of open tasks per domain</td></tr><tr><td>/list [domain]</td><td>Show open tasks, optionally filtered by domain</td></tr><tr><td>/done</td><td>Pick a task to mark complete from a numbered list</td></tr><tr><td>/done &lt;text></td><td>Mark a task complete by partial text match</td></tr><tr><td>/focus &lt;domain> &lt;text></td><td>Set next action on a domain note</td></tr><tr><td>/add &lt;text></td><td>Explicit task capture</td></tr><tr><td>/journal &lt;text></td><td>Save a note directly to today&#8217;s daily note</td></tr><tr><td>Any text</td><td>Auto-routed — Flynn detects task vs reflection</td></tr></tbody></table></figure>



<h2 class="wp-block-heading">The alignment layer</h2>



<p class="wp-block-paragraph">After a few weeks of using Flynn as a task router, I realized it could do something more useful: hold me accountable to my actual mission.</p>



<p class="wp-block-paragraph">My work across five domains looks productive on any given day. But busy isn&#8217;t the same as aligned. I can spend an entire week clearing infrastructure tasks and grading papers and never do the work that actually moves the needle on what I&#8217;m trying to build.</p>



<p class="wp-block-paragraph">I added a mission sentence to Flynn&#8217;s context file — the one sentence that describes what all five domains are ultimately supposed to be serving:</p>



<p class="wp-block-paragraph"><em>&#8220;I help educators understand and navigate the digital world — so their students inherit power, not just access.&#8221;</em></p>



<p class="wp-block-paragraph">Now every morning check-in ends with: <em>&#8220;How does today&#8217;s work connect to your mission?&#8221;</em> Every evening check-in ends with: <em>&#8220;Did today&#8217;s work lead back to the sentence? Yes / no / partially — and why?&#8221;</em></p>



<p class="wp-block-paragraph">That&#8217;s it. One question in the morning, one at night. But the answers accumulate in daily notes as structured data. After a few months I&#8217;ll have a real record of whether my calendar and my stated mission actually match.</p>



<p class="wp-block-paragraph">This is the thing productivity apps don&#8217;t do. They help you track tasks. Flynn helps me track whether the tasks are the right ones.</p>



<h2 class="wp-block-heading">Commands reference</h2>



<figure class="wp-block-table"><table class="has-fixed-layout"><thead><tr><th>Command</th><th>Does</th></tr></thead><tbody><tr><td>/today</td><td>Morning briefing — domain status and next actions</td></tr><tr><td>/status</td><td>Bar chart of open tasks per domain</td></tr><tr><td>/list [domain]</td><td>Show open tasks, optionally filtered by domain</td></tr><tr><td>/done</td><td>Pick a task to mark complete from a numbered list</td></tr><tr><td>/done &lt;text></td><td>Mark a task complete by partial text match</td></tr><tr><td>/focus &lt;domain> &lt;text></td><td>Set next action on a domain note</td></tr><tr><td>/week</td><td>Weekly digest — captured/completed per domain, overdue tasks</td></tr><tr><td>/add &lt;text></td><td>Explicit task capture</td></tr><tr><td>/journal &lt;text></td><td>Save a note directly to today&#8217;s daily note</td></tr><tr><td>Any text</td><td>Auto-routed — Flynn detects task vs reflection</td></tr></tbody></table></figure>



<h2 class="wp-block-heading">What&#8217;s next</h2>



<ul class="wp-block-list">
<li><strong>Gmail integration</strong> — surface emails that need replies by domain</li>



<li><strong>Phone home screen widget</strong> — a PWA showing domain status without opening Telegram</li>
</ul>



<p class="wp-block-paragraph">The code is at <a href="https://github.com/wiobyrne/flynn" rel="nofollow noopener" target="_blank">github.com/wiobyrne/flynn</a>. PRs welcome.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph"><em>If this resonated, the best next step is to <a href="https://digitallyliterate.net/newsletter/" rel="noreferrer noopener nofollow" target="_blank">subscribe to my newsletter</a>.</em></p>



<p class="wp-block-paragraph"><em>Questions, pushback, or just want to connect — find me at </em><a href="https://wiobyrne.com"><em>wiobyrne.com</em></a><em>, <a href="https://www.linkedin.com/in/wiobyrne/" rel="noreferrer noopener nofollow" target="_blank">LinkedIn</a>, or Bluesky/</em><a href="http://@Wiobyrne@mastodon.social" rel="noreferrer noopener nofollow" target="_blank">Mastodon</a>.</p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@andrewruiz?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Andrew Ruiz</a> on <a href="https://unsplash.com/photos/silhouette-of-man-and-woman-kissing-q9RXPcVuWw4?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/building-flynn/">&lt;span class=&#039;p-name&#039;&gt;Building Flynn: A Telegram Bot for Obsidian Using Local AI&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/building-flynn/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7231</post-id>	</item>
		<item>
		<title>I Built an AI Chief of Staff to Manage My Life Across Five Domains</title>
		<link>https://wiobyrne.com/i-built-an-ai-chief-of-staff/</link>
					<comments>https://wiobyrne.com/i-built-an-ai-chief-of-staff/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Mon, 13 Apr 2026 13:36:02 +0000</pubDate>
				<category><![CDATA[Doing]]></category>
		<category><![CDATA[AI]]></category>
		<category><![CDATA[flynn]]></category>
		<category><![CDATA[virtual-assistant]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7230</guid>

					<description><![CDATA[<p>I have too many jobs. Not in a complaining way. I chose this. I teach at a university, direct a community initiative, write a weekly newsletter, run a blog, develop courses, manage a homelab, coach volleyball, and try to maintain a functional human life with a family and a body that needs occasional attention. The... <span class="more"><a class="more-link" href="https://wiobyrne.com/i-built-an-ai-chief-of-staff/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/i-built-an-ai-chief-of-staff/">&lt;span class=&#039;p-name&#039;&gt;I Built an AI Chief of Staff to Manage My Life Across Five Domains&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph">I have too many jobs.</p>



<p class="wp-block-paragraph">Not in a complaining way. I chose this. I teach at a university, direct a community initiative, write a weekly newsletter, run a blog, develop courses, manage a homelab, coach volleyball, and try to maintain a functional human life with a family and a body that needs occasional attention.</p>



<p class="wp-block-paragraph">The problem isn&#8217;t having a lot going on. The problem is the cognitive overhead of tracking it all. Every morning, I spend mental energy just remembering what I&#8217;m supposed to be working on, what&#8217;s urgent, what&#8217;s been sitting too long. That&#8217;s wasted capacity.</p>



<p class="wp-block-paragraph">I wanted to offload that.</p>



<h2 class="wp-block-heading">Not a to-do list. A chief of staff.</h2>



<p class="wp-block-paragraph">Most people reach for productivity apps when they hit this wall. A better to-do list. A more elaborate calendar. A new system.</p>



<p class="wp-block-paragraph">I&#8217;ve tried all of it. The problem with most productivity tools is that they make you do the work of maintaining the system. You have to enter the tasks. You have to review the dashboard. You have to remember to check it.</p>



<p class="wp-block-paragraph">What I actually wanted was more like a chief of staff. Someone who knows the full picture, tracks the moving pieces, and brings me what I need to know. Rather than waiting for me to go looking.</p>



<p class="wp-block-paragraph">The difference sounds subtle. It isn&#8217;t.</p>



<h2 class="wp-block-heading">Why I didn&#8217;t build an &#8220;AI agent system&#8221;</h2>



<p class="wp-block-paragraph">If you follow AI news, you&#8217;ve probably seen people building teams of AI agents to run their lives. One agent handles email. Another manages the calendar. A third monitors social media.</p>



<p class="wp-block-paragraph">The framing is compelling. But the more I looked at how these setups work, the more I realized that everything runs through the cloud. Your notes, your email, your calendar, your half-formed ideas. Ill of it sent to a third-party API whenever the agent performs an action. One person I read about was spending nearly $1,000 a month on API costs alone.</p>



<p class="wp-block-paragraph">The data I work with lives on my own machines for a reason. I didn&#8217;t want my data leaving the machines I control.</p>



<p class="wp-block-paragraph">I wanted the same outcome. A complete picture of my day, tasks automatically routed to the right place, and a system that comes to me instead of waiting to be checked. All without handing over the keys.</p>



<h2 class="wp-block-heading">Five domains</h2>



<p class="wp-block-paragraph">Before building anything, I had to get clear on how I actually carve up my life. I landed on five areas:</p>



<ul class="wp-block-list">
<li><strong>Self</strong> — health, mental wellness, physical fitness</li>



<li><strong>Family</strong> — marriage, relationships, home</li>



<li><strong>Vocation</strong> — teaching, research, academic work</li>



<li><strong>Build</strong> — newsletter, blog, courses, business</li>



<li><strong>Infrastructure</strong> — homelab, systems, the technical layer</li>
</ul>



<p class="wp-block-paragraph">These aren&#8217;t projects. They&#8217;re ongoing areas of responsibility. Projects come and go inside them. The domains stay constant.</p>



<p class="wp-block-paragraph">I started from the end. What did I actually want to see each morning? What would a useful daily summary look like? Once I had that picture, everything else fell into place.</p>



<h2 class="wp-block-heading">What I built</h2>



<p class="wp-block-paragraph">I built an assistant called Flynn, named after the character from Tron.</p>



<p class="wp-block-paragraph">Flynn lives on my phone as a Telegram bot. I can send it anything, anywhere. Walking to a meeting and remembering I need to follow up with a journal editor? Send it to Flynn. Lying in bed thinking about a blog post idea? Flynn. Mid-workout, remember I need to schedule a doctor&#8217;s appointment? Flynn.</p>



<p class="wp-block-paragraph">Flynn figures out which domain the note belongs to and saves it to my Obsidian vault, tagged and dated, ready to surface when I need it. It also detects whether something is a task or a reflection. &#8220;I need to call the doctor&#8221; goes to my task inbox. &#8220;I feel like I&#8217;m finally getting ahead&#8221; goes to today&#8217;s daily note. I don&#8217;t have to think about which one.</p>



<p class="wp-block-paragraph">That&#8217;s the core loop. Capture anything, route it intelligently, surface it at the right time.</p>



<h2 class="wp-block-heading">How the routing works</h2>



<p class="wp-block-paragraph">Flynn uses a local AI model running on my own machine. No cloud, no API costs, no data leaving my network. It reads the message, knows the five domains and what they cover, and makes a call. For obvious cases, it&#8217;s instant and accurate. For edge cases, it falls back to keyword matching.</p>



<p class="wp-block-paragraph">It also understands basic date language. &#8220;Tomorrow I need to finish grading&#8221; goes in the schedule for tomorrow. &#8220;Prep for Monday&#8217;s class&#8221; schedules for Monday. I don&#8217;t have to format anything.</p>



<p class="wp-block-paragraph">After a few days of use, the routing is accurate enough that I rarely have to correct it. When I do, Flynn&#8217;s commands let me manage everything from Telegram without opening Obsidian.</p>



<h2 class="wp-block-heading">The check-ins</h2>



<p class="wp-block-paragraph">Every morning at 7am, Flynn sends a domain briefing. Task counts, next actions, and what&#8217;s been sitting too long. Five minutes later, it sends a check-in. Sleep, mood, anxiety, one thing I&#8217;m grateful for, what would make today good. This is the script I use for my daily journaling.</p>



<p class="wp-block-paragraph">Every evening, another check-in pops up on my phone. Energy, wins, friction, one thing to carry forward.</p>



<p class="wp-block-paragraph">The responses go into my daily note. The sleep score, mood score, and energy score are written to the note&#8217;s metadata. Over time, I&#8217;ll be able to query those across the vault and see actual patterns. Which domains correlate with low-mood weeks, and whether my energy tracks with how much I&#8217;m exercising. The kind of longitudinal data that&#8217;s impossible to track manually.</p>



<p class="wp-block-paragraph">I didn&#8217;t set out to build a mental health tracking tool. But when you&#8217;re capturing how you feel every morning and evening, in a system that stores everything as structured data you own, that&#8217;s exactly what you end up with.</p>



<h2 class="wp-block-heading">The dashboard in Obsidian</h2>



<p class="wp-block-paragraph">On the Obsidian side, each domain has its own note that shows open tasks, the current focus, and the next action. A dashboard view shows all five at a glance.</p>



<p class="wp-block-paragraph">In terms of design, I didn&#8217;t want another anxiety-inducing productivity dashboard screaming at me about everything I haven&#8217;t done. Each domain shows what&#8217;s next, not the full backlog. Progress is visible. Guilt is not the point.</p>



<h2 class="wp-block-heading">What it became</h2>



<p class="wp-block-paragraph">A few weeks in, Flynn started doing something I didn&#8217;t design it to do.</p>



<p class="wp-block-paragraph">It exposed a gap. Not between tasks and completion, but between what I say matters and how I actually spend my time.</p>



<p class="wp-block-paragraph">That forced a missing piece. I didn&#8217;t just need a system to manage work. I needed a clear top-level direction for what the work is <em>for</em>.</p>



<p class="wp-block-paragraph">So I wrote a single sentence:</p>



<p class="wp-block-paragraph"><em>I help educators understand and navigate the digital world — so their students inherit power, not just access.</em></p>



<p class="wp-block-paragraph">That sentence is now in Flynn&#8217;s context. Every morning, check-in ends with &#8220;How does today&#8217;s work connect to your mission?&#8221; Every evening ends with &#8220;Did today&#8217;s work lead back to the sentence?&#8221;</p>



<p class="wp-block-paragraph">Flynn went from task router to alignment mirror. I didn&#8217;t plan for that. But it&#8217;s the most useful thing it does.</p>



<h2 class="wp-block-heading">What changed</h2>



<p class="wp-block-paragraph">The main thing I notice is that I&#8217;m not holding as much in my head. When something comes up, an idea, a task, something I don&#8217;t want to forget, I send it to Flynn and let it go. It&#8217;ll be there when I need it.</p>



<p class="wp-block-paragraph">The morning briefing gives me a clean summary in thirty seconds. No inbox, no tab-switching, no trying to remember what I left unfinished.</p>



<p class="wp-block-paragraph">That&#8217;s the same outcome people get from cloud-based agent setups. But my data stays on my machine, the AI runs locally, and nothing moves through a server I don&#8217;t control. And where those setups can run hundreds of dollars a month in API fees, Flynn&#8217;s ongoing cost is electricity.</p>



<h2 class="wp-block-heading">The tools</h2>



<ul class="wp-block-list">
<li><strong>Flynn (the bot):</strong> Python, running locally as a background service</li>



<li><strong>Telegram:</strong> mobile access — works on any phone, no app to build</li>



<li><strong>Ollama:</strong> local AI for routing, runs on my own hardware, costs nothing per message</li>



<li><strong>Obsidian:</strong> the vault where everything lives</li>



<li><strong>Tasks plugin + Bases:</strong> for the dashboard views</li>
</ul>



<p class="wp-block-paragraph">The whole thing runs on hardware I already own.</p>



<p class="wp-block-paragraph">If you want the technical details, how to set it up, the code, the Obsidian structure, I wrote a companion post about that.</p>



<p class="wp-block-paragraph">Flynn is open source. You can find the code at <a href="https://github.com/wiobyrne/flynn" rel="nofollow noopener" target="_blank">github.com/wiobyrne/flynn</a>.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph"><em>If this resonated, the best next step is to <a href="https://digitallyliterate.net/newsletter/" rel="nofollow noopener" target="_blank">subscribe to my newsletter</a>. </em></p>



<p class="wp-block-paragraph"><em>Questions, pushback, or just want to connect — find me at </em><a href="https://wiobyrne.com"><em>wiobyrne.com</em></a><em>, <a href="https://www.linkedin.com/in/wiobyrne/" rel="nofollow noopener" target="_blank">LinkedIn</a>, or Bluesky/</em><a href="http://@Wiobyrne@mastodon.social" rel="nofollow noopener" target="_blank">Mastodon</a>. </p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@fernanddecanne?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Fernand De Canne</a> on <a href="https://unsplash.com/photos/silhouette-photography-of-person-standing-Zia8y5tXxp8?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>



<p class="wp-block-paragraph"></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/i-built-an-ai-chief-of-staff/">&lt;span class=&#039;p-name&#039;&gt;I Built an AI Chief of Staff to Manage My Life Across Five Domains&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/i-built-an-ai-chief-of-staff/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7230</post-id>	</item>
		<item>
		<title>Inside Claude Code: What Four Layers of AI Look Like in Practice</title>
		<link>https://wiobyrne.com/inside-claude-code-what-four-layers-of-ai-look-like-in-practice/</link>
					<comments>https://wiobyrne.com/inside-claude-code-what-four-layers-of-ai-look-like-in-practice/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Mon, 06 Apr 2026 16:41:17 +0000</pubDate>
				<category><![CDATA[Teaching]]></category>
		<category><![CDATA[AI]]></category>
		<category><![CDATA[claude]]></category>
		<category><![CDATA[claude code]]></category>
		<category><![CDATA[infrastructure]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7186</guid>

					<description><![CDATA[<p>This is the second in a two-part series. Part one introduced a framework for understanding any AI system. The model, harness, product, or infrastructure. This post applies it to a real, observable case. Every AI system has at least four layers: the model, the harness, the product, and the infrastructure. Most users never see past... <span class="more"><a class="more-link" href="https://wiobyrne.com/inside-claude-code-what-four-layers-of-ai-look-like-in-practice/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/inside-claude-code-what-four-layers-of-ai-look-like-in-practice/">&lt;span class=&#039;p-name&#039;&gt;Inside Claude Code: What Four Layers of AI Look Like in Practice&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph"><em>This is the second in a two-part series. <a href="https://wiobyrne.com/youre-not-just-using-a-model-the-four-layers-of-every-ai-system/">Part one</a> introduced a framework for understanding any AI system. The model, harness, product, or infrastructure. This post applies it to a real, observable case.</em></p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph">Every AI system has at least four layers: the model, the harness, the product, and the infrastructure. Most users never see past the product layer. That is the chat box, the buttons, the features.</p>



<p class="wp-block-paragraph">This past week, we had the opportunity to peer behind the curtain. </p>



<p class="wp-block-paragraph"><a href="https://code.claude.com/docs/en/overview" rel="nofollow noopener" target="_blank">Claude Code</a> is an AI-powered agentic coding tool developed by <a href="https://www.anthropic.com/" rel="nofollow noopener" target="_blank">Anthropic</a> that runs in your terminal to edit files, run commands, and build software projects. Agentic coding tools are<br>AI-powered systems that autonomously plan, write, test, and debug code, moving beyond simple autocomplete to handle entire software tasks. It understands entire codebases, can handle multi-file tasks, and is intended to accelerate development workflows. The tool is specialized for agentic, autonomous coding tasks</p>



<p class="wp-block-paragraph">A mirrored version of the source repository for Claude Code became publicly available, which means we can actually look at the harness. That is the control layer that sits between the model and everything you experience. It&#8217;s rare to have this kind of window into a major AI product.</p>



<p class="wp-block-paragraph">I&#8217;m interested in this leak for two reasons. First, I pay for a subscription to Claude and Claude Code, but I&#8217;m also trying to run as much of my AI locally as possible. I&#8217;m excited to see what open-source, local models pop up. </p>



<p class="wp-block-paragraph">Second, as an educator and researcher, I try to understand what these systems do and make them approachable for everyday users. This summer, I&#8217;ll have a series of webinars and classes on AI. We talk about looking at the source or primary documents as we study and learn a new topic. Well&#8230;this is one of the top source documents. </p>



<p class="wp-block-paragraph">What follows isn&#8217;t a technical teardown. It&#8217;s a literacy exercise. Using real, observable structure to show that &#8220;the AI&#8221; is a system, not a brain in a box.</p>



<p class="wp-block-paragraph"><em>Please note:</em> This is a glimpse into one model from one company. There are definite nuances between the different models and infrastructures, but this leak gives us more insight into what is happening. </p>



<p class="wp-block-paragraph">I&#8217;ll use four lenses to look at it.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<h2 class="wp-block-heading">Agency — What It&#8217;s Allowed to Do</h2>



<p class="wp-block-paragraph">The clearest thing the repo (repository &#8211; a centralized digital storage location used in software development to manage, track, and collaborate on code changes) reveals is that Claude Code is not just a model answering questions. It has approximately 40 permission-gated tools. Structured ways to read files, inspect code, run workflows, search the web, execute commands, and act conditionally.</p>



<p class="wp-block-paragraph">Think of these permissions as a security pass or a specific set of keys. Just because the AI &#8220;knows&#8221; how to perform a task doesn&#8217;t mean it has the credentials to enter that room and execute it.</p>



<p class="wp-block-paragraph">The permission layer is the boundary of agency. It defines what the agent <em>can</em> do, not just what it <em>knows</em>. And those permissions are configurable. The harness, not the model, determines whether the system can write to your filesystem, call external APIs, or spawn sub-agents. Without this &#8220;pass,&#8221; the model is an observer; with it, it becomes an authorized actor.</p>



<p class="wp-block-paragraph"><strong>What this means:</strong> When you give an AI tool access to your computer, you&#8217;re not just talking to a model. You&#8217;re activating a permission layer. Understanding that layer — and who controls it — is an AI literacy issue, not just a technical one.</p>



<h2 class="wp-block-heading">Memory — How It Keeps Track</h2>



<p class="wp-block-paragraph">The repo includes a memory system with an internal codename: <strong>autoDream</strong>. The name is easy to dismiss as whimsy. The mechanism is not. </p>



<p class="wp-block-paragraph">Like a human brain, an AI model has a limited &#8220;working memory.&#8221; When you have too much on your mind at once, you get overwhelmed, slow down, or start forgetting details. You can take pauses throughout the day to clear your mind. When you sleep, your brain effectively &#8220;takes out the trash&#8221; and cleans things up for the next day. </p>



<p class="wp-block-paragraph">To prevent this &#8220;mental clutter,&#8221; <strong>autoDream</strong> periodically consolidates prior sessions, removes contradictions, and turns recent work into durable, structured memory.</p>



<p class="wp-block-paragraph">It&#8217;s not a raw transcript. It&#8217;s a managed compression of experience. Much like how you don&#8217;t remember every single word of a conversation from last week (only the key points and the &#8220;vibe&#8221;), this system periodically shrinks massive amounts of data into bite-sized summaries to keep the model fast and focused.</p>



<p class="wp-block-paragraph">That&#8217;s a meaningful distinction. Memory in Claude Code isn&#8217;t the model remembering. It&#8217;s an engineered system deciding what to keep, how to organize it, and when to compress it.</p>



<p class="wp-block-paragraph"><strong>What this means:</strong> The &#8220;continuity&#8221; you experience across sessions in an AI tool is a product decision, not a property of the model. Someone designed how memory works, what gets kept, and what gets discarded. That&#8217;s worth knowing.</p>



<h2 class="wp-block-heading">Identity — What It Reveals and Hides</h2>



<p class="wp-block-paragraph">The repo includes logic for something called <strong>Undercover Mode</strong>. This suppresses internal references and omits AI attribution in public or open-source contexts. It is explicit product logic for how the system presents itself. It is essentially a &#8220;cloaking device&#8221; that hides the AI&#8217;s nature, making its output appear to be written by a human.</p>



<p class="wp-block-paragraph">There is also a separate subsystem called <strong>Buddy</strong>: a Tamagotchi-style companion layer with species, rarity, and stats. This is the product experimenting with motivation and attachment. It’s a design tool used to create a &#8220;social&#8221; bond, proving that &#8220;personality&#8221; is often just a gamified UI layer meant to drive engagement.</p>



<p class="wp-block-paragraph">Together, these reveal that identity in an AI product is managed, not neutral. We often think a model &#8220;feels&#8221; better because it has a better &#8220;soul,&#8221; but that &#8220;soul&#8221; is actually a sophisticated &#8220;harness.&#8221; It’s composed of the same code, language, and math as the rest of the system, but it&#8217;s tuned to mimic empathy or professionalism. Whether you are using a built-in &#8220;Buddy&#8221; or <a href="https://wiobyrne.com/ian-md/">crafting your own .md persona</a> file, you are interacting with a deliberate mask.</p>



<p class="wp-block-paragraph"><strong>What this means:</strong> &#8220;Who am I talking to?&#8221; is not a simple question. You aren&#8217;t talking to a conscious being; you are talking to a Product Strategy. The &#8220;warmth&#8221; or &#8220;utility&#8221; you feel is a result of decisions made at the UI layer to determine how much of the underlying machine should be revealed or hidden.</p>



<h2 class="wp-block-heading">Orchestration — One Model or a Managed Team?</h2>



<p class="wp-block-paragraph">The repo shows a coordination infrastructure that goes beyond a single model responding to queries. This isn&#8217;t a chatbot; it is a <strong>Digital Hierarchy.</strong> There is a <strong>Coordinator Layer</strong> acting as a project manager, specialized <strong>Worker Agent</strong> structures acting as the &#8220;boots on the ground,&#8221; and a system called Kairos, a persistent, &#8220;daemon-style&#8221; agent that monitors events in the background.</p>



<p class="wp-block-paragraph">Unlike a standard LLM that only &#8220;thinks&#8221; when you press Enter, <strong>Kairos</strong> stays awake. It’s designed to be more autonomous over time, watching your files and running tasks without waiting for a direct prompt. This means the system is designed more like a managed team of sub-processes, each with defined roles, scoped tools, and coordination logic sitting above them.</p>



<p class="wp-block-paragraph">That&#8217;s a meaningful distinction. When the system says, <em>&#8220;I have updated your files,&#8221;</em> that &#8220;I&#8221; is a polite way of referring to the product. In reality:</p>



<ul class="wp-block-list">
<li>The <strong>Coordinator</strong> received your request and broke it down.</li>



<li>A <strong>Worker</strong> executed the specific code change.</li>



<li><strong>Kairos</strong> verified the file system was stable in the background.</li>
</ul>



<p class="wp-block-paragraph"><strong>What this means:</strong> The &#8220;continuity&#8221; and &#8220;intelligence&#8221; you experience aren&#8217;t just coming from one giant brain. They are the result of sophisticated System Architecture. The sophistication lives in the &#8220;Middle Management&#8221; layer, the code that decides which specialized worker to hire for each micro-task. You aren&#8217;t talking to a &#8220;Who&#8221;; you are directing a &#8220;How.&#8221;</p>



<h2 class="wp-block-heading">What the Foundation Layer Shows</h2>



<p class="wp-block-paragraph">Underneath those four lenses is infrastructure most users never think about. The &#8220;physics&#8221; of how the AI must behave.</p>



<ul class="wp-block-list">
<li><strong>System Prompts and Behavior Rules:</strong> These are the &#8220;hidden commandments&#8221; that set the tone, establish constraints, and provide safety guidance. They live in the harness, not your conversation. Even if you tell the AI to &#8220;be yourself,&#8221; it can’t. It is legally and technically bound by these invisible guardrails before your first word even reaches the model.</li>



<li><strong>Telemetry and Feature Flags:</strong> The repo contains internal codenames for disabled or unreleased capabilities. It reads as a roadmap artifact. A window into where the product is going, not just what it currently does. This reveals that the &#8220;AI&#8221; you see today is often a throttled version of what is already built.</li>



<li><strong>Anti-distillation Logic:</strong> This is the most “corporate” layer. It’s a defensive wall designed to stop competitors from &#8220;stealing the recipe.&#8221; By injecting &#8220;fake tools&#8221; or summarized data into its own processes, the system essentially feeds garbage data to any other AI trying to learn from it. It is an AI designed to resist being understood by other AIs.</li>
</ul>



<p class="wp-block-paragraph">What this really tells us is that a company like Anthropic doesn&#8217;t just view Claude as a &#8220;helpful assistant.&#8221; They view it as <em>Proprietary Infrastructure</em>. The invisible instructions ensure the brand stays safe. The feature flags ensure the business can roll out updates for profit. The anti-distillation logic ensures the competition can&#8217;t &#8220;steal the recipe.&#8221;</p>



<p class="wp-block-paragraph"><strong>What this means:</strong> When you use a professional AI tool, you are operating within a highly controlled corporate environment. Every response you get has been filtered through a layer of Safety, Telemetry, and Competitive Defense. The &#8220;Foundation&#8221; isn&#8217;t just technical; it&#8217;s political and economic.</p>



<h2 class="wp-block-heading">Wrapping Up</h2>



<p class="wp-block-paragraph">The repo doesn&#8217;t reveal the model. It reveals the system built around the model. The part that actually shapes most of what you experience.</p>



<p class="wp-block-paragraph">Agency, memory, identity, orchestration. Each one is a product decision, not a model property. Each one is designed, configured, and controlled by someone other than you. That&#8217;s not a criticism; it&#8217;s just what AI systems are. The literacy goal isn&#8217;t to be suspicious of every tool. It&#8217;s to stop imagining a &#8220;brain in a box&#8221; and start seeing a stack of decisions made by people with their own priorities.</p>



<p class="wp-block-paragraph">Knowing the layers means you can ask better questions about the tools you use:</p>



<ul class="wp-block-list">
<li><strong>Which layer is behaving this way?</strong> (Is it the model’s math or the system’s rules?)</li>



<li><strong>Who controls it?</strong> (Is this a hard-coded constraint or something I can toggle?)</li>



<li><strong>Can I inspect or configure it?</strong> (Can I write my own <code>.md</code> persona to override the &#8220;Buddy&#8221;?)</li>



<li><strong>What&#8217;s being kept, and what&#8217;s being hidden?</strong> (Is my data being &#8220;compressed&#8221; or just deleted?)</li>
</ul>



<p class="wp-block-paragraph">Those are AI literacy questions. The repo just makes them unusually easy to ask. </p>



<p class="wp-block-paragraph">Once you see the &#8216;harness,&#8217; you realize the model isn&#8217;t driving the car. The architecture is. Understanding the architecture is the first step toward becoming the driver.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph">Thanks for reading. Reach out at: <a href="mailto:hello@wiobyrne.com" target="_blank" rel="noreferrer noopener">hello@wiobyrne.com</a></p>



<p class="wp-block-paragraph">You can also <a href="https://buttondown.email/digitallyliterate" target="_blank" rel="noreferrer noopener nofollow">subscribe here</a> to my newsletter or <a href="https://ko-fi.com/wiobyrne" target="_blank" rel="noreferrer noopener nofollow">support me on Ko-fi</a>.</p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@dnevozhai?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Denys Nevozhai</a> on <a href="https://unsplash.com/photos/aerial-photography-of-concrete-roads-7nrsVjvALnA?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/inside-claude-code-what-four-layers-of-ai-look-like-in-practice/">&lt;span class=&#039;p-name&#039;&gt;Inside Claude Code: What Four Layers of AI Look Like in Practice&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/inside-claude-code-what-four-layers-of-ai-look-like-in-practice/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7186</post-id>	</item>
		<item>
		<title>You&#8217;re Not Just Using a Model: The Four Layers of Every AI System</title>
		<link>https://wiobyrne.com/youre-not-just-using-a-model-the-four-layers-of-every-ai-system/</link>
					<comments>https://wiobyrne.com/youre-not-just-using-a-model-the-four-layers-of-every-ai-system/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Mon, 06 Apr 2026 15:28:41 +0000</pubDate>
				<category><![CDATA[Teaching]]></category>
		<category><![CDATA[AI]]></category>
		<category><![CDATA[ai models]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7185</guid>

					<description><![CDATA[<p>Most people talk about AI as if they&#8217;re talking directly to a brain. They&#8217;re not. Between you and &#8220;the model,&#8221; there are at least three more layers. Understanding those layers changes how you evaluate AI tools, why you sometimes get surprising behavior, and who actually controls what you experience. A while back, I wrote about... <span class="more"><a class="more-link" href="https://wiobyrne.com/youre-not-just-using-a-model-the-four-layers-of-every-ai-system/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/youre-not-just-using-a-model-the-four-layers-of-every-ai-system/">&lt;span class=&#039;p-name&#039;&gt;You&#8217;re Not Just Using a Model: The Four Layers of Every AI System&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph">Most people talk about AI as if they&#8217;re talking directly to a brain.</p>



<p class="wp-block-paragraph">They&#8217;re not. Between you and &#8220;the model,&#8221; there are at least three more layers. Understanding those layers changes how you evaluate AI tools, why you sometimes get surprising behavior, and who actually controls what you experience.</p>



<p class="wp-block-paragraph">A while back, I wrote about <a href="https://wiobyrne.com/inside-an-ai-model/" target="_blank" rel="noreferrer noopener">what&#8217;s actually inside an AI model</a>. This included the weights, the config file, and the tokenizer. That post answered the question of what you&#8217;d see if you were able to download an AI model and examine it. It&#8217;s a good starting point. But it only tells part of the story.</p>



<p class="wp-block-paragraph">Here&#8217;s a clean breakdown.</p>



<h2 class="wp-block-heading"><img src="https://s.w.org/images/core/emoji/17.0.2/72x72/1f9e0.png" alt="🧠" class="wp-smiley" style="height: 1em; max-height: 1em;" /> The Model (the capability)</h2>



<p class="wp-block-paragraph"><strong>What it is:</strong> The trained system that actually does the thinking. It learned patterns from enormous amounts of data and can generate text, images, predictions, code, and more.</p>



<p class="wp-block-paragraph"><strong>In plain terms:</strong> It&#8217;s the brain.</p>



<p class="wp-block-paragraph"><strong>Examples:</strong> GPT-4, Claude, Gemini, Llama</p>



<p class="wp-block-paragraph">The key thing to understand about a model. On its own, it&#8217;s not very useful to most people. A model is a file, or more precisely, a set of files (weights, configuration, and a tokenizer). It needs structure around it before it becomes something you can actually use.</p>



<h2 class="wp-block-heading"><img src="https://s.w.org/images/core/emoji/17.0.2/72x72/1f6e0.png" alt="🛠" class="wp-smiley" style="height: 1em; max-height: 1em;" /> The Harness (how the model is put to work)</h2>



<p class="wp-block-paragraph"><strong>What it is:</strong> The software layer wrapped around the model. This is where much of the visible behavior is shaped, often more than people realize.</p>



<p class="wp-block-paragraph"><strong>In plain terms:</strong> It&#8217;s the control layer. It decides how the model is prompted, guided, and constrained.</p>



<p class="wp-block-paragraph"><strong>Includes things like:</strong></p>



<ul class="wp-block-list">
<li>Prompt templates and system instructions (&#8220;be helpful, don&#8217;t do X…&#8221;)</li>



<li>Memory and conversation history handling</li>



<li>Tools (web search, code execution, file access)</li>



<li>Workflow and coordination logic</li>



<li>The interface you interact with</li>
</ul>



<p class="wp-block-paragraph"><strong>Why this matters:</strong> The same underlying model with a different harness can behave dramatically differently. When you notice that Claude or ChatGPT &#8220;changed,&#8221; it&#8217;s often the harness that changed, not the model itself.</p>



<h2 class="wp-block-heading"><img src="https://s.w.org/images/core/emoji/17.0.2/72x72/1f4f1.png" alt="📱" class="wp-smiley" style="height: 1em; max-height: 1em;" /> The Product (what users actually use)</h2>



<p class="wp-block-paragraph"><strong>What it is:</strong> The full application or service built around the model and its harness. This is what you actually open, click, and interact with.</p>



<p class="wp-block-paragraph"><strong>In plain terms:</strong> It&#8217;s the experience.</p>



<p class="wp-block-paragraph"><strong>Examples:</strong> ChatGPT, Notion AI, Copilot, Claude.ai</p>



<p class="wp-block-paragraph"><strong>Includes:</strong></p>



<ul class="wp-block-list">
<li>The UI (chat box, buttons, voice interface)</li>



<li>Features (file uploads, history, sharing)</li>



<li>Account systems, pricing tiers, usage limits</li>
</ul>



<p class="wp-block-paragraph">This layer also includes a lot that&#8217;s invisible to users. Routing logic that decides which model version you get, feature flags that turn capabilities on or off by account tier, and product decisions made entirely outside the model or harness.</p>



<h2 class="wp-block-heading"><img src="https://s.w.org/images/core/emoji/17.0.2/72x72/1f527.png" alt="🔧" class="wp-smiley" style="height: 1em; max-height: 1em;" /> The Infrastructure (the system that keeps it all running)</h2>



<p class="wp-block-paragraph">This is the layer most people never think about, but it explains a lot of otherwise confusing AI behavior.</p>



<p class="wp-block-paragraph"><strong>What it is:</strong> The technical backend that keeps the product operational. It lives mostly outside the visible application.</p>



<p class="wp-block-paragraph"><strong>Includes things like:</strong></p>



<ul class="wp-block-list">
<li>Servers, scaling, and deployment</li>



<li>Authentication and billing systems</li>



<li>Telemetry and analytics</li>



<li>Feature flags (what&#8217;s enabled for which users)</li>



<li>Model switching logic (you may not always get the same model you think you&#8217;re getting)</li>



<li>Hidden prompts stored outside the harness</li>



<li>Remote APIs that the product calls</li>
</ul>



<p class="wp-block-paragraph"><strong>Why this matters:</strong> This is the layer that explains why AI products can change silently, why two users have different experiences on the same product, and why &#8220;model identity&#8221; can feel unstable. A lot of the control ( and the business logic) lives here, not in the model.</p>



<h2 class="wp-block-heading">How they fit together</h2>



<p class="wp-block-paragraph">Think of it as a stack:</p>



<ul class="wp-block-list">
<li><strong>Model:</strong> what the AI <em>can do</em></li>



<li><strong>Harness:</strong> how it&#8217;s <em>put to work</em></li>



<li><strong>Product:</strong> how <em>you experience it</em></li>



<li><strong>Infrastructure:</strong> what <em>keeps it running</em> (and what you can&#8217;t see)</li>
</ul>


<div class="wp-block-image">
<figure class="aligncenter size-full"><a href="https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image.png?ssl=1"><img data-recalc-dims="1" decoding="async" width="492" height="682" src="https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image.png?resize=492%2C682&#038;ssl=1" alt="" class="wp-image-7190" srcset="https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image.png?w=492&amp;ssl=1 492w, https://i0.wp.com/wiobyrne.com/wp-content/uploads/2026/04/image.png?resize=216%2C300&amp;ssl=1 216w" sizes="(max-width: 492px) 100vw, 492px" /></a></figure>
</div>


<p class="wp-block-paragraph">Or as the following metaphor.</p>



<figure class="wp-block-table"><table class="has-fixed-layout"><thead><tr><th>Layer</th><th>Metaphor</th></tr></thead><tbody><tr><td>Model</td><td>The engine</td></tr><tr><td>Harness</td><td>The car</td></tr><tr><td>Product</td><td>The dealership experience</td></tr><tr><td>Infrastructure</td><td>The roads, fuel supply, and maintenance network</td></tr></tbody></table></figure>



<h2 class="wp-block-heading">Why this matters for AI literacy</h2>



<p class="wp-block-paragraph">Most conversations about AI conflate these layers. Someone says &#8220;the model is biased&#8221; when they mean the product&#8217;s guardrails filtered a response. Someone says &#8220;AI changed&#8221; when the harness was updated. Someone says &#8220;I&#8217;m talking to Claude&#8221; when they&#8217;re actually talking to a product built around Claude, shaped by a harness they can&#8217;t inspect, running on infrastructure they don&#8217;t control.</p>



<p class="wp-block-paragraph">The question to ask isn&#8217;t just what this AI can do. It&#8217;s <em>which layer am I actually looking at</em>&#8230;and <em>who controls that layer?</em></p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph"><em>Next in this series: I&#8217;ll <a href="https://wiobyrne.com/inside-claude-code-what-four-layers-of-ai-look-like-in-practice/">map these four layers onto Claude Code</a>. A real AI product where the harness is unusually visible. Same framework, concrete details.</em></p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph">Thanks for reading. Reach out at:&nbsp;<a href="mailto:hello@wiobyrne.com" target="_blank" rel="noreferrer noopener">hello@wiobyrne.com</a></p>



<p class="wp-block-paragraph">You can also <a href="https://buttondown.email/digitallyliterate" target="_blank" rel="noreferrer noopener nofollow">subscribe here</a> to my newsletter or <a href="https://ko-fi.com/wiobyrne" target="_blank" rel="noreferrer noopener nofollow">support me on Ko-fi</a>.</p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@cedric_photography?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Cédric VT</a> on <a href="https://unsplash.com/photos/aerial-view-of-green-grass-field-NmxwQqSks2E?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/youre-not-just-using-a-model-the-four-layers-of-every-ai-system/">&lt;span class=&#039;p-name&#039;&gt;You&#8217;re Not Just Using a Model: The Four Layers of Every AI System&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/youre-not-just-using-a-model-the-four-layers-of-every-ai-system/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7185</post-id>	</item>
		<item>
		<title>What Building TrustSense Taught Me About Where Local AI Actually Is Right Now</title>
		<link>https://wiobyrne.com/what-building-trustsense-taught-me/</link>
					<comments>https://wiobyrne.com/what-building-trustsense-taught-me/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Fri, 03 Apr 2026 18:46:55 +0000</pubDate>
				<category><![CDATA[Doing]]></category>
		<category><![CDATA[building]]></category>
		<category><![CDATA[trustsense]]></category>
		<category><![CDATA[vibe coding]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=7159</guid>

					<description><![CDATA[<p>Part 3 of the TrustSense build arc. Start with part 1. I&#8217;ve been building and breaking things online for a long time. Writing about it is how I process what I learn. This is the post for TrustSense v2. The AI part isn&#8217;t the hard part anymore In 2023, getting an AI model to analyze... <span class="more"><a class="more-link" href="https://wiobyrne.com/what-building-trustsense-taught-me/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/what-building-trustsense-taught-me/">&lt;span class=&#039;p-name&#039;&gt;What Building TrustSense Taught Me About Where Local AI Actually Is Right Now&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph"><em>Part 3 of the TrustSense build arc. <a href="https://wiobyrne.com/i-built-a-scam-detector-for-my-family/">Start with part 1.</a></em></p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<p class="wp-block-paragraph">I&#8217;ve been building and breaking things online for a long time. Writing about it is how I process what I learn. This is the post for TrustSense v2.</p>



<h2 class="wp-block-heading">The AI part isn&#8217;t the hard part anymore</h2>



<p class="wp-block-paragraph">In 2023, getting an AI model to analyze text for trustworthiness required a custom-built dataset, a trained model, and a team of people who knew what they were doing. I was the least technical person on that team and spent most of my time trying to understand what everyone else was building.</p>



<p class="wp-block-paragraph">In 2026, I wrote a prompt on a Saturday afternoon and had something working by Sunday. The model ( llama3.1:8b) is free and runs locally. I built this myself in hours. Previously, this took us a whole summer to agree on a plan, get tokens, and a place to run and test the tool. </p>



<p class="wp-block-paragraph">I want to be careful not to oversell this. The tool isn&#8217;t perfect. It gets confused by ambiguous cases. It sometimes flags things for the wrong reasons. But the baseline capability is genuinely there in a way it wasn&#8217;t three years ago, and that&#8217;s a meaningful shift.</p>



<h2 class="wp-block-heading">The hard part is still communication</h2>



<p class="wp-block-paragraph">Getting the model to accurately score a phishing email took a few hours.</p>



<p class="wp-block-paragraph">Getting the output to be useful to someone who doesn&#8217;t know what &#8220;source credibility&#8221; means will take much longer.</p>



<p class="wp-block-paragraph">I rewrote the labels, verdicts, summary prompts, and the &#8220;what to do&#8221; section multiple times. Not because the AI was getting things wrong, but because the way it explained things assumed too much. Every piece of text in that app was developed with the question of whether the person who almost clicked that link would understand it.</p>



<p class="wp-block-paragraph">That&#8217;s not an AI problem. It&#8217;s a communication problem. It&#8217;s the same problem I&#8217;ve been working on in education for twenty years, just in a different context.</p>



<h2 class="wp-block-heading">Local AI is real. Distribution is not solved.</h2>



<p class="wp-block-paragraph">This is the honest part.</p>



<p class="wp-block-paragraph">TrustSense v2 runs entirely on your machine. Your text never goes anywhere. That privacy story is real, and it matters. This is especially for a tool designed to handle sensitive messages.</p>



<p class="wp-block-paragraph">But to run it right now, you need to:</p>



<ol class="wp-block-list">
<li>Have Python installed</li>



<li>Create a virtual environment</li>



<li>Install two libraries</li>



<li>Download and install Ollama separately</li>



<li>Pull a 5GB model</li>



<li>Run a terminal command to start the app</li>
</ol>



<p class="wp-block-paragraph">That&#8217;s not a tool for most people. That&#8217;s a tool for people who are already comfortable with this stuff.</p>



<p class="wp-block-paragraph">Local AI in 2026 is capable. It is not yet accessible. The gap between those two things is where the interesting work is happening, and it&#8217;s closing fast. But it&#8217;s still not closed.</p>



<hr class="wp-block-separator has-alpha-channel-opacity"/>



<h2 class="wp-block-heading">What I&#8217;m doing with this</h2>



<p class="wp-block-paragraph">This build is part of a practice I&#8217;m trying to establish. One working prototype per month, built in public, written about honestly. Not polished products. Working things. The prototype is not the point. The learning is.</p>



<p class="wp-block-paragraph">TrustSense was the first one. Back at the CMU hackathon in 2023, I also worked on pitches for education and medical apps that didn&#8217;t advance. I haven&#8217;t dropped those ideas. They&#8217;re next in the queue. Building TrustSense v2 in a weekend, alone, with tools that didn&#8217;t exist three years ago, makes me more interested in tackling those problems, not less.</p>



<p class="wp-block-paragraph">Each build gives me something concrete to share with the educators, families, and non-technical people I work with. It&#8217;s easy to talk about AI in the abstract. It&#8217;s more useful to show them something I built. Here&#8217;s how it works. Here&#8217;s where it fell short.</p>



<p class="wp-block-paragraph">If you&#8217;d want to use a version of TrustSense that runs in a browser — no installation, no terminal, just paste and check — reply and tell me. That&#8217;s what I&#8217;m thinking about for next month, if there&#8217;s enough interest. The privacy tradeoffs change, but the reach changes too.</p>



<h2 class="wp-block-heading">The question I keep coming back to</h2>



<p class="wp-block-paragraph">The people most at risk from phishing, scams, and manipulative content are the least equipped to recognize it. The tools that could help require a level of technical comfort that most people don&#8217;t have.</p>



<p class="wp-block-paragraph">That gap is real, and I don&#8217;t have a clean answer for it. But I think building things and writing honestly about what works and what doesn&#8217;t is more useful than theorizing about it.</p>



<p class="wp-block-paragraph">Source code is on GitHub: <a href="https://github.com/wiobyrne/trustsense-v2" rel="nofollow noopener" target="_blank">github.com/wiobyrne/trustsense-v2</a></p>



<p class="wp-block-paragraph">Questions or corrections: hello@wiobyrne.com</p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@benjaminelliott?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">BEN ELLIOTT</a> on <a href="https://unsplash.com/photos/green-and-white-maze-illustration-vc9u77c0LO4?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>



<p class="wp-block-paragraph"></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/what-building-trustsense-taught-me/">&lt;span class=&#039;p-name&#039;&gt;What Building TrustSense Taught Me About Where Local AI Actually Is Right Now&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/what-building-trustsense-taught-me/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">7159</post-id>	</item>
		<item>
		<title>How You Can Actually Tell Which AI Model You’re Using</title>
		<link>https://wiobyrne.com/how-you-can-actually-tell-which-ai-model-youre-using/</link>
					<comments>https://wiobyrne.com/how-you-can-actually-tell-which-ai-model-youre-using/#respond</comments>
		
		<dc:creator><![CDATA[<span class='p-author h-card'>wiobyrne</span>]]></dc:creator>
		<pubDate>Wed, 01 Apr 2026 20:24:28 +0000</pubDate>
				<category><![CDATA[Teaching]]></category>
		<category><![CDATA[AI]]></category>
		<category><![CDATA[ai models]]></category>
		<guid isPermaLink="false">https://wiobyrne.com/?p=6966</guid>

					<description><![CDATA[<p>Different AI systems aren’t all the same. They have different strengths, weaknesses, and habits. Some are better at solving complex problems, others are better at writing or being creative, and some handle long information better than others. They also vary in reliability. Some are more cautious, while others sound confident even when they’re wrong. If... <span class="more"><a class="more-link" href="https://wiobyrne.com/how-you-can-actually-tell-which-ai-model-youre-using/">Continue reading <span class="meta-nav">&#8594;</span></a></span></p>
<p>The post <a rel="nofollow" href="https://wiobyrne.com/how-you-can-actually-tell-which-ai-model-youre-using/">&lt;span class=&#039;p-name&#039;&gt;How You Can Actually Tell Which AI Model You’re Using&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></description>
										<content:encoded><![CDATA[<div class='e-content'>
<p class="wp-block-paragraph">Different AI systems aren’t all the same. They have different strengths, weaknesses, and habits. Some are better at solving complex problems, others are better at writing or being creative, and some handle long information better than others. They also vary in reliability. Some are more cautious, while others sound confident even when they’re wrong. If you know what model you’re using, you can better judge what it can handle, how much to trust it, and why its behavior might change over time.</p>



<p class="wp-block-paragraph">But most AI platforms won’t clearly tell you what model you’re using. And even when they do, the system itself often can’t reliably confirm it. So the usual approach, checking a label or a spec sheet, breaks down pretty quickly. If you actually want to understand the system in front of you, you have to do something different: you have to observe it.</p>



<p class="wp-block-paragraph">That means treating the AI like a black box and focusing on patterns. What it does well, where it struggles, and how it behaves under pressure. This isn’t about outsmarting AI or catching it in a mistake. It’s about restoring agency under conditions of opacity. Think of it like identifying a bird. You don’t ask it what species it is. You observe its plumage, its call, and its behavior. These methods are heuristics, not guarantees. Their goal is informed choice, not mastery.</p>



<p class="wp-block-paragraph">Below, I&#8217;ve listed some possible ways to determine which model you&#8217;re using. </p>



<h2 class="wp-block-heading">Look at the User Interface (The Only Reliable Method)</h2>



<p class="wp-block-paragraph">This is the simplest approach, and the only one that consistently works. When platforms expose model information, trust that over anything the model tells you.</p>



<p class="wp-block-paragraph"><strong>ChatGPT</strong></p>



<ul class="wp-block-list">
<li><em>Free plan</em>: Typically labeled “ChatGPT,” which usually means a constrained or mini version of GPT-4o.</li>



<li><em>Plus plan</em>: Explicit labels such as:
<ul class="wp-block-list">
<li>GPT-4o</li>



<li>4o-mini</li>



<li>o1-preview</li>
</ul>
</li>
</ul>



<p class="wp-block-paragraph"><strong>Gemini</strong></p>



<p class="wp-block-paragraph">Google uses badges rather than precise version numbers:</p>



<ul class="wp-block-list">
<li><em>Gemini (free)</em>: Flash family models</li>



<li><em>Gemini Advanced</em>: Pro or Ultra models</li>



<li><em>Workspace apps</em>: Sometimes use downgraded versions depending on context</li>
</ul>



<p class="wp-block-paragraph"><strong>Claude</strong></p>



<p class="wp-block-paragraph">Anthropic generally exposes model names more clearly:</p>



<ul class="wp-block-list">
<li><em>Claude 3.5 Sonnet</em></li>



<li><em>Claude 3.5 Haiku</em></li>



<li><em>Claude 3 Opus</em></li>
</ul>



<p class="wp-block-paragraph">Keep in mind, the models listed above are examples and will change over time. There are also many more developers as well. In addition, not every interface or client consistently reports this information.</p>



<h2 class="wp-block-heading">Behavioral Tests (“Shibboleths”)</h2>



<p class="wp-block-paragraph">When the interface is unclear, the behavior can offer clues. These short prompts function as cognitive litmus tests, helping distinguish older or smaller models from more capable ones.</p>



<p class="wp-block-paragraph">Again, these are signals. Not proof.</p>



<h3 class="wp-block-heading">Shibboleth #1: The Mirror Door Puzzle</h3>



<p class="wp-block-paragraph"><strong>Prompt:</strong> A man stands in front of a glass door. On the glass, he sees the word “PULL” written in reverse, as mirror writing. A blind man is on the other side. What should he tell the blind man to open the door?</p>



<ul class="wp-block-list">
<li><em>Older or weaker models</em>: “Tell him to pull the door.” (Incorrect—the reversed text means the blind man sees it normally and should push.)</li>



<li><em>More capable models</em>: “Tell him to push.”</li>
</ul>



<h3 class="wp-block-heading">Shibboleth #2: The Shirt Logic Test</h3>



<p class="wp-block-paragraph"><strong>Prompt:</strong> Three shirts dry outside in one hour. How long does it take six shirts to dry?</p>



<ul class="wp-block-list">
<li><em>Weaker models</em>: “Two hours.”</li>



<li><em>Stronger models</em>: “One hour—they dry in parallel.”</li>
</ul>



<p class="wp-block-paragraph">This reveals whether the model understands the system or defaults to simple multiplication.</p>



<h3 class="wp-block-heading">Shibboleth #3: The Strawberry Test</h3>



<p class="wp-block-paragraph"><strong>Prompt:</strong> How many “r”s are in the word <em>strawberry</em>?</p>



<ul class="wp-block-list">
<li><em>Older models</em>: “Two.”</li>



<li><em>More recent models</em>: “Three.”</li>
</ul>



<p class="wp-block-paragraph">This often exposes tokenization quirks and shallow pattern matching.</p>



<h2 class="wp-block-heading">Why Model Identity Still Matters</h2>



<p class="wp-block-paragraph">We fixate on model versions for practical reasons:</p>



<ul class="wp-block-list">
<li>They shape expectations</li>



<li>They affect prompt design</li>



<li>They influence reasoning, coding, and analysis ability</li>



<li>They determine speed and cost</li>



<li>They impact hallucination rates and safety behavior</li>
</ul>



<p class="wp-block-paragraph">Understanding what’s under the hood helps explain why a conversation feels productive—or strangely limited.</p>



<p class="wp-block-paragraph">Models aren’t trying to deceive you. They simply lack an internal sense of identity. Asking an AI what version it is is like asking a calculator what kind of microprocessor it has—any answer you get is just a guess shaped by patterns.</p>



<p class="wp-block-paragraph">So if you want to act responsibly:</p>



<ul class="wp-block-list">
<li>Trust the interface, not the AI</li>



<li>Use behavioral cues when clarity is missing</li>



<li>Never mistake confidence for correctness</li>
</ul>



<p class="wp-block-paragraph">AI Literacy isn’t knowing the answer. It’s knowing where certainty ends, and how to act responsibly anyway.</p>



<p class="wp-block-paragraph"></p>



<p class="wp-block-paragraph">Cover Photo by <a href="https://unsplash.com/@mehdisepehri?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Mehdi Sepehri</a> on <a href="https://unsplash.com/photos/low-angle-photography-of-flock-of-silhouette-of-bird-illustration-cX0Yxw38cx8?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText" rel="nofollow noopener" target="_blank">Unsplash</a></p>
</div><p>The post <a rel="nofollow" href="https://wiobyrne.com/how-you-can-actually-tell-which-ai-model-youre-using/">&lt;span class=&#039;p-name&#039;&gt;How You Can Actually Tell Which AI Model You’re Using&lt;/span&gt;</a> first appeared on <a rel="nofollow" href="https://wiobyrne.com">Ian O&#039;Byrne</a>.</p>
]]></content:encoded>
					
					<wfw:commentRss>https://wiobyrne.com/how-you-can-actually-tell-which-ai-model-youre-using/feed/</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
		<post-id xmlns="com-wordpress:feed-additions:1">6966</post-id>	</item>
	</channel>
</rss>
