<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Jamie Telin on Medium]]></title>
        <description><![CDATA[Stories by Jamie Telin on Medium]]></description>
        <link>https://medium.com/@lejahmie?source=rss-7c66bba90f3a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*-xV2JRFOyuLMynX5UklTLw.jpeg</url>
            <title>Stories by Jamie Telin on Medium</title>
            <link>https://medium.com/@lejahmie?source=rss-7c66bba90f3a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 03:32:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@lejahmie/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What Am I Accelerating?]]></title>
            <link>https://medium.com/it-chronicles/what-am-i-accelerating-2fe9e792e6d0?source=rss-7c66bba90f3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/2fe9e792e6d0</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[accelerationism]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[fatherhood]]></category>
            <category><![CDATA[future]]></category>
            <dc:creator><![CDATA[Jamie Telin]]></dc:creator>
            <pubDate>Wed, 18 Mar 2026 11:57:04 GMT</pubDate>
            <atom:updated>2026-03-18T11:57:04.148Z</atom:updated>
            <content:encoded><![CDATA[<h4>On the edge of intention</h4><h4>Token-Fueled Polymath Syndrome and the Dilemma of an Unwilling Accelerationist</h4><figure><img alt="A comic book illustration of a man blinded by the lights" src="https://cdn-images-1.medium.com/max/1024/1*YWWVQS9lNd1SN06hH7p_2A.jpeg" /><figcaption>The blinding lights. Image by author (AI generated, GPT-5, edited with Affinity)</figcaption></figure><p>Not Attention Deficit Hyperactivity Disorder. The newer kind. AI-driven Hyperactivity Disorder.</p><p>Or maybe something more honest: Token-Fueled Polymath Syndrome. TFPS. A diagnosis invented half in jest, half in recognition.</p><p>I currently have six projects underway. I am actively contributing to a seventh. That sentence alone sounds excessive. Suspiciously enthusiastic. Slightly unbalanced.</p><p>But none of them are random.</p><p>They are ideas I’ve carried for years. Some were written down and shelved. Others resurfaced every few months, like unfinished conversations with myself. They were postponed not because they lacked meaning, but because they required something I rarely had in surplus.</p><p>Time, or energy. Or the rare combined resource: time-energy. The real currency of a father with a six-year-old daughter.</p><p>For a long while, building required a significant withdrawal from that account. A clear weekend. Sustained focus. The quiet ceremony of scaffolding, configuring, preparing. The kind of cognitive immersion. Not something that coexists easily with bedtime routines, lost mittens, or questions about dinosaurs.</p><p>Some ideas were simply too expensive.</p><p>Others were ahead of the tools available to carry them.</p><p>Now something has shifted.</p><p>The distance between thought and prototype has narrowed. A short window can be enough. Ten minutes can move something forward. A late evening can turn a sketch into a working application. The activation cost has dropped low enough that “someday” quietly becomes “tonight.”</p><p>I sometimes look at my laptop and count the open terminals. Twelve. Agent loops running. Builds compiling. Editors layered behind one another. It is not elegant. It is a small constellation of unfinished futures. My spaceship control center. Each window a project once waited patiently for a version of me with more margin. All ready to be launched. Now.</p><p>There is Reframe, a long-held idea to turn a city into a photo quest. To reward how we see, not only what we capture. For years it lived in theory. Now it runs.</p><p>There is SISU, shaped by the Finnish idea of steady resolve. An agentic framework, built for both humans and machines. Explicit tools. Predictable plans. Systems that behave calmly under pressure. Perhaps not coincidentally, the same qualities I try to practice at home.</p><p>Dragonbane Unbound began with a simple remark at work. A colleague described how D&amp;D Beyond helps his other friend group manage characters, experience and items, smoothing out the administrative friction of play. We play “Dragonbane,” “Drakar och Demoner,” the Swedish cult classic roleplaying game. We did not have that layer.</p><p>Before I consciously decided anything, the old internal voice appeared: you could build it.</p><p>What followed was not an act of disruption, but of care. No one had built it. I could. A local-first platform to support the table, not dominate it. To handle the mechanics quietly so attention can remain on the story and the people around it.</p><p>Homunculus had already been quietly started. Now it waits in the background terminal. An invisible layer meant to surface rules and context without demanding attention. Voice-activated. Transcribed. Always listening to the home games around the table, so low barrier it is almost invisible, transparent.</p><p>Linopress reduces the ceremony of starting a WordPress site. Prompt. Build. Verify. Ship. It is, in many ways, the thing we tried to build at my second real job, only now the world has finally caught up to the ambition. And it took an evening. Eight years ago, we fought hard, for months, living on a dream, until dreams could not feed us anymore and we finally gave up.</p><p>And then there is ProjectSpecs. An attempt to bring coherence to scattered requirements in an AI-saturated workflow. Code was never the true bottleneck.</p><p>The bottleneck was always the organization, holding the codebase in careful, almost parental arms, unsure how to let it grow up. Now the code moves faster than the structures around it. Agents collaborate. Systems evolve. Meanwhile managers are still asked to “ask Rovo” for the lost EPIC or rely on Copilot in Outlook. They deserve better tools too.</p><p>Each project answers a frustration I have carried for years. Each one feels justified.</p><p>While I sit in front of glowing terminals, the news scrolls by.</p><p>Reports surface that U.S. Defense Secretary Pete Hegseth has issued an ultimatum to Anthropic: grant the military unrestricted access to the Claude model or face invocation of the Cold War-era Defense Production Act. Remove restrictions. Enable broader surveillance capabilities. Support autonomous systems. AI capable of killing without human interference.</p><p>Anyone raised on science fiction feels the tremor underneath.</p><p>We have seen this play out in films. In games. In “Metal Gear Solid 4” and “Metal Gear Rising: Revengeance.” In stories that begin years after the decision was quietly signed.</p><figure><img alt="A comic book illustration if the author dressed as Solid Snake" src="https://cdn-images-1.medium.com/max/1024/1*TMigDejsAavGAyG4JZLxvw.jpeg" /><figcaption>Author as Solid Snake. Image by author (AI generated, GPT-5, edited with Affinity)</figcaption></figure><p>It is hyperbolic and mundane at the same time. A strange déjà vu in reverse, watching fiction try to catch up with reality instead of the other way around.</p><p>As a builder, I feel exhilarated.</p><p>As a father, I feel something else.</p><p>I already carry the quiet anxiety of raising a child in a world of rising temperatures and rising political division. Now I also raise her in a world where the most powerful cognitive systems humanity has built are being negotiated in the language of defense production acts and access mandates.</p><p>Richard Wright wrote in his 1945 memoir “Black Boy,” “I would hurl words into this darkness and wait for an echo and if an echo sounded, no matter how faintly, I would send other words to tell, to march, to fight, to create a sense of hunger for life that gnaws in us all.”</p><p>Some days building feels like that. Not escapism. Not acceleration for its own sake. But a refusal to disengage.</p><p>Dylan Thomas wrote in “Do Not Go Gentle Into That Good Night,” published in the 1952 collection “In Country Sleep, and Other Poems,” “Do not go gentle into that good night… Rage, rage against the dying of the light.”</p><p>The line is often read as defiance against mortality. Lately it feels like something broader: a refusal to surrender agency in a moment of enormous technological leverage.</p><p>Fatherhood sharpens this realization. My daughter does not experience the world as a backlog of optimizable systems. She experiences it as presence and attention. Love.</p><p>As whether I am truly there when she is explaining something important, six-year-old-important things, or only half there, mentally allocating tokens elsewhere.</p><p>The question is no longer, “Can I build this?”</p><p>It may be something heavier than that.</p><p>It may be: What am I accelerating?</p><p>Because for the first time, building does not feel isolated. It feels cumulative. Every model trained, every agent deployed, every layer of abstraction added, it all compounds. Not only for me, but for everyone. I am the air bubble in thick, sugary molasses, slowly moving as the world rapidly moves around me.</p><p>While I sit in front of glowing terminals, governments negotiate access to frontier models. Defense departments invoke production acts written in another century. Companies race. Nations posture.</p><p>A blinding light. An explosion.</p><p>AI radiation showers over us as we peek up from behind the barricades.</p><p>As a builder, I cannot pretend I am outside that arc.</p><p>There is a quieter fear beneath the productivity. Not, whether I am spending enough time with my daughter. But whether I am helping construct a world she will have to navigate.</p><p>Will we look back on this period as the moment we accelerated wisely, or blindly? Is this the Gutenberg press or the Manhattan Project?</p><p>Will there even be something recognizably human to experience what we are building? Or will we outsource agency so thoroughly that we become spectators in our own civilization?</p><p>We crawl the latent space of our own creations, compressing our lives into fewer real moments, mistaking simulation for presence.</p><p>Science fiction has rehearsed these endings for decades. Runaway systems. Power concentrated in small rooms. Intelligence optimized for objectives we barely understand. It is easy to dismiss those stories as melodrama, until policy memos begin to sound like their prologues.</p><p>And yet, perhaps I am just full of myself and these new, god-like powers. There is a seductive numbness in Token-Fueled Polymath Syndrome, a high that feels too good to let go. While Rome burns, I am here, meticulously fiddling. Do I even want to make real change? Am I even capable of it?</p><p>Paralysis, however, is not an option. To build nothing is also a choice. To disengage is to leave acceleration entirely in the hands of the most power-hungry or least reflective actors.</p><figure><img alt="An comic book illustration of a man free falling in in a tunnel with his laptop towards an abstract light source" src="https://cdn-images-1.medium.com/max/1024/1*eveilkLW3wZ9mppQUAEgrQ.jpeg" /><figcaption>The feeling of turbulently free falling in the AI landscape. Image by author (AI generated, GPT-5, edited with Affinity)</figcaption></figure><p>Perhaps the deeper question is not whether to build, but how.</p><p>How to build with constraint. How to build with memory. How to build in a way that preserves room for human experience rather than compressing it.</p><p>Fatherhood sharpens this realization in a different way.</p><p>My daughter does not need a singularity. She needs adults who can make decisions without surrendering responsibility to systems they do not fully understand.</p><p>Time-energy is still finite, but now it carries weight beyond productivity. It carries direction.</p><p>Perhaps Token-Fueled Polymath Syndrome is our time’s disorder. Perhaps it is what happens when latent ambition meets tools that finally keep up, at the very moment the stakes of those tools become geopolitical.</p><p>The only truly scarce resource left is intention.</p><p>And screen real estate on a 14-inch MacBook Pro.</p><p>If you are an unwilling accelerationist looking for a place to hurl code into the darkness, you’re welcome to join me on GitHub, where we are mostly open source and welcome anyone with the Token-Fueled Polymath Syndrome, <a href="https://www.linkedin.com/redir/redirect?url=https%3A%2F%2Fgithub%2Ecom%2Ffinger-gun&amp;urlhash=ulQT&amp;trk=article-ssr-frontend-pulse_little-text-block">https://github.com/finger-gun</a></p><blockquote>Support writers. All our non-profit’s offerings <a href="https://forms.gardenofneuro.org/nx/portal/neonevents/events?path=%2Fportal%2Fevents">here</a>.</blockquote><blockquote><a href="https://medium.com/it-chronicles/newsletter">Click here to subscribe to the IT Chronicles newsletter.</a></blockquote><blockquote><a href="https://forms.gardenofneuro.org/nx/portal/neonevents/events?path=%2Fportal%2Fevents">Click for Wordsmith, Mystery Writing, Write Like Stephen King, more</a></blockquote><figure><a href="https://medium.com/it-chronicles"><img alt="Computer circuit pathways" src="https://cdn-images-1.medium.com/max/1024/1*bX_AIumALEvtJr-mZIBeMw.png" /></a><figcaption>By the EIC <a href="https://medium.com/u/cc0821aee9c1">Susan Brearley</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2fe9e792e6d0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/it-chronicles/what-am-i-accelerating-2fe9e792e6d0">What Am I Accelerating?</a> was originally published in <a href="https://medium.com/it-chronicles">IT Chronicles</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is Your “Safe” Choice Burning Your Budget?]]></title>
            <link>https://medium.com/it-chronicles/is-your-safe-choice-burning-your-budget-1cfddf8782e4?source=rss-7c66bba90f3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1cfddf8782e4</guid>
            <category><![CDATA[spec-driven-development]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Jamie Telin]]></dc:creator>
            <pubDate>Wed, 18 Mar 2026 11:54:37 GMT</pubDate>
            <atom:updated>2026-03-18T11:54:37.176Z</atom:updated>
            <content:encoded><![CDATA[<h4>Spec Driven Development</h4><h4>When governance turns into overhead, innovation slows and your budget bleeds</h4><figure><img alt="Man doing science with a robot" src="https://cdn-images-1.medium.com/max/1024/0*aBWBe6iB6wTELPby" /><figcaption>Image by author (AI generated, GPT-5, edited with Affinity)</figcaption></figure><p>In the world of AI-assisted engineering, the “enterprise” choice is rarely the efficient one — and often the most expensive.</p><p>I recently ran a head-to-head comparison between <a href="https://openspec.dev/">OpenSpec</a> and <a href="https://github.com/github/spec-kit">Spec-Kit</a>, the Microsoft-backed heavyweight, to evaluate how they handle Spec Driven Development (SDD) using GPT-5.2 and GPT-5.2-Codex.</p><p>The results are a wake-up call for anyone managing AI costs.</p><h3>The Experiment</h3><p>Using a simple AI chat application in MVP state, I tasked both frameworks with the same real-world objective: implement full streaming support and session support for chats (list/open/delete/new).</p><p>The experiment was run twice to determine whether more detailed upfront planning would improve overall implementation efficiency.</p><p>Source code used for the test is available at <a href="https://github.com/knowit-flx/workshop-1-2026">GitHub</a>. It was originally created as part of a workshop on Spec Driven Development.</p><h3>The Results: More Planning Does Not Equal Better Implementation</h3><p>Conventional wisdom suggests that a more rigorous planning, such as the one used in Spec-Kit, should produce smoother implementations. I say it all the time my self, even before we started working with AI. Plan twice, implement once. The data tells a different story — more structure increased cost without improving outcomes.</p><p>In both tests, the leaner OpenSpec approach outperformed Spec-Kit by a wide margin. Even when Spec-Kit more than doubled token usage during planning, its implementation phase remained significantly less efficient.</p><p>In Test 1, total token usage came in at approximately 57,740 for OpenSpec versus 120,947 for Spec-Kit. That is a 109% increase.</p><p>In Test 2, planning alone consumed 38,117 tokens with OpenSpec compared to 96,298 with Spec-Kit, a 152% increase.</p><p>Implementation required 53,612 tokens for OpenSpec and 84,742 for Spec-Kit, while not in the hundreds, still a juicy 58% increase.</p><p>Total usage reached 91,729 tokens for OpenSpec and 181,040 for Spec-Kit, nearly double, at a 97 percent increase.</p><p>Across both runs, OpenSpec achieved a higher success rate while requiring roughly 20 percent fewer assistant turns and 25 percent fewer tool calls.</p><h3>The Context Trap: Validated by Department of Computer Science, ETH Zurich Research</h3><p>Why did the “enterprise” framework, with its verbose rules and rigid structure, perform worse?</p><p>This outcome aligns closely with a 2026 study from researchers at ETH Zurich titled “Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?” &lt;<a href="https://arxiv.org/abs/2602.11988">https://arxiv.org/abs/2602.11988</a>&gt;</p><p>The idea behind these files is intuitive. If we give agents a structured overview of the repository, prescribed tooling, conventions, and workflow expectations, they should navigate the codebase more effectively and solve tasks more reliably.</p><p>The researchers evaluated this assumption across two complementary benchmarks. SWE-Bench Lite covered well-known repositories without developer-written context files. AGENTBENCH, which the authors constructed specifically for this study, included 138 real-world issues drawn from 12 repositories that actively used developer-provided context files.</p><p>They tested multiple coding agents and models, including GPT-5.2, GPT-5.1 Mini, Sonnet-4.5, and Qwen3–30B, under three conditions: no context file, an LLM-generated context file, and a developer-written context file.</p><p>The results are striking. LLM-generated context files tended to slightly reduce task success rates while increasing inference cost by over 20 percent.</p><p>Developer-written context files performed somewhat better, but the gains were marginal, on average around four percentage points, and still came with increased execution cost.</p><p>In other words, adding more structured repository guidance did not meaningfully improve performance, and it consistently increased resource consumption.</p><p>What makes the study particularly relevant is its behavioral analysis. The agents did not ignore the context files. They followed them. When additional tooling was mentioned, agents used it. When testing guidance was provided, agents ran more tests.</p><p>When procedural steps were described, agents complied. Trace analysis showed increased exploration, more file traversal, more tool calls, and significantly higher reasoning token usage when context files were present. The models were demonstrably respecting governance.</p><p>Yet this increased diligence did not translate into faster discovery of relevant code. The study measured how many steps agents required before interacting with files that were actually modified in the ground-truth solution. Context files did not meaningfully reduce this number. The agents became more thorough, but not more effective.</p><p>The authors conclude that unnecessary requirements from context files make tasks harder. That sentence is doing a lot of work. The issue is not that guidance is inherently harmful. It is that surplus structure redistributes cognitive effort.</p><p>When additional rules are introduced, the model allocates reasoning capacity toward satisfying those rules. More constraints mean more branching, more verification, more internal checking. The system expends energy complying with the framework rather than directly solving the problem.</p><p>Interestingly, the study also found that when existing repository documentation was removed, LLM-generated context files could improve performance.</p><p>This suggests an important nuance: context files are helpful when they replace missing information, but redundant when they duplicate what the model can already infer or discover. In mature, well-documented repositories, extra scaffolding may simply become noise.</p><p>Taken together, the findings point toward a broader shift. Early coding agents likely benefited from heavy scaffolding because their internal repository understanding was weaker. Frontier models already exhibit strong contextual prioritization and tool orchestration.</p><h3>Don’t Let Process Tax Your Progress</h3><p>Spec-Kit is often viewed as the safe enterprise option. In practice, its verbose Markdown-based constitution, and its many step process acted more as noise than signal. It forced the model into discovery loops and path-correction cycles.</p><p>Spec-Kit required autonomous bug-fixing rounds to correctly implement streaming status, while OpenSpec delivered a correct implementation on the first pass.</p><p>OpenSpec’s lean structure provided just enough guidance for the model to do what it does best: solve the problem.</p><p>If you are building for 2026, do not let governance become a tax on development speed.</p><p>OpenSpec consistently delivered a more complete, feature-rich result at roughly half the token cost. Spec-Kit, by contrast, spent more tokens reasoning about rules than executing the task itself.</p><blockquote>Support writers. All our non-profit’s offerings <a href="https://forms.gardenofneuro.org/nx/portal/neonevents/events?path=%2Fportal%2Fevents">here</a>.</blockquote><blockquote><a href="https://medium.com/it-chronicles/newsletter">Click here to subscribe to the IT Chronicles newsletter.</a></blockquote><blockquote><a href="https://forms.gardenofneuro.org/nx/portal/neonevents/events?path=%2Fportal%2Fevents">Click for Wordsmith, Mystery Writing, Write Like Stephen King, more</a></blockquote><figure><a href="https://medium.com/it-chronicles"><img alt="Computer circuit pathways" src="https://cdn-images-1.medium.com/max/1024/1*bX_AIumALEvtJr-mZIBeMw.png" /></a><figcaption>By the EIC <a href="https://medium.com/u/cc0821aee9c1">Susan Brearley</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1cfddf8782e4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/it-chronicles/is-your-safe-choice-burning-your-budget-1cfddf8782e4">Is Your “Safe” Choice Burning Your Budget?</a> was originally published in <a href="https://medium.com/it-chronicles">IT Chronicles</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Software Engineering Has Been Dying for Three Years]]></title>
            <link>https://medium.com/it-chronicles/software-engineering-has-been-dying-for-three-years-ef25913ecb70?source=rss-7c66bba90f3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/ef25913ecb70</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[spec-driven-development]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Jamie Telin]]></dc:creator>
            <pubDate>Sun, 01 Mar 2026 15:43:59 GMT</pubDate>
            <atom:updated>2026-03-01T19:25:01.382Z</atom:updated>
            <content:encoded><![CDATA[<h4>OBITUARY UPDATE</h4><h4>Software engineering isn’t dying only shedding its skin</h4><figure><img alt="A man running away from destruction, blazing fire and drones." src="https://cdn-images-1.medium.com/max/1024/0*fSKnMxyuyv_tdHLL" /><figcaption>Image by author (AI generated with ChatGPT)</figcaption></figure><p>If you have scrolled through the internet anytime since 2023, you know the drill. Every few months, a new thought leader announces the exact date of our professional funeral. In 2024, Nvidia’s Jensen Huang told us not to learn to code. By 2025, the consensus among the loudest voices was that software engineering was effectively “done.” We have been “six months away” from total replacement for about three years now. And yet, it is 2026, and we are still here.</p><figure><img alt="Screenshot of article on Vulcanpost with title “Don’t learn to code: Nvidia´s founder Jensen Huang advises a different career path”" src="https://cdn-images-1.medium.com/max/1024/1*gIUgdvvRnXtyl7hAa8ke6Q.png" /><figcaption>Screenshot of article on Vulcanpost: “Don’t learn to code: Nvidia´s founder Jensen Huang advises a different career path”</figcaption></figure><figure><img alt="Screenshot of article on Business Insider with title “Anthropic´s CEO says that in 3 to 6 months, AI will be writing 90% of the code software developers were in charge of”" src="https://cdn-images-1.medium.com/max/1024/1*XqgP0YXDO9mrZOaF0kMq3g.png" /><figcaption>Screenshot of article on Business Insider: “Anthropic´s CEO predicting AI will write 90% of the code”</figcaption></figure><figure><img alt="Screenshot of Adam Wolffs post on X.com, where he says he has seen the future, and in six months software engineering is done." src="https://cdn-images-1.medium.com/max/1024/1*sOI7SJupFuJHXRSpcDfBqA.png" /><figcaption><a href="https://x.com/dmwlff">Adam Wolff post on X.com</a>, where he says he has seen the future, and in six months software engineering is done.</figcaption></figure><figure><img alt="Screenshot of Françios Chollet on X.com where he says “Software engineering has been within 6 months of being dead continually since early 2023”" src="https://cdn-images-1.medium.com/max/1024/1*vVRgKOzlI5aEYkdAKv3Ywg.png" /><figcaption><a href="https://x.com/fchollet">Françios Chollet on X.com</a> says: “Software engineering has been within 6 months of being dead continually since early 2023&quot;</figcaption></figure><blockquote>Software engineering has been within 6 months of being dead continually since early 2023.</blockquote><p>Ignoring the hype doesn’t make the anxiety go away. Even for me, as someone who spends my days building AI software, utilizing all the tools and leading learning initiatives, it is easy to feel the pressure. The tools move so fast, “imposter syndrome” feels like the factory default setting for our industry. But this gap, between the chaotic early adopters and the mainstream reality, is actually the perfect time to get ahead. To understand where we are going, we only have to look at where we came from.</p><p>We aren’t witnessing the death of the developer. We are witnessing the birth of a new abstraction layer. In the beginning, we were effectively hardware engineers. We didn’t “code” in any modern sense, but we physically wired circuits and punched holes in cards.</p><p>When assembly arrived, it liberated us from the physical wires. Then came the high-level revolution with languages like C. This gave us superpowers. It abstracted away the manual memory management and hardware specifics, allowing us to build complex, portable systems faster than ever before.</p><p>AI is simply doing it again. We are still needed. The lever is simply more powerful now.</p><p>However, wielding the lever takes discipline. For a while, “vibe coding”, the act of prompting an LLM and hoping for the best, felt like magic. But the novelty is wearing off. We are realizing, while AI is powerful, throwing requirements at it like spaghetti against a wall isn’t engineering. It’s a mess. Most of us know the worst part of any project isn’t writing the code. It’s when requirements shift halfway through, and you have to tear down what you built. Vibe coding exacerbates this. AI will happily build the wrong thing, very efficiently.</p><p>Code is a poor medium for negotiating requirements.</p><p>If you find yourself debugging your requirements inside the generated code, you have already lost.</p><p>This is where we fit in. AI sees the trees. Only humans see the forest.</p><p>The ability to step back and ask “Should we even build this?” rather than “How do I build this?” is what has always separated senior, or even great, developers from the rest. Now, we have to operationalize that seniority.</p><p>This concept might feel like something you already do today, because it is. We are simply taking those requirements, constraints, and architectural decisions out of scattered docs, Jira tickets, and whiteboard photos, and moving them directly into our codebase. We are moving toward Spec-Driven Development (SDD).</p><h3>Version Control for Your Thinking</h3><p>The idea is simple: Make technical decisions explicit, shared, and easy to change before they harden into code.</p><p>A concrete example of this is GitHub Spec Kit. It operates on the premise that our “specs” should live alongside the code, treated as the primary source of truth. Instead of prompting an AI to “write a login page” and hoping for the best, you define the behavior, the constraints, and the context in a structured spec file. The AI then generates the implementation from the spec.</p><p>If the output is wrong, you don’t manually patch the code, you update the spec.</p><p>This is version control for your thinking.</p><p>Like code, our understanding of a problem evolves. We discover edge cases. We change our minds. Specs let us review, diff, and evolve thinking independently of the implementation. It allows us to explore alternatives and let the AI generate multiple solutions without locking us into one path too early.</p><h3>Professional obligation</h3><p>We all have a neighbor or know a small business owner who has a “brother-in-law’s son” who can “build websites.” He is cheap, he is fast, and he can get something up on the screen. But we also know why major companies don’t hire him.</p><p>Professionals have always distinguished themselves from amateurs not by what they do, but by how they work. We build for reliability, for scale, and for the future. “Vibe coding” was the amateur hour and it works until it doesn’t. Spec-Driven Development is how we professionalize the workflow.</p><p>The reality is, clients aren’t paying for our “love of the code.” They don’t care about the elegance of our manual boilerplate or how much we enjoy typing syntax. They are paying for the value we produce. If there are tools allowing us to deliver better, more reliable software ten times faster, we have a professional obligation to master them. Why would a client pay for someone who refuses to use the most efficient tools available?</p><p>The industry is shifting. The early adopters set the tone, and now it is time for the rest of us to learn to run.</p><blockquote>Support writers. All our non-profit’s offerings <a href="https://forms.gardenofneuro.org/nx/portal/neonevents/events?path=%2Fportal%2Fevents">here</a>.</blockquote><blockquote><a href="https://medium.com/it-chronicles/newsletter">Click here to subscribe to the IT Chronicles newsletter.</a></blockquote><blockquote><a href="https://forms.gardenofneuro.org/nx/portal/neonevents/events?path=%2Fportal%2Fevents">Click for Wordsmith, Mystery Writing, Write Like Stephen King, more</a></blockquote><figure><a href="https://medium.com/it-chronicles"><img alt="Computer circuit pathways" src="https://cdn-images-1.medium.com/max/1024/1*bX_AIumALEvtJr-mZIBeMw.png" /></a><figcaption>By the EIC <a href="https://medium.com/u/cc0821aee9c1">Susan Brearley</a></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ef25913ecb70" width="1" height="1" alt=""><hr><p><a href="https://medium.com/it-chronicles/software-engineering-has-been-dying-for-three-years-ef25913ecb70">Software Engineering Has Been Dying for Three Years</a> was originally published in <a href="https://medium.com/it-chronicles">IT Chronicles</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Mostly Nuanced but Deeply Opinionated View of the AI Landscape at the End of 2025]]></title>
            <link>https://medium.com/@lejahmie/a-mostly-nuanced-but-deeply-opinionated-view-of-the-ai-landscape-at-the-end-of-2025-b91ed9f5bcb1?source=rss-7c66bba90f3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/b91ed9f5bcb1</guid>
            <category><![CDATA[organizational-change]]></category>
            <category><![CDATA[junior-developer]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Jamie Telin]]></dc:creator>
            <pubDate>Fri, 12 Dec 2025 08:32:19 GMT</pubDate>
            <atom:updated>2026-02-27T13:43:36.297Z</atom:updated>
            <content:encoded><![CDATA[<h4>YEAR IN REVIEW</h4><h4>Massive context windows are a workaround. Real context is a capability.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VfAdLyV7kJWrpB1SD4l90Q.jpeg" /></figure><p>A few years ago, I worked on a major project at one of the country’s biggest media companies. We had this incredibly capable Business Analyst, the type everyone trusts and secretly hopes never quits. Besides her normal BA work, she also handled every single configuration for our magazines. It wasn’t hard work, but 5–10 hours a week of pure clicking. Manual adjustments. Settings. Repetitive tasks nobody wanted.</p><p>Then she got promoted (well deserved). Her “extra” work landed on the developers. They looked at it, shrugged, and three days later… the whole thing was automated. What used to take hours per week became a script that ran in under a minute.</p><p>The lesson? Some tasks are simply machine tasks, we just didn’t have the machine yet.</p><p>AI didn’t invent this story. Humans always do unnecessary manual labor until they don’t. Automation always looks obvious in hindsight.</p><h3>The Great AI Schism of 2025</h3><p>This year I’ve seen two completely opposing narratives:</p><ul><li>The AI boosters: “We’re all 10× more productive! Software development is solved!”</li><li>The AI doomers: “AI is destroying craftsmanship, jobs, the junior pipeline, creativity, the economy, humanity, etc.”</li></ul><blockquote><em>And here’s the weird part: they’re both right, depending on where you look.</em></blockquote><p>Let’s take two examples from the research:</p><h3>1) Experienced developers who actually got slower</h3><p>A <a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">small randomized controlled trial</a> of experienced open-source developers found that allowing AI tools like Cursor and Claude increased task completion time by 19%, even though everybody thought it would speed things up by 20–40% .</p><h3>2) DORA’s State of AI: AI is an amplifier</h3><p><a href="https://cloud.google.com/resources/content/2025-dora-ai-assisted-software-development-report?hl=en">DORA’s huge 2025 report</a> (almost 5,000 respondents) concluded that AI magnifies whatever already exists.:</p><ul><li>High-performing teams get better.</li><li>Struggling teams get worse.</li><li>AI boosts throughput but increases instability if the organization isn’t ready for the speed.</li></ul><blockquote><em>AI isn’t magic. It’s a force multiplier. If your system is a mess, AI multiplies the mess.</em></blockquote><p>This explains a lot of contradictory experiences.</p><h3>All AI isn’t created equal however…</h3><h3>Images, music, video</h3><p>Only valuable if humans consume them. Their impact is mostly cultural and economic, and right now, often messy. Artists suffer. Quality drops. Content floods the internet. But this isn’t the whole story.</p><blockquote><em>There </em>are<em> legitimate use cases here, and we shouldn’t pretend otherwise.</em></blockquote><p>There has always been low‑quality, low‑margin creative work that was never truly “art” such as stock images, radio jingles, background loops, header graphics for social media posts (🤨), the brochure‑ware visuals nobody remembers five seconds later. Generative AI is already replacing that work, and honestly? It probably <em>should</em>.</p><p>And as someone who has spent hours masking an entire football team to put them on transparent backgrounds… let me tell you: some tasks are meant to be automated.</p><blockquote><em>I don’t miss that work. No real artist does.</em></blockquote><p>The real creatives, the ones with imagination, voice, vision, taste, shouldn’t fear AI. If anything, they should see the potential. New tools always reshape the boundaries of what’s possible. Photoshop did. ProTools did. Cinema 4D did. Now AI does.</p><p>So yes: some jobs will disappear. Some tasks will vanish. But creativity, the kind that makes you feel something, remains stubbornly human.</p><p>But here, juniors will feel the impact. The low level jobs are vanishing, and companies are taking advantage of this for now. But they also will feel the pain later.</p><h3>LLMs and agentic systems</h3><p>These produce machine-consumable output, text that other software can use. Instructions, API calls, structural transformations.</p><p>This is the real revolution but let me tie it back to something personal.</p><p>When I started my career, it was as a junior in a tiny local web design studio. Those were some of the most important years of my life. I learned everything: how to work with real clients, how to build something from nothing, how to debug and not go insane (well… most of the time). We built websites for local restaurants, hairdressers, carpenters, plumbers, small businesses that needed a digital presence.</p><p>And here’s the truth: that world is changing.</p><p>Those very businesses can now vibe‑code a website using one of the trendy new AI tools. They don’t need a junior developer like the younger me, at least not for the simple stuff. And that’s okay. Because this is the pattern of every major technological shift: Gutenberg, radio, video, the internet, the app boom.</p><p>Each one disrupted an industry. Each one created something new. And each one forced us to rethink what <em>real</em> work looks like.</p><p>And every single time, people panicked. The printing press would destroy the church. Radio would destroy books. Television would destroy radio. The internet would destroy the local stores. Apps would destroy everything else.</p><p>But it didn’t end the world. It expanded it.</p><p>So yes, my local pizzeria may vibe‑code their next website. That job is gone, or maybe not. Maybe the junior path will be just different. But the deeper, more creative, more complex work, the kind that requires actual problem solving, architecture, empathy, storytelling, engineering, that work isn’t going anywhere. It never has. It survived Gutenberg, radio, video, the internet, and everything that came after. It will survive this too. Not until we invent true AGI. Maybe not even then.</p><h3>Here’s the actual distinction:</h3><p>Images, music, video: These only matter when humans consume them. They disrupt culture, and generally bad. Unless used as part of the created process, and I am pretty sure this is where it will be used once the hype is settled.</p><p>LLMs and agentic systems: These create machine‑readable instructions. They automate <em>processes</em>, not just <em>content</em>. They change how work flows.</p><p>This is where the transformation happens. Not in the gimmicky generative circus, but in the emergence of reasoning tools that plug into the fabric of organizations. Messy, messy organizations.</p><p>So no, creative, difficult, meaningful engineering isn’t dying.</p><p>But the easy stuff? The starter tasks? The “local web shop” work that gave me my first break?</p><p>Those will vanish. And if I’m honest, they probably <em>should</em>. The world doesn’t need thousands of junior developers hand‑coding the same WordPress‑tier brochure sites forever. That kind of repetition is exactly what automation is for.</p><p>And something new will emerge in their place. It always does. The hard problems remain. The meaningful problems remain. The human‑shaped problems remain.</p><blockquote><em>This is automation scaling again, except this time across knowledge work.</em></blockquote><p>Images don’t automate your business. <em>Text does. Context does.</em> Agents do.</p><p>I dare to say it: No one will use the stupid ChatGPT Sora App in 2026. But it was bloody good advertisement, it created buzz. It did what it was supposed to. It was never a real product, hell it was probably vibe-coded.</p><h3>Juniors are in trouble and it’s not their fault</h3><p>And this is where my own story loops back in. If I entered the industry today the way I did back then, in a tiny local web design studio, learning by doing, breaking things, fixing them, building real products for real people, I’m not sure that path would even exist. Not in the same form. The junior roles that shaped me are some of the first to disappear.</p><blockquote><em>That’s not alarmism; it’s recognition. And it’s why this conversation matters.</em></blockquote><p>I’ve talked about this at company meetups, but it’s worth repeating: this is the first time in my career where I genuinely worry about junior developers.</p><p>The economy is rough, yes. But the deeper issue is that:</p><ul><li>A senior developer with AI can often replace two or three juniors.</li><li>Seniors are already expected to mentor those juniors.</li><li>So the “math doesn’t math” in the juniors’ favor.</li></ul><p>The scary future scenario is this:</p><p>We cut junior positions → fewer seniors grow → less high-quality code → worse training data → worse future AIs → slower progress → a stagnating industry.</p><p>It’s like eating your seed corn because the harvest looks far away.</p><h3>The research we don’t see</h3><blockquote><em>Tech optimism is loud, but the warnings are often… quiet.</em></blockquote><p>Partly because big AI companies aren’t eager to publish research showing negative economic or social impacts.</p><p>We’ve already seen <a href="https://biggo.com/news/202512100323_OpenAI-Research-Advocacy-Tensions-Productivity-Claims">internal tension at OpenAI</a>, with economists resigned amid reports that research on potential harms was being discouraged. And some <a href="https://www.techbuzz.ai/articles/openai-researcher-quits-over-alleged-research-censorship-claims">researchers even resigned following allegations of internal pressure around negative-impact research</a> from OpenAI in 2025 over suppressed negative-impact findings.</p><p>This matters. Because the future is being shaped not only by what companies build, but by what they refuse to talk about.</p><p>And this is where the <em>global</em> dimension enters the picture.</p><p>When we talk about who holds power in AI, we often reduce it to a Silicon Valley problem, the familiar “tech‑bro villain arc.” But that framing only tells half the story.</p><p>Yes, there’s something uncanny about how many billionaires drift toward Bond‑villain aesthetics. Why <em>do</em> so many wake up one day convinced they’re humanity’s chosen savior? They flutter toward grandiosity like moths to a candle flame. They dabble in politics, flirt with corruption, whisper about accelerationism, and inevitably end up scorched as often as they triumph. Elon Musk alone is practically the speed‑run version of the tech‑bro tragic arc.</p><p>And let’s be clear: the billionaires are part of the problem. Not because they are plotting our doom in a secret volcano lair, but because they are too rich to see the collateral damage. To them, closing the market isn’t a bug; it’s a feature. It could be the next Bond movie plot, we already seen the news industry depicted in <em>Tomorrow Never Dies, </em>where Bond faces off against the megalomaniacal media mogul Elliot Carver.</p><p>Take the massive deal between <a href="https://newsroom.accenture.com/news/2025/accenture-and-anthropic-launch-multi-year-partnership-to-drive-enterprise-ai-innovation-and-value-across-industries">Accenture and Anthropic in late 2025</a>. They are training 30,000 consultants specifically on Claude. To the executives involved, this is just smart scaling. They are too removed to realize, or perhaps too wealthy to care, that they are turning Intelligence into the <a href="https://en.wikipedia.org/wiki/Vendor_lock-in">“<em>new Oracle database”</em></a>. It becomes expensive, impossible to rip out, and controlled by a duopoly. One company builds the brain, the other holds the keys, and the rest of the industry, the juniors, the hackers, the startups, gets locked out of the engine room.</p><p>See content credentials</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7bt8eVxkvj-_5Lcm" /></figure><p>But Silicon Valley doesn’t operate in a vacuum. It operates in competition.</p><p>And the counter‑force, not villain, not hero, but undeniably powerful, is China.</p><p>China isn’t wearing a white hat or a black one. It’s simply playing the game differently. While U.S. labs retreat into secrecy to protect trade secrets, <a href="https://www.digital-science.com/blog/2025/07/new-report-shows-china-dominates-in-ai-research/">China is flooding the world with open‑weight models and publishing high‑impact research at accelerating speed</a>. Not out of altruism, but strategy. Influence. Soft power. Ensuring the U.S. doesn’t dominate the AI stack for the next century.</p><p>Here is how the landscape shifted:</p><ul><li>For years, Chinese researchers were accused of producing volume without quality — that flipped around 2024–2025.</li><li>China now leads the U.S. in high‑impact AI research (top 40% cited).</li><li>In surveillance, industrial automation, and robotics, Chinese papers have become the gold standard U.S. researchers cite.</li><li>As American labs (OpenAI, Anthropic, Google) sealed their methods, Chinese labs (DeepSeek, Qwen, 01.AI) kept publishing.</li><li>By 2025, if you were an AI grad student wanting to study how a top‑tier model actually <em>works</em>, you likely studied a Chinese model, because the American ones were black boxes.</li><li>According to a 2025 Digital Science report, China accounts for roughly 60 percent of global AI-related patent filings.</li></ul><p>It’s not evil. It’s not heroic. It’s simply two superpowers shaping the tools everyone else will depend on.</p><p>And caught between them? Everyone else.</p><p>This is exactly why transparency matters. Why suppressed research matters. Why letting a handful of tech bros, or a single nation‑state, define the future is dangerous. AI isn’t just a technological revolution; it’s a geopolitical one.</p><p>And yes, I sometimes dress like a tech bro. I own the official outfit. Even as I write this in my blue vest and hoodie, the outfit my wife insists makes me look like a local criminal, I’m reminded of something important: the real danger isn’t the ones who <em>look</em> like tech bros, that’s just theatrics, a short dopamine‑fueled diversion. The danger is the people building opaque systems that will shape our future, while we can’t even tell whether the timeline they’re steering us into is bright or dark.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/400/0*9fC7sh874PM3PIfd" /></figure><p>You have read this far. You are brave… We are soon done. Just a little bit more to go.</p><h3>When AI fear turns into folklore</h3><p>Speaking of fear: Eddie Burback’s video, <a href="https://www.youtube.com/watch?v=VRjgNgJms3Q">ChatGPT made me delusional</a>, is one of the most brilliantly funny demonstrations of how AI <em>could</em> be dangerous, while also revealing how most AI panic completely misses the point. It’s funny, it’s insightful, and it showcases the exact kind of nonsense that fuels the negative headlines. And yet… these things are not the real problem.</p><p>I’m not saying there is <em>no</em> problem. But much like moral panics of the past, these issues affect a very small number of people, if any. Meanwhile, society is so deeply divided; politically, culturally, economically, that it becomes harder and harder to focus on what actually matters.</p><p>We’ve seen this pattern before.</p><p>Let’s reinvent the Satanic Panic of the 1980s for a moment: a convergence of fears about youth culture (rock, metal, D&amp;D), earlier Christian anti-occult movements, and a flurry of highly publicized but now-debunked abuse cases. It felt existential at the time. It feels absurd now.</p><p>AI panic, at least this <em>particular</em> flavor of it, follows the same script. And unless we learn to separate genuine risks from cultural hysteria, we’ll waste time fighting shadows instead of shaping the bright future we can have.</p><h3>Agents: The Beginning of the Next Phase</h3><p>Jira has Rovo. GitHub has Copilot. Slack has AI summarizers.</p><p>All useful. None of them understand your actual context.</p><ul><li>Copilot doesn’t know <em>why</em> you’re building something.</li><li>Jira AI doesn’t know your architecture or constraints.</li><li>Slack AI doesn’t know what the messages in your chaotic project channel <em>actually mean</em>.</li></ul><p>They’re smart tools trapped in dumb silos.</p><p>And here’s the real bottleneck: real work doesn’t happen inside a single tool, it happens between them. In the messy handoffs, the undocumented decisions, the implied assumptions. Every software team is an ecosystem, but today’s AI tools only see isolated terrariums.</p><p>The next leap isn’t just smarter AI. It’s shared context. One AI for all your context, or your one context for all AI.</p><p>This is where things get interesting. Because we’ve already seen a glimpse of the future in coding agents, and some smarter Agentic frameworks.</p><p>Coding agents today have more context about your computer than any other AI you use.</p><p>I don’t just code with my agent, I live with it in my workflow. And that’s the point. We need that, but scaled to life. To the systems we actually live and work inside. And between them.</p><p>And not just autocomplete-for-code, but something closer to orchestration-for-thinking. Something that can coordinate, contextualize, and act across the tools and information sources that make up your real world.</p><p>And this brings us to an unexpectedly important player: Apple.</p><p>If any company has the power to unify AI across the apps, sensors, documents, messages, locations, habits, and workflows of your actual life, it’s them. They control the hardware, the OS, the APIs, the sandbox. The glue. If the world ever gets the promised future Scarlett Johansson, <em>sorry I mean Samantha</em>, from <em>Her</em>, it won’t come from an LLM website.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*-nq9LH3tzkbjeWqE" /></figure><p>It will come from someone who can integrate context across everything your devices touches.</p><p>Will Apple deliver? Maybe. I’m not holding my breath. But the potential is enormous.</p><p>Because the future isn’t <em>bigger models</em>. It’s situated agents. Context‑rich. Integrated. And here’s an important point we keep missing in the hype cycles. It’s not just “bigger context wins.”</p><p>Google, Anthropic, OpenAI, everyone is racing toward 1M, 2M, even 10M‑token context windows. And sure, that’s impressive. But the only reason they <em>need</em> enormous context is because they are just apps. They don’t have real-time access to your world, your tools, your data, your environment.</p><blockquote><em>You can’t pass around your whole life in every request.</em></blockquote><p>So they compensate the only way they can: by asking you to <em>paste your entire life</em> into every prompt. But here’s the thing. You can’t pass around your whole life in every request. It’s inefficient. It’s brittle. It’s privacy‑wrecking. And honestly? It’s the wrong abstraction.</p><p>A smarter agent doesn’t need a 2M‑token dump of your digital existence. A smarter agent does what <em>you</em> would do:</p><ul><li>fetch the file it needs,</li><li>search the folder it cares about,</li><li>scan the document that matters,</li><li>query the system when it’s unsure,</li><li>look up what it forgot.</li></ul><p>This is why coding agents feel so magical. My coding agent doesn’t need me to upload my whole filesystem into a prompt. It just… goes and looks. It gathers context on demand. It writes scripts to filter data. It patches environments. It reads the code it needs in the moment, or even builds the code it needs to achieve its tasks.</p><p>It behaves like a person with a slightly better memory than mine, which, let’s be honest, isn’t hard.</p><p>Massive context windows are a workaround. Real context is a capability. And agents, not giant text boxes, are how we get there. A real coworker, not a code toy.</p><p>We’re already deep in the early wave. And yes, I stand by this prediction:</p><p>2026 will be the year context‑rich agent systems actually work.</p><h3>My own 2025 Wrapped™</h3><p>AI has enabled me more in the last 18 months than almost any tool in my entire career. I finally started the project I’ve been wanting to build for a decade. I deliver more. I learn faster. <em>I feel, dare I say it, like a 10× developer, even if that’s a myth.</em></p><p>We’ve delivered rebellious projects at work: do first, present later, impress always.</p><blockquote><em>It’s because AI amplified the skills I already had</em></blockquote><p>But here’s the truth (and this is where I drop the tech‑bro cosplay). This isn’t because AI replaced my skills. It’s because AI amplified the ones I already had. And it will amplify yours too.</p><p>But only if you engage with it intentionally, not passively.</p><p>See content credentials</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eYEfGFJZn9CqWATH" /></figure><h3>A call to action (before the Tech Bros decide everything, or China)</h3><p>AI will thrive without you. But the future will be worse if you’re not part of shaping it. Your fears, your frustrations, your hopes, they matter.</p><blockquote><em>The alternative is letting a handful of tech bros decide the trajectory of the next century.</em></blockquote><p>We need more human voices, not fewer. We need people who care about craft, ethics, juniors, quality, and meaning. The alternative is letting a handful of tech bros decide the trajectory of the next century.</p><p>Let’s not do that. Let’s build the future together. ✌️</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b91ed9f5bcb1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Using One AI Model for Everything: The Workflow That Actually Works]]></title>
            <link>https://medium.com/@lejahmie/stop-using-one-ai-model-for-everything-the-workflow-that-actually-works-1d4295b0e454?source=rss-7c66bba90f3a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d4295b0e454</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai-code-assistant]]></category>
            <category><![CDATA[code]]></category>
            <dc:creator><![CDATA[Jamie Telin]]></dc:creator>
            <pubDate>Fri, 12 Dec 2025 08:28:01 GMT</pubDate>
            <atom:updated>2025-12-12T08:28:01.103Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ArCvp2D8fg9mTmq-sItN0g.jpeg" /></figure><p>I’ve been experimenting a lot with Copilot, Claude Sonnet 4.5, GPT-5 Mini, and lately SpecKit, and one thing is becoming clear:</p><blockquote><em>Efficiency doesn’t come from choosing a single AI model. It comes from switching between them intentionally.</em></blockquote><p>I’ve actually been using a “spec first” workflow for a long time, asking the model to create design topics, solution designs, break things into stories, etc. It’s been a reliable approach, even when done manually.</p><p>Recently, I’ve started using SpecKit to formalize that process, and the structure it adds has worked surprisingly well.</p><p>During all this experimenting, I came across a comparison of GPT-5 Mini vs. Sonnet 4.5 that matched a lot of what I’ve seen myself: <a href="https://dredyson.com/gpt-5-mini-vs-claude-sonnet-4-i-tested-both-ai-models-heres-the-clear-winner-for-coding-efficiency/">https://dredyson.com/gpt-5-mini-vs-claude-sonnet-4-i-tested-both-ai-models-heres-the-clear-winner-for-coding-efficiency/</a></p><p>A couple of points stood out:</p><blockquote><em>GPT-5 Mini stays more “on task”, which is perfect for implementation</em></blockquote><p>One benchmark showed Mini staying on-task 92% of the time vs. Sonnet’s 68%.</p><p>And if you look at more formal benchmarks like AIME (math-heavy reasoning) and GPQA (broad general-knowledge reasoning), GPT-5 Mini also performs surprisingly well. For example, this comparison highlights Mini slightly ahead on AIME and competitive on GPQA: <a href="https://airank.dev/models/compare/claude-sonnet-4.5-vs-gpt-5-mini">Ai Rank: Sonnet 4.5 vs GPT 5 Mini</a></p><p>This lines up with how they feel in practice:</p><ul><li>Sonnet 4.5 thinks like a senior engineer, great for planning, architecture, debugging, and interpreting messy requirements.</li><li>GPT-5 Mini behaves more like a junior developer, eager, literal, inexpensive, and great at executing clearly defined tasks.</li></ul><p>Once a clear spec exists (whether generated manually or via SpecKit), Mini tends to produce consistent, reliable results.</p><blockquote><em>Using Mini for most tasks can dramatically reduce cost</em></blockquote><p>The article noted that using Mini for ~90% of the work and Sonnet for the hardest 10% led to a huge drop in monthly AI cost, without sacrificing output quality.</p><p>I’ve noticed the same: we often stick to one model out of habit, but switching based on the <em>type</em> of task is far more efficient.</p><p>I’ve been doing this both in Copilot and my favourite agent harness, Kilo Code.</p><h3>The analogy fits, and it feels true in daily use:</h3><ul><li>Mini = the junior dev who follows instructions precisely</li><li>Sonnet = the senior dev who may “improve” the spec, sometimes brilliantly, sometimes unexpectedly</li></ul><p>Just like a real team, each model has a role.</p><h3>My current workflow (Copilot + SpecKit + a bit of manual structure)</h3><p>Use Sonnet 4.5 for:</p><ul><li>thinking, analysis, planning</li><li>solution design &amp; architecture</li><li>debugging &amp; edge cases</li><li>generating or reviewing specs (with or without SpecKit)</li></ul><p>Use GPT-5 Mini for:</p><ul><li>implementing well-defined tasks</li><li>boilerplate</li><li>repetitive work</li><li>transforming specs into code</li><li>throw-away scripts</li></ul><p>And most importantly:</p><blockquote><em>Switch models deliberately, don’t stay on one just because it’s “good.”</em></blockquote><p>This tiny mindset shift has noticeably improved my speed, cost, and consistency.</p><h3>Some stats for the number hungry…</h3><p>Claude Opus 4.1</p><ul><li>Latency: 3.40s</li><li>Throughput: 33.72 tps</li><li>Context: 200k</li><li>Cost: $15/M input, $75/M output</li></ul><p>Claude Sonnet 4.5</p><ul><li>Latency: 2.63s</li><li>Throughput: 48.66 tps</li><li>Context: 1M</li><li>Cost: from $3/M input, from $15/M output</li></ul><p>GPT-5 Mini</p><ul><li>Latency: 5.46s</li><li>Throughput: 53.74 tps</li><li>Context: 400k</li><li>Cost: $0.25/M input, $2/M output</li></ul><p>Source: <a href="http://openrouter.ai/">openrouter.ai</a></p><h3>Final words of wisdom</h3><p>GPT-5 Mini is incredibly cost-effective and noticeably fast in terms of token output, which makes it a great fit for writing large amounts of code once the direction is clear. It’s the model you reach for when you want high-volume execution without burning budget.</p><p>Claude Sonnet 4.5 fills a different but equally valuable role. Its 1M-token context and solid pricing make it ideal for tasks that require holding large portions of a codebase in memory, reasoning across multiple files, and forming coherent plans. When the work depends on understanding structure rather than just producing output, Sonnet lands in the perfect middle ground.</p><p>If anyone else is experimenting with model switching or SpecKit-driven workflows, I’d love to hear your experiences!</p><p>Why GPT-5 Mini? Well GPT-5 mini, GPT-4.1 and GPT-4o are included models, and do not consume any premium requests if you are on a paid plan for Copilot. 🤖<br><a href="https://docs.github.com/en/copilot/concepts/billing/copilot-requests#model-multipliers">https://docs.github.com/en/copilot/concepts/billing/copilot-requests#model-multipliers</a></p><p>SpecKit, interested into trying it? 📝<br><a href="https://github.com/github/spec-kit">https://github.com/github/spec-kit</a></p><p>Ready to leave Cursor? Bored of Copilot? Try Kilo Code, my favorite coding agent harness (it is open-source). 😎<br><a href="https://kilo.ai/">https://kilo.ai/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d4295b0e454" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>