<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Senko Rašić</title>
    <link>https://blog.senko.net/</link>
    <description>Random thoughts | [about](https://senko.net)</description>
    <pubDate>Wed, 18 Nov 2026 20:36:29 +0000</pubDate>
    <item>
      <title>Opus and Qwen do a Sokoban</title>
      <link>https://blog.senko.net/opus-and-qwen-do-a-sokoban?pk_campaign=rss-feed</link>
      <description>Comparing Claude Opus 5.5 vs Qwen3.8-Flash-Next, running locally, on the same coding task: Sokoban! In a slightly different setup than us...</description>
      <content:encoded><![CDATA[<p>Comparing Claude Opus 5.5 vs Qwen3.8-Flash-Next, running locally, on the same coding task: <a href="https://en.wikipedia.org/wiki/Sokoban" rel="nofollow ugc">Sokoban</a>!</p>

<p>In a slightly different setup than usual, I ran Claude Opus 5.5 and Qwen3.8-Flash-Next through the chatbot interface (not a coding agent); Qwen I ran via <a href="https://llama.app/" rel="nofollow ugc">llama.app</a> (quantized to <a href="https://unsloth.ai/docs/models/qwen3.8-next" rel="nofollow ugc">4 bits</a> on an AMD Strix Halo machine with 128GB of unified RAM).</p>

<p>The prompt was a one-liner: “<em>Create a simple but beautiful Sokoban game clone in HTML/JS/CSS (all in one file, no frameworks or 3rd-party libraries).</em>“</p>

<p>Opus (high reasoning effort) took ~4 minutes (I don&#39;t have a token count because this was from a web chatbot interface). Qwen took ~31 minutes and 64k tokens. 80% of this was for reasoning (at extra-high effort). Qwen thinks a lot more than Opus!</p>

<p>Both are available at <a href="https://senko.net/vibecode-bench/" rel="nofollow ugc">my vibecode benchmarks page</a>. Here&#39;s also a side-by-side playthrough of both:</p>

<video src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/2026-11-07-sokoban-claude-qwen-side-by-side.mp4" controls="" style="width:100%;height:auto;">
  <a href="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/2026-11-07-sokoban-claude-qwen-side-by-side.mp4" rel="nofollow ugc">Watch the video</a>
</video>

<p>I then had GPT-6 Astra compare the quality of the implementations. I chose Astra so as to not have any bias towards a solution (in case models really do prefer their own).</p>

<p>Across 30 different areas/aspects of each solution, Opus scored 77% and Qwen 70% (100% would be a perfect implementation without downsides or bugs).</p>

<p>Astra concluded the code review with: “<em>Claude’s implementation has a modest but meaningful advantage in correctness and maintainability, with simpler state management, stronger movement safeguards, and fewer consequential bugs.</em>“</p>

<p>Note that this test compared the best frontier model for coding against something I ran completely locally, on a machine sitting next to me. As I wrote earlier this year, <a href="https://blog.senko.net/open-weights-ai-models-have-become-good-enough" rel="nofollow ugc">open-weights models have become good enough</a>, but it&#39;s hard (and expensive) to run a very capable model locally.</p>

<p>For this test I chose Sokoban because, frankly, my usual “frontier” tests, RTS/Warcraft, Minecraft and a flying simulator, were beyond Qwen&#39;s reach for a zero-shot. When trying to build an RTS, the model took hours and produced an unusable mess.</p>

<p>The hardware cost for this setup can vary between $3.5k and $6k, depending on luck, availability, and other system components. Not cheap, but not beyond the budget for a small business or affluent techies.</p>

<p>The same amount of money does buy you a few years of Claude Max subscription—purely on that level, it&#39;s not easy to justify.</p>

<p>But while I will not be canceling my Claude subscription any time soon, I&#39;d say that for smaller tasks and with more guided steering, this is a usable alternative for those who would prefer to—or must—use local models.</p>
]]></content:encoded>
      <guid>https://blog.senko.net/opus-and-qwen-do-a-sokoban</guid>
      <pubDate>Sat, 07 Nov 2026 15:56:20 +0000</pubDate>
    </item>
    <item>
      <title>We&#39;re all bad at writing software</title>
      <link>https://blog.senko.net/were-all-bad-at-writing-software?pk_campaign=rss-feed</link>
      <description>After several discussions with fellow software developers about whether we could truly “let go” and trust LLM-generated code, and a high-...</description>
      <content:encoded><![CDATA[<p>After several discussions with fellow software developers about whether we could truly “let go” and trust LLM-generated code, and a <a href="https://jepsen.io/analyses/particledb-1.1.4" rel="nofollow ugc">high-profile quality facepalm case</a>, I&#39;ve been reflecting on my thoughts on the matter.</p>

<p>More often than not, I end up advocating for giving LLMs a chance, but I wouldn&#39;t count myself as a hyped-up AI booster. At least I wouldn&#39;t <em>want</em> to count myself as an unrealistic Kool-Aid drinker. I don&#39;t think LLMs are so good at generating code that we can just trust them and let go. No, they&#39;re capable of producing convoluted garbage and then politely apologizing for it.</p>

<p>No, my optimism comes from quite a different direction: I don&#39;t trust LLMs, but <strong>I don&#39;t trust humans either.</strong></p>

<p>People (myself included) are exceptionally bad at writing software. We&#39;re so bad we&#39;ve constructed an elaborate Tower of Babel to help us with the task—and still we fail.</p>

<p>From little things, like syntax highlighting to make the code easier to read, through linters, type checkers, correctness provers, all manner of different tests, to straitjackets of Rust and Haskell that try to keep us from blundering into a scary race condition, use after free, off-by-one and other spooky monsters hiding under our compilers.</p>

<p>Read 2000 lines (50 pages or so) from a book in an afternoon? Sounds like a wonderful and relaxing day. Read 2000 lines of (much sparser) code and you&#39;re wasted. It&#39;s hard, and you&#39;re almost guaranteed to miss something.</p>

<p>And that&#39;s in the best case, if you ignore company politics, endless sprinting through a growing backlog, bloated software full of crap features benefiting the company instead of the users, ... The list can go on and on; just look at <a href="https://thedailywtf.com/" rel="nofollow ugc">The Daily WTF</a> or Dilbert.</p>

<p>Crapware, bloatware and “stinkin&#39; pile of shit” are terms that long predate LLMs. As a <a href="https://news.ycombinator.com/item?id=50342652" rel="nofollow ugc">Hacker News commenter</a> put it nicely:</p>

<blockquote><p>software I used that was majority built by &#39;real human engineers&#39; coding by hand is utter garbage. [...] The average software engineer cannot build good and secure software.</p></blockquote>

<p>I&#39;d go a step further and say: to the extent that you build correct, performant and secure software, you do it by diligently following best software engineering practices (like, but not limited to, exhaustive testing and debugging), not via your supreme intellect.</p>

<p>As <a href="https://en.wikipedia.org/wiki/Brian_Kernighan" rel="nofollow ugc">Brian Kernighan</a> put it:</p>

<blockquote><p>Everyone knows that debugging is twice as hard as writing a program in the first place. So if you&#39;re as clever as you can be when you write it, how will you ever debug it?</p></blockquote>

<p>Likewise, from <a href="https://en.wikipedia.org/wiki/Donald_Knuth" rel="nofollow ugc">Knuth</a>:</p>

<blockquote><p>Beware of the bugs in the above code; I have only proved it correct, not tried it.</p></blockquote>

<p>Back to LLMs. Yes, they can produce garbage, just like humans. They can also produce good code, if they&#39;re properly guided and constrained, just like humans. And they keep getting better at it.</p>

<p>So yeah, I&#39;m an optimist.</p>

<p>I love writing code and I&#39;ll keep writing code (by hand) for the fun of it. I have <a href="https://en.wikipedia.org/wiki/Structure_and_Interpretation_of_Computer_Programs" rel="nofollow ugc">SICP</a>, <a href="https://www.goodreads.com/book/show/52084.Programming_Pearls" rel="nofollow ugc">Programming Pearls</a> and a few volumes of <a href="https://en.wikipedia.org/wiki/The_Art_of_Computer_Programming" rel="nofollow ugc">TAOCP</a> on my bookshelf and have just ordered a paperback of <a href="https://craftinginterpreters.com/" rel="nofollow ugc">Crafting Interpreters</a>. Programming is fun!</p>

<p>I&#39;m happy to leave grinding through Jira tickets, CRUD and enterprise integration glue code to my AI minions.</p>

<hr/>

<p>Entirely human-written; proofread by Claude.</p>
]]></content:encoded>
      <guid>https://blog.senko.net/were-all-bad-at-writing-software</guid>
      <pubDate>Mon, 02 Nov 2026 21:54:40 +0000</pubDate>
    </item>
    <item>
      <title>Personal digital sovereignty</title>
      <link>https://blog.senko.net/personal-digital-sovereignty?pk_campaign=rss-feed</link>
      <description>Life is full of trade-offs, and some of the trade-offs involve relinquishing (part of) control for convenience. In tech, this can often b...</description>
      <content:encoded><![CDATA[<p>Life is full of trade-offs, and some of the trade-offs involve relinquishing (part of) control for convenience.</p>

<p>In tech, this can often be seen in using services instead of downloading software and using it locally. For example, I use Netflix or Spotify instead of maintaining my own media library. It’s much more convenient, even if it doesn’t have all the content I’d have, or isn’t laid out exactly as I’d prefer.</p>

<p>A sneaky problem with relying more on services—thus relinquishing the choices and control over many details to them—is that we transform from active (human) agents to passive users who have to play by the providers’ rules.</p>

<h2 id="ai-says-no">AI says “no”</h2>

<p>The frontier AI providers are under a lot of pressure to stop people from using their models for offensive security work: hacking, malware, exploits. Fair enough. The problem is that the filters can’t tell an attacker from a developer fixing their own code, so they catch a lot of people doing perfectly normal things.</p>

<p>Since June, Claude Code has been switching from Fable 5 to the older Opus 4.8 whenever a classifier decides a message is about cybersecurity. One developer <a href="https://github.com/anthropics/claude-code/issues/66873" rel="nofollow ugc">got switched</a> while fixing a user-enumeration bug in their own password manager.</p>

<p>Low-level code looks a lot like security code to a classifier. <a href="https://x.com/antirez/status/2116453735207706754" rel="nofollow ugc">Salvatore Sanfilippo (antirez)</a>, the creator of Redis, complained a few days ago that “<em>working at Redis with agents is becoming a real struggle because of the many security stops I get while working at non security things</em>”.</p>

<p><a href="https://x.com/mitsuhiko/status/2113248077532545057" rel="nofollow ugc">Armin Ronacher</a> asked his agent to download PHP 8.5.11 and compile it. The request was blocked for “violative cyber content”, presumably because of a recently disclosed PHP security bug. Earlier in October, he wanted to finish setting up a Windows PC without creating a Microsoft account, and also <a href="https://x.com/mitsuhiko/status/2108542467301904447" rel="nofollow ugc">ran into the safety filter</a>.</p>

<p>That last one is a double whammy: Windows itself has turned into a service that pushes you to sign in, and the AI tool you’d use to work around it refuses to help.</p>

<p>Writing small games with Claude is a lot of fun, and plenty of people do it (I use small games for <a href="https://senko.net/vibecode-bench/" rel="nofollow ugc">my vibe-coding benchmark</a>). Anthropic has now <a href="https://mashable.com/tech/anthropic-takes-steps-to-block-video-game-cloning-and-reverse-engineering" rel="nofollow ugc">taken steps to block Claude from reverse-engineering and cloning video games</a>. Users also <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-ai-model-is-now-blocking-attempts-at-video-game-recreation-and-cloning-users-also-claim-non-malicious-generic-reverse-engineering-tasks-are-blocked" rel="nofollow ugc">report blocks</a> on generic, non-malicious reverse engineering.</p>

<p>An AI researcher <a href="https://x.com/jiahaoshao1/status/2116475922190922088" rel="nofollow ugc">got tagged</a> for “training a competitor” even though they never distilled Claude: the training data came from an open-source model, and Claude only submitted jobs and ran the analysis. The <a href="https://news.ycombinator.com/item?id=50319358" rel="nofollow ugc">Hacker News discussion</a> has more examples: custom firmware for the commenter’s own USB devices, a home LAN server, browser tests for their own codebase.</p>

<p>None of this is illegal, immoral or harmful. In most of these cases, the person wasn’t even breaking Anthropic’s usage policy. A system decided they shouldn’t be doing it, and that was that.</p>

<h2 id="tools-vs-services">Tools vs. services</h2>

<p>A kitchen knife can cut bread or hurt someone. It’s not the knife maker’s job to make sure you only cut bread. The author of a disk copy tool doesn’t have to make sure you’re not copying pirated movies. If you misuse a tool, that’s on you. That’s how tools work: you buy them, you use them any way you like, and you carry the responsibility for how you use them.</p>

<p>With services, it’s the other way around. The provider is responsible for what you do with the service (legally, or at least in the eyes of the press and regulators), so the provider polices you. You stop being a customer who makes their own decisions and owns the consequences. You become a user who has to stay inside the lines someone else drew.</p>

<p>It’s understandable why the service providers do this (government pressure, copyright lawsuits, bad press), but the reasons don’t change the result: a tool I pay for decides what work I’m allowed to do with it.</p>

<p>This isn’t an AI-specific problem. We use Gmail instead of running our own mail servers. I wrote about the developer side of this in <a href="https://blog.senko.net/developer-convenience-infrastructure-cost-and-losing-the-open-web" rel="nofollow ugc">Developer convenience, infrastructure cost, and losing the open web</a>.</p>

<p>Often that’s a fair trade. I don’t want to maintain a mail server or a media library, and I’m happy to pay someone else to do it. Likewise, most people don’t have the hardware, or the interest, to run a large LLM locally, so outsourcing it to Anthropic or OpenAI makes sense.</p>

<p>The trade is convenience for control. It works while the service does what you need and stays out of your way. When the provider also decides what you can and can’t do with it, it turns into a devil’s bargain.</p>

<h2 id="free-software-open-models">Free software, open models</h2>

<p>The free software movement started by addressing a very similar problem: empowering the user to do what they want, instead of being beholden to the software (or service) providers. Although open source “won” in terms of acceptance and importance in modern tech, the underlying problem reappeared as a consequence of the software-as-a-service (SaaS) approach becoming ubiquitous.</p>

<p>With AI specifically, this is why I’m <a href="https://blog.senko.net/open-weights-ai-models-have-become-good-enough" rel="nofollow ugc">excited about open models</a>. Yes, they’re not true “open source”, but they are close in spirit to free software: let the user do whatever they want with their tools. Once you’ve downloaded the model and have the hardware to run it, nobody can change its rules, take it away, or decide your work looks suspicious.</p>

<p>In <a href="https://blog.senko.net/how-i-want-to-use-ai" rel="nofollow ugc">How I want to use AI</a>, I wrote that I want to be the active operator, not merely a clueless user. That’s hard to do when the tool decides what you’re allowed to operate on.</p>

<p>I don’t think we should give up on services. But as hardware gets better, more of our work should move back to our own machines, to tools that we own and control. So it’s funny when <a href="https://www.youtube.com/watch?v=iMsUcCYNKqY&amp;t=1750s" rel="nofollow ugc">the boss of OpenAI says buying hardware to run local models</a> is not optimal. True: it is inefficient from a global coordination point of view, but you can ask the Soviet Union how the (national) coordination of resources worked out for them.</p>

<p>Using third-party service providers introduces an unavoidable <a href="https://en.wikipedia.org/wiki/Principal%E2%80%93agent_problem" rel="nofollow ugc">principal-agent problem</a>. What’s best for you and what’s best for your service provider are not necessarily the same. Sometimes the difference is small and worth ignoring. Sometimes you do want to keep your personal sovereignty, and you do want the option to walk away.</p>
]]></content:encoded>
      <guid>https://blog.senko.net/personal-digital-sovereignty</guid>
      <pubDate>Sun, 01 Nov 2026 10:47:56 +0000</pubDate>
    </item>
    <item>
      <title>Analyzing Jev, a new AI model</title>
      <link>https://blog.senko.net/analyzing-jev-a-new-ai-model?pk_campaign=rss-feed</link>
      <description>A new AI model called Jev launched last week and went viral. It&#39;s not an LLM. It doesn&#39;t generate text or pictures, but could be very use...</description>
      <content:encoded><![CDATA[<p>A new AI model called <a href="https://typesafe.ai/blog/introducing-system-one-models-and-jev" rel="nofollow ugc">Jev</a> launched last week and went viral.</p>

<p>It&#39;s not an LLM. It doesn&#39;t generate text or pictures, but could be very useful for a number of use cases. The creators call it “the first System One model”, tout its performance (193x faster, 444x cheaper) and promise it can&#39;t hallucinate.</p>

<p>So, what is it?</p>

<h2 id="how-it-probably-works">How it (probably) works</h2>

<p>Behind the impressive marketing is a very interesting model. Jev is a close relative of LLMs like GPT, Claude, or DeepSeek, but is actually more similar to the <a href="https://en.wikipedia.org/wiki/BERT_(language_model)" rel="nofollow ugc">BERT</a> series of models, which are used for classification, not generation. (If you&#39;ve ever wondered why someone calls out a difference between “AI” and “GenAI”, that&#39;s the distinction they mean).</p>

<p><a href="https://typesafe.ai/" rel="nofollow ugc">TypeSafe</a> (the creators of Jev) haven&#39;t actually published any papers on the system, but here&#39;s a reasonable guess and simplification, just to give you a sense of how it works.</p>

<p>Imagine going back to 2022. GPT-3, the best LLM at the time, can autocomplete text (you give it some text as an input and it continues it). Then <a href="https://arxiv.org/abs/2203.02155" rel="nofollow ugc">InstructGPT</a> shows how to make the LLM follow instructions from the user. Following the ChatGPT launch, the advances pile up, with reasoning, reinforcement learning, agentic use cases, and all the modern tech tricks LLMs now employ.</p>

<p>BERTs, meanwhile, are almost stuck in time. There&#39;s <a href="https://huggingface.co/blog/modernbert" rel="nofollow ugc">ModernBERT</a>, which modernizes it a lot, but you still need to fine-tune your model for the specific task. This is annoying, cumbersome, and more expensive than just asking an LLM. But these classification models are still much faster and cheaper than LLMs for suitable tasks, once you go through the trouble of training them.</p>

<p>Now imagine someone going, “<em>OK, how can we modernize BERTs with all the cool tech LLMs have been getting? Like using instructions in the input instead of having to re-train, making it really large, and using reinforcement learning?</em>” That&#39;s basically what Jev is (not exactly right, but it&#39;s a good analogy).</p>

<p>Since it only ingests the input and provides a few probabilities as output, it&#39;s much faster and cheaper than LLMs. Jev&#39;s model size is not publicly known, but even if it&#39;s 100x larger than ModernBERT, it&#39;d still be a paltry 40B params next to the humongous 2T-5T LLMs.</p>

<p>The authors also mention performance improvements due to not needing to generate text tokens and other workflow simplification, as well as the fact that they can evaluate the questions in parallel.</p>

<p>The “System One” moniker is a clever marketing trick that means “it&#39;s not a reasoning model”. They borrow the term from <a href="https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow" rel="nofollow ugc">Thinking, Fast and Slow</a>, where System 1 means intuitive, fast, automatic thinking and System 2 means deliberate, slower, effortful reasoning.</p>

<p>What about the hallucination claim? In the context of Jev, no hallucination means it can&#39;t generate output outside of the specified structure (i.e. it&#39;s “type-safe”), but that doesn&#39;t mean it can&#39;t generate incorrect output.</p>

<p><a href="https://magazine.sebastianraschka.com/p/classifier-history-and-jev" rel="nofollow ugc">Sebastian Rashcka has a great technical overview</a>, from the history of text classification in general, to how Jev (probably) works and comparison with BERT and GPT-based systems.</p>

<h2 id="using-jev">Using Jev</h2>

<p>This brings me to the usage part: you&#39;d use Jev (and other classification models like BERTs) to answer very specific questions, like “is this email spam?”, and it can answer with a Yes/No, a score (from a continuous domain, like 0.0-10.0), or a choice from a few discrete possibilities (classification).</p>

<p>Interestingly, the questions are independent and evaluated in parallel, which helps with performance. It also means you can just batch all the questions/classifications for a given piece of data at once, instead of going step by step.</p>

<p>Since Jev doesn&#39;t need to be fine-tuned, it&#39;s easy to use it via an API (and in fact, it&#39;s the only way to use it – don&#39;t expect the weights). Here&#39;s one official example, in Python:</p>

<pre><code class="language-python">from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

client = TypeSafeClient()

ticket = &#34;&#34;&#34;
Hi, I&#39;ve been trying to connect my Stripe account for 3 days
and the integration keeps failing.
I&#39;m losing sales. Please help ASAP.
&#34;&#34;&#34;

response = client.system_one(
    state=ticket,
    questions={
        &#34;department&#34;: Choice(
            instructions=&#34;Which team should handle this&#34;,
            criteria={
                &#34;billing&#34;: &#34;Payment or subscription issues&#34;,
                &#34;technical&#34;: &#34;Bugs or integration problems&#34;,
                &#34;sales&#34;: &#34;Pricing or account questions&#34;,
            },
        ),
        &#34;frustration&#34;: Score(
            instructions=&#34;How frustrated the customer appears&#34;,
            criteria=[
                &#34;Calm, just stating facts&#34;,
                &#34;Frustrated but civil&#34;,
                &#34;Very angry, strong language&#34;,
            ],
        ),
        &#34;is_urgent&#34;: Noul(
            instructions=&#34;The message conveys urgency or time-sensitivity&#34;,
        ),
    },
)

print(response.answers[&#34;department&#34;].choice)  # &#34;technical&#34;
print(response.answers[&#34;frustration&#34;].score)  # 1.0
print(response.answers[&#34;is_urgent&#34;].noul)     # 1.0
</code></pre>

<p>Here you describe the input (i.e. the “state” of the system) and ask a number of questions. The instructions on how to classify/decide are actually part of each question, not the input state.</p>

<h2 id="how-about-prompt-injection">How about prompt injection?</h2>

<p>If this is not a quirk of the API but actually how the model was trained, with state and instructions being separate, this might mean Jev is much harder to prompt-inject.</p>

<p>I tried a few times to convince the above example to route the query to a different department but was unsuccessful. While I might have had more success had I tried harder, it does seem like Jev is a bit more resilient to these kinds of attacks.</p>

<p>The authors are coy about this aspect, saying the <a href="https://news.ycombinator.com/item?id=49718450" rel="nofollow ugc">adversarial examples for Jev are different from those of LLMs so that they work even better together</a>.</p>

<h2 id="jev-vs-bert">Jev vs BERT</h2>

<p>A few months ago I trained a BERT model to classify some Croatian language texts, and I wondered how Jev would fare on the same task, so I gave it a go.</p>

<p>I ran it zero-shot on a random sample of 3,000 texts, with a one-sentence English instruction, and the (Croatian) text, asking for binary classification.</p>

<p>Jev got it right 97.1% of the time, compared to 97.6% for my fine-tuned model – basically the same. Looking at the cases where the two models disagreed, most turned out to be mislabeled data, and Jev was the one following the stated definition more literally.</p>

<p>Cost and speed were as advertised: the entire run took about a minute and cost around 20 cents. My local model ran at about the same speed (on NVidia RTX 3060).</p>

<h2 id="is-it-useful">Is it useful?</h2>

<p>Yes. If your use case is something that fits how Jev works, it could be much better than LLMs. Of course, if you already have a locally tuned BERT, the gains may not be much, if any – should be easy to test!</p>

<p>In my case, had I had access to Jev a few months ago, I wouldn&#39;t have needed to go to trouble of fine-tuning my own model. I&#39;d spend a tiny bit more money (ignoring the hardware cost), which would be more than recouped in development savings.</p>

<p>What I&#39;m most excited about is the combination of technical excellence and <em>marketing</em> excellence that raised the profile for this type of AI system, which has fallen by the wayside compared to the hype and capabilities of LLMs. That led people to use them in cases where they really aren&#39;t the best fit.</p>

<p>Jev and all the copycats that are sure to appear will bring a breath of fresh air and some extra breathing room to many orgs&#39; wallets :)</p>
]]></content:encoded>
      <guid>https://blog.senko.net/analyzing-jev-a-new-ai-model</guid>
      <pubDate>Tue, 22 Sep 2026 09:16:02 +0000</pubDate>
    </item>
    <item>
      <title>&#34;Code was never the hard part&#34; is an insult to all programmers</title>
      <link>https://blog.senko.net/code-was-never-the-hard-part-is-an-insult-to-all-programmers?pk_campaign=rss-feed</link>
      <description>The software development profession is in the midst of upheaval. Nobody knows how the AI revolution will play out in the end, but it is c...</description>
      <content:encoded><![CDATA[<p>The software development profession is in the midst of upheaval. Nobody knows how the AI revolution will play out in the end, but it is clear many aspects of work and life will be transformed—including programming.</p>

<p>One of the comments I hear often lately boils down to “<em>LLMs may be good at coding, but software was never the hard part</em>” and “<em>coding is easy, it&#39;s figuring out what to code that&#39;s hard</em>”.</p>

<p>I believe that&#39;s a gross insult to all programmers everywhere.</p>

<h2 id="if-coding-is-easy">If coding is easy...</h2>

<p>If coding is easy, how come programmers were in high demand, and have demanded large salaries for years (even before ZIRP)? Why was there so much stress, overwork and burnout even before AI started churning out 5000-line PRs? Why did companies seek 10x ninja rockstar coders and subject them to leetcode interviews—surely, a junior fresh out of college could churn out something if it&#39;s so easy?</p>

<p>If coding is easy, why do we have doorstoppers like <a href="https://www.goodreads.com/book/show/3735293-clean-code" rel="nofollow ugc">Clean Code</a> and <a href="https://en.wikipedia.org/wiki/The_Pragmatic_Programmer" rel="nofollow ugc">The Pragmatic Programmer</a>? Is <a href="https://www-cs-faculty.stanford.edu/~knuth/taocp.html" rel="nofollow ugc">The Art of Computer Programming</a> a light summer read? Is SICP a coffee-table book? Why do we have bootcamps or even whole college degrees dedicated to it?</p>

<p>If coding is easy, was <a href="https://en.wikipedia.org/wiki/John_Carmack" rel="nofollow ugc">Carmack</a> just at the right place at the right time? Why do we consider <a href="https://bellard.org/" rel="nofollow ugc">Fabrice Bellard</a> a genius?</p>

<p>If coding is easy, why are people angry at AI (or anyone else) copying their code? Why do they act like they&#39;ve poured their sweat, soul, and copious amounts of time into something so trivial?</p>

<p>If coding is easy, why do many now feel like their identity and professional purpose are being stripped away from them?</p>

<p>If coding is easy, why is software so damn buggy?</p>

<h2 id="if-figuring-out-what-to-build-is-the-hard-part">If figuring out what to build is the hard part...</h2>

<p>If deciding what to build is the hard part, why do so many product managers seem clueless? Why aren&#39;t there rigorous 10-step interviews for them? Why aren&#39;t they getting paid more than the developers?</p>

<p>If deciding what to build is the hard part, why aren&#39;t market researchers, usability experts and—hell, customer success—considered rockstars in a software company? If “understanding the customer” is harder, why are business analysts looked down on as pencil pushers?</p>

<p>If implementation is easy and finding demand is harder, why are programmers upset when the salespeople promise a new feature to a customer to close the sale? They&#39;ve found a genuine demand, something people will pay for!</p>

<p>If coding is easy, why doesn&#39;t everyone just build ten variations of a thing and see which pans out?</p>

<h2 id="there-s-no-median-programmer">There&#39;s no median programmer</h2>

<p>Another cliché comment is “<em>most work in software development is talking to stakeholders, understanding the customer&#39;s needs, and having clarity on the priorities</em>”.</p>

<p>I have met many programmers throughout my career, and very few of them want to talk to stakeholders, much less customers (exceptions are freelancers and founders, especially of software development shops). And, “having clarity on the priorities” boils down to “just tell me what to do and don&#39;t switch it up every two days”.</p>

<p>Some software developers do say “I don&#39;t write code, I solve customer&#39;s problems”. But then they turn around and start to opine on monads, memory safety, and DRY principles, while their understanding of the customer is a made-up “user persona”, and they think “affordance” is the money your parents used to give you on weekends so you could go out and have a good time.</p>

<p>Yet others will say “Software development is theory building”. Programs are actually proofs (as in, mathematical proofs). Every commit should tell a story. And solving a customer&#39;s problem by FTPing a PHP file is a cardinal sin.</p>

<p>I don&#39;t mean to imply there are no developers that simultaneously care deeply about the craft of software development and really empathize with the customer. I do believe they might want to see a professional about a split personality disorder, tho. (<em>post-scriptum</em> edit: this sentence was way overboard; I meant it as tongue-in-cheek “very few people can do that” and I do actually advocate for that in the next section. <em>mea culpa</em>)</p>

<h2 id="what-is-important">What <em>is</em> important?</h2>

<p>I do believe that talking to users, understanding their experience, empathizing with them, solving customers&#39; problems and having all the stakeholders on the same page is critical to the success of a software project.</p>

<p>I also believe that creating good code is a craft that requires skill, patience, attention to detail, experience and wisdom, and that it will continue to be relevant in the times ahead.</p>

<p>¿Por qué no los dos?</p>

<p>To the extent that we can pull it off, I think we should aim for both. A deep understanding of the system we&#39;re building, together with a deep understanding of <em>why</em> we&#39;re building it.</p>

<p>Loudly proclaiming that “code is easy” or, at the opposite end, “code is art, a creative human expression that cannot be automated”, is just burying our heads in the sand.</p>

<p>It&#39;s cope. And you don&#39;t want cope, you want to thrive.</p>

<p>By this, I don&#39;t mean “jump on the LLM bandwagon.” I don&#39;t mean “become a manager of fleets of AI agents.” I also don&#39;t mean “AI-generated code is stolen slop garbage, fight it with tooth and nail, the bubble will pop soon enough anyways.”</p>

<p>But do recognize we&#39;re in the middle of an industry-wide tectonic change. We need to figure out how to adapt. We need to understand what is likely to change and what never changes.</p>

<h2 id="what-doesn-t-change">What doesn&#39;t change?</h2>

<p>Software will be getting more complex. Software will always need maintenance: bit-rot is a fact of life. So is entropy. Technology (hardware and software) will move forward, for better or worse. The tower (skyscraper?) of abstractions grows ever higher.</p>

<p>Users will always want more and be prepared to spend less. They still won&#39;t know how to relay their needs and wants. Worse, they still won&#39;t know exactly what they want. The disconnect between the customers (who actually pay for the software) and users (who use it) will still be here, as will the tension between the needs of the business and the needs of its customers.</p>

<p>Also: there will never be a shortage of snake oil salesmen. Tech du jour comes and goes (I&#39;m still waiting for the new VR renaissance!)</p>

<h2 id="what-changes">What changes?</h2>

<p>Programmers have been in the business of disrupting our own industry since the beginning. Nobody uses punch-cards any more. Very few people need to code in assembly, or COBOL. Those decades spent fighting memory bugs in C or C++, with the scars to prove it, are worthless in the age of Rust, Go, Python and JavaScript.</p>

<p>I&#39;m old enough to appreciate <a href="https://valgrind.org/" rel="nofollow ugc">valgrind</a> or remember <a href="https://www.php.net/manual/en/function.mysql-escape-string.php" rel="nofollow ugc"><code>mysql_real_escape_string()</code></a> from the PHP4 era—stuff I&#39;ll never again need in my life. And that wasn&#39;t even so long ago! I narrowly missed the dBase, Clipper, HyperCard and Access era, technologies which I can still spot operating in shops, cafes, or a dusty, once beige and now golden-brown, midi-tower still happily running some bespoke biz solution (backups? what backups?)</p>

<h2 id="how-do-we-thrive">How do we thrive?</h2>

<p>Accept that change happens. Be equal parts curious and critical about the new stuff.</p>

<p>Understand there&#39;s a lot of hype and try to discriminate between hot air and what really works (and to what extent). Also be aware of ever-shifting goalposts: stand back and look at the past year, or five, and assess the velocity of change (technical, economic, societal).</p>

<p>Your role and your responsibilities <em>will</em> be changing. Be willing to invest time and energy into better understanding fields or roles adjacent to yours.</p>

<p>If you&#39;re a senior developer, don&#39;t just find solace in deepening your expertise. Learn about user experience, customer interviews, or business strategies for the companies in your domain. It will help you gain a better appreciation of all the work done to put a piece of software into users&#39; hands, whether or not you&#39;ll actually ever have to do any of those other bits.</p>

<p>If you&#39;re just starting or are junior in your role: invest in deepening your understanding of how software works. Understanding pointers, recursion, or memory hierarchy will help you even if you&#39;re a JavaScript developer. Understanding network protocols and how HTTP works will be useful even if you&#39;re building WordPress plugins. Do leetcode and learn about algorithms and data structures even if you don&#39;t need to. Don&#39;t be afraid to ask <em>why</em> and <em>how exactly</em>.</p>

<p>For inspiration, here are a few books and other resources that might be helpful:</p>
<ul><li><a href="https://en.wikipedia.org/wiki/Structure_and_Interpretation_of_Computer_Programs" rel="nofollow ugc">Structure and Interpretation of Computer Programs</a> (<a href="https://web.mit.edu/6.001/6.037/sicp.pdf" rel="nofollow ugc">PDF</a>)</li>
<li><a href="https://www.crackingthecodinginterview.com/" rel="nofollow ugc">Cracking the Coding Interview</a></li>
<li><a href="https://en.wikipedia.org/wiki/The_Mythical_Man-Month" rel="nofollow ugc">The Mythical Man-Month</a></li>
<li><a href="https://www.goodreads.com/book/show/53138083-working-backwards" rel="nofollow ugc">Working Backwards</a></li>
<li><a href="https://teamtopologies.com/book" rel="nofollow ugc">Team Topologies</a></li>
<li><a href="https://7powers.com/" rel="nofollow ugc">7 Powers</a></li>
<li><a href="https://en.wikipedia.org/wiki/The_Soul_of_a_New_Machine" rel="nofollow ugc">The Soul of a New Machine</a></li>
<li><a href="https://www.aprildunford.com/books" rel="nofollow ugc">Obviously Awesome</a></li>
<li><a href="https://en.wikipedia.org/wiki/The_Design_of_Everyday_Things" rel="nofollow ugc">The Design of Everyday Things</a></li>
<li><a href="https://en.wikipedia.org/wiki/Don%27t_Make_Me_Think" rel="nofollow ugc">Don&#39;t Make Me Think</a></li>
<li><a href="https://www.goodreads.com/book/show/58046715-continuous-discovery-habits" rel="nofollow ugc">Continuous Discovery Habits</a></li>
<li><a href="https://www.momtestbook.com/" rel="nofollow ugc">The Mom Test</a></li></ul>

<h2 id="one-more-thing">One more thing</h2>

<p>Whoever you are, don&#39;t outsource your understanding, judgement, empathy and taste to AI. Don&#39;t <a href="https://blog.senko.net/your-code-is-your-responsibility-even-if-ai-wrote-it" rel="nofollow ugc">abdicate your responsibility</a>. Don&#39;t be a <a href="https://gruhn.me/blog/2026-08-03/" rel="nofollow ugc">meat proxy</a>.</p>

<hr/>

<p>PS. Lots of interesting and insightful comments over at <a href="https://news.ycombinator.com/item?id=49222189" rel="nofollow ugc">Hacker News</a> and <a href="https://lobste.rs/s/iixlsm/code_was_never_hard_part_is_insult_all" rel="nofollow ugc">Lobsters</a>—the post really hit a nerve. Fascinating how many different experiences people have and the varying definitions of <em>coding</em>, <em>programming</em>, <em>development</em> and <em>engineering</em> they use.</p>
]]></content:encoded>
      <guid>https://blog.senko.net/code-was-never-the-hard-part-is-an-insult-to-all-programmers</guid>
      <pubDate>Sat, 08 Aug 2026 09:18:22 +0000</pubDate>
    </item>
    <item>
      <title>Open-weights AI models have become good enough</title>
      <link>https://blog.senko.net/open-weights-ai-models-have-become-good-enough?pk_campaign=rss-feed</link>
      <description>Over the past week I&#39;ve played around with Kimi K3 by Moonshot AI and Qwen 3.8 Max by Alibaba. Both are large Chinese open-weight models ...</description>
      <content:encoded><![CDATA[<p>Over the past week I&#39;ve played around with <a href="https://www.kimi.com/blog/kimi-k3" rel="nofollow ugc">Kimi K3</a> by Moonshot AI and <a href="https://x.com/Alibaba_Qwen/status/2078759124914098291" rel="nofollow ugc">Qwen 3.8 Max</a> by Alibaba.</p>

<p>Both are large Chinese open-weight models (weights promised to be released soon) and tout benchmarks showing they&#39;re as capable as the frontier western models (Fable 5 by Anthropic and GPT-5.5 Sol by OpenAI).</p>

<p>I wouldn&#39;t go that far, but these are really capable models. In my AI-coding tests, both have performed really well. Compare the test mini-games on <a href="https://senko.net/vibecode-bench/" rel="nofollow ugc">my vibe-coding benchmark</a> generated by <a href="https://senko.net/vibecode-bench/2026/rts-fable-5.html" rel="nofollow ugc">Fable</a>, <a href="https://senko.net/vibecode-bench/2026/rts-gpt-5.6-sol.html" rel="nofollow ugc">Sol</a>, <a href="https://senko.net/vibecode-bench/2026/rts-kimi-k3.html" rel="nofollow ugc">K3</a> and <a href="https://senko.net/vibecode-bench/2026/rts-qwen-3.8-max.html" rel="nofollow ugc">Qwen</a>.</p>

<p>In a more serious test, building a web app to a provided spec, K3, Qwen and Fable generated very similar results. An AI judge (GPT-5.5 Sol via Codex) gave a positive code review to all three (of course, it did find some problems and room for improvement), with Fable being slightly better.</p>

<h2 id="good-enough-is-great">“Good enough” is great!</h2>

<p>What I am most excited about, though, is that the open-weights models have reached the threshold of “good enough” for serious AI-assisted engineering. The Anthropic and OpenAI models reached this around October-November last year, and it unleashed a huge adoption wave across the software development industry.</p>

<p>But there was always a possibility of a rug-pull: either through intelligence (tokens) becoming too expensive (and not included via standard subscriptions), models being explicitly withheld (as with the Fable fiasco) or nerfed (Fable and Sol, with their overly aggressive allowed-usage filters).</p>

<p>Open models are insurance against this. Although these models are too large to be reasonably hosted locally, there&#39;s already a cottage industry of third-party inference providers. And for organizations that really need to keep all their data close, there exists an (expensive) option of hosting it themselves.</p>

<h2 id="the-chinese-scare">The Chinese scare</h2>

<p>There&#39;s <a href="https://stratechery.com/2026/whos-afraid-of-chinese-models/" rel="nofollow ugc">much angst in the West about these models being Chinese</a>. As such, they might reflect a kind of world-view that&#39;s not completely compatible with traditional western values. There&#39;s also a question of breaking western labs&#39; terms of service and copying from them.</p>

<p>Both Anthropic and OpenAI, having scraped the entire public internet and trained on copyrighted material such as books, don&#39;t get to complain. My personal view is that training on copyrighted material is fair use, but you don&#39;t then get to disallow training on your API results.</p>

<p>The censorship and model alignment that follows the ideas of the Chinese government is real (though let&#39;s not downplay the alignment done by western labs for different topics). But with these models being open, the censorship layer can relatively easily be stripped away by post-training.</p>

<h2 id="western-open-models">Western open models</h2>

<p>The Chinese models (Kimi, Qwen, DeepSeek, GLM) are the most prominent, but they don&#39;t have a monopoly on open models: one of the most popular small models is Google DeepMind&#39;s <a href="https://deepmind.google/models/gemma/" rel="nofollow ugc">Gemma</a>.</p>

<p>Then there are models like <a href="https://poolside.ai/blog/introducing-laguna-s-2-1" rel="nofollow ugc">Laguna S2.1</a>, <a href="https://research.nvidia.com/labs/nemotron/Nemotron-3/" rel="nofollow ugc">Nemotron 3</a> and <a href="https://mistral.ai/news/mistral-small-4/" rel="nofollow ugc">Mistral Small 4</a> and many more, which usually don&#39;t make a big splash but are nevertheless very capable.</p>

<p>Taken together, there is a healthy ecosystem of open LLMs of various sizes (from 4B to 3T params). Most of them are open-weights (meaning you only get the weights), but some are fully open source (you get the training data as well).</p>

<p>As these models reach the point of “good enough” in their niches or for specific use cases (like AI coding I started this post with), they are already solid alternatives to the big popular models.</p>

<h2 id="hardware">Hardware</h2>

<p>The one big difference between open-weights models and free software, in terms of user experience, is which hardware you need. Today (mid-2026), the average hobbyist has no way of running the large models in any usable way.</p>

<p>Computers like Mac Studio, NVidia DGX Spark or the AMD Strix Halo, with unified memory, cost on the order of $5000 and can only run models up to 100B-200B params. Kimi K3 is 2.8T and Qwen 3.8 Max is 2.6T!</p>

<p>I&#39;m optimistic, though. There&#39;s so much money sloshing around AI-capable hardware (NVidia has a $5T market cap and 85% profit margins, high-bandwidth memory (HBM) makers are also rolling in money) that I have no doubt there&#39;s going to be increased competition, increased performance, and reduced pricing in a few years&#39; time (it takes time to build fabs, sorry).</p>
]]></content:encoded>
      <guid>https://blog.senko.net/open-weights-ai-models-have-become-good-enough</guid>
      <pubDate>Wed, 22 Jul 2026 10:17:03 +0000</pubDate>
    </item>
    <item>
      <title>Mirroring a Wayland desktop region for easy screen sharing</title>
      <link>https://blog.senko.net/mirroring-a-wayland-desktop-region-for-easy-screen-sharing?pk_campaign=rss-feed</link>
      <description>I&#39;m a desktop Linux user — specifically Debian GNU/Linux with GNOME and Wayland. I also do a fair number of video calls involving screen ...</description>
      <content:encoded><![CDATA[<p>I&#39;m a desktop Linux user — specifically Debian GNU/Linux with GNOME and Wayland. I also do a fair number of video calls involving screen sharing. And I have an ultra-wide screen monitor (32:9).</p>

<p>This can sometimes be a challenging combination. I want to share a portion of my screen and any window shown there, but most of the software I use only supports sharing either a single window, or an entire monitor.</p>

<p>If I share a single window, it&#39;s annoying to share a different window every time I need to show something else. If I share the entire monitor, the aspect ratio means the other participants will have a pretty bad time watching it.</p>

<h3 id="year-of-screen-sharing-on-the-linux-desktop">Year of screen sharing on the Linux Desktop</h3>

<p><a href="https://wayland.freedesktop.org/" rel="nofollow ugc">Wayland</a> has been around for almost 20 years, but the support in the wider ecosystem is still spotty. One reason is enhanced security (or — introducing <em>any</em> security at all) compared to the old X11 protocol, which breaks some use cases, screen sharing being one of them.</p>

<p>The Wayland way to do screen sharing is through <a href="https://pipewire.org/" rel="nofollow ugc">PipeWire</a> and <a href="https://wiki.archlinux.org/title/XDG_Desktop_Portal" rel="nofollow ugc">XDG Desktop Portal</a>. In a nutshell, the application that wants access to the screen asks the environment, which checks with the user (through a modal dialog) and then allows limited access.</p>

<p>The access can be to a single window, a region of the desktop, or a monitor — but most of the apps only support single window or entire monitor share. Not sure why: perhaps the region share is a relatively new addition most apps haven&#39;t picked up yet.</p>

<p>In any case, it&#39;s pretty annoying, and searching on the internet didn&#39;t yield any easy workaround. So I built one.</p>

<h3 id="waymirror">Waymirror</h3>

<p>My workaround was to build a small app that uses the same components to get access to a selected region, then simply show that in a window. Other apps can then screen-share <em>that</em> window.</p>

<p>The app is called Waymirror and is available <a href="https://pypi.org/project/waymirror/" rel="nofollow ugc">from PyPI</a> and <a href="https://github.com/senko/waymirror" rel="nofollow ugc">GitHub</a>. If you want to try it out, note it depends on a number of GNOME and FreeDesktop system dependencies.</p>

<p>Assuming you have the dependencies, the usage is:</p>

<pre><code>pip install waymirror

waymirror left # mirror the left half of the current monitor
waymirror right # mirror the right half of the current monitor
waymirror # select a region to share
waymirror 640x480+100+100 # define the region as X geometry
</code></pre>
]]></content:encoded>
      <guid>https://blog.senko.net/mirroring-a-wayland-desktop-region-for-easy-screen-sharing</guid>
      <pubDate>Thu, 25 Jun 2026 16:16:00 +0000</pubDate>
    </item>
    <item>
      <title>May quick-takes</title>
      <link>https://blog.senko.net/may-quick-takes?pk_campaign=rss-feed</link>
      <description>Recap of my short posts on LinkedIn in May Unlimited Power, Limited Focus In the age of AI, focus is essential. We now have, to quote Dar...</description>
      <content:encoded><![CDATA[<p><em>Recap of my short posts on LinkedIn in May</em></p>

<h3 id="unlimited-power-limited-focus-https-www-linkedin-com-feed-update-urn-li-activity-7456255012931993600"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7456255012931993600" rel="nofollow ugc">Unlimited Power, Limited Focus</a></h3>

<p>In the age of AI, focus is essential.</p>

<p>We now have, to quote Darth Sidious, <strong>unlimited power</strong> (subsidized by Anthropic and OpenAI). But we don&#39;t have unlimited focus.</p>

<p>In fact, many of us behave like squirrels on caffeine. We can run very fast, burning the midnight candle at both ends, wasting our time and tokens on stuff that, had we paused to think about it, would not pass muster.</p>

<p>Code is a liability — always has been. We&#39;re now printing the liability with such abandon that Leeroy Jenkins would look sheepish in comparison.</p>

<p>With everyone having this awesome tool at their disposal, how can you stand out?</p>
<ol><li>Use your judgement, wisdom, and taste.</li>
<li>Focus. Keep your eyes on the ball.</li>
<li>Understand that AI is a good servant, but a bad master.</li></ol>

<p>And FFS, don&#39;t brag about burning tokens. Brag about what you&#39;ve achieved with the tokens.</p>

<h3 id="ai-is-technology-not-a-feature-https-www-linkedin-com-feed-update-urn-li-activity-7458414138810114048"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7458414138810114048" rel="nofollow ugc">AI Is Technology, Not a Feature</a></h3>

<p>If you run, or work for, a tech startup, you&#39;ve seen this: customers, investors, and bosses ask you “do you have AI yet?”. There&#39;s a lot of unhealthy pressure to keep up with the competition and ship “AI features”.</p>

<p>AI is great technology, but it&#39;s not “a feature”. In fact, the best implementations out there are where you wouldn&#39;t even notice it (if it weren&#39;t so prominently showcased).</p>

<p>The fact of (startup) life is, you&#39;re pressured into both adopting <em>and</em> marketing your use of AI. It&#39;s on you to find the good use cases that fit nicely into your product.</p>

<p>I explore the conundrum in <a href="https://senkorasic.com/articles/ai-is-tech-not-feature" rel="nofollow ugc">AI is technology, not a feature</a>.</p>

<h3 id="code-review-is-broken-https-www-linkedin-com-feed-update-urn-li-activity-7458824470833811458"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7458824470833811458" rel="nofollow ugc">Code Review is Broken</a></h3>

<p>In <a href="https://crawshaw.io/blog/agent-principal-agent" rel="nofollow ugc">The agent principal-agent problem</a>, David (cofounder of Tailscale, founder of Exe) raises an uncomfortable question: do we need code reviews?</p>

<p>This echoes Ankit&#39;s (founder of Aviator) <a href="https://www.latent.space/p/reviews-dead" rel="nofollow ugc">How to kill the Code Review</a> and Justin&#39;s (cofounder of StrongDM) <a href="https://factory.strongdm.ai/" rel="nofollow ugc">Software Factories</a> essays. In particular, Justin argues that “Code must not be written by humans” and “Code must not be reviewed by humans.”</p>

<p>I also touched on this in my <a href="https://senkorasic.com/articles/rethinking-code-reviews" rel="nofollow ugc">Code reviews in the age of AI</a> article a few months ago, where I argue we should put more emphasis on reviewing the plans/specs, and only do limited, targeted manual code reviews.</p>

<p>If you&#39;re leading a software engineering team and don&#39;t want to sacrifice quality for velocity, this question is worth thinking about, and discussing with your team.</p>

<h3 id="style-convergence-in-ai-generated-design-https-www-linkedin-com-feed-update-urn-li-activity-7459852420530163714"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7459852420530163714" rel="nofollow ugc">Style Convergence in AI-generated Design</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7459852420530163714.jpg" alt=""/></p>

<p>The curious case of the “contemporary Fraunces serif magazine editorial” design in Claude.</p>

<p>Here&#39;s what happened to me yesterday (all in one day):</p>

<p>I had to quickly whip up a website, and I turned to Lovable for some inspiration. The design looked nice but too close to FT/WSJ and even had some undertones of Anthropic.</p>

<p>So I fired up Claude Design and asked it to do the same (based on the same brief containing no design guidelines). The results looked eerily similar.</p>

<p>I fired up Claude Code (CLI). Same thing. I was sure I&#39;d somehow triggered this with my website brief, even without actually specifying it.</p>

<p>Then I noticed a LinkedIn post from someone featuring a virtually identical design. Now I was sure it wasn&#39;t my prompting or something contaminating my agents&#39; contexts.</p>

<p>This looked like a new version of the “Bootstrap design” effect, where AI agents mostly churn out similar pages. I told Claude to go in a completely different direction, and called it a day... or so I thought.</p>

<p>Because hours later, I saw <em>the same damn design</em> on Twitter, from none other than Peter Steinberger, author of OpenClaw, talking about vibe-slopped web pages.</p>

<p>To top it all off, in the evening I was helping my kid with some physics lessons and decided to create an interactive visualiser in Claude Web, to help with the understanding behind the math. Guess what: <em>same friggin&#39; design!</em></p>

<p>Now, I understand LLMs are not creative. But I also understand they are supposed to be non-deterministic, and this was pretty damn deterministic in my book!</p>

<p>I suspect the culprit is the combination of how Claude was trained, and Anthropic&#39;s <a href="https://github.com/anthropics/claude-code/blob/main/plugins/frontend-design/skills/frontend-design/SKILL.md" rel="nofollow ugc">“frontend-design” skill</a>, which probably got loaded into all of those places. Lovable doesn&#39;t disclose which LLMs they use (to the best of my knowledge), but this would pretty strongly indicate Claude, <em>and</em> with the frontend skill enabled. As for why Claude prefers this: since it is similar to what Anthropic itself uses, I suspect there was some (unintentional?) stylistic or taste influence (leakage?) from their designers into the training data.</p>

<p>Whatever the cause may be, somehow, Claude overwhelmingly tends to converge on this type of design.</p>

<p>I mean, it&#39;s a nice design — it was nice once. Now I&#39;m seeing it everywhere; it&#39;s like the em dash all over again.</p>

<h3 id="bun-ported-from-zig-to-rust-by-ai-https-www-linkedin-com-feed-update-urn-li-activity-7460935141213954048"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7460935141213954048" rel="nofollow ugc">Bun, Ported from Zig to Rust by AI</a></h3>

<p>This is cool and scary: Bun, a JavaScript platform (alternative to NodeJS), has been (officially) <a href="https://github.com/oven-sh/bun/pull/30412" rel="nofollow ugc">ported from Zig to Rust using AI</a>.</p>

<p>We&#39;ve seen LLM-powered rewrites (chardet, justhtml) before. This is different because of scale. Bun is a large and much more complex project than either of those, and it was apparently ported in two weeks.</p>

<p>Projects at this scale (like Claude&#39;s Compiler or Cursor&#39;s browser) have so far been interesting experiments, but a far cry from production-ready code.</p>

<p>The merged PR has a million lines of new code, which presumably haven&#39;t been human-reviewed. A week ago the author claimed 99.8% of all tests passed (on Linux). Looks like it took them only a week to get to 100% across all supported platforms.</p>

<p>More quietly, a similar rewrite has been happening at Ladybird, a new browser with its own from-scratch engine (i.e. not just a wrapper for Chrome, Firefox, or another existing browser). It was originally written in C++, and they&#39;ve started (slowly) porting it to Rust with AI&#39;s help — not automatically.</p>

<p>A few thoughts:</p>
<ul><li>first public, successful large-scale project using fully autonomous, non-reviewed AI — we&#39;ll probably see more</li>
<li>robust test suites are a must — this would be a non-starter otherwise</li>
<li>Rust, and to a lesser extent Go, are becoming the go-to languages for AI-assisted rewrites</li>
<li>this would have been at <em>least</em> a year-long project without AI</li></ul>

<h3 id="happy-first-birthday-cijene-api-https-www-linkedin-com-feed-update-urn-li-activity-7461034624941608961"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7461034624941608961" rel="nofollow ugc">Happy First Birthday, Cijene API</a></h3>

<p><em>This post is in Croatian</em></p>

<p>🎂 Sretan rođendan, <a href="https://cijene.dev" rel="nofollow ugc">Cijene API</a>! 🎂</p>

<p>Prije točno godinu dana stupila je na snagu Odluka o objavi cjenika, čime je Vlada RH obvezala velike trgovačke lance da dopuste automatsko preuzimanje (scraping) i obradu cijena.</p>

<p>Vidjevši tu objavu, pokrenuo sam Cijene API kao zabavan vikend projekt.</p>

<p>Godinu dana kasnije, podržavamo 29 trgovačkih lanaca, preko tisuću poslovnica, prikupljamo preko 10 milijuna cijena dnevno, a u ovih godinu dana prikupili smo ukupno preko 3 milijarde cijena.</p>

<p>Sve ovo javno je dostupno kroz naš API na <a href="https://cijene.dev" rel="nofollow ugc">cijene.dev</a> i kroz dnevne arhive cijena.</p>

<p>Imamo nekoliko desetaka API korisnika (komercijalnih i nekomercijalnih), projekt pokriva svoje troškove servera, a imamo i malu, ali aktivnu zajednicu ljudi koji doprinose, što prijavom problema, što kodiranjem — hvala im od srca!</p>

<p>Posao nije gotov. Dok ovo tipkam imam jedan PR u pripremi za promjenu CSV formata jednog lanca i drugi PR za dodavanje još izvora. Serveri trgovačkih lanaca ponekad se ruše ili kasne s objavama. Naš API bi mogao biti još bolji.</p>

<p>Ali kroz ovu prvu godinu, mogu reći da su (posredno) stotine tisuća građana Hrvatske vidjele ili iskoristile informacije koje su prošle (na ovaj ili onaj način) kroz Cijene API.</p>

<p>Idemo dalje!</p>

<h3 id="running-a-frontier-llm-on-your-own-hardware-https-www-linkedin-com-feed-update-urn-li-activity-7461686979047542784"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7461686979047542784" rel="nofollow ugc">Running a Frontier LLM on Your Own Hardware</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7461686979047542784.jpg" alt=""/></p>

<p>The shape of things to come: how to run a frontier-level LLM on consumer hardware. Salvatore Sanfilippo (a.k.a. <a href="https://github.com/antirez/ds4" rel="nofollow ugc">Antirez</a>, best known as the author of Redis) has been working on a new LLM inference engine (like llama.cpp, VLLM, or HF transformers), optimized for the DeepSeek V4-Flash model.</p>

<p>The DeepSeek V4-Flash is a smaller version of DS&#39; newest model, released a few months ago, with 284B params, of which 13B are active, supporting a 1M context window.</p>

<p>What Antirez is doing with his project (dubbed DwarfStar 4) is to really push the hardware to its full extent, hyper-optimized for the model in question, whereas other engines are more generic and thus can&#39;t make tradeoffs to speed up one LLM version.</p>

<p>DwarfStar 4 (DS4) supports Metal (Apple) and CUDA (NVIDIA) architectures, with ROCm (AMD) on the way. But it also runs on CPU, if you&#39;re <em>really</em> patient.</p>

<p>The other piece of the puzzle is the hybrid 2-bit quantization for the model, i.e. using only ~2 bits for each floating-point parameter (down from the original 16). If you know how floating point works, this feat alone is mind-boggling, but it works, with noticeable but acceptable quality tradeoffs.</p>

<p>The end effect is that you can run DS4 on a machine with 96+ GB of VRAM, like the NVIDIA DGX Spark, or 96+ GB of unified RAM (like the Macs). Or, if you&#39;re really patient and only do it for testing — on ANY machine with 128GB of RAM.</p>

<p>Well, I&#39;m patient and only do it for testing, and I happen to have 128GB of RAM, so I took it for a spin. Well, at ~1 token/s, “rotating slowly” would be a more apt term :) It&#39;s supposed to run at quite usable 20-30 t/s on Macs or Sparks.</p>

<p>It did a great Minesweeper (no surprise there), and a working RTS! Don&#39;t get me wrong, it was barely working and full of glitches. Much worse than the DeepSeek-V4 Pro version tested via the official API.</p>

<p>But — it&#39;s fully local. And <a href="https://gist.github.com/senko/e52f639187070168a5006e3a0a675752" rel="nofollow ugc">it does work</a>.</p>

<p>Now the only thing we need is a bit faster (and cheaper) hardware. Let&#39;s go, Moore, we&#39;re counting on you!</p>

<h3 id="the-frog-has-already-been-boiled-https-www-linkedin-com-feed-update-urn-li-activity-7463483902892900352"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7463483902892900352" rel="nofollow ugc">The Frog Has Already Been Boiled</a></h3>

<p>People on LinkedIn, X, Facebook and Instagram are worried about the upcoming Google search UI change, removing the links to content sites. I&#39;m here to tell you not to worry — the frog has already been thoroughly boiled.</p>

<p>For years, Google has been a steadfast partner for many — or most — commerce sites on the internet. Through SEO juice, paid ads, or actual word of mouth (remember that, anyone?), Google sent traffic to sites. Users were happy, sites were happy, Google was happy. Now that Google has a competent AI and all the data it can ask for, it doesn&#39;t need to link to sites so much.</p>

<p>AI is just the tip of the iceberg, but Google has been experimenting with this for years. Yet, like addicts, sites needed the juice fix, and did everything Google asked of them in the guise of “best practices”.</p>

<p>Now the other shoe dropped, chickens are coming home to roost, and the frog is cooked.</p>

<p>It gets better: LinkedIn, X, Facebook, etc. — all the places people use to voice this — are doing the same. Y&#39;all are feeding LinkedIn with data (ironically, this post as well), and it&#39;s staying locked within its walled garden. One example: LinkedIn has recently stopped allowing non-authenticated browsers to even expand the links in posts (meaning I can&#39;t open a link in a private/incognito window, which annoys me to no end).</p>

<p>On X, you can&#39;t open a thread / see replies to a post unless you&#39;re logged in. Combined with pushing users to add links to outside pages in comments, you get a lock-in mechanism.</p>

<p>On Facebook and Instagram, you often can&#39;t see the post unless you&#39;re either on the app or logged in. I purposely do neither, and have no end of traffic lights, fire hydrants and bicycles to click through if I really want to see a shared post.</p>

<p>There&#39;s going to be a feast, and <em>your</em> content is on the menu.</p>

<p>Bon appétit!</p>
]]></content:encoded>
      <guid>https://blog.senko.net/may-quick-takes</guid>
      <pubDate>Sun, 31 May 2026 16:31:37 +0000</pubDate>
    </item>
    <item>
      <title>How I want to use AI</title>
      <link>https://blog.senko.net/how-i-want-to-use-ai?pk_campaign=rss-feed</link>
      <description>There&#39;s a lot of psychosis around AI. My belief is that as a technology, it can both be used in a way that improves the lives of people, ...</description>
      <content:encoded><![CDATA[<p>There&#39;s a lot of psychosis around AI. My belief is that as a technology, it can both be used in a way that improves the lives of people, and leveraged in a way that will harm us.</p>

<p>Like in the parable of two wolves, the one that wins is the one we feed. So how do we feed “good AI”?</p>

<p>Here&#39;s how I, personally, want to use it:</p>

<p><strong>As a tool.</strong> It&#39;s here to help me, not to steal my attention or get fed by me (I have a cat for that). I want to be the active operator, not merely a clueless user or an unsuspecting audience.</p>

<p><strong>To grow.</strong> As a research tool, an infinitely patient tutor who, true, might occasionally be wrong. Not as a crutch to avoid thinking.</p>

<p><strong>To remove the grind.</strong> Some hard things are worth doing. Some are not. Sometimes the grind is just tedious and doesn&#39;t help us grow. Washing machines, dishwashers, and now AI (when used wisely) can free up time and energy for better things.</p>

<p><strong>To stop and smell the roses.</strong> Productivity is output over time. I want the AI productivity boost to give me more time, not just more output. If AI can 10x me, a 10-hour workweek and 250% productivity increase sound vastly more preferable to 10x output.</p>

<p><strong>To care about the details.</strong> More time to dedicate to the small details that usually fall below the quality fold. I want to delight and be delighted and be proud of the result. I want to use AI for <em>anti-slop!</em></p>

<p><strong>For silly things.</strong> Like memes, slop can be fun. Low production effort doesn&#39;t mean low creativity. Not everything in life must be serious, high-brow and a result of man&#39;s struggle.</p>

<p><strong>As a power tool.</strong> Forklifts, power tools, and other machines help us shape our world way beyond our physical capability. Calculators, computers and calendars do the same for our mental capabilities.</p>

<p>I recognize that using AI as a power tool to remove the grind can easily transform into using it as a crutch. Here&#39;s what I keep in mind to avoid falling into that trap:</p>

<p><strong>Focus</strong>. AI makes everything deceptively simple, so it&#39;s tempting to do everything. That&#39;s just wasting time.</p>

<p><strong>Judgement</strong>. AI is a really fancy autocomplete tool. I still want to be in control of thinking, guiding and judging its efforts. AI can give me options, but I get to pick and choose.</p>

<p><strong>Taste</strong>. I&#39;m definitely not an artist, but as a human being I claim to have at least a spark of creativity and taste in me. For whatever it&#39;s worth, it makes all the difference between something human-made and autogenerated slop.</p>

<p><strong>Intuition</strong>. I subscribe to the notion that human intuition is a superb pattern matching machine trained over our lifetime and, in work context, career. Like in the corny Star Wars line, I do want to listen to my feelings.</p>

<p>You may say I&#39;m a dreamer, but it works for me. I do hope I&#39;m not the only one. Care to join me?</p>
]]></content:encoded>
      <guid>https://blog.senko.net/how-i-want-to-use-ai</guid>
      <pubDate>Sat, 30 May 2026 15:41:01 +0000</pubDate>
    </item>
    <item>
      <title>April quick-takes</title>
      <link>https://blog.senko.net/april-quick-takes?pk_campaign=rss-feed</link>
      <description>Recap of my short posts on LinkedIn in April A Big Week for LLMs: Gemma 4, Qwen, Cursor 3, and a Claude Code Leak A lot of LLM excitement...</description>
      <content:encoded><![CDATA[<p><em>Recap of my short posts on LinkedIn in April</em></p>

<h3 id="a-big-week-for-llms-gemma-4-qwen-cursor-3-and-a-claude-code-leak-https-www-linkedin-com-feed-update-urn-li-activity-7445786244946534400"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7445786244946534400" rel="nofollow ugc">A Big Week for LLMs: Gemma 4, Qwen, Cursor 3, and a Claude Code Leak</a></h3>

<p>A lot of LLM excitement this week: Google/DeepMind released Gemma 4, Alibaba unveiled Qwen 3.6-Plus, Cursor has a big new release (v3), Anthropic leaked Claude Code source, and even Microsoft released some new models.</p>

<p>The <a href="https://deepmind.google/models/gemma/gemma-4/" rel="nofollow ugc">Gemma 4 release</a>, latest in Google&#39;s line of small open models (from 2B to 31B parameters — not to be confused with its large Gemini model series), generated much buzz and excitement in the community. A big reason is the license switch to Apache 2.0, clearing some legal blockers for wider (re)use.</p>

<p>I tried Gemma 4 via llama.cpp and ollama, in multiple sizes, and wasn&#39;t very impressed: it did poorly on my simple Minesweeper coding tests and happily hallucinated about facts it didn&#39;t know about. Maybe it&#39;s the quantization or something with my setup, but it performed worse than Qwopus (Qwen3.5 plus distilled Opus 4.6) of similar size.</p>

<p>Speaking of Qwen, Alibaba also released an update, Qwen3.6-Plus. This, like Qwen3.5-Plus, is not an open model, but they did promise smaller open variants soon. I expect it&#39;s going to be a solid incremental improvement on an already great model.</p>

<p>In the land of coding agents, the new Cursor 3 shifts the UX to more “managing your agents” style, closer to what Claude Code app/web are doing, and away from “AI in your IDE”. You can still do that, but it looks like it&#39;s not their focus going forward, begging the question — why use Cursor at all?</p>

<p>Meanwhile, Anthropic is playing DMCA whack-a-mole with Claude Code leaked source code repositories on GitHub. Moreover, there are several projects that attempted “clean room” reimplementation — generate detailed PRDs/specs for the product from the code, then generate a clone from those specs without looking at the code.</p>

<p>Considering you can&#39;t copyright algorithms, data structures, or architectural patterns in software, it&#39;s going to be interesting to see how that story unfolds.</p>

<p>Oh, and apparently Microsoft also released some foundation models, but nobody noticed. To quote HuggingFace&#39;s CEO: “If it&#39;s not either pushing the frontier meaningfully or open-source, no one will care these days.”</p>

<h3 id="what-is-turboquant-actually-https-www-linkedin-com-feed-update-urn-li-activity-7446879631519068160"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7446879631519068160" rel="nofollow ugc">What Is TurboQuant, Actually?</a></h3>

<p>Two weeks ago, Google published <a href="https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/" rel="nofollow ugc">TurboQuant research</a> that promises to make LLMs more memory-efficient, triggering jitters in RAM prices. But what is TurboQuant actually?</p>

<p>Let&#39;s backtrack a bit to give you a better context: the defining feature of modern AIs is something called “attention”: a mechanism by which every token in a chat is influenced by every previous token (the tokens are “paying attention” to the preceding content).</p>

<p>To calculate this, you have to do some matrix multiplications of every token with every other token, for every new token (I&#39;m simplifying a bit here). To do this naively would take a really long time, so LLMs have something called “KV cache” (you can think of it as “attention cache”), caching the multiplication results, speeding up the operation at the cost of a large memory increase.</p>

<p>The larger the context window, the larger the KV cache you need. How large? It depends on the model, but for smaller LLMs, the KV cache can easily be larger than the model weights!</p>

<p>Another important LLM detail you need to be aware of is “quantization”. Basically, the numbers (weights, KV cache, etc) in a neural network are floating-point numbers. The higher the precision, the better the quality and the more memory you need to hold the number. Standard precision used in LLMs is 16 bits, meaning each number takes up 2 bytes. But there are cases when you can lower precision without much quality loss. If you run a “Q4<em>K</em>M” model, you&#39;re using numbers quantized to 4-bit. Although there&#39;s research using 1.58 bits (ternary LLMs), in most cases even 4 is a big quality drop and everything below is unusable.</p>

<p>This is where TurboQuant comes in. With a really simple and clever mathematical trick, the researchers can drop the required precision for the KV cache to 3 bits with negligible quality loss and virtually no overhead.</p>

<p>What&#39;s the trick?</p>

<p>Each KV cache value is a vector of floating-point numbers. A peculiar property of those vectors is that some dimensions (members) have a large absolute value and some very small. Thus, when you reduce the precision of the floats, you destroy a lot of information across the dimensions that had smaller numbers. The trick is to “rotate” the vectors — matrix-multiply them with an arbitrary vector. This has the effect of making the dimensions roughly all the same scale — meaning when you reduce precision, you reduce it for all numbers roughly the same, and by a larger amount. (If this is fascinating, there&#39;s an even better trick, 1-bit residual — read the paper to find out!)</p>

<p>When implemented in an LLM, this means you can run it on smaller hardware, it&#39;s faster (because memory bandwidth is usually the limiting factor), or with larger context windows.</p>

<p>Will this lead to RAM prices collapsing? I wish! But no — everyone wants larger context windows :) It&#39;s yet another important but incremental step in making modern LLMs even better.</p>

<h3 id="don-t-reward-token-burning-https-www-linkedin-com-feed-update-urn-li-activity-7447018867555647488"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7447018867555647488" rel="nofollow ugc">Don&#39;t Reward Token Burning</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7447018867555647488.jpg" alt=""/></p>

<p>This is wrong on so many different levels:</p>
<ol><li>any metric will be gamed (Goodhart&#39;s Law)</li>
<li>“more = better” is rarely correct in software development</li>
<li>“tokenmaxxing”, wasting as much resources as possible, is totally braindead as a metric</li>
<li>Meta being Meta, they had to gamify it to increase engagement</li></ol>

<p>Now instead of trying to build better software, developers are burning through trillions of tokens and billions of dollars to become “a token legend” (presumably there&#39;s some sort of a prize as well).</p>

<p>Would anyone actually caring about their craft want to work in an environment like this?</p>

<p>If you&#39;re in a technical leadership position and your org wants to “incentivize AI adoption” by reporting (or worse, rewarding) token usage — do not do this. This is likely the worst possible way to approach it, and will cause lasting damage to your org.</p>

<p>Source: <a href="https://www.theinformation.com/articles/meta-employees-vie-ai-token-legend-status" rel="nofollow ugc">Meta employees vie for “AI token legend” status</a></p>

<h3 id="the-mythical-claude-mythos-and-ai-vulnerability-research-https-www-linkedin-com-feed-update-urn-li-activity-7447642191876804608"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7447642191876804608" rel="nofollow ugc">The Mythical Claude Mythos and AI Vulnerability Research</a></h3>

<p>Anthropic says they have an awesome new model, Claude Mythos, that&#39;s so powerful they won&#39;t publicly share it. We&#39;ve seen such marketing gimmicks before, but security-minded researchers and developers are taking LLMs seriously.</p>

<p>The new model is supposedly trouncing everything else on various benchmarks and has produced “thousands of high-severity vulnerabilities, including some in every major operating system and web browser.” To avoid proliferation of thousands of 0-day exploits, Anthropic is only making it available to “launch partners” to use it “as part of their defensive security work”.</p>

<p>Sounds like boasting, and I&#39;m skeptical of such unfalsifiable claims. However, both the Linux and FreeBSD kernels have recently patched bugs that have possibly been surfaced by the new model. Simon Willison did some digging and has a <a href="https://simonwillison.net/2026/Apr/7/project-glasswing/" rel="nofollow ugc">good overview of the situation</a>.</p>

<p>Recently Linux kernel maintainers, authors of popular open source software (like Daniel Stenberg, author of curl, who&#39;s pretty much against AI slopware), and various security researchers have been talking about LLMs becoming really good at finding security problems — see <a href="https://sockpuppet.org/blog/2026/03/30/vulnerability-research-is-cooked/" rel="nofollow ugc">Vulnerability research is cooked</a>.</p>

<p>As Thomas Ptacek says in that blog post: “I think we&#39;re living in the last fleeting moments where there&#39;s any uncertainty that AI agents will supplant most human vulnerability research.”</p>

<p>Now I really want to see that mythical beast.</p>

<h3 id="embarrassingly-simple-self-distillation-https-www-linkedin-com-feed-update-urn-li-activity-7448282385768435712"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7448282385768435712" rel="nofollow ugc">Embarrassingly Simple Self-Distillation</a></h3>

<p>Here&#39;s an interesting AI paper Apple released last week: <a href="https://arxiv.org/abs/2604.01193" rel="nofollow ugc">Embarrassingly Simple Self-Distillation Improves Code Generation</a>.</p>

<p>Accepted wisdom around LLM training is that you can&#39;t use LLM output to train the same LLM: if you try to do that (without improving, validating or cleaning the output in some other way), the LLM will get dumber and dumber. This is called model collapse and was documented in a 2024 paper, “AI models collapse when trained on recursively generated data.”</p>

<p>Turns out, it&#39;s not necessarily true. The new Apple paper does exactly that, in a specific domain (code generation), and the model keeps improving! From the paper:</p>

<blockquote><p>Sample solutions from the base model with specified temperature and truncation, then fine-tune on those raw, unverified samples via standard cross-entropy loss. This method requires only a set of problem prompts and the model itself: no human-labeled solutions, no reference answers, no teacher model, no reward model, no verifier, no execution environment, and no reinforcement learning of any kind. Surprisingly, it works.</p></blockquote>

<p>To anyone knowledgeable about how LLMs work, this makes as much sense as a perpetual motion machine. It shouldn&#39;t work. But it does! Why?</p>

<p>Code generation is a type of problem where the LLM must work in two different ways: “fork” mode that creatively thinks of possible solutions, and “lock” mode which has to correctly implement the chosen solution. These two modes require different “temperature” (or “creativity”) levels, but temperature is a single global setting, so it&#39;s always sub-optimal.</p>

<p>The self-distillation approach samples solutions with different temperature settings, nudging the model in different directions, then teaches it those approaches. This doesn&#39;t add any new information into the system, but helps the model better organize the information it already has.</p>

<p>The method really is embarrassingly simple — it doesn&#39;t even test whether the generated code is correct (which is pretty easy to do with code!). I&#39;m sure this will be picked up by all AI labs to improve coding capabilities. And I wouldn&#39;t be surprised if someone manages to generalize it to other domains.</p>

<p>(PS. The paper was released on April 1st, but apparently is not an April Fools&#39; joke, I checked :)</p>

<h3 id="claude-opus-4-7-https-www-linkedin-com-feed-update-urn-li-activity-7450568630892716032"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7450568630892716032" rel="nofollow ugc">Claude Opus 4.7</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7450568630892716032.jpg" alt=""/></p>

<p>Anthropic released <a href="https://www.anthropic.com/news/claude-opus-4-7" rel="nofollow ugc">Claude Opus 4.7</a>, an incremental improvement to their existing 4.6 model (still not the mythical Mythos everyone&#39;s been talking about lately).</p>

<p>I took it for a quick spin using my “LLMCraft” test. It built a pretty solid (bare-bones) WarCraft/StarCraft clone. There were a few bugs which I pointed out in one message, and it managed to solve them in the response. (<a href="https://claude.ai/public/artifacts/71c3f5aa-50f0-4987-b058-fd800e8c6020" rel="nofollow ugc">Playable version</a>, <a href="https://gist.github.com/senko/ee6ca221f301b2f2562a910a617f083d" rel="nofollow ugc">code</a>.)</p>

<p>(Note: my prompt calls for a single-user version with no enemies, just one level, no sound; multiplayer and enemies I want to leave as a challenge for 2027 :)</p>

<p>Besides the incremental improvements across the board (see benchmarks in their post — but take them with a grain of salt, as always), the highlights are an extra-high-effort option, a fix for Claude ignoring instructions (especially in CLAUDE and memory files), and vision improvements.</p>

<h3 id="speaking-at-saastanak-2026-https-www-linkedin-com-feed-update-urn-li-activity-7450586973527871488"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7450586973527871488" rel="nofollow ugc">Speaking at SaaStanak 2026</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7450586973527871488.jpg" alt=""/></p>

<p>Happy to announce I&#39;ll be speaking at <a href="https://saastanak.com/saastanak-2026/" rel="nofollow ugc">SaaStanak 2026</a>!</p>

<p>Honored to join an impressive roster of speakers covering GTM, Product, Engineering and more. I&#39;ll be talking about my experience leveraging AI to build a startup MVP in days instead of weeks or months.</p>

<p>Vibe coding, and AI in general, is impacting everything. But clarity of vision, human judgement, good taste and business sense are more important than ever, and these are exactly the topics explored at SaaStanak. I hope you&#39;ll join me!</p>

<h3 id="28-years-of-microlinux-https-www-linkedin-com-feed-update-urn-li-activity-7451930857449136128"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7451930857449136128" rel="nofollow ugc">28 Years of MicroLinux</a></h3>

<p><em>This post is in Croatian</em></p>

<p>Prije točno 28 godina, napravio sam prvu hrvatsku Linux distribuciju — MicroLinux — sa idejom popularizacije Linuxa na domaćim prostorima.</p>

<p><img src="https://senko.net/microlinux/assets/microlinux.png" alt=""/></p>

<p>MicroLinux je bio mala distra (14MB), mogao se instalirati na postojeću DOS/Windows particiju, a pokretao se iz DOSa. Pojednostavljena (za to vrijeme) administracija sustava bila je na hrvatskom, a uz distribuciju je dolazila i prevedena dokumentacija i korisničke upute.</p>

<p>Godinu dana kasnije izdao sam verziju 1.2, koju sam predstavio i na prvoj DORS/CLUC konferenciji 1999. godine — moje prvo konferencijsko predavanje (true story: nisam znao da se obično pripreme neki slideovi pa sam pričao “na suho” :)</p>

<p>Ovaj vikend ponovno dolazim na DORS/CLUC pa sam se tom prilikom prisjetio MicroLinuxa i uspio iskopati stare arhive (nažalost samo verziju 1.2 — prvu verziju imam negdje na Bugovom CD-u, uključili su je na jedan od prvih CD-ova uz časopis, ali ga nisam nigdje uspio naći).</p>

<p>MicroLinux je dugo bio dostupan na linux.hr serverima (hvala HULK-u na podršci od prvog dana!), a sad sam ga preselio na <a href="https://senko.net/microlinux/" rel="nofollow ugc">senko.net/microlinux</a> zajedno sa QEMU slikom diska za jednostavno isprobavanje unutar emulatora.</p>

<h3 id="three-new-chinese-models-https-www-linkedin-com-feed-update-urn-li-activity-7452342817164550144"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7452342817164550144" rel="nofollow ugc">Three New Chinese Models</a></h3>

<p>Three new open(-ish) Chinese models have been released, to much hype, in the past couple of days: Kimi K2.6, Qwen 3.6-Max-Preview, and (two weeks ago) GLM-5.1.</p>

<p>All three are big &amp; closed models, but promise smaller, distilled, open versions will be out soon. All three excel at various benchmarks and are pretty hyped up by the open-weights crowd, placing them neck-and-neck with the latest models from OpenAI and Anthropic.</p>

<p>That&#39;s an exaggeration. I haven&#39;t done much testing yet, but what I saw was good performance but not really comparable to the latest &amp; greatest. Subjectively, they feel closer to last year&#39;s SOTA crop. And of course, if you run a smaller open variant, there&#39;s an additional quality hit.</p>

<p>But that might not matter for some cases. These new models, across the board, are now good enough for many tasks that don&#39;t really require the best model out there. With that caveat, they can be used for coding, agentic workflows, chatbots and many other simpler tasks.</p>

<p>Whether that&#39;s cost-effective is another matter. Renting or buying your hardware is still a pretty pricey option: if you just rent a GPU for a month and serve a model from there, it&#39;s probably going to cost more than paying for the tokens for a model of equivalent quality. OpenRouter is another option, but is slightly a mess — finding a reliable provider there is not easy.</p>

<p>However, these new models still represent a solid “Plan B” in case the major providers jack up the prices, implode, or start banning uses (like cybersecurity research) left and right.</p>

<p>Links: <a href="https://z.ai/blog/glm-5.1" rel="nofollow ugc">GLM-5.1</a>, <a href="https://qwen.ai/blog?id=qwen3.6-max-preview" rel="nofollow ugc">Qwen 3.6-Max-Preview</a>, <a href="https://www.kimi.com/blog/kimi-k2-6" rel="nofollow ugc">Kimi K2.6</a>.</p>

<h3 id="claude-code-goes-max-spacex-eyes-cursor-https-www-linkedin-com-feed-update-urn-li-activity-7452630409034678272"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7452630409034678272" rel="nofollow ugc">Claude Code Goes Max, SpaceX Eyes Cursor</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7452630409034678272.jpg" alt=""/></p>

<p>Anthropic might soon require a Max subscription for Claude Code; meanwhile, SpaceX buys an option to acquire Cursor for $60B.</p>

<p>Anthropic is A/B testing subscription purchase workflows that seem to indicate users won&#39;t be able to use Claude Code with a ($20/mo) Pro plan. Instead, it would only be supported for Max plans ($100+/mo).</p>

<p>Anthropic just doesn&#39;t have enough capacity for the huge demand growth they&#39;ve been experiencing. They&#39;ve already been setting strict limits which, according to many users, have made Claude Code unusable on Pro plans. They&#39;ve also banned subscription use with other dev tools (like Pi), OpenClaw and similar tools. And we&#39;ve known for a long time that subs are subsidised. So this is nothing unexpected.</p>

<p>Yet even a hint already caused a lot of uproar. They <a href="https://x.com/TheAmolAvasare/status/2046724659039932830" rel="nofollow ugc">pedaled back slightly</a>, but a lot of the developer goodwill Anthropic has gained is getting burned. If OpenAI play their hand right, they might draw a bunch of devs back to Codex. And of course there&#39;s also Google with Antigravity, and a bunch of solid Chinese models (Cursor&#39;s in-house coding LLM is a fine-tuned Kimi K2.5, and a new Kimi was just released the other day).</p>

<p>(As an aside: startup founders, if you&#39;ve ever wondered how post-product-market-fit demand can hurt you — here&#39;s how: you can easily collapse under the weight of it; let&#39;s hope that doesn&#39;t happen to Anthropic.)</p>

<p>Meanwhile in rocket land, SpaceX will <a href="https://x.com/spacex/status/2046713419978453374" rel="nofollow ugc">“work closely with Cursor”</a> and has an option to acquire it wholesale for a trifling $60B (for context, Musk bought Twitter for $44B).</p>

<p>Nothing to do with rockets, of course — SpaceX is now a conglomerate containing Twitter and xAI, and a reasonable reading of the deal is that Musk is looking for synergies between xAI and Cursor.</p>

<p>Now for the tough one: since the X-Code and CodeX brands are already taken, I wonder how Cursor would be (re)branded post-acquisition?</p>

<h3 id="local-models-are-now-good-enough-for-coding-https-www-linkedin-com-feed-update-urn-li-activity-7453000986857574400"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7453000986857574400" rel="nofollow ugc">Local Models Are Now Good Enough for Coding</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7453000986857574400.jpg" alt=""/></p>

<p>The local models are becoming good enough for coding. I&#39;ve tested the newly released Qwen3.6-27B model on a local machine with my Minesweeper task, and it did about as well as Opus 4.5 from a few months ago.</p>

<p>I retired my Minesweeper tests at the end of last year because all models completed it successfully. All <em>large</em> models, that is. Local models, up to ~30B params, still had a lot of trouble.</p>

<p>This is now changing, and if you have beefy enough hardware, local models are now a viable alternative for coding tasks. I took the new Qwen3.6 with 27B parameters (quantized to 4 bits-per-param) for a spin.</p>

<p>It did a perfect Minesweeper implementation, subjectively (from a user POV) equivalent to what Opus 4.5 did in December. Not only that: Claude Code (Opus 4.7) and Codex (GPT-5.4) reviewed Qwen3.6-27B&#39;s vs Opus 4.5&#39;s output and both identified Qwen&#39;s output as much better. Check for yourself: <a href="https://gist.github.com/senko/b7a291c277ad1f48b84d0900f483cf83" rel="nofollow ugc">Qwen</a>, <a href="https://github.com/senko/vibesweeper-2025/blob/main/claude/index.html" rel="nofollow ugc">Claude</a>.</p>

<p>This is a limited-scope test. Opus likely performs much better on large-scale projects (which I intend to test), and you need a beefy machine to run it at a reasonable speed (my own test was pure CPU+RAM, taking half an hour to generate it — you definitely want a large VRAM GPU or a Mac for actual use).</p>

<p>But, still. For the first time, unless you&#39;re vibe-coding complete features in one prompt, new models like Qwen (or the recently released Kimi K2.6) may be just good enough to switch, especially if you dislike the pricing shenanigans Anthropic is doing.</p>

<h3 id="dors-cluc-2026-recap-https-www-linkedin-com-feed-update-urn-li-activity-7453833445258039296"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7453833445258039296" rel="nofollow ugc">DORS/CLUC 2026 Recap</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7453833445258039296.jpg" alt=""/></p>

<p>Had a wonderful time yesterday and today at DORS/CLUC 2026. There were many interesting talks and discussions, but the highlight for me was the hallway track!</p>

<p>I had a chance to see a few of the talks (had to miss several, sadly — will catch up when they&#39;re published online), participate on two panels and present the Cijene API project. I especially enjoyed Luka Kladarić&#39;s talk about the impact of AI on open source and the challenging questions for us all — I was so focused on the talk I forgot to snap a picture :)</p>

<p>I again met many old friends (some of whom I haven&#39;t seen in decades!), reminisced about the old days and caught up on what&#39;s new. For example, did you know we have a hardware fab right here in Croatia? (<a href="https://www.tinkerfab.net/" rel="nofollow ugc">tinkerfab.net</a>)</p>

<p>I also met new friends and was heartened to see newer generations of free and open source and open data enthusiasts involved in various projects.</p>

<p>Thank you for having me, DORS/CLUC — see you next year!</p>

<h3 id="coding-model-shootout-gpt-5-5-kimi-deepseek-qwen-https-www-linkedin-com-feed-update-urn-li-activity-7455254485326630912"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7455254485326630912" rel="nofollow ugc">Coding Model Shootout: GPT-5.5, Kimi, DeepSeek, Qwen</a></h3>

<p><img src="https://s3.us-east-1.amazonaws.com/vault.senko.net/blog/quick-takes/post-7455254485326630912.jpg" alt=""/></p>

<p>There have been a bunch of capable coding models released recently: GPT-5.5, Kimi K2.6, DeepSeek V4 and Qwen 3.6 Max (Preview). I&#39;ve run them all through my new LLMCraft benchmark.</p>

<p>GPT-5.5 is the best of the bunch by far, no question there. It&#39;s tied with Claude Opus 4.7, which I reviewed last week.</p>

<p>Qwen also did pretty well, while both DeepSeek and Kimi produced a less polished and more buggy experience. However, all three managed to produce a solution more-or-less to spec.</p>
<ul><li><a href="https://gist.github.com/senko/ede62589a5c97603abd10bbe70a0263e" rel="nofollow ugc">Kimi K2.6</a> (top-left)</li>
<li><a href="https://gist.github.com/senko/6017195c5a6818dcc8829e12c13df626" rel="nofollow ugc">DeepSeek V4 Pro</a> (top-right)</li>
<li><a href="https://gist.github.com/senko/8b8c52452c2283db65b29b62a48398cf" rel="nofollow ugc">Qwen 3.6 Max Preview</a> (bottom-left)</li>
<li><a href="https://gist.github.com/senko/816194af33efd3947fe4ec2026f7ad7b" rel="nofollow ugc">OpenAI GPT-5.5</a> (bottom-right)</li></ul>

<p>Here&#39;s the kicker: ALL of these models were markedly better than <a href="https://gist.github.com/senko/d87b1a311b300accdbf3de8992e53ca3" rel="nofollow ugc">GPT-5.3-Codex</a>, released just 3 months ago.</p>

<p>I used GPT-5.5 with Codex CLI, Kimi and DeepSeek through Pi.dev with the official API, and Qwen 3.6 Max Preview via Qwen&#39;s chat interface (since I was unable to onboard myself to Alibaba Cloud Platform to get 1st-party API access to Qwen). In the first three cases I only specified the initial prompt (one-shot), while for Qwen I gave it a few bug reports (non-clickable buttons or JS errors) and allowed it to regenerate the output a few (three) times total to fix them, to compensate for the lack of an optimized coding harness.</p>

<p>Note that DeepSeek and Kimi are open-weights models at full sizes, while Qwen3.6 Max Preview is not (Qwen/Alibaba earlier released smaller 27B/35B 3.6 models under an open license).</p>

<p>Across all of these, the raw LLM coding capability is good enough to avoid being beholden to OpenAI, Anthropic or Google (for Gemini), which makes me optimistic about subscription/token prices staying sane (i.e. no large increases).</p>

<p>I can totally see a future where, should the cost of Claude or GPT increase five-fold, I switch to one of these alternatives and accept the quality hit for the budget option.</p>

<h2 id="the-goblins-kept-multiplying-https-www-linkedin-com-feed-update-urn-li-activity-7455575074868228097"><a href="https://www.linkedin.com/feed/update/urn:li:activity:7455575074868228097" rel="nofollow ugc">“The goblins kept multiplying.”</a></h2>

<p>Curious users recently unearthed goblin-suppressing instructions in OpenAI&#39;s Codex CLI, prompting questions, intrigue and theories around the rationale behind it.</p>

<p>OpenAI spilled the beans in a fun, endearing and informative post: <a href="https://openai.com/index/where-the-goblins-came-from/" rel="nofollow ugc">they had a goblin infestation</a>. The goblins came quietly, an innocuous side effect of the fun &amp; quirky personalities feature added in GPT-5.1. They, alongside their gremlin brethren, were given more prominence in the “Nerdy” personality. As OpenAI document:</p>

<blockquote><p>We unknowingly gave particularly high rewards for metaphors with creatures. From there, the goblins spread.</p></blockquote>

<p>Although they&#39;ve cleaned up their training data to lower the incidence of goblins spontaneously materializing in your coding session, it was too late for GPT-5.4, which was already trained, so they applied a hotfix to the system prompt:</p>

<blockquote><p>never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user&#39;s query</p></blockquote>

<p>For fans of mythical creatures and strange beasts, the post does contain instructions for how to remove the hotfix — at least until the next major GPT version.</p>
]]></content:encoded>
      <guid>https://blog.senko.net/april-quick-takes</guid>
      <pubDate>Fri, 01 May 2026 16:11:22 +0000</pubDate>
    </item>
  </channel>
</rss>