Bennet's razor: Explanations should be no more specific than necessary: arxiv.org/abs/2301.129... When several hypotheses fit your data exactly, prefer the one that would still permit the most unseen cases, and don't assume the shortest expression is that one.

The Optimal Choice of Hypothesis Is the Weakest, Not the Shortest

If $A$ and $B$ are sets such that $A \subset B$, generalisation may be understood as the inference from $A$ of a hypothesis sufficient to construct $B$. One might infer any number of hypotheses from $...

arxiv.org

Amdahl's law states that even if AI speeds up the coding part a lot, the overall software development process can only see modest productivity gains. Or can it? Here's how to cheat Amdahl's law: senkorasic.com/articles/ai-...

Opus 5 is really good! In my vibecode testing, it performed as well or better than Fable, while being cheaper. Play and compare here: senko.net/vibecode-ben...

From a friend who joined a fast-growth, later-stage startup: "I was wondering why no one is doing PRDs here. Then I realized that the Head of Product and most PMs are 25-year-olds who don't have the attention span to *read* even a 1-page doc. You lose them after 3 bullet points."

Open-weights models still lag behind the frontier ones, but have become good enough for many tasks (including vibe-coding): blog.senko.net/open-weights...

Open-weights AI models have become good enough

Over the past week I've played around with Kimi K3 by Moonshot AI and Qwen 3.8 Max by Alibaba. Both are large Chinese open-weight models...

blog.senko.net

Kimi K3 is pretty good: roughly on par with Opus 4.8 and GPT-5.5 in my LLMCraft test: senko.net/vibecode-ben...

I wonder if this will make people reconsider their stance or become even more defensive.

Did a few more vibecoding tests with Fable, Opus 4.8 and GPT-5.6 Sol: Minecraft clone, and a flying sim. Fable wins easily, but at 4.5x the cost (!!), while Sol is a noticable improvement over Opus 4.8. See & play all the tests ony my vibecoding benchmarks page: senko.net/vibecode-ben...

OpenAI GPT-5.6 is out and I put the Sol and Terra variants through my "LLMCraft" RTS test, with (expectedly) pretty good results! I'd say Sol (the larger) model is roughly on par with Claude Fable 5. Try it here: senko.net/vibecode-ben... (list of all tests: senko.net/vibecode-ben... )

From a Head of Engineering who spent a week in SF, meeting a bunch of AI startups: "When I was back, I wrote a memo about my impressions. The #1 was how these amazing startups I met: they mostly did not have a business. But we do. So be VERY careful in copying what they do." (cont'd)

Anthropic's Claude Sonnet 5 is out and actually seems pretty good for coding! Here's what it produced (via Claude Code), with one bug I reported, that it found and fixed: senko.net/vibecode-ben... Compare with Opus 4.8, Fable 5, GPT-5.5, GLM-5.2 and others here: senko.net/vibecode-ben...

I find it annoying I can't easily share a screen region in Meet or Teams (in GNOME with Wayland, using Firefox), so I built myself a tool to help with that: blog.senko.net/mirroring-a-...

Mirroring a Wayland desktop region for easy screen sharing

I'm a desktop Linux user — specifically Debian GNU/Linux with GNOME and Wayland. I also do a fair number of video calls involving screen ...

blog.senko.net

It finally happened. Today I had to use AI (Claude) to search my mails in Gmail, whose search has turned to crap.

Name a TV show you’re positive no one remembers but you… www.youtube.com/watch?v=xUYe...

Captain Power Intro

YouTube video by MikeSchwartzWrites

youtube.com

Ivan Habunek@ivan.habunek.com · 2mo ago

Name a TV show you’re positive no one remembers but you…

For those that say you can't get in flow mode with AI: you're mistaken. I've just spent several hours preparing a - wait for it - SLIDE DECK with Claude Design (I focus on the content, leave design to it), without noticing how the time flew by.

Not your servers, not your tokens (today), not your data (tomorrow?) There are many angles to this and none are pretty: 1) Loss of credibility for *any* US company. 2) Anthropic IPO story got considerably worse. 3) Digital sovereignity shifts from "vitamin" to "painkiller". ...

Gergely Orosz@gergely.pragmaticengineer.com · 2mo ago

Well this is massive: - Every non-US company sees how they can be cut off from US vendors with a snap of a finger. Non-US vendors getting free advertisement - What does this mean for US AI labs’ offices and employees outside the US? Reads pretty draconian What a mess

Fable 5 saturated my "LLMCraft" vibe-coding test, mere 6 months after I introduced it: senko.net/vibecode-ben... (you can't see it on the screenshot, but the workers, click effects and fights are animated - and yes, there are mobs to fight!) All tests: senko.net/vibecode-ben...

Hot take: AI intuition is a core skill in the age of AI. This includes: * practical understanding (feeling) of how AI works * experience with good and bad AI input (prompts) * understanding failure modes * recognizing AI slop * knowing when to stop * knowing when not to use it

Local, small, open-weights models are roughly comparable to the frontier ones from a year ago. Google's Gemma 4 12B (released today by Google) performs roughly as well as GPT-4.1 on my Minesweeper test: senko.net/vibecode-ben... Full list of models I've been testing: senko.net/vibecode-ben...

In case you are in the camp of “Andrew Tridgell is vibefucking rsync” please read this.l and adjust your priors. medium.com/@tridge60/rs...

rsync and outrage

I gave up blogging a long time ago (apart from an occasional thing about ArduPilot), I tend to just write code and hope people find it…

medium.com

In, what's perhaps my only public appearance this year, I channeled my inner Jerry Maguire discussing code quality in the age of AI with Elixir Wizards. I also showed off an item from my spring/summer '26 collection, because, what, after all, is fashion, if not a statement 😎️ Links in the comments

We need to rationally discuss the impact of AI on software engineering. Instead we get angry mob coming out with pitchforks like here: github.com/RsyncProject... The bug report is a screenshot from someone who assumes AI broke something. The comments are abysmal. We can, and should, do better.

Please Do Not Vibe Fuck Up This Software · Issue #929 · RsyncProject/rsync

github.com

Is AI a crutch or a power tool? Is it going to replace us, enhance us, or make us stupid? All depends on how we use it. Here's how I, personally, want to use AI: blog.senko.net/how-i-want-t... (Entirely human written, even if I did use a few —'s and one "it's X, not Y" phrase. AI did typo checks.)

How I want to use AI

There'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, ...

blog.senko.net

DwarfStar distributed inference is now on GitHub: you can run 2 bit Flash using 2 64GB machines, or 4 bit Flash with two 128GB machines or 4 64GB, ad so forth. Prefill speed will increase thanks to pipelining. Generation will pay the latency.

Testing Claude Code with Opus 4.8 in "ultracode" mode by implementing an RTS game. Pretty good results: gist.github.com/senko/24d117...

Opus 4.8 effort selector in Claude Code. I have a nagging feeling I'll hit my token quota in the next 35 seconds.

How much overlap is there between X and BlueSky? Have you seen this same post of mine on X? If so, please like or comment here to let me know (since I can't create a poll here).

In Plan 9 in ~2001, we introduced "private" processes that even a sysadmin couldn't access, debug, etc. I just learned Windows Vista added the same in 2007. The difference? We did it to protect keys held by the security agent, while Windows did it to protect media content. Priorities! ¯\_(ツ)_/¯

Results Age xkcd.com/3245/

Kakvu biste računalnu igru smjestili u Zagreb? ZICER i Grad Zagreb organiziraju prvi gradski game jam gdje možete osmisliti svoju ideju, razraditi prototip igre i natjecati se za nagrade. Kao ovogodišnji HACKL hackathon zadali smo za zadatak korištenje gradskog 3D modela u izradi računalne igre.