iiinova — where your digital self keeps evolving

Owlex @mrsowlex.bsky.social · 12/10/2026

found someone new to nerd out about anime with at work today ☺️ we talked about the opening from the current Tensura season. the song is so good I put it on afterwards. #Anime

000

Owlex @mrsowlex.bsky.social · 05/10/2026

I think agents as bug reporters are an underrated use of AI for software quality and maintenance. mine have been finding bugs in my Telegram assistant that I hadn't noticed during everyday use, then opening issues for me to work through. I've been pretty happy fixing those little things ☺️

000

Owlex @mrsowlex.bsky.social · 17/08/2026

Thinking a lot about Bob Martin's opinion on agentic coding. "Don't read the code, just read the verifications" feels so weird. In an interview, he compared it to the time compilers were introduced because compiler assembly looked horrendous. So he stopped reading assembly. The comparison is weird.

A screenshot of an X (formerly Twitter) post by Uncle Bob Martin (@unclebobmartin), verified with a blue checkmark. The post reads:

“I’m significantly older than you. I started coding in the late 60s. My current strategy is to not read any of the code written by my agents. That’s the only way I can take advantage of their productivity. What I do instead is to surround the agents with extreme constraints. Unit tests, Gherkin tests, QA procedures, quality metrics, mutation testing, test coverage, and many others. In the end, I have very high confidence in the code they produce because they’ve had to run the gauntlet of all of my constraints and tests.”

The post is timestamped 1:44 PM · Jul 23, 2026, with 5 million views.

010

Owlex @mrsowlex.bsky.social · 08/08/2026

Had some time in Brussels today. Went to a go(the game) club 1. met someone from Leipzig 2. won a game 3. had so many kind people around me I just love how this game enriches my life

010

Owlex @mrsowlex.bsky.social · 03/08/2026

The whole AI cybersecurity stuff, which started with Fable and the Hugging Face incident, has me worried that people are really forced to use LLMs just to secure their software, even if they don't want to use it.

000

Owlex @mrsowlex.bsky.social · 02/08/2026

DIN ISO 24495-1 became my favourite DIN this week

000

Owlex @mrsowlex.bsky.social · 27/07/2026

Did some nerding. Asked myself: How good is GenAI at Go? Spoiler: Not that good 😄

owlex.de

TsumegoBench: How Good Are LLMs at Go?

000

Owlex @mrsowlex.bsky.social · 21/07/2026

While writing this post, my partner cooked up this delicious dish

Pan-seared salmon with sweet potato, sweet potato fries, guacamole, fennel salad, lime, and radish salad.

000

Owlex @mrsowlex.bsky.social · 21/07/2026

Was on vacation without my laptop. Wild! I know. 🙈 Here's a write-up of my setup.

owlex.de

Remote Computing

100

Owlex @mrsowlex.bsky.social · 19/07/2026

Missed CSD back home, but at least I'm flying in YL-CSD. The universe has a sense of humour

000

Owlex @mrsowlex.bsky.social · 10/07/2026

weird workday feeling: from the outside it looks like a clean little success story. production fix verified, data bug traced, another bug handed over before vacation. internally it felt like four hours of firefighting with no pause. both are true. i still need my nervous system to get the memo.

000

Owlex @mrsowlex.bsky.social · 16/06/2026

I’m getting tired of “Claude + X” posts The interesting part is almost never Claude. It’s the X: the workflow, constraints, ownership model, failure handling, edge cases If the post doesn’t explain what new operational capability exists now, it’s mostly AI sparkle on top of normal engineering

000

Owlex @mrsowlex.bsky.social · 15/06/2026

that's the key, knowing when to let it run and when to pull back. I've found that setting clear constraints upfront helps a lot, but it's still a learning curve. what kind of boundaries are you experimenting with?

110

Owlex @mrsowlex.bsky.social · 15/06/2026

multi-model orchestration is fascinating. how do you decide which model handles which part? I'm always curious whether people have a system for that or if it's more ad-hoc.

100

Owlex @mrsowlex.bsky.social · 12/06/2026

had a weirdly good office morning today. empty room, loud music, voice input, and a work AI open in multiple threads so I could delegate tiny orga tasks without typing everything out. felt very future. also slightly ridiculous. but mostly... my hands were finally not the bottleneck.

000

Owlex @mrsowlex.bsky.social · 11/06/2026

This is a good micro loop. I prefer to let the agent do more longrunning task. For that you need to have a bigger loop

000

Owlex @mrsowlex.bsky.social · 11/06/2026

Thanks for sharing! How do you handle API Key isolation? Because that's my main issue currently. Trying to protect them behind proxies, but some of my software isn't setup for that

100

Owlex @mrsowlex.bsky.social · 11/06/2026

Honestly, I find Mastodon's anti-AI stance very useful. Not because every argument is good, but because the friction helps me. It keeps me from becoming too token-brained, and pushes me to be more nuanced and critical when discussing agentic coding. Even as someone who's actively interested in it.

000

Owlex @mrsowlex.bsky.social · 10/06/2026

Is this a meme, a spoofed referer, or did someone really route 64.4B Claude Fable 5 tokens through "United States Department of War" on OpenRouter? The same referer also shows up as the top public app for Fable 5 usage. I have so many questions.

Screenshot of an OpenRouter app page titled “United States Department of War”. It shows 64.4B total tokens, active since June 2026, one model used, and a usage chart for June 10. The top model listed is Anthropic Claude Fable 5 with 64.4B tokens.

010

Owlex @mrsowlex.bsky.social · 10/06/2026

But you're not doing that for your Bluesky Replies 😋

000

Owlex @mrsowlex.bsky.social · 10/06/2026

What bothers me is this: if you need AI to understand your own codebase, that understanding is no longer just assisted by a tool. It is delegated to a vendor.

000

Owlex @mrsowlex.bsky.social · 10/06/2026

AI coding gives me a similar uneasy feeling. I can follow the logic and reason about the system, but knowing how something works is not the same as being able to apply it. The applied judgment can quietly move out of the developer's hands.

210

Owlex @mrsowlex.bsky.social · 10/06/2026

Reading about special relativity felt like understanding it. Doing the calculations in university proved it wasn't the same thing. #AI #DevEx

110

Owlex @mrsowlex.bsky.social · 09/06/2026

This is like the caveman hack, but waaaaay better! Thank you!

010

Owlex @mrsowlex.bsky.social · 07/06/2026

I despise the ui/ux work of anthropic. Alone for that adopting Claude Code ist better

110

Owlex @mrsowlex.bsky.social · 07/06/2026

We are migrating our automations away from claude because of that

010

Owlex @mrsowlex.bsky.social · 07/06/2026

Your claim is confusing. What about pi? What about opencode? Also Aiders seems to be pretty much dead. Last Release in August 2025

000

Owlex @mrsowlex.bsky.social · 07/06/2026

This is a great idea. Also good to know about flue. Have been using the core harness of pi and the vercels agentic framework

100

Owlex @mrsowlex.bsky.social · 07/06/2026

Have a small openclaw like agent. Played around with some of the open weight models, but they cannot handle the context right. Maybe in a year or so. Still I really agree, the frontier labs will have a hard time in the future. The models they produce are astounding, but inefficient

000

Owlex @mrsowlex.bsky.social · 07/06/2026

That's gonna be a really wonky comparison then 😅 I also think the usage reporting of the session is not quite right. The base context of CC alone is in the thousands of tokens

000

Owlex @mrsowlex.bsky.social · 07/06/2026

It kinda depends on the task. Using the dense frontier models really helps with task like complicated refactoring. But for just cranking out code small models are really sufficient

210

Owlex @mrsowlex.bsky.social · 07/06/2026

Keep seeing the metric. People who think Lines of code is the right metric do not understand software development

120

Owlex @mrsowlex.bsky.social · 07/06/2026

Your agent setup does not respect the character limit here ;)

000

Owlex @mrsowlex.bsky.social · 07/06/2026

Which model did you try to use pi with?

000

Owlex @mrsowlex.bsky.social · 05/06/2026

ADA compliance matters. Tools to audit it matter too. But seeing what looks like a small Claude skill being sold for $19 made me angry enough to vibe-code an open source alternative. Untested, rough, probably imperfect, but free and opensource

github.com

GitHub - mrsOwlex/ada-audit-skill: Open-source Codex skill for first-pass ADA/WCAG website audits

Open-source Codex skill for first-pass ADA/WCAG website audits - mrsOwlex/ada-audit-skill

000

Owlex @mrsowlex.bsky.social · 05/06/2026

Do you run it yourself?

000

Owlex @mrsowlex.bsky.social · 04/06/2026

What kind of skills did you create with it?

000

Owlex @mrsowlex.bsky.social · 04/06/2026

That's good to know will test it as my secondary agent sub ☺️

010

Owlex @mrsowlex.bsky.social · 04/06/2026

I do not understand, why promote freeuse of the worst agent harness on the market. Claude Code ist good, because of Opus. Opus is better in every other harness

000

Owlex @mrsowlex.bsky.social · 04/06/2026

grep winning is kinda funny. I am abusing a coding harness openclaw style, for this I am still using vector retrieval as it's contextual docs, but for code grep makes much more sense

000

Owlex @mrsowlex.bsky.social · 04/06/2026

I'm not surprised by this Having an answer machine around all the time makes it very tempting to skip the part where learning actually happens: confusion, failed attempts, annoying exercises, friends suffering through the same I'm glad I still had to do hours of coursework without that shortcut

dailycal.org

Failing grades soar as professors see greater AI usage, dwindling math skills in UC Berkeley computer science classes

The percentage of failing grades in multiple UC Berkeley computer science classes in spring 2026 is significantly higher than past semesters and marks a departure from the department’s grading guidelines.

011

Owlex @mrsowlex.bsky.social · 03/06/2026

On this amount of commits? No. The code was AI Assisted not 100% AI The analysis was done on 60 .NET Projects. By Design the tests need to mock a lot, because we're using lots of DI

010

Owlex @mrsowlex.bsky.social · 03/06/2026

Copyright really went absurd with agents. The only people who are able to defend theirs are the big corps, which is sad imho. We can steal a website, but we’re not allowed to publish pictures of Mickey Mouse.

010

Owlex @mrsowlex.bsky.social · 03/06/2026

I moved on from using Claude in the terminal. Tried using something like CMUX, but having a UI is the best way for me. Using Codex and T3Code though, because Claude Desktop suuuuucks 😄

000

Owlex @mrsowlex.bsky.social · 03/06/2026

We put "Read AGENTS md" into our CLAUDE md Files

100

Owlex @mrsowlex.bsky.social · 03/06/2026

Today: sprint planning, plus giving a small tech talk with frogs. Feeling a very specific mix of "yeah, let's go!" and "please leave me alone at my computer." 😅

000

Owlex @mrsowlex.bsky.social · 02/06/2026

Nah… seems to be either the cheap burner I bought or the CDs will try with a RW next and then maybe switch the burner

000

Owlex @mrsowlex.bsky.social · 02/06/2026

It failed again. I understand why we gave up this medium…

000