https://blog.shibuyaltd.co.uk/20+ years in software engineering and the AI made me surprisingly reflective. So, here we are. 2026-04-15T14:18:26+01:00 Zsolt Kulcsar https://blog.shibuyaltd.co.uk/ Jekyll © 2026 Zsolt Kulcsar /assets/img/favicons/favicon.ico /assets/img/favicons/favicon-96x96.png 2026-04-09T00:00:00+01:00 2026-04-14T11:36:35+01:00 https://blog.shibuyaltd.co.uk/posts/code-stats/ Zsolt Kulcsar

TL;DR: I built a small project in a few days to compare code metrics across AI-assisted and non-AI-assisted open-source projects. You can explore it at csq.shibuyaltd.co.uk. Preamble A few days ago, the complete source code of Claude Code leaked in a funny way. I jumped on the opportunity and took a look, but I wasn’t that interested in the obvious questions like: “Is it secure?” or “What are...

2026-03-30T00:00:00+01:00 2026-03-30T00:00:00+01:00 https://blog.shibuyaltd.co.uk/posts/mirage-reasoning/ Zsolt Kulcsar

This morning an email popped up in my inbox, with the subject “The mirage of visual understanding in current frontier models”. This sounds interesting enough, I started to read. Quickly I ended up on the research paper’s page on arxiv: MIRAGE: The Illusion of Visual Understanding it was released on 26th of March. Abstract (at least, elements of it) We report three findings … First, Frontie...

2026-03-29T00:00:00+00:00 2026-03-29T00:00:00+00:00 https://blog.shibuyaltd.co.uk/posts/ai-coding-example/ Zsolt Kulcsar

I see a lot of backlash from software engineers against AI assisted coding, and for a while I was in that camp, but as I gained more experience with the tools I warmed up to them. Over time I realised that the Genie allows me to concentrate on the problem and its solution rather than bogging down in the code. I’m not saying that we can dismiss bad code and bad engineering. After all in the last...

2026-03-18T00:00:00+00:00 2026-03-22T19:53:48+00:00 https://blog.shibuyaltd.co.uk/posts/your-llm-doesnt-write/ Zsolt Kulcsar

I’ve taken the liberty of reusing the title of a Substack post. It’s a great post and worth reading, although it makes a different point from the one I’m making here. When one is prompting the language model to generate code, one relies on the collective knowledge slurped up in the model’s training data. One might get the feeling that the piece of code written in front of one’s eyes is unique,...