Week 26, 2025 - Markdown Collapse
This week on Leadtime.tech I addressed another reader submission — how to decide the necessary resources to operate a third-party platform. I usually write about the people management aspect of an EM’s life, so it was a fun detour to think about something closer to technical strategy. Read it here if you missed it.
Two weeks ago I shared my frustration about the way Meilisearch, the tool that helps search content across all my sites on peterszasz.com, approaches updates. Yesterday I had some time and wrote a bash script that does everything: checks for a new version, manages the update via data dump and import, creates a db backup, etc. It’s published here if somebody needs it.
I never wrote more than a screen-full of bash in my life, and that includes the years I was paid to code. The fact that I can do this — something that for sure could be improved a million ways, but does the job well enough — in an hour or two of vibe coding, is amazing.
🤔 An article that made me think
LLMs Will Not Replace You
I usually link AI-pragmatic or even AI-optimistic articles, because I’m generally excited about the opportunities of smart LLM use. This is an exception from the AI-sceptical side, with a lot of good content. There’s a high-level explanation of how LLMs work, and then conclusions about why they are doomed to fail in what some companies try to implement: replacing workers and automating processes with AI.
The most interesting and novel idea for me was that Model Collapse (the degradation of LLM output quality due to increasingly being trained on more and more AI-generated source, and therefore gradually losing its ability to create content that can be seen as human-like) is not just a theoretical danger in the future, but something that already started to happen because there’s no reliable way to distinguish human content from LLM-generated. This has an interesting consequence that’s opposite to the statement “...and this will only get better from here”: no, in fact, according to David Haney, we’re reaching the peak now, and newer models will gradually be worse due to every training material already "poisoned" with AI-generated content.
Meanwhile, I imagine there are a bunch of pragmatic engineers out there who, ignoring all the hype, scepticism, theories, and doomsday scenarios, just experiment with LLMs and include in their toolbox whatever brings the most value for them, today.
🖥️ Something Cool: MARP - Markdown Presentation Ecosystem
Not new in any sense, but I just recently discovered this tool when I needed to do a simple slideshow for a client. I wanted something that forces me to focus on the content by removing the ability to get lost in the format. The idea is that you write your presentation in a simple Markdown file, and then use Marp to create a self-contained HTML file that can be presented with any browser.
marp --html -w slides.md
Works as advertised, and they even have a VS Code extension for in-editor preview.
That’s it for today, be proud this weekend,
Péter