Martin Klepsch

Martin Klepsch

@martinklepsch.org

Web, AI and wherever curiosity will take us :) Oh and Clojure/Script of course

167 followers182 following183 posts

I’ve fully come around to file system based routing. yes it’s weird to have some of these [slug] filenames but wow, it’s just so much less boring glueing stuff together

i think the thing about JS that bothers me the most is the laizzes-faire control flow, return anywhere, crazy syntax constructs for if expressions, … if it only was s-expressions 🥲

its becoming annoyingly commonplace that model choosers switch to 'auto' and then don't even clearly label what model was used and why 🥲 like how am I supposed to build ~intuition~

The thing with RLS is, you only really need it if you want to connect clients to the DB(?), but with Astro on the server that need kind of disappeared for me. A few actions with server-side validation is all it takes.

Can't remember where I read this but thought it was an interesting take: o1-pro, gpt-4.5 and other expensive models are effectively a "drop" to see if the market can find emergent, useful behaviors. Presumably because they haven't found them internally yet.

Pricing these models high will allow OpenAI to deprecate those models without much backlash in the case that no use cases justifying the compute requirements are found. All this makes it even more fascinating that Sonnet 3.5 is still everyone's darling when it comes to coding

I stopped myself from vibe coding my way to some organized markdown files and instead tried Claude code. It's cool! (and slow & expensive) But mostly cool. More things should be agents.

Fully back on the Perplexity train. Having access to R1, deep research, great search grounding all in one package is such a killer. Never been so fun to follow your curiosity ✨

I want to package a small node server + react app as an NPM package that you can run with `npx`. I started with `vite-express` but when running the package it won’t serve the static files… Anyone think of an example I could look at?

A super easy change I've made in how I use LLMs: editing my initial message over providing "corrections" in subsequent messages. It means less context to get confused over for the LLM and at the same time forces me to "craft a prompt" vs. "no, do it this way". Learning more!