Opened the PR for scaffolding production holm (FastAPI) + HTMX + Tailwind + BasecoatUI apps Testing instructions are in the PR Waiting a bit with the merge and release, in case there's feedback This should be the easiest way to build web apps with Python github.com/volfpeter/holm/pu…

Seems like Fable 5.1 is not capable of following simple "do not stage, do not commit" instructions

After the recent renderer optimization in the htmy Python lib, I started exploring whether an optional, simple Rust accelerator would be beneficial No surprise, it is: - 10x+ improvement if there are only tags in the component tree - 4-6x improvement if there are many async or custom Python components in the tree This will make htmy comparable to Jinja, "only" 2-2.5x slower (incl. component tree creation in Python) I "just" need to finalize and release this and then the holm init script, which also keeps improving...

I've released version 0.13.2 of the htmy Python library The important bit is a 1.5-1.6x renderer performance improvement I also did a security review and pushed a small fix for the built-in i18n module See volfpeter.github.io/htmy/

THE ANNOUNCEMENT: We’re going to make the prophecy of The Year of Linux on the Desktop come true. All the pieces are now in place. Time to go all in! omarchy.org/news/2026/08/oma…

Good episode, liking the podcast more with every new episode I'm always amazed how naive extremely smart engineers can be (speaking of jagged intelligence...) Finishing the OpenAI/HF presentation with an AI "helps defense more than offense" is a perfect example of this But thinking alignment is a) a purely model/technical issue and b) solvable is similarly naive IMO Agents are some combo of human + AI + environment, meaning alignment must be social + intelligence + tech problem Even if intelligence + tech could be solved to ~95%, the social part remains and is way harder

Making progress on the `holm new` script Think of it as create-next-app for the Python/FastAPI/HTMX stack It scaffolds: - HTMX, TailwindCSS, BasecoatUI, JS config - holm/FastAPI app with a vendored pure-Python BasecoatUI component library - Convenient local dev setup The created app is ready to run Deployment to Vercel or @FastAPIcloud also works out of the box (0 config, thanks FastAPI) Feedback welcome, try the early prototype with: `uvx --from "git+github.com/volfpeter/holm@pr…" holm new`

htmui is now on PyPI and has a CLI with an init script You can run it with uvx/pipx to vendor all UI components into your FastAPI/htmy app Next step: create-next-app-like script for holm to fully bootstrap a holm app! Web dev with Python/FastAPI should be as easy as it is with NextJS HTMX is a great transport layer Demo: htmui.vercel.app/

Strange experience: I trust the work of some engineers more during review when they heavily used AI than when they didn't The entire review/fix process is smoother too, it's easier to address detailed comments AI raises the floor, you should climb too, not just accept the lift

Unprecedented scientific progress! Oh wait... On more a serious note, the field is shifting similarly to engineering (especially in open source): lot more noise, more reviews, more focus on understanding and providing direction, etc. Wonder how hard the more rigid institutions will find it to adapt

A post of ours on vendor lock in and LLMs. We don't like the growing trend of AI companies quietly hiding your data while stripping away your control. We think that's bad for users and bad for the ecosystem. Here's are our thoughts: earendil.com/posts/session-p…

Can't help it, I get very frustrated every time I need to install and configure a Windows machine I get why people buy Apple products, but there are very few reasons for using Windows Terrible piece of anti-user junk

> when an AI agent autonomously hacks into a company’s computers, determining who is liable is much murkier No, it's not - same as any other penentration testing/hacking sw There is always someone in charge of the tool Intent and negligence can be debated, but you shouldn't leave these tools unsupervised

It's so easy to turn Grok 4.5 into a caveman Much easier than other models, which is good if you don't use other models, and inconvenient if you do

many apps focus on facilitating parallel work because of how slow the smarter models are more context switching, less focused work my hope is eventually we'll have models at least as good as GLM 5.2, but at some hundreds of TPS, hopefully at similar cost - it would be a real transformation

Occasionally, mostly during planning and review In most cases I'm in context switching state