@mitsuhiko.at
Creator of Flask • earendil.com ♥︎ writing and giving talks • Excited about AI • Husband and father of three • Inhabits Vienna; Liberal Spirit • “more nuanced in person” • More AI content on https://x.com/mitsuhiko
More stuff: https://ronacher.eu/
@mitsuhiko.atWe did not look at Ghostty in particular, but Jake Stewart's gist was an early inspiration for this.
@mitsuhiko.atPython 3.15 is the first release where parts of the core team started to suffer AI psychosis. Welcome aboard!
@mitsuhiko.atStill worth completing. On some monitors it's really hard and those are the most important data points.
@mitsuhiko.atYou can build yourself a wrapper around context! that does Serde(...) for all values. I was thinking of adding one, but I was not sure if that is needed.
@mitsuhiko.atIn the process I learned there is no good open source contrast algorithm so we're doing a survey on contrast to make one. Would be cool if you could take some minutes out of your day and do it. For science! contrastsurvey.earendil.com
@mitsuhiko.atOur designer Max Blazek wrote about how we built the new Pi theme and what all of this has to do with colors and contrast. You should read it, it's super cool. earendil.com/posts/system...
@mitsuhiko.atIf you are using MiniJinja, I finally released 3.0 which gets rid of Serde as a default dep, supports free threading, fixes tons of legacy cruft, runs quite a bit faster and is just overall nicer. If you want to play with it, there is a WASM demo here: mitsuhiko.github.io/minijinja/
@mitsuhiko.atBecause all harnesses are trained on JavaScript for codemode. Also JavaScript is trivally embeddable thanks to quick-js and Python is not. You need to ship too much stdlib to make it work at all (eg: asyncio or a mock of it)
@mitsuhiko.atI did. It's complicated. You're quite far off what the models expect (they are trained for codemode on JavaScript) and it really struggles with async and gather in monty.
@mitsuhiko.atI wrote a bit about what Codemode actually is and how it works. lucumr.pocoo.org/2026/10/6/co...
@mitsuhiko.atThe goal is for sure to ensure Gondolin works locally and we will continue to support it for that. We just need to re-think this a bit given what Gondolin can do vs. what we have to do on E2B or other platforms.
@mitsuhiko.atBut nobody needs to do that any more
@mitsuhiko.atIn Österreich kann man dieser Tage erleben was passiert wenn der Gesetzgeber etwas will, das von selber nicht passiert, er ein paar schlecht überlegte Gesetze macht und dann … insert pikachu meme here.
@mitsuhiko.atDefinitely see this different. CI/CD has been held back and up by bash scripts. The better comparison is the work of durable workflow engines over the last decade.
@mitsuhiko.atWe released Pi 1.0! earendil.com/posts/pi-1-0/
@mitsuhiko.atIt's on pause at the moment. I think we will pick it back up, but given the current market dynamics it's too expensive to run yourself vs things like e2b. Only useful local for which we use it.
@mitsuhiko.atNo, but the question is what part of it you mean precisely. We're intentionally not supporting parts of MCP (eg: elicitations) and given the lack of stateless servers to test against, we also don't do that yet. That's pretty common today for MCP clients.
@mitsuhiko.atI think it works decently well but it definitely over attributes once it made a determination.
@mitsuhiko.atThat's not too surprising as the structure no longer matches the original AI generated text.
@mitsuhiko.atLobsters broke their Anubis rollout and now you can't upvote or message the moderators. Thankfully codex does not give a shit and found a way to solve Anubis and send a message to the moderators.
