Tweet épinglé
New blog post, wherein I beat a dead horse for the last time.
mariozechner.at/posts/2025-11-…
English
Mario Zechner
104.2K posts
Mario Zechner
@badlogicgames
Old man yelling at Claudes. The fucking bluecheck is temporary... https://t.co/mnOoWUqt4g https://t.co/8i5vIRDt6P
0xa000 Inscrit le Eylül 2010
1.2K Abonnements31.7K Abonnés
People of pi, you have 2 hours to clone this with a pi extension. Go!
Thariq@trq212
We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord. Use this to message Claude Code directly from your phone.
English
@badlogicgames didn't see this in 5.3-codex or other openai models before, this thing is the weirdest in the family.
English
People of pi. Do you feel experimental? Want to try a new edit tool? Stuff this into your ~/.pi/agent/extensions folder. Use it with your preferred model(s) for a while. Report back if it works.
Example: GPT 5.4 prefers rewriting entire files sometimes over doing multiple small edits. This mostly solves this for me.
gist.github.com/badlogic/30c35…
English
@badlogicgames I think it just doesn't like the edit tool, maybe OpenAI RLed apply_patch usage way too much.
English
Mario Zechner retweeté
Smart mode in Amp now lets you use Opus 4.6 wth 300k input tokens of context, up from 168k. Also, tokens above 200k are now charged at the same (not higher) per-token rate.
Why not 1M tokens in smart? Quality/cost aren't right. If truly needed, use (hidden) large mode.
English
OSS weekend starting Friday morning CET. Then I can clean out the issue and PR trackers over the weekend.
English
People of pi. Cloudflare release.
Also, breaking changes for extension authors! Easy to migrate, just point your clanker at the CHANGELOG.md.
I rewrote the keybinding mess into something that's mostly sane. Now you (your clanker) also have to suffer a little.
English
English
@bleuonbase @mitchellh I will have to clean it up to make it public since im pretty sure my token is hardcoded 😂
English
I'm dying for pi-mono-style minimal library that handles the hard parts of email (auth, syncing with local state, etc.) and gives me an opinionated way to add agentic loops on top of that. I want to build my own agents and logic and guardrails, I don't trust vendors right now.
English
who here will be a AI Engineer London in April? I'm ready to have more pub visits.
English
Keep seeing people, like @tobi @0xSero , rave about pi-mono, and with the rebirth of GUIs, thinking to make a Pi desktop app. Is there already one that I don't know about?
@badlogicgames looking to make a pi-core-server similar to codex-app-server, please lmk if complete crap idea :) will be clanker slop
English
@badlogicgames @tobi @0xSero Makes sense, can you share any high-level constraints you already expect for integrating the future pi server, even if the design isn’t settled yet? For example:
- preferred integration boundary: package APIs vs subprocess/JSON mode
- whether sessions/threads will be first-class
English
@zeeg @0xblacklight longer context is basically a hack in all current implementations. don't see this being fixed anytime soon. so longer context with the current hack won't help at all.
English
@0xblacklight Yeah I agree it’s both recall and quantity and neither problem improved much. Codex skill calling is fairly impressive on recall but it still struggles more as time goes on just like everything else
English
1) not surprising whatsoever
2) this is exactly what I keep saying about models not being powerful enough today
the fact that they can do so much with lossy compression is amazing, but there's no magic here
imo (for transformers) context windows need to be 1-2 orders of magnitude larger for the future people keep saying is reality, and even then the compute is probably not worth it
Lossfunk@lossfunk
🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵
English
Mario Zechner retweeté
We've been experimenting with pi-autoresearch to optimize the new canvas rendering engine. 10x improvement on the slowest test in couple hours. 🤯
While it did take some shortcuts that degraded rendering quality, it also came up with several great ideas that were easy to cherry pick.
This is huge. HT to @karpathy for the original inspiration and @davebcn87 for the π extension.
tobi lutke@tobi
And the most important part: we open sourced the /autoresearch plugin for pi. Just tell it what you want, it will do the rest. github.com/davebcn87/pi-a…
English
Mario Zechner retweeté
Small Pi subagents update: new option to run subagents from a fork of your current place in the convo (not just a fresh blank context), so they can inherit the context you already built up.
pi install npm:pi-subagents
github.com/nicobailon/pi-…
English
@paxaral @badlogicgames @sudbalaji I _think_ the issue is it messes with cache. Similar to why inserting an updating time stamp in the system prompt is a not advisable.
English
this is immensely illegal
David Cramer@zeeg
how dirty of a hack do i go with to auto inject new tools mid-turn in Pi 😅
English