Thomas Mustier

we must ensure all humanity can access the same intelligence supporting open models is simpler than people think just ask your coding agent: "read this, then help me find, redact & prep session transcripts to share on @huggingface: github.com/huggingface/hu…"

@cremieuxrecueil i assume it's deterministic

Will teachers be able to use this to teach biology or is that banned?

@Howaboua I am so here for this please please do it

I've done so much harness engineering at this point I am seriously debating just building my own harness from scratch(-ish), lol. Is that a good idea? Probably not. Have I made: - an entire conversion of Pi tools (even created my own extension-ish system inside of an extension)? - a server mode for Pi that works.... while Pi doesn't have one yet? - a custom Claw system with tools and everything that I am using every day? - a GUI app where I have to add things like secret markers as a workaround (easier) to trigger UI instead of allowing the agents to "just do the thing"? Well, yeah, I have. Are these all perfect? Not necessarily lol. But they're mine. OOOF. Think gippity 5.6 is finally giving me real psychosis. And I have too many resets to burn.

@pidotdev my workflow involves playing zelda while waiting for fable x.com/thomasmustier/…

Change the harness not your workflow. Here we ask Pi to build a custom workflow extension with custom TUI to streamline the commit and push process. Pi customizes itself when asked to. How have you extended Pi?

@pidotdev @nicopreme heard your name being called ;)

@lefthanddraft just claude in a blazer

I really hate this on multiple levels. It feels deceptive to take a model, give it a corporate brand identity and then try to hide the underlying model. This is like turning up to work and seeing your roommate wearing a false moustache pretending to be someone else.

@kenwheeler most using @pidotdev will have done

i’m actually curious, how many of you have actually written a tool in the literal sense for an agent

@gavinpurcell five nights at freddy’s

well, i connected GPT-5.6 Sol to Blender and... let's just say the world of 3D graphics (at least via Blender's MCP) is safe for now... but still... this whole scene was generated by GPT-5.6 with very little human intervention feels like *EARLY* pixar, like too early lol

@SaanyaOjha or 3. just straight up source theft

Zero data retention matters, but it answers only the narrow privacy question. There are really two risks: 1. Learning leakage: the vendor retains or benefits from customer interactions 2. Learning evaporation: the enterprise generates corrections, evaluations and decisions, but squanders this valuable output by allowing each interaction to disappear into ephemeral chat history

@joon_h_lee congrats on this def not easy

@joon_h_lee this is super cool, how did you keep the cache warm with delegation? and did you see a material time difference vs either of fable or opus standalone?

There are many agent harnesses, but this one is open source and it is yours. Thank you People of Pi for 70000 stars on GitHub⭐️

@thorstenball that’s so fetch

You know what time it is

@FaYiMoRa @pidotdev they’re bundled as separate extensions btw, you can run `pi config` to pick and choose

@FaYiMoRa @pidotdev yeah i have gone back and forth on this ultimately my objective is to display a <1 screen tldr of everything the agent did, which makes it pretty dense bg colour is one way to break it up but haven’t managed to make it nice yet, still iterating though

And Pi out of the box sends less than 1.2k tokens ;) systima.ai/blog/claude-co…

@joshua_soileau @emollick @nikitabier waddaya say 👀

@thomasmustier @emollick I wish you could share “mute lists” to share these in bulk

At this point, I’ve given up hope that X’s bot-reply problem is solvable. The “the part no one says” writing is annoying, but worse is that they all make similar points. Proposal: X should measure semantic distance in latent space and surface replies that offer actual variation.

@injaneity @pidotdev

pi-computer-use v0.4.3 is here! an update to bring us to parity with codex (who to their credit, stepped up their game massively in their last update): - a ghost cursor now shows what computer use is working on - parallel tasks allow subagents to coordinate across different apps safely - better batching improves the accuracy of repeated / simple groups of actions we now see an almost 50% improvement in latency across a variety of tasks, and better handling for non-trivial use cases (like microsoft apps) try it out here: github.com/injaneity/pi-c…

@badlogicgames @trashpandaemoji i would love to know what you learn from people telling you what they learn !!!

are you a junior "AI native" developer? i'd love to know how you work and learn! post below, hit my DMs, or write an old man email to contact@mariozechner.at.

I've been seeing more and more people trying to advertise things one can do with agents that were necessary a year-ish ago. Skills with personas, agent councils with personas, overcomplicated workflows, repos with 50+slopskills and stuff. Makes me wonder if this is because of non-heavy AI users discovering the things we discovered a year ago, and going through the psychosis of "I JUST DISCOVERED THIS, THIS IS THE BEST YOU SHOULD TRY THIS". Or if this is a cold calculation: things like councils of personas look cool to newbies - they are something a newbie can comprehend. And there are more newbies starting to use AI. So the heavy users figured out that stuff from ca a year ago can be used to farm engagement, even if it's suboptimal, and they are bloody well aware that it's suboptimal. Food for thought.