As President Trump meets President Xi this week, a call to the American AI community:
If your startup, lab, non-profit or company benefits from open international AI - especially Chinese (Deepseek, Qwen, Kimi, GLM,โฆ), please share!
Open source is the most important driver of competition, jobs and wealth creation in AI today. Letโs support and promote it at critical times like this week!
Multi-Agent Orchestration is one of the biggest hype in AI right nowโฆ but it often just doesnโt work in practice. Hereโs a thread with the strongest, data-backed critiques + links (2025โ2026)
My dev workflow has totally shifted over the last month. I used to rely on planning mode / architecture mode (since the days of aider by @paulgauthier ๐) to map out core changes.
Now, I have the agent run experiments in a gitignored folder, benchmarking scripts and throwing ideas around.
Only when I have something finalized does it move into my repo in a "plan" (just to help guide my human review later). The weirdest part? Most of the code clanker writes isn't even part of my main codebase anymore.
Itโs all prototyped first.
This makes me wonder what to do with the (repo)/runs/experiments/ folder. There's obvious value there, but it's pure chaos, the exact kind of chaos clankers love to create. Still, there are good experiments and unfinished ideas that might see the light of day later.
I've been thinking about how to actually manage this runs/experiments/ folder. Should it be a git submodule or something similar? Maybe create a dedicated pi extension for it. Or maybe I should just use simple branches since Git already does the job. But honestly, I have less and less attachment to these files living on GitHub anyway.
WDYT @badlogicgames
The more I replace plans with prototypes, the better the outputs
Who'd have thought that low fidelity prototypes were better than walls of spec
Oh yeah, the entire industry for 20 years
Stop going against decades of knowledge because someone in SF shipped it as a 'mode'
People of https://t.co/TgG5bkXUdV. If you've installed pi 0.73.1, you will now be notified to do a `pi update` for 0.74.0.
Starting today, all pi packages on NPM are in the earendil-works namespace, instead of mariozechner.
Update imports from @mariozechner to @earendil-works in extensions or SDK-based apps!
The pi GitHub repository has also moved to https://t.co/0h7Q5NmJCB
Update all the things!
People of https://t.co/TgG5bkXUdV. Update to the latest pi now (0.73.1), which is the last version coming from the `@mariozechner` namespace on NPM.
Next release in 10 minutes (0.74.0) will be in the `@earendil-works` namespcae on NPM. Update your extension imports from mariozechner of earendil-works.
They will continue to work for a while. They will no longer work after the big refactor is done (2 weeks i hope)
i had this model in nvidia's pre-brief last week. tested all 5 modalities through their hosted nim endpoint, text + image + audio + video + tool calling all verified end to end before the lifted tonight.
now unsloth dropped the ggufs day-zero, which means it runs local on consumer hardware. 30b moe with 3b active params, 256k context, ~25gb ram per their spec. omni model means one model handles every modality, no router orchestration needed.
spark landed yesterday with 128gb unified memory, beast 5090 mobile has 24gb and my 3090 is still up. cross-tier benchmark series queued, same model on three nvidia silicon classes, one set of numbers.
multimodal local agentic workloads on consumer hardware is the lane this opens. data lands soon.