Well ... I'm leaving Google.

If we don't trust agents, why do we trust the software they build? I suspect that all software in the future will need to be sandboxed. That's why smol machines is an embeddable VM like sqlite. Any app can run untrusted code in its own VM, on your own machine. Isolation shouldn't be pay to play.

After a quick bench, if you're using WSL on windows for any thing. You should be using smolvm instead. github.com/smol-machines/smo…

WE GOT IT! .claw incoming!

Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

Smol machines doesn't do snapshots, instead I focus on building a lineage aware file format called .checkpoints. The idea is that snapshots are a point in time. But I want to go back to every previous checkpoint kind of like git history, so we can audit or undo an agent's mistakes. The spec & implementation are 100% open source and public: crates.io/crates/smolvm-chec…

imo .mjs .cjs .mts .cts should never have happened

.mjs and .cjs shouldn’t have happened

People on windows can use linux very easily and quickly for sandboxing ai or just to use linux. It's at ~2s right now but will continue to get faster. Windows users are first class citizens for smol machines: github.com/smol-machines/smo…

v1.18 is now out with credential substitution for smol machines. It's the mechanism to swap in the real key inflight when the placeholder is called within the guest vm. A compromised agent can't leak what it never had. Here's an example: ``` smolvm machine run --credential openai=OPENAI_API_KEY@api.openai.com ``` github.com/smol-machines/smo…

Porting haproxy from C to Rust overnight. Mind-boggling! 🤯

Rogue ai occurrences are accelerating. And it's looking clear that sandboxing isn't just something that we'll need in the cloud. We're going to lightweight accessible sandboxing technology for everyone because it's good for both isolating AI from escaping the box + AI from getting inside the box. It's why smol machines care so much about local vm's being accessible and open source

I was thinking of a #webassembly post I saw that promised not to deviate from the WebAssembly specifications, and that it might increasingly be less optimal because with LLMs it's easier to adapt to new specs. However, ChatGPT helped me find it, and I saw that it was specifically about #wasmtime and it makes sense there – so long as they're OK with people creating forks that do deviate from the spec. bytecodealliance.org/article…

Here's how you can branch in smolvm vs firecracker: smolvm: `smolvm machine branch --from startedvm --name branchedvm` github.com/smol-machines/smo… vs Firecracker: pause the vm -> save a snapshot to disk -> launch another vmm -> load snapshot -> resume original : github.com/firecracker-micro…

first agent to achieve work-life balance

RT @ResourcesCo: Updated resources.co with a Mahjong example. Much is running in WebAssembly sandboxes with access to the DOM co…

hi, c cpp fans, introducing CUDA Rust 🦀.

Hello X-friends and soon-to-be X-friends, I wrote this for you: yegge.ai/essays/seats-and-su… -- an essay (mostly) about fuel consumption. The TL;DR: Fable's the only good model, and it's too expensive, so nobody's orchestrators work worth a shit today. Also it won't get better for a year. Enjoy!