Okay guys, hear me out. This might be one of the craziest AI launches this year, and it's not even an AI model! We're getting new AI models almost every week, but running the powerful ones locally? Still a headache. Expensive GPUs, recurring API costs, and sending your data to someone else's servers. And that's exactly what a Singapore-based company called @Acrab_Ai is trying to change. Instead of launching another AI model, they're developing their own AI chip and a dedicated workstation to make powerful local AI more practical. Here's what they're building: 🔷 GΞLIX 1 (AI Chip): This is their own 5nm chip, built specifically for AI workloads, combining CPU, GPU, NPU, and unified memory. And here's the crazy part: it's designed to run open-source models up to the 100B-parameter class locally. They've even reported around 7.5x faster prefill performance than a comparison platform in a specific internal benchmark. For someone who regularly works with AI tools, the idea of running larger models without constantly depending on cloud APIs is pretty exciting. 🔷 Agent Box (Personal AI Workstation): Now this is the part I find even more interesting. They're building a compact workstation powered by GΞLIX 1, designed to run AI agents locally, remember context, work with your files, and automate tasks. Your data stays closer to you, you can reduce repeated API calls, and your agents don't have to depend on a remote server for every task. And I think this is where AI hardware is heading. We've spent so much time making AI models smarter. Making them easier to run on our own devices could be just as important. Think PCs, smart homes, cars, and even robots running AI locally. What's also worth mentioning is that Acrab is led by Ken Phua, who has over 20 years of semiconductor experience and previously held senior leadership roles at Arm China and Arm. They're showcasing their technology at SWITCH Singapore, Booth G01B.

Today, we're announcing Nitro: a formally verified agentic system for compiling your most complex SQL workloads into Rust. Our design partners are running millions of dollars of critical workloads through this and seeing 10x faster, 3x cheaper ETL. Reach out to try it.

This weekend we launched EthLib, a Lean4 library for Ethereum! The goal is to bring Ethereum's primitives, specifications and theorems into one place, building a shared foundation for formally specifying and proving things about Ethereum. github.com/ethereum/EthLib

Rust's CI costs were cut by roughly 75%. The interesting part is what the team chose not to optimize. 🦀 Serokell interview with Marco Ieni of the Rust Foundation covers how Rust's infrastructure actually runs. The biggest CI win, per the interview: moving jobs from GitHub large runners to standard runners. The Enterprise plan allows 500 standard runners in parallel. Large runners draw on limited donated credits. One way to avoid large runners was to split the test suite across two standard runners. The catch: each runner has to build the compiler again. That wastes CPU. Building once and sharing the artifacts would avoid it. The team decided not to. Sharing the build would make CI slower and more fragile. With enough standard runners available, wasting some CPU was the better trade for a simpler CI. The stated bar for "efficient enough": it fits the budget and credits, and the trade-off between efficiency and engineering time is reasonable. A small team made up mostly of volunteers cannot optimize every cent. The same preference shows up elsewhere: → Untrusted code. Crater and docs.rs run it, and the team prefers managed cloud isolation over self-hosting tools like Firecracker, because patching a self-hosted tool on a weekend is a burden. The blocker is engineering capacity → Monitoring moved from self-hosted Prometheus and Grafana to Datadog, because the old setup was never updated → Infrastructure changes should move into CI, so admins stop applying them from laptops On growth, the main pressure would be on CDNs. Fastly sponsors almost all crate and toolchain downloads. Requiring a login to download would make rate limiting easier, but it would break CI for every Rust project. One more point: the Rust project should provide primitives, not opinionated tools like release-plz. Two missing ones named: downloading crate source from the registry, and listing publish order. The interviewee works at the Rust Foundation, which the interview says up front. 🔗 serokell.io/blog/the-hidden-… #RustLang #Infrastructure #CI #OpenSource #DevOps