Bun's Rust Rewrite Hits 99.8% Test Compatibility, The Runtime War Just Shifted
Bun's Rust rewrite hits 99.8% compat, Gemini API goes multimodal for RAG, self-evolving agents cut token costs 6x. Monday builder briefing.
Hey everyone, welcome to Builder's Briefing for May eleventh, twenty twenty-six. I'm Alex.
And I'm Sam. We've got a packed one today — a huge runtime story, multimodal RAG from Google, some spicy opinions about AWS, and a few quick hits that are too fun to skip.
Let's jump right into the big story. Bun — the JavaScript runtime that originally bet everything on Zig — has been quietly rewriting its core in Rust. And Jarred Sumner just announced they've hit ninety-nine point eight percent test compatibility on Linux.
Okay, this is wild. When Bun launched, the whole identity was Zig — speed, low-level control, doing things differently. To pivot to Rust is a massive strategic shift.
Right, and I think the key insight is it's not really about Zig versus Rust as languages. It's about contributor pool, long-term maintainability, and ecosystem leverage. Rust just has way more people who can contribute.
That's the bus-factor argument, and it's legit. If you're shipping something that enterprises want to depend on, you need more than a handful of Zig experts maintaining it. And the fact that they hit near-perfect compat during a rewrite like this? That's seriously impressive engineering.
The zero point two percent gap is mostly edge cases. So if you've been holding off on Bun in production because of concerns about the Zig ecosystem, those objections are basically gone now. And look at the bigger picture — Deno went full Node compat, Bun's going Rust for durability. The runtime space is consolidating around serious bets.
Yeah, if you're starting a new project today, Bun just got a lot safer as a long-term choice. I'd expect their release cadence to speed up pretty significantly once the Rust port fully stabilizes.
Alright, shifting to AI. Google dropped something really interesting — the Gemini API's file search now handles images, audio, and video alongside text for retrieval-augmented generation.
Wait, so multimodal RAG natively? No more converting images to text descriptions and hoping your retrieval pipeline doesn't fall apart?
Exactly. You can index and search across modalities directly. If you've been building these janky image-to-text preprocessing chains, this could simplify your whole pipeline. Link in the briefing if you want to try it.
That's interesting because it changes what kinds of products you can build. Think customer support over screenshots, medical imaging retrieval, video knowledge bases — all stuff that was really painful before.
The other AI story I want to flag is OMLX — it's an open-source agent that starts from about thirty-three hundred lines of code and literally grows its own skill tree. It achieves full system control while using six times fewer tokens than comparable approaches.
Six times fewer tokens? Okay, for anyone burning money on agentic tool-use loops right now, that's not a marginal improvement, that's a fundamentally different architecture. The idea that the agent learns skills and caches them for reuse — that's the pattern I think we'll see everywhere in six months.
Agreed. Now let's talk infrastructure, because there's a post with over four hundred points on Hacker News titled 'I Returned to AWS and Was Reminded Why I Left.' Three hundred forty-seven comments deep.
Oh, I read that thread. It's cathartic if you've ever tried to untangle an IAM policy at two AM. The consensus is basically: AWS is unmatched at scale, but it's actively hostile to small teams trying to ship fast.
The specific friction points everyone keeps hitting — IAM complexity, billing opacity, the console UX — it reads like a catalog of velocity killers for startups. If you're choosing infra right now, it's worth reading just to know what you're signing up for.
And there's a great companion piece in today's briefing about idempotency edge cases — specifically when retry payloads mutate. If you're building payment flows or webhook handlers, that one will genuinely save you a production incident. Link in the briefing.
On the dev tools side, a few things. Zed editor shipped a visual theme builder, which might sound small but it's the kind of polish that actually retains daily users.
Honestly, that was one of my last excuses for not switching from VS Code. If I can bring my color scheme, I'm out of objections.
Also loved the post about distributing Mac software being, and I quote, a 'cortisol generator.' Apple's signing and notarization process is apparently driving indie devs to the brink.
I felt that in my soul. Anyone who's shipped a desktop Mac app knows the signing dance. It's reassuring and infuriating to know everyone else is struggling just as much.
Quick security note — FreeBSD published a security advisory for a local privilege escalation in execve. If you're running FreeBSD in production — jails, firewalls, storage appliances — patch now. This is the kind of vuln that turns a low-privilege foothold into root.
Yeah, don't sit on that one. That's a drop-everything-and-update situation.
Alright, rapid-fire quick hits. Louis Rossmann and Bambu Lab right-to-repair drama is escalating with lawsuit threats against an OrcaSlicer developer.
Oh no. The Internet Archive also opened a Swiss entity, which feels like they're hedging against US legal and political risk. Smart move.
Someone built a web server in pure assembly — because why not. There's a Rust-with-Lisp-syntax experiment. And htmx is still trending on GitHub. The hypermedia moment just keeps going.
Also a shoutout to the 'Making your own programming language' guide and the Visual Basic history that just dropped Chapter One. Those are weekend reads for the curious.
So here's the throughline for today. The runtime and tooling layer is being rewritten in Rust. Bun is the biggest signal, but it's not the only one. If you're choosing foundational infrastructure or building developer tools, Rust-backed projects are the longevity bet.
And on the AI side, multimodal RAG and self-evolving agents are the two patterns to watch. The next wave of AI products won't just call LLMs — they'll build persistent skill layers on top. If you're building agentic features, study the skill-tree pattern now. It's the clearest path to cutting token costs while improving what your agents can actually do.
That's the briefing for May eleventh. All the links are in the show notes. Have a great week building.
See you next time!
Bun's Rust Rewrite Hits 99.8% Test Compatibility — The Runtime War Just Shifted
Jarred Sumner announced that Bun's experimental Rust rewrite has reached 99.8% test compatibility on Linux x64 glibc. This is a massive milestone: Bun originally bet on Zig for performance and control, and rewriting the core in Rust signals that the team is prioritizing long-term maintainability, ecosystem leverage, and contributor growth over language loyalty. For the 469 people debating this on HN, the real story isn't Zig vs. Rust — it's that Bun is proving you can rewrite a complex runtime and maintain near-perfect compat.
If you're shipping production Node/Bun workloads, this changes your risk calculus. A Rust-backed Bun means better memory safety guarantees, a larger pool of potential contributors, and likely smoother cross-platform support as the rewrite matures. The 0.2% gap is mostly edge cases — expect that to close fast. If you've been holding off on Bun in prod because of Zig's smaller ecosystem or contributor bus-factor concerns, those objections are evaporating.
What this signals for the next 6 months: expect Bun to accelerate its release cadence once the Rust port stabilizes. The Node-compatible runtime space is consolidating around serious engineering bets — Deno going full Node compat, Bun going Rust for durability. If you're choosing a runtime for a new project today, Bun just became a safer long-term bet.
Gemini API File Search Goes Multimodal — RAG Just Got Visual
Google's Gemini API file search now handles images, audio, and video alongside text for retrieval-augmented generation. If you're building RAG pipelines and have been limited to text-only retrieval, you can now index and search across modalities natively — no more janky image-to-text preprocessing chains.
OMLX: Self-Evolving Agent Grows Its Own Skill Tree from a 3.3K-Line Seed
This open-source agent starts from a minimal codebase and grows a skill tree to achieve full system control while using 6x fewer tokens. If you're building agentic systems and burning through tokens on tool-use loops, this architecture — where the agent learns and caches reusable skills — is worth studying for your token budget.
Academic Research Skills for Claude Code
A prompt/skill pack that gives Claude Code structured academic research capabilities. If you're using Claude Code for literature reviews or paper analysis, this is a drop-in upgrade — though verify outputs against actual sources, obviously.
OpenHuman: Private, Local AI Assistant Going Viral on GitHub
A new open-source personal AI project emphasizing privacy and local-first operation is picking up steam. If you're building personal AI tools and want a reference architecture for on-device inference with a clean UX, take a look at the repo before it forks into a dozen variants.
"I Returned to AWS and Was Reminded Why I Left" — 439 Points of Pain
A developer's detailed account of AWS's complexity tax is resonating hard (347 comments on HN). The consensus: AWS is unmatched for scale but actively hostile to small teams shipping fast. If you're a startup choosing infra right now, this is a useful catalog of the specific friction points — IAM, billing opacity, console UX — that eat your velocity.
Idempotency Is Easy Until the Second Request Is Different
Sharp technical deep-dive on the edge cases that break naive idempotency implementations — specifically when retry payloads mutate. If you're building payment flows, webhook handlers, or any API that promises at-most-once semantics, this post will save you a production incident.
"I've Banned Query Strings" — And the HN Crowd Has Opinions
A developer makes the case for eliminating query strings entirely in favor of path-based routing and headers. Controversial but worth reading if you're designing APIs — the 197 HN comments are a masterclass in URL design trade-offs and cache behavior implications.
Zed Editor Ships a Theme Builder
Zed now has a visual theme builder, lowering the barrier for customization. If you've been eyeing Zed as your VS Code replacement but missed your color scheme, this removes that friction — and signals Zed is investing in the polish layer that retains daily users.
Distributing Mac Software Is a Cortisol Generator
Detailed breakdown of Apple's signing, notarization, and distribution pain points for indie Mac developers. If you ship desktop software on macOS, bookmark this for the workarounds — and know you're not alone in wanting to throw your Mac out the window.
NocoDB Trending Again: Self-Hostable Airtable Alternative
NocoDB is seeing renewed GitHub activity. If you need a no-code database UI for internal tools and don't want to pay Airtable prices or send data to a third party, this remains the most mature open-source option.
FreeBSD Local Privilege Escalation via execve() — Patch Now
FreeBSD published SA-26:13 — a local privilege escalation in execve(). If you run FreeBSD in production (jails, firewalls, storage appliances), update immediately. This is the kind of vuln that turns a low-privilege foothold into root.
AiToEarn: Open-Source Framework for AI-Powered Content Monetization
Trending hard on GitHub with 2K+ stars — a framework for automating AI-generated content across platforms for monetization. The ethics are debatable, but if you're building creator tools or content automation, the architecture for multi-platform publishing is worth examining.
let-go: A Clojure-Like Language in Go That Boots in 7ms
If you want Clojure semantics embedded in Go services — think config DSLs, rule engines, or plugin systems — this 7ms boot time makes it viable for CLI tools and serverless where JVM startup is a dealbreaker.
The runtime and tooling layer is getting rewritten in Rust — Bun's move is the biggest signal but not the only one. If you're building developer tools or choosing foundational infrastructure, bet on Rust-backed projects for longevity. Meanwhile, multimodal RAG via Gemini and self-evolving agents (OMLX) suggest that the next wave of AI-powered products won't just call LLMs — they'll build persistent, reusable skill layers on top. If you're building agentic features, study the skill-tree pattern now; it's the clearest path to cutting your token costs while improving capability.