WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-05-12 1 00:00:00.000 --> 00:00:03.023 Hey everyone, welcome to the Builder's Briefing for May twelfth, twenty-twenty-six. I'm Alex, joined as always by Sam. 2 00:00:03.023 --> 00:00:09.068 Hey hey. Good lineup today — we've got a new tool that's basically a spell-checker for AI-generated React code, a really strong push toward local AI, and some security stuff that should make you go audit your Obsidian plugins right now. 3 00:00:09.068 --> 00:00:16.497 Yeah, let's jump right in. So the big story — Million.js shipped a tool called react-doctor, and it racked up seventeen hundred GitHub stars basically overnight. It's a static analysis layer purpose-built to catch the specific anti-patterns that AI coding agents produce in React codebases. 4 00:00:16.497 --> 00:00:18.341 Okay, so this isn't just another ESLint config. What makes it different? 5 00:00:18.341 --> 00:00:26.129 It's designed around the failure modes of tools like Copilot, Cursor, and Claude. Think inline object creation causing unnecessary re-renders, unstable keys, misused effects — the stuff that looks totally fine to a human reviewer because the AI wrote it so confidently. But it tanks performance at scale. 6 00:00:26.129 --> 00:00:32.687 That's interesting because I've absolutely seen that in my own work. The AI writes something that's plausible, it passes review, it even works — and then six months later you're wondering why your app is so slow. It's like a type-checker for AI code smell. 7 00:00:32.687 --> 00:00:40.295 Exactly how they describe it. Zero-config for standard React projects, already supports Next.js and Remix. Slots right into your CI pipeline. And I think the bigger signal here is we're entering this AI code quality tooling era. The first wave was generating code. The second wave is verifying it. 8 00:00:40.295 --> 00:00:45.469 Right, and I'd bet we see equivalents for Vue, Svelte, maybe even backend frameworks within months. If you're building dev tools, the meta-layer that validates AI output is a wide-open market right now. 9 00:00:45.469 --> 00:00:51.950 And this pairs perfectly with a piece that was trending today — James Shore arguing that the real metric for AI code generation isn't speed to first commit, it's total cost of ownership. Optimize for readability and changeability, not just does it work. 10 00:00:51.950 --> 00:00:59.763 A hundred percent. Also love that another experienced dev publicly posted about going back to writing code by hand. The pattern I'm seeing is senior devs segmenting — AI for boilerplate, hand-writing the core logic. If you're a team lead, that's your cue to define which code paths get AI and which don't. 11 00:00:59.763 --> 00:01:06.782 Alright, shifting to the local AI story, because this was everywhere today. A manifesto-style post called 'Local AI Needs to Be the Norm' hit over nine hundred points on Hacker News. The argument is that defaulting to cloud AI is an architectural mistake for most use cases. 12 00:01:06.782 --> 00:01:13.673 And what's wild is, the hardware case is actually there now. There were practical benchmarks posted for running models on the M4 with twenty-four gigs of unified memory. Concrete numbers on what actually runs well — gives you a real model-size ceiling to design around. 13 00:01:13.673 --> 00:01:19.283 And then there's Litter — a fully local AI meeting assistant written in Rust. It does live transcription with speaker diarization and summarization, all on-device using Whisper and Ollama. No audio leaving your machine. 14 00:01:19.283 --> 00:01:26.072 That's a killer reference architecture if you're building collaboration tools and want to avoid sending audio to third-party APIs. Especially for anything in health, finance, enterprise — the privacy-sensitive stuff where local inference is basically a requirement. 15 00:01:26.072 --> 00:01:33.167 Also worth calling out — there was a deep technical walkthrough on training an LLM in pure Swift, optimizing matrix multiplication from gigaflops to teraflops on Apple hardware using Metal. If you're building ML tooling in the Apple ecosystem, link in the briefing, go read it. 16 00:01:33.167 --> 00:01:37.061 Okay, dev tools. The headline for me — Nvidia shipped CUDA-oxide, an official compiler that lets you write Rust and compile it directly to CUDA kernels. 17 00:01:37.061 --> 00:01:38.393 Wait, official from Nvidia? Not a community project? 18 00:01:38.393 --> 00:01:42.620 Official. If you've been waiting to ditch C++ for Rust in GPU-accelerated workloads, the door is now open. This is a huge deal for the Rust ML and compute ecosystem. 19 00:01:42.620 --> 00:01:49.075 Between that and Meetily — which is a Rust-based drop-in Spark replacement claiming to unify batch, streaming, and AI workloads — Rust is just eating infrastructure. If you're running Spark and frustrated by JVM overhead, that one's worth benchmarking. 20 00:01:49.075 --> 00:01:54.199 Oh and a fun one — Ratty, a terminal emulator that renders inline 3D graphics. Niche, but it signals that terminals are becoming rich application surfaces. Worth watching if you build CLI-first tools. 21 00:01:54.199 --> 00:01:59.117 Quick security roundup. A malicious Obsidian plugin was caught deploying a remote access trojan. If your team uses Obsidian — and a lot of dev teams do — go audit your installed plugins today. 22 00:01:59.117 --> 00:02:03.779 Broader lesson there: any extensible tool with a plugin marketplace is an attack surface. Treat community plugins like third-party dependencies, because that's exactly what they are. 23 00:02:03.779 --> 00:02:09.517 Also, Gitleaks is trending again on GitHub. If you haven't added secret scanning to your CI, it takes five minutes to add to a GitHub Action. And it catches the API keys your AI coding agent just helpfully committed for you. 24 00:02:09.517 --> 00:02:11.746 Ha! There it is again — the verification layer. That ties right back to the hero story. 25 00:02:11.746 --> 00:02:17.177 And one more — Google now requires a QR code scan plus sending an SMS to register a new Gmail account. If your product relies on Gmail signups for onboarding, expect higher friction and possibly lower conversion. 26 00:02:17.177 --> 00:02:25.041 So here's the takeaway for today. The AI-assisted coding stack is clearly splitting into two layers — generation and verification. React-doctor, Gitleaks, the maintenance cost argument — they all point the same direction. Shipping AI-generated code without a quality gate is becoming the new technical debt. 27 00:02:25.041 --> 00:02:33.648 And on the where-to-run question, the local-first case just got a lot stronger. The M4 benchmarks, Litter's architecture, CUDA-oxide — on-device compute is production-ready for most workloads under thirteen billion parameters. If your AI features don't need frontier-scale models, you're probably overpaying and over-exposing user data. 28 00:02:33.648 --> 00:02:36.440 Well said. Invest as much in your validation pipeline as your generation pipeline. That's the move right now. 29 00:02:36.440 --> 00:02:37.516 Build the guardrails before you need them. 30 00:02:37.516 --> 00:02:41.000 That's the Builder's Briefing for May twelfth. All the links are in the briefing notes. We'll see you tomorrow — go ship something good.