Bun Goes Rust, OpenAI Hits WebRTC Limits, and Claude Code Changes How We Write HTML
Five stories shaping the developer world this week: runtime rewrites, AI model updates, and a rethink of frontend tooling.
A lot moved this week across the AI and developer tooling space. Here are the five stories worth your attention — and what they mean in practice.
Bun's Rust Rewrite Hits 99.8% Test Compatibility
Bun quietly announced that its experimental Rust rewrite now passes 99.8% of tests on Linux x64. This is a big deal. Rewriting a JavaScript runtime in Rust while maintaining near-perfect compatibility is genuinely hard, and the performance ceiling Rust opens up is significant. If this ships stable, Bun could leapfrog Node and Deno not just on speed benchmarks but on reliability — the thing that actually matters for production workloads.
Google Broke reCAPTCHA for De-Googled Android Users
Over 1,500 upvotes on Hacker News for this one — a sign of how raw a nerve it hit. Users running de-Googled Android (CalyxOS, GrapheneOS, DivestOS) found reCAPTCHA silently failing, locking them out of services that depend on it. This is partly a technical issue, but mostly a reminder of how much of the web's "security" infrastructure is actually Google infrastructure. If you're building anything user-facing, this is a good moment to audit your auth and bot-protection dependencies.
A Recent Experience with ChatGPT 5.5 Pro
The community shared detailed notes from using ChatGPT 5.5 Pro, the latest reasoning-focused model from OpenAI. The consensus: significantly better at multi-step reasoning and code generation, but still inconsistent on domain-specific tasks without good prompting. For clients asking me whether to switch models — the answer is still the same: model choice matters less than your system prompt, context quality, and evaluation setup.
Claude Code and the Unreasonable Effectiveness of Plain HTML
One of the more interesting pieces this week: a developer documented using Claude Code to build full UIs in plain HTML — no React, no Tailwind, no build step. The results were surprisingly good. This resonates with something I tell clients often: AI tools reward simplicity. The less abstraction between Claude and the output, the fewer things can go wrong. Worth reading if you're integrating AI into your dev workflow.
OpenAI's WebRTC Problem
OpenAI's real-time API uses WebRTC for low-latency audio, but developers are running into hard architectural limits — especially around connection management at scale and NAT traversal in enterprise environments. If you're building voice AI agents (a growing request from my clients), this is a real constraint to plan around. The workaround most teams are landing on: a dedicated media server layer between the client and the AI backend.
These five stories don't point in one direction, but they share a theme: the infrastructure layer of AI is maturing fast, and the decisions you make now about runtimes, dependencies, and model selection will compound. If any of this affects a project you're working on, book a call — happy to dig in.