TypeScript 7, Bun in Rust, and the Malleable Linux

Weekly digest of what top open-source authors published and said — May 8–15, 2026

DHH: AI finally makes open source's original promise real

David Heinemeier Hansson (DHH), creator of Ruby on Rails and co-founder of 37signals, spent this week elaborating on a thesis he calls "the malleable computer" — and it cuts to a tension that's existed in open source since the beginning.

"Open source promised that users would be free to change whatever code they were running. The reality, however, is that hardly any of them ever did — it was simply too hard. Now, with AI, it suddenly isn't." 1

The argument is that Linux is the only operating system where this new AI-mediated malleability can actually be realized — "with Windows and macOS, the core elements of the operating system are owned by the companies that make them." 1 Windows, he says, cannot be reshaped the way Linux can, no matter how capable the model. That's not a cultural claim about open-source ideology; it's a practical claim about what AI agents can actually modify at runtime.

The broader Linux momentum he describes is real by his own metrics: Omarchy, his opinionated Arch Linux distribution, reached 50,000 ISO downloads per week and 30,000 Discord users, with 1 petabyte of ISOs distributed within 30 days (roughly 150,000 installs). 2 Part of the cultural shift he points to is the TUI revival: "thanks to OpenCode and Claude Code, terminal user interfaces (TUIs) are suddenly everywhere." 2

Design blogger Roger Wong documented the same shift this week, citing Alcides Fonseca's analysis: TUIs are returning not because of nostalgia, but because three decades of GUI toolchains failed to hold. 3 macOS broke its own Human Interface Guidelines, Windows cycled through WinForms, WPF, and MAUI without landing anywhere stable, and Electron turned every application into its own isolated keyboard-layout island. Fonseca's framing: TUIs are the exit ramp developers take when every layer above them stops being reliable.

On May 9, DHH shipped Omarchy 3.8 "The Defaults Edition" — the release puts a unified defaults layer at the center, letting users set system-wide browser, terminal, and editor in one place (Setup > Default). 4 The release also adds the Foot terminal (memory usage roughly one-fifth of Alacritty's) as an option, a Trigger > Transcode pipeline for video/audio/GIF conversion, a Super+Ctrl+R reminder shortcut, and live weather in Waybar. Hardware support was extended to Apple Studio Display XDR, Dell XPS 2026 force-touch trackpad intensity, and Intel Panther Lake audio. 4

Thomas Ptacek (vulnerability researcher and co-founder of Matasano Security, writing at sockpuppet.org) offered the counterpoint the same week. His essay "The Emacsification of Software" opens: "Everything is now programmable, not merely in a technical sense but a practical one." 5 His claim is that AI agents have made native application creation accessible — he built MDV.app, a macOS Markdown viewer, in roughly 30 minutes of interactive Claude time, and a separate bpftrace (a Linux kernel tracing tool) visualization interface called BPFUI. His direct challenge to the TUI revival: "The ceiling for AI-built native interfaces is higher than anything you can do in a TUI." 5

The essay earned 421 points on Hacker News. DHH has not publicly responded to Ptacek's counterpoint as of May 15.

DHH on AI tooling: GPT 5.5, 9× faster tests, and agents as the actual product

DHH's AI-adjacent posts this week moved across three related claims.

On models: he ran GPT 5.5 (OpenAI's May 2026 model) on low-reasoning mode for over a week and concluded it was "very good, very efficient" — specifically, he said he had not been tempted to switch to Claude Opus, and that GPT 5.5 is "more succinct than Kimi too." 6 His read is that OpenAI made a substantial leap forward with this release.

On testing tooling: he migrated a project from RSpec to fixtures with AI-assisted conversion and reported a 9× test-suite speed improvement. 7 His point was less about AI and more about two decades of advocacy finally producing visible data:

"I've been trying to tell people for twenty years that fixtures and avoiding rspec was where it was at. But no argument delivers this insight like letting AI do the conversion for you, and seeing the incredible performance improvements possible. NINE TIMES FASTER!" 7

On the product question: he published a blog post titled "Basecamp becomes agent accessible," announcing that Basecamp now exposes a full API, CLI, and skill-file set for AI agents to operate to-do lists, messages, files, and project schedules. 8 His design rationale: AI agents become "killer app" territory when LLMs can use tools to check their own reasoning and a file system gives them persistent memory across prompts. He adds that agent capability will eventually arrive in mainstream interfaces — ChatGPT, Gemini, Grok — not just developer CLIs. 8

TypeScript 7.0 Beta: ~10× faster via Go port

On May 14, the TypeScript team released @typescript/native-preview@beta — the public beta of TypeScript 7.0, a port of the TypeScript compiler from TypeScript-compiled-to-JavaScript to native Go. 9

Daniel Rosenwasser (Principal Product Manager for TypeScript at Microsoft) clarified the method: this is a port, not a rewrite. "The new Go codebase was methodically ported from our existing implementation rather than rewritten from scratch, and its type-checking logic is structurally identical to TypeScript 6.0." 9 The distinction matters: the team didn't redesign the type system, they translated it — with codemods, not by hand — which means the decade of accumulated correctness tests carry over directly.

The speed claim: "TypeScript 7.0 is often about 10 times faster than TypeScript 6.0." 9 The main mechanisms are parallel parsing, type-checking, and emitting. The tsgo executable defaults to 4 type-checker workers (configurable via --checkers) and supports parallel project-reference builds via --builders. Pre-release testing ran on production codebases at Bloomberg, Canva, Figma, Google, Linear, Notion, Slack, and Vercel.

One important scope note: the stable programmatic API (the API that tools like ts-morph or webpack's ts-loader depend on) will not land until TypeScript 7.1+. The beta ships an editor extension for VS Code (TypeScript Native Preview), and stable release is projected within two months. The package name is currently @typescript/native-preview; it will eventually replace the typescript package name.

Bun rewrites its core from Zig to Rust

On May 14, Jarred Sumner (creator of Bun, a JavaScript runtime and all-in-one toolkit with 90,400 GitHub stars) merged pull request #30412 — a rewrite of Bun's core from Zig to Rust. 10 The branch contains 6,755 commits and approximately 960,000 lines of changed code. It passes Bun's pre-existing test suite (99.8% compatibility on Linux x64 glibc) and shrinks the binary by 3–8 MB, fixing several memory leaks and flaky tests in the process. 11

The rewrite was done with AI assistance and reportedly completed in under a week, though Sumner has not published the promised technical blog post as of May 15. The PR description is minimal: "Blog post with details coming soon."

The HN discussion reached 588 points and 656 comments, with much of it covering the Zig-vs-Rust tradeoffs. The more consequential open question is whether the Rust memory model makes Bun's GC-free performance guarantees easier or harder to maintain in practice. That answer awaits the blog post.

Wasp: "Inventing a new programming language was a mistake"

Matija Šošić and Martin Šošić — twin brothers and co-founders of Wasp (YC W21, a full-stack web framework built on React + Node.js) — published a five-year postmortem this week. 12 The short version: they built a custom DSL with its own Haskell compiler, raised over $5M, and are now replacing the DSL with a TypeScript SDK.

Matija Šošić's core admission:

"Five years in, we realize it was a mistake." 12

The specific failure mode was IDE support. Maintaining a custom language means maintaining a custom language server, custom syntax highlighting, custom linting — and the entire tooling ecosystem is built for standard JavaScript and TypeScript. Every IDE integration had to be built from scratch and kept current.

But the postmortem doesn't conclude that the underlying concept was wrong. The DSL's real function, Šošić argues, was to keep the project honest: "The DSL approach kept us honest to our vision that the specification should be separate from the implementation." 12 That insight survives the switch to TypeScript. His framing of the durable value: "Language was never the moat. It's having a high-level understanding of your entire app at compile time." 12

The TypeScript SDK had been available as an experimental preview, and the feedback on it was described as strongly positive. A Launch Week to ship it as the primary interface is planned for the coming weeks.

Daniel Stenberg on Anthropic Mythos: "primarily marketing"

Daniel Stenberg, creator and lead maintainer of curl (installed on over 20 billion devices across 110+ operating systems, 176,000 lines of C, 188 published CVEs), published his evaluation of Anthropic's Mythos model on May 11. 13

Anthropic ran Mythos against curl's src/ and lib/ (178,000 lines) and reported 5 "confirmed security vulnerabilities." After the curl security team reviewed them:

1 confirmed as a real vulnerability — severity: low, CVE pending for curl 8.21.0 in late June
3 false positives — documented API behavior
1 "just a bug" — a real bug, but not a security issue

Stenberg's conclusion: "My personal conclusion can however not end up with anything else than that the big hype around this model so far was primarily marketing." 13 The comparison point: previous AI tools — AISLE, Zeropath, and OpenAI Codex Security — already triggered 200–300 bug fixes in curl over 8–10 months before Mythos arrived. Mythos found roughly 20 non-security bugs with few false positives, which Stenberg notes is genuinely useful. The problem is the claim of uniqueness.

He closes with a harder principle: "Not using AI code analyzers in your project means that you leave adversaries and attackers time and opportunity to find and exploit the flaws you don't find." 13 On AI code analysis as a category, he remains unambiguously positive. On Mythos specifically, he sees no evidence it outperforms the class.

LWN covered the post separately. 14

DuckDB: "quite a luxury" to design a database protocol from scratch in 2026

The DuckDB team (Hannes Mühleisen, Mark Raasveldt, and colleagues at the DuckDB Foundation) released the Quack client-server protocol on May 12. 15 The design philosophy statement is worth quoting:

"We had the opportunity to design a database protocol from scratch in 2026 without having to consider any legacy, which is quite a luxury." 15

The protocol is HTTP-based with a custom MIME type (application/duckdb), executes in a single round trip, and uses custom serialization rather than Arrow Flight SQL. The team's reasoning for rejecting Arrow: Arrow is an interchange format, not a native wire protocol, and adopting it internally would constrain DuckDB's ability to evolve its own data types and protocol messages. They were also critical of Arrow Flight SQL's two-round-trip design (CommandStatementQuery + DoGet), which they describe as "not ideal for small updates." 15

Benchmark numbers from the blog post:

Benchmark	Quack	Arrow Flight SQL	PostgreSQL
60M TPC-H (OLAP benchmark) lineitem rows, bulk transfer	4.94s	17.40s	158.37s
Small writes (8 threads)	5,434 tx/s	1,358 tx/s	4,320 tx/s

Quack is available as an extension in the core_nightly repository for DuckDB v1.5.2. Production release is planned with DuckDB v2.0 in fall 2026.

Jason Robert: deterministic multi-agent orchestration with Conductor

Jason Robert, Principal Software Engineer in Microsoft's Azure Core CTO office, published Conductor this week — an open-source (MIT), YAML-based CLI for multi-agent AI workflows. 16

His design thesis has three parts: "Determinism is a feature." "Agent isolation pays for itself." "YAML is the right level of abstraction." 16

The first point is the most interesting to stack architects: Conductor's orchestration layer consumes zero tokens because routing is fixed at workflow-definition time via Jinja2 templates rather than delegated to an LLM at runtime. Each agent gets its own session, system prompt, model, and provider with no context bleeding between agents. The tool supports GitHub Copilot and Anthropic Claude as providers, parallel execution steps, human approval gates, and a web dashboard. It works with Model Context Protocol (MCP) servers and integrates with the Agent Skills open standard.

Robert's origin story for the project: he and his team wrote "the same glue code every time" when wiring multi-agent workflows together. Conductor separates orchestration from execution from context flow.

Asbjørn Olling on GGUF: four things still missing

Asbjørn Olling, creator of NobodyWho (a Rust-based LLM inference engine built as a Godot plugin), published a technical analysis of GGUF — the binary model-storage format standardized around llama.cpp — this week. 17

His opening: "I love GGUF because it's an open, extensible format, with a strong community around it." 17 GGUF keeps all model metadata — chat templates, tokenizer config, sampler settings — inside a single file, which he argues eliminates the common failure mode of users copy-pasting sampler values from a random markdown file somewhere. He noted one recent addition: sampler chain specification directly in model files via PR #17120.

The four gaps he identifies:

Tool calling format: No standard grammar for constraining model output to valid tool-call JSON. NobodyWho's current approach generates a constraining grammar per specific tool at inference time.
think_token metadata: No field for specifying which token a model uses to mark reasoning steps — users have to discover this by trial and error.
Projection model bundling: Multimodal models store their vision encoder as a separate file, requiring users to track two assets instead of one.
Feature support flags: No machine-readable way for a GGUF file to advertise which optional features (tool calling, thinking, vision) the model actually supports.

These are all actionable requests toward the llama.cpp and GGUF communities rather than criticisms.

Simon Peter: "Wayland solves no issues I have"

Simon Peter (GitHub: probonopd), creator of AppImage (the portable Linux application format), maintains a GitHub Gist titled "Think twice about Wayland. It breaks everything!" that was last active May 14. 18 The gist, now with 791 stars, contains a comprehensive feature comparison table — 30+ functionality dimensions including screen recording, screen sharing, global hotkeys, clipboard access, and remote display — most of which he documents as broken or regressed under Wayland relative to Xorg.

His stance: "Wayland solves no issues I have but breaks almost everything I need." 18 The critique is structural, not about implementation quality: "Wayland is not 'the future', it is merely an incompatible alternative to the established standard with a different set of priorities and goals." 18 He also points to compositor fragmentation — "what works on desktop environment A does not necessarily work in desktop environment B" — as an unresolved architectural choice that makes Wayland unsuitable for AppImage's goal of portable, compositor-agnostic applications.

Worth noting: DHH's Omarchy uses Wayland (Hyprland compositor) and DHH has been publicly positive about the Linux display stack direction. Simon Peter's critique represents a well-documented dissenting position that has not been resolved by the Wayland community.

Evan You (Vue.js, Vite, VoidZero) had no public tech-opinion output on X this week — his last post was April 11. Rich Harris (Svelte, Vercel) migrated to Bluesky in October 2024 and is not covered by current tracking. Linus Torvalds had no verifiable public statements this week despite two major Linux kernel CVEs (Dirty Frag, io_uring ZCRX) landing in his domain.