Published: August 2, 2026
Andrew Kelley, creator of the Zig programming language, published a blunt critique of the Anthropic-owned Bun runtime’s Zig-to-Rust rewrite on July 9, 2026, writing that he now finds himself “battling the marketing arm of Anthropic” and that Bun’s explanation of the rewrite reads as if “the marketing department of a trillion dollar company has a lot of money riding on this article.” Bun, a JavaScript and TypeScript runtime created by Jarred Sumner and acquired by Anthropic in 2026, merged an AI-driven rewrite of its codebase from Zig to Rust in May 2026 and published its rationale roughly two months later, in early July. The dispute reached 436 points and 202 comments on Hacker News by July 13, driven by a follow-up essay from software engineer Ray Myers.
What Andrew Kelley said about the Bun rewrite
Kelley’s post, “My Thoughts on the Bun Rust Rewrite,” rebuts 4 specific claims in Bun’s official explanation. First, Kelley wrote that Bun’s post “implies you were diligently fuzzing your Zig code, while during our calls the Bun team told us that they were not fuzzing anything.” Second, he attributed Bun’s reported performance gains to link-time optimization, “which Zig has supported for all of Bun’s existence.” Third, he argued the binary-size engineering Bun cited “had nothing to do with the rewrite.” Fourth, he challenged Bun to publish post-rewrite compilation times, noting the Zig compiler — roughly 600,000 lines of code — builds from scratch in 16 seconds on his machine.
Kelley also disclosed a business rupture behind the technical dispute. Bun had donated $60,000 per year to the Zig Software Foundation; according to Kelley, the donation “silently stopped” after the Anthropic acquisition, and the Bun team stopped attending its monthly meeting with the foundation. Kelley wrote that Bun’s codebase had long invited criticism — “Jarred was already writing slop well before he had access to LLMs” — and that the Zig team welcomed the rewrite: “It Tastes Like It’s Not My Problem Anymore.”
The use-after-free the AI did not catch
Ray Myers, a former chief architect at a coding-agent startup, amplified Kelley’s critique in a July 12 essay arguing that Anthropic is an “unreliable narrator” on AI coding capability. According to Myers, the rewrite — a file-by-file port of the codebase to unsafe Rust, described by Bun as almost entirely AI-written — shipped a use-after-free bug that the AI failed to catch, the exact class of memory error the migration was meant to eliminate. Myers wrote that management approved the Rust rewrite in part “because it was a great marketing opportunity to showcase their new Fable model.” He added that while researching the story he caught 50 bugs in the Zig version of Bun using his own hybrid analysis approach.
Myers credited one part of Bun’s explanation: the decision to port to unsafe Rust file-by-file to minimize migration risk, which he called “a sensible move explained well.” His objection targets the narrative, not the technique — Anthropic, which has raised $132 billion and is approaching an IPO Myers says would value it above $1 trillion, needs the story that AI can replace software engineers, and a shipped memory bug undercuts that story. Neither Anthropic nor Bun has publicly responded to Kelley’s or Myers’s posts as of August 2, 2026.
What this means for businesses evaluating AI coding claims
Buyers should treat vendor-published AI case studies as marketing exhibits, not evidence, and weigh independent tests before acting on them. The Bun rewrite is the highest-profile agentic coding showcase to date, and its loudest critics are not AI skeptics — Kelley leads a major systems language, and Myers builds coding agents for a living. Their shared complaint is selective disclosure: published upsides, unpublished trade-offs, and a two-month gap before any rationale appeared. The same diligence applies to choosing agents for business work; our comparison of the 12 best AI agents for business tasks evaluates vendors on documented behavior rather than launch-day claims.
Our Take
Strip away the language-war drama and one detail remains: an AI rewrote a million-line codebase into a memory-safe language and still shipped a memory-safety bug, because the port used unsafe Rust. That is not a scandal — it is a measurement. Agentic rewrites are real, fast, and imperfect, and the imperfection surfaced only because outsiders went looking. Ask every AI coding vendor the question Bun’s blog post skipped: what are the cons?
For Context
WithO2 has covered the stakes behind Anthropic’s capability narrative: the company’s $1.2 trillion secondary valuation that passed OpenAI explains why marketing wins matter, and Claude Fable 5’s move to usage credits covers the model at the center of the Bun rewrite. For independent evidence on what coding agents actually deliver, see Databricks’ head-to-head test of AI coding agents on real code.
Related reading: our GitHub Copilot vs Cursor vs Claude Code comparison ranks the assistants these capability claims are meant to sell, and Kimi K3’s open-weight challenge to Claude Fable 5 tracks the competitive pressure behind the marketing push.
