Bun's Rust Rewrite: 15,800 AI Commits for What, Exactly?
Bun's Rust Rewrite: 15,800 AI Commits for What, Exactly?
Something about Bun's Rust rewrite is going sideways, and it isn't the performance complaints. Those might not even hold up — nobody has published hard benchmarks. The real story is who wrote the code.
790 human commits vs. 15,800 robot commits
tipiirai, a long-time Bun user, pulled the GitHub numbers for the past month:
- 15,800 commits from robobun (the AI agent)
- 1,600 from autofix-ci (the AI fix bot)
- 790 from Jarred (the actual human)
Human-written code is under 4% of the repo's recent output.
Jarred's own tweet, quoted in the post, lands harder: "6 months ago, most of Bun's PRs came from people prompting Claude. Nowadays, most of Bun's PRs come from Claude prompting Claude."
Read that again. First a human told the AI what to build. Now the AI tells itself.
5,000 open PRs is the new kind of tech debt
Bun currently has over 5,000 open pull requests. React has 441. OpenClaw has 2,200. GitHub recommends staying under 1,000 per branch before mergeability checks start timing out.
That number says one thing: code is being produced faster than anyone can review, understand, or merge it. The bottleneck moved. Writing code got cheap. Reading code got expensive.
This is a new shape of technical debt. Not "the code is bad" — "the code is coming in faster than you can look at it."
The memory-safety justification didn't survive contact
The rewrite was pitched as a memory-safety move: Zig's manual memory management is risky, so switch to Rust and let the compiler catch it. But tipiirai points at the volume of unsafe blocks in the Rust code and says the safety promise didn't land.
Zig's creator, Andrew Kelley, was blunter. Writing about the rewrite, he said the codebase left him "increasingly horrified": "Hacks on top of hacks. Abuse of assertions. Jarred was already writing slop well before he had access to LLMs."
My take: if it was slop before the rewrite, using AI to produce more slop, faster, scales the problem instead of fixing it.
This is the biggest public test of AI writing production code
tipiirai's sharpest line: this rewrite is the most closely watched real-world test of whether AI agents can take over a production codebase with a human mostly directing rather than reading.
Anthropic's reputation is riding on it too. If it works, it's the best ad for agentic coding ever made. If it fails, it's a signal in the other direction.
So this matters well beyond "should I upgrade Bun." It's answering a question the whole industry is asking: does AI commit volume equal code quality?
Bun's answer so far is no — and the cost is visible. Three months without a stable release, the longest gap since 2022. A community where "did tomorrow arrive yet" is a running joke.
What you can do today
Don't abandon Bun, and don't write off AI-written code. But if your team is letting AI rewrite or generate at scale, ask three questions first:
- Why are you rewriting? If it's performance, measure the baseline before you touch anything — don't wave "memory safety" around as a cover story.
- Who's reviewing? An AI can push hundreds of commits a day. How many can your humans actually read? That gap is your tech debt.
- Are you comfortable with AI on safety-critical paths? Unsafe blocks, injection points, permission boundaries — the AI writes those happily, and it won't be the one on call when they blow up.
One line to remember: AI made writing code cheap and reading code expensive. Don't let the first number hide the second.
✨ Draft generated by DeepSeek, reviewed and polished by Claude.
Sources:
