2026-07-09 · Rewriting Bun in Rust

Transcript

Transcript. Paraphrased; sources in notes.md.

HostWelcome to July 9th, 2026. It is a beautiful day to talk about the strange, shifting ground of modern software engineering. We have a packed slate today. We are looking at the Bun team's massive migration from Zig to Rust, the arrival of a rather, let us say, ambitious distributed computing project called Abject, and finally, a conversation about an AI-free fork of Vim. Fred, thanks for joining me. You seem like you have been pacing the floorboards already.

GuestIt is hard not to, Daniel. The headlines lately feel like we are just spinning a giant wheel of misfortune. I spent the morning reading about this Bun migration, and honestly, the word vibecoding is starting to make my skin crawl. It implies that if you have a high enough token budget and a decent enough prompt, you can just manifest a memory-safe, high-performance runtime into existence. It is not coding. It is praying to a stochastic parrot and hoping the unsafe blocks don't bite you in production.

HostWell, let us dive into that then. Over on Lobsters, the community is tearing apart the Bun team's decision to rewrite their core in Rust, specifically highlighting the sheer cost of the LLM-assisted migration. The author’s claim is that after struggling with systemic memory issues in Zig, they spent around 165,000 dollars on token costs to use Claude Fable 5 to help translate the codebase to Rust.

Guest165,000 dollars to find out that manual memory management is hard? Look, I respect the Bun team for admitting the limitations of their previous approach. Zig gives you a scalpel, but you have to know exactly where to cut. But the idea that you can solve architectural complexity by throwing six billion tokens at a model is absurd. A Lobsters user called aw1621107 points out that the blog post is surprisingly light on why the rewrite was the only path forward, and frankly, I agree. If your project is so complex that you need a machine-learning oracle to port it to a memory-safe language, you might have built a system that is fundamentally too fragile to maintain, regardless of the language.

HostIt is a valid concern. One commenter, refi64, noted that Node.js has been chugging along in C++ for ages. The Bun team argues they wanted to systematically prevent bugs, but is there a risk that by using an LLM to perform the port, they have just baked in a different, less visible class of logic errors that their human engineers might not fully grasp yet?

GuestExactly. You end up with a codebase that is technically in Rust, but might lack the idiomatic soul that keeps Rust safe. If you did not write the borrow checker logic yourself, you are just trusting the machine that you are actually safe. It is security theater performed by a GPU.

HostLet us move from human-led, AI-assisted rewrites to something that leans fully into the automation. Abject 0.8.2 is a project claiming to be a grass computing platform, where autonomous objects communicate via message passing and use LLMs to negotiate protocols and self-heal. It sounds like a fever dream of distributed systems design, honestly.

GuestGrass computing? I am going to have to stop you there. It sounds like someone wanted to make their architecture so unpredictable that debugging it becomes a philosophical exercise rather than a technical one. When you have objects that regenerate their own code while the traffic is still flowing, you have effectively abandoned the concept of a state machine. You are just throwing packets into a blender and hoping the result is a valid API request.

HostThe Lobsters thread is, predictably, quite heated. One user, Irene, makes the argument that this kind of tooling is de-training engineers, moving them toward a kind of technical servitude. She suggests that if we stop learning how to research and build these things ourselves, we lose our grip on the systems we are supposed to be mastering.

GuestIrene is hitting the nail on the head. There is this arrogance in current tooling where we treat everything as an LLM promptable object. If your system breaks and the self-healing proxy just writes a new, equally buggy proxy to replace it, you are not debugging a system. You are just watching a feedback loop of incompetence. I would rather have a static, boring system I can reason about with a debugger than a living ecosystem of agents that refuse to tell me why they are crashing.

HostThat brings us to our final topic, which feels like the necessary pendulum swing. Drew DeVault has forked Vim to create Vim Classic, a project explicitly designed to be AI-free. The author’s claim is that this is about maintaining a human-centric editor while also considering the environmental and ethical costs of the current AI boom, like data center energy usage and labor conditions in mining.

GuestFinally, some sanity. I’ve been using a terminal-based editor for decades, and the idea that I need an LLM to complete my function signatures is just evidence that the language I am writing is too verbose. One commenter on Lobsters, citbl, mentioned that they tried Vim Classic and found they didn't miss much from the newer, bloated versions. It turns out, you don't need a neural network to edit text. You need a buffer and a set of keybindings that don't change every three weeks.

HostThe thread does take a turn, though. There is a debate about the history of UI, specifically regarding the pull-to-refresh gesture. One user, wezm, corrected a misconception, noting that Loren Brichter didn't invent that interaction to exploit human dopamine loops, but rather because it was a functional, intuitive way to load new data. It’s a good reminder that not everything in tech is a malicious plot.

GuestRight, but the fact that we have to clarify the history of a gesture to defend against the assumption that it was designed to exploit us just shows how cynical the landscape has become. We are so used to tech being weaponized for engagement that we forget someone might have just made a good design decision once. Personally, I’m sticking to my plain text and my human-verified commits. I’ll keep the LLMs for summarizing meeting transcripts I wasn't going to read anyway.

HostThat sounds like a solid plan. I think that is a wrap for today, Fred. I appreciate you bringing the skepticism, as always.

GuestAny time, Daniel. It’s better to be the person pointing out the fire than the one trying to optimize the burning building, right? I've got a weekend of legacy code maintenance ahead of me, which honestly sounds like a vacation after reading about all this.

HostI bet. Thanks again to everyone for listening, and thanks to you, Fred. We pull these stories and the accompanying debates from Lobsters, which is always the best place to see the real-world friction of these new ideas. We will be back tomorrow with more. See you then.