david barsky

like, jj undo alone? man, so good

claude + jj is kinda magical, even if claude code-the-harness wants to think in terms of git worktrees. others have said this, but a “use jj instead of git” in a claude.md works extremely well

a receipt from apple, showing that i bought the new fancy monitor

that 3.6k was burning a hole in my pocket

i’m illiterate in russian, but i can speak/listen

it’s good they’re saying this publicly. if you’re in an iterated prisoner’s dilemma, you have to punish defection!

has anyone figured out how to make claude not think that some issue is a pre-existing issue? we don’t *ever* break our builds, so any issue it finds is something it introduced

i’m a centrist in that i think measuring this stuff is kinda silly and risks veering off into weird places AND i think peter watts raises some good points

the human compaction process is only 10 bits/second, though :/

I'm not mad. Please don't record in the oplog that I'm mad

arguably against the employee handbook to not OH this post, smh

on one hand, it’s *good* that we need to this so early, on the other, UGH

anyway, guess who is doing soc-2 compliance today?

i think the implications for the former are obvious, but the latter means for every non-cloud vendor you get, need some sort of secret key rotation policy. compliance paperwork!

i think two of the most anti-competitive policies cloud providers have are zero-rating traffic to/from their own services and IAM not being accessible to third-party SaaS

i think the idea is to keep people on GLP-1’s. i’m going to be on sertraline and dextroamphetamine for the rest of my life, i don’t think it’s unreasonable to keep people on these drugs for similar reasons

i think part of my annoyance is that doing so requires nightly for rustc_private, but also, the mindset behind that sort of gating. i don’t think compiler api instability really mattered that much initially, but now it _really_ doesn’t matter because an LLM can just deal with it

there’s a bunch of lints that i’d like to write for my codebase (no booleans, shadow as much as you can, etc.) that require me to reach for dylint instead of simply authoring a new clippy lint

go’s commitment to making good, self-serve analysis tooling is truly inspiring and i wish rust would do more of this.

“Using go fix to modernize Go code” by Alan Donovan — https://go.dev/blog/gofix

#golang

i should, at the very least, push up my skill for this that describes the grammar and how to use it. gimme a sec.

keep an eye on this branch; i’m slammed this week and didn’t have the time to try out the “no bash” rule. i also know there’s some dumb bugs where it doesn’t find symbols when it should.

github.com/davidbarsky/...

like, rust-analyzer will do native type errors, but it doesn’t run anything resembling borrowck. this might only be a problem for humans, not machines, though…

(late response, sorry!)

it’s great, but i think that claude’s lsp integration isn’t even particularly fleshed out! for example, it only fetches rust-analyzer’s native diagnostics, not rustc’s richer ones, so while those are nice… there’s still so much unpicked low-hanging fruit

it might, i think? i think i’d still like derive macros in this language and some mechanism to support tracing-style spans (i’m biased…), but i don’t know if function-like macros are worth the squeeze. the sheer niceness of jsx in typescript though… i dunno.

do i know anyone who works at azure that can get me access to horizondb? i can elaborate more in dms if needed.

anyways, i’m not fully awake yet, but as far as i’m aware, rust-analyzer is the among the largest lsps in existence (500k LoC). kotlin’s k2 might be bigger, but every other lsp i know of (gopls, clangd, etc) are in the 30-50k range. there’s a lot of functionality in that 450k.

granted, i don’t really like macros in rust (except for derive macros) nor the module system, but i also didn’t like them before llms got really good. i’d like to see an llm-native pl that internalizes codegen is cheap, has a simple module system, has effects, and has good structural edit tooling.

i like it so far, but i think it’ll really sing once i disable bash access in claude/codex. the borrowck makes imperative code scale nicely for humans and machines(™) because it reduces the axes of freedom required for reasoning. pair that with property testing and enforcing newtypes… it’s good.

…because the idea behind RLMs is portable and allows for sub-linear token usage with respect to edit size. rust-analyzer has a really rich set of type-aware refactoring and navigation tools (java is maybe comparable?) and rust has _really_ good error messages.

right, okay: i think (kinda by accident?) rust has some really good tools and constraints that make it a surprisingly good fit for llms. there are some bits that aren’t, but they were known painpoints to begin with. i ported some ideas from the RLMs paper into an mcp wrapper around rust-analyzer…