Watch the video, transcript and sources
+ Theo releases Rust TypeScript port
+ Haiku 5.5: cheap tier, real limits
- Infrastructure: few regular maintainersYou expect an AI compiler rewrite to collapse. Theo's Rust port beats Go on his T three Code benchmark, and he says he never read its code. T S C R S is an experimental, open source TypeScript compiler, type checker and language server written in Rust. Theo Browne's project ports Microsoft's Go implementation, so it inherits an existing design and tests. It checks the types in your app and powers editor diagnostics. Wednesday brought Theo's port and a cheaper Haiku. Today's Hacker News board surfaced a maintainer census. The theme is abundant code, scarce ownership.
Theo says months of OpenAI attempts stalled around eighty four percent compatibility. His chart values their tokens above four hundred thousand dollars at API prices. That's usage valued at list prices. The imaginary invoice practically writes its own headline. He says Opus five point five replaced that earlier Rust attempt and produced a working first version in ten hours. The eventual run lasted two weeks, with roughly twenty four thousand dollars of API equivalent usage through his Claude subscriptions. Starting over means replacing the unsuccessful port. Microsoft's compiler algorithms and test corpus were already there. This is unusually friendly terrain for an agent, because the original implementation can keep telling it exactly where it is wrong.
Then the readme draws a Slop Line. Everything below it was written by agents. The project reports more than one hundred eighty thousand passing ported tests, with its own attribution label attached. First question answered. Agents produced a port with an upstream oracle. The maintenance question begins where conversion ends, and a green test suite has a date on it. The flashy speedup is twelve and a half times faster on T three Code with Effect diagnostics, against the older JavaScript compiler plus its plugin. The comparison with the Go compiler plus Effect is closer to twice as fast.
Remove Effect, and Bun wins that same project's table. The winning horse changes with the race. The old Effect plugin also reports different errors, so the stopwatch comes with a semantic footnote. These are the project's measurements on an M four Pro, using five timed runs after a warmup. Its preview binaries also lack optimizations used for the measured build. Put the machine and build settings beside the multiplier before borrowing it for your sales deck. The readme names a September development revision. The tooling manifest still labels a June pin current, so verify the binary you're comparing. Version drift can turn an extra error into a migration exercise.
The release supports Linux on Intel compatible machines and Apple silicon Macs. Windows users get to watch the launch through a window. The known problems list also includes monorepo emit differences and builds reading stale dependency output when references are missing. I'd compare diagnostics and exercise project references before changing a release gate. Keep a long editor session in the test plan. The readme acknowledges growing memory during edits, which your ten second benchmark will politely miss. The interesting deployment angle is WebAssembly. The repo includes a browser example and a small module with type checking in memory. That could bring checking closer to users, but the documented browser build is single threaded and lacks the native language server mode.
Virtual machine savings depend on what an application moves into the browser. An infrastructure bill needs a workload study, even when the compiler has excellent taste in programming languages. Meanwhile, Haiku five point five offers very cheap tokens, with a threshold worth remembering. Prompts up to one hundred thousand tokens get the cheapest tier. Above that, input and output prices both jump fivefold. The large context window contains a pricing trapdoor. Artificial Analysis finds strong small model performance, but at maximum effort Haiku uses roughly three times Luna's output tokens per intelligence test task. Cheap tokens can still arrive in large bags. Their provisional cost figures also exclude that long prompt price step.
On Terminal Bench, Artificial Analysis reports thirty three percent, below Anthropic's roughly thirty nine. Different evaluation setups make those different measurements. Pick an effort setting with your own task success and latency in view, then calculate the bill from actual usage. And the maintainer census on Hacker News counted regular committers across twenty three infrastructure projects. Nearly half had only one or two. Its definition requires ten commits in a year, so it captures contribution concentration while missing some review and support work. The report attributes most recent time zone database changes to Paul Eggert. The official release notes explain why this keeps moving. Governments change clocks, and somebody translates that into software before everyone's calendar develops an opinion.
That answers our open question. Somebody still has to follow upstream changes, reproduce failures and maintain the port's compatibility promise. The agents inherited years of human decisions. I'd like the experiment even more with a visible plan for who owns the next update.
Verdict: NEEDS REVIEW — Match versions; test the maintenance path
Sources
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.
YouTube · Newsletter · thedailydiff.dev · forward this to the intern who deployed on Friday.