Chad Smith

@headinthebox If trump were a Haskell engineer this is what he'd sound like

I gotta tell you, folks, Claude Code — very overrated, very brain-dead software. Nobody’s talking about it, but I will. It’s a disaster. You’re typing, you’re focused, beautiful code, really tremendous code — and suddenly, boom — it hijacks you. Total focus theft. You’re thrown into another IntelliJ window like you’ve been kidnapped. Nobody knows how. Nobody understands it. You’re typing, typing, typing — and your characters? Gone. Disappeared. Maybe they’re in another window, maybe they’re in another dimension. Nobody knows. The engineers? They have no idea. Zero clue. This thing just spews output wherever it wants. No respect for the user. No respect for workflow. It’s chaos. Absolute chaos. Very disruptive. Very irritating. Frankly, unacceptable. And they call this intelligent? I don’t think so. I’ve seen smarter autocomplete in Notepad.

@olofj The age of new languages with LLVM backends has arrived LLM-friendly frontend, LLVM backend github.com/magpie-lang/ma… TypeScript frontend, LLVM backend github.com/cs01/ChadScript

Today's experiment just for fun, tasking Claude with: Design a language that is optimal *for you*. Not for humans. Minimal token cost when generating code, and something you will enjoy coding in yourself. First attempt was something fairly Rust-like. Too alike, honestly. So I reminded it that there were no constraints, and didn't have to be easy to read for humans. That made for some interesting choices. It called the second design RADICAL.md. I called it Ore. It's about halfway through implementing it now (with an LLVM backend). Not looking to do anything with it, just playing around. Still, fun times to be had by tinkering with it. github.com/olofj/ore-lang…

Just when I got claude code cooking overnight for the first time, it shipped 27 PRs in about 11 hours, then BAM login issue and I still can't get back in. It would have taken me so much longer to pair program and fix all these legitimate quality issues. Total game changer.

@unifiedh Cool project! Can it self-host?

I built the world's first programming language designed for LLMs and AI Agents. Introducing Magpie. Check it out yourself at: magpie-lang.com Details below.

Incredible analysis in seconds that would take a person hours to weeks. "The main debugging challenge was an intermittent SIGSEGV in the native compiler's findI64EligibleVariables. GDB revealed rax contained ASCII string bytes ("block\0re") being dereferenced as a struct — the function was receiving a BlockStatement* instead of an ObjectArray* for the statements parameter. Root cause: func.body || { statements: [] } produces an opaque i8* in the native compiler, and .statements on an opaque pointer doesn't generate a GEP."

@GergelyOrosz Frameworks can be rewritten, and even languages. Typescript to native github.com/cs01/ChadScript Microsoft attempted this in 2019 but never finished. With llms the vast quantity of tedious code to self host a compiler can be done by a guy who doesn't know compilers in his spare time

We will see much, much more of this happening. AI is changing open source incredibly rapidly. Rewriting an open source project to a new language/framework used to be a massive effort: AI is making it trivial as Cloudflare just showcased with their 1-week NextJS rewrite

@t64_bv First draft of jsx support has landed github.com/cs01/ChadScrip…

Great project. Any interest in collaborating on Rezi/Zireael integration into a TypeScript to LLVM IR compiler that can link directly against your C library? It would be a native executable with no JavaScript runtime whatsoever. I'm adding JSX support to it now. github.com/cs01/ChadScript

I’ve been working on a side project — a TypeScript TUI framework that lets you write React-style components for the terminal. But rendering doesn’t run in JS. It runs in a native C engine that diffs framebuffer state and only writes changed cells. So you get: React/TS-style components Partial redraws Much lower overhead than JS-only renderers Works with Node and Bun Meet Rezi 👇 github.com/RtlZeroMemory/… #typescript #bun #nodejs #tui

@t64_bv I'm not able to but mine should be open now

@cs01_software Hello. DM me

@enggirlfriend It can compile itself which afaik usually takes years for a compiler to achieve

@enggirlfriend The closest I can think of given the current LLM training set is faster impls of existing languages. Many new languages will emerge now that they are nearly trivial to build. I've never written llvm ir and I built this in a few weeks with Claude. github.com/cs01/ChadScript

i keep thinking a new programming language interface will be invented soon that’s simpler and easier for humans to read/write. and the code we write now will be abstracted away for LLMs to do the heavy lifting for. kinda like how today we don’t see or touch assembly

@t64_bv Nice work. I like the way AI lets us build ergonomic tools with faster runtimes.

@clattner_llvm @AnthropicAI "The Claude C Compiler doesn’t mark the end of software or compiler engineering. If anything, it opens the door wider." Indeed! I was able to build a self hosting typescript to llvm compiler with no llvm experience github.com/cs01/ChadScript

The Claude C Compiler is the first AI-generated compiler that builds complex C code, built by @AnthropicAI. Reactions ranged from dismissal as "AI nonsense" to "SW is over": both takes miss the point. As a compiler🐉 expert and experienced SW leader, I see a lot to learn: 👇

@RhysSullivan what if it's written in typescript and compiles to the same backend rust uses cs01.github.io/ChadScript

you'll know if something is written in rust because people will tell you it is you'll know if something is written in typescript because you'll actually be using it

@nbaschez @simonw I do something similar where I tell it to occasionally share "teaching moments" to help keep me on the loop. It shares new concepts relevant to the work. Great JIT learning for me about the code or concepts.

my current favorite trick for reducing "cognitive debt" (h/t @simonw) is to ask the LLM to write two versions of the plan: 1. The version for it (highly technical and detailed) 2. The version for me (an entertaining essay designed to build my intuition) Works great

> Rust is nowhere near optimal for LLMs as a target language True, typescript is. But it targets an interpreted language. Unless someone wrote a compiler to generate llvm ir from it to get a compiled binary that runs at the speed of Rust with the syntax of typescript. github.com/cs01/ChadScript

I think it must be a very interesting time to be in programming languages and formal methods because LLMs change the whole constraints landscape of software completely. Hints of this can already be seen, e.g. in the rising momentum behind porting C to Rust or the growing interest in upgrading legacy code bases in COBOL or etc. In particular, LLMs are *especially* good at translation compared to de-novo generation because 1) the original code base acts as a kind of highly detailed prompt, and 2) as a reference to write concrete tests with respect to. That said, even Rust is nowhere near optimal for LLMs as a target language. What kind of language is optimal? What concessions (if any) are still carved out for humans? Incredibly interesting new questions and opportunities. It feels likely that we'll end up re-writing large fractions of all software ever written many times over.

@DavidKPiano Writing code was definitely hard

Okay let's be real, writing code was never the hard part. Understanding the problem, making the tradeoffs, knowing what (not) to build, etc... LLMs made the typing 10x faster but it didn't make the thinking 10x easier. Also let's not pretend we were writing code in notepad before LLMs. We had frameworks, generators, libraries, intellisense, snippets, autocomplete, StackOverflow (RIP), and an entire ecosystem of abstractions designed to make writing code not *that* tedious.

Been heavily using Opus 4.6. > It is much more terse and to the point. > It never tells me I'm absolutely right (end of an era). > It is extremely good at coding, root causing multiple layers of code generation. Uses gdb for segfaults effectively. >It has corrected me on desgin decisions. > And it's only going to get better.

@ChShersh John Carmack said he thinks AGI will be less than 10,000 lines of code, which is apparently 40x larger than what it takes to build GPT.

Can’t believe a 243 LOC Python script is taking away my job