Anthropic spent $20,000 letting 16 Claudes build a C compiler from scratch.

The GitHub issues are better than the compiler. And the media coverage is better than both.

The Setup

Here’s the story: A researcher at Anthropic stuck Claude Opus 4.6 in a bash loop, spun up 16 parallel instances in Docker containers, pointed them at a shared git repo, and said “build a C compiler in Rust. Keep going until it’s perfect.”

Nearly 2,000 Claude Code sessions and 2 billion input tokens later, the agents produced a 100,000-line compiler that can build the Linux kernel across x86, ARM, and RISC-V. It compiles PostgreSQL, Redis, SQLite, FFmpeg, and - the true benchmark - DOOM.

What I Loved About The Build

Some things I loved about this project:

The agents coordinated by taking “locks” on tasks via text files. When two agents tried to grab the same task, git’s merge conflicts sorted it out. No orchestrator. No Jira board. Just vibes and version control.

At one point, Claude accidentally ran pkill -9 bash, killing its own process and ending the infinite loop. Even AI needs a coffee break.

When the agents hit a wall compiling the Linux kernel (all 16 kept fixing the same bug and overwriting each other), the researcher used GCC as an oracle - randomly compiling most files with GCC and only some with Claude’s compiler, then binary-searching for which files broke. Brilliant debugging of the debugger.

The GitHub Issues Are Art

Now, the GitHub issues. They are art.

Issue #1: “Hello world does not compile.” The README example. On day one. $20K well spent.

Issue #221: “Issue - This repo exists.” Steps to reproduce: “Fall prey to the propaganda of non-deterministic ML’s marketed as AI. Spend $20,000. Have your first issue be ‘Hello world does not compile.’”

Issue #228: Someone fed it trigraphs and inline assembly, then wrote: “Seriously? A $20B SoTA model vibe-coded in Rust can’t handle Hello World?” Alternative suggestion: “Write it in Python. I hear Anthropic’s models are pretty good at that language.”

Issue #232: Someone hunted for traces of chibicc (a small educational C compiler) in Claude’s output. Found plenty of matching quirks. The AI didn’t have internet access during development… but it had chibicc in its training data. Their motivation? “Speed-ran my hobby of past three years in two weeks. I got the perfect villain motive to discredit this blasphemy.”

Issue #230: “Respond to issues, please.” (Anthropic has not responded to issues.)

Issue #217: Title: “wow it’s shit.” Body: emoji coffins and “AI AI AI AI AI” on repeat.

And a GitHub commenter summed up the training data debate perfectly: “If I went to the supermarket, stole a bit of every bread they had, and shoved it together, no one would say I made bread from scratch.”

Same Project, Different Media Narratives

But here’s where it gets really interesting - watch how the media covered the exact same project:

The Register ran TWO pieces on the compiler. First: “Claude Opus 4.6 spends $20K trying to write a C compiler.” Then the opinion follow-up: “Anthropic’s AI-built C compiler is not all that impressive… It is not the moment when software engineering as we know it flips over and dies.” Their forums went even harder - one commenter called it “open fraud,” another compared it to an Obfuscated C Contest entry.

India Today and Moneycontrol covered the compiler but framed it through the SaaS panic lens - “After SaaS scare, Anthropic launches new Claude AI with agent teams that build C compilers on their own.” Context: the Nifty IT index had just crashed 6% (its worst day in six years), Infosys dropped 7%, TCS nearly 7%. A $20K compiler demo lands differently when your portfolio is bleeding.

InfoQ played it straight, noting the online debate ranged from “look! a working compiler!” to “it’s not a working compiler, though. And it doesn’t matter how few hours it took, because it doesn’t work.”

Medium, naturally, went full breathless: “16 AI Agents Just Built a C Compiler From Scratch. I Barely Understood Half the Article - And That’s the Point.” The author admits they’ve never written a compiler or used Rust, but assures us this should “hit you in the gut.”

Analytics Vidhya called it a “major milestone in the era of AI development” and reassured readers: “this wasn’t autocomplete on steroids.”

Chris Lattner (creator of LLVM and Swift) at Modular wrote the most measured take: “this is real progress, a milestone for the industry. We’re not in the end of times, but this also isn’t just hype.” He pointed out the compiler’s architecture basically mirrors LLVM - because that’s what was in the training data. “Some have criticized CCC for learning from this prior art, but I find that ridiculous - I certainly learned from GCC when building Clang!”

Techzine noted that Ars Technica cut through the noise: “most of the work was not in the programming itself, but in designing test harnesses, CI pipelines, and feedback mechanisms.” The unsexy truth behind the sexy headline.

And on Hacker News, someone distilled the two camps perfectly: “Pro-LLM: Look! A working compiler! Anti-LLM: It’s not a working compiler though. And it doesn’t matter how few hours it took, because it doesn’t work.”

One independent blogger actually tried using the compiler on real projects and concluded: “It is a neat party trick.”

The Real Takeaway

The honest takeaway? This is genuinely impressive AND genuinely flawed - which makes it the most realistic AI demo I’ve seen. The compiler outputs less efficient code than GCC with optimizations disabled. It can’t do 16-bit x86 on its own. The Rust code quality is “reasonable” by the author’s own charitable assessment.

But a fleet of AI agents wrote a working compiler that boots Linux in two weeks, completely autonomously, for the cost of one staff engineer’s biweekly paycheck. The researcher describes feeling “uneasy” about what this means.

I think the GitHub issues and the media coverage tell the real story of where we are: somewhere between “this changes everything” and “Hello world does not compile.”

The truth depends on which headline you read first.