August 3, 2026
Code, Chaos, and Comment Wars
What's the largest software project AI can complete on its own?
AI built a whole program alone, and the comments instantly turned into a brag-fest
TLDR: A new test shows AI can now rebuild some full software tools on its own, but it still stumbles on tiny hidden gotchas and is far from perfect. In the comments, people swung between amazement, suspicion about fairness, and full-on flexing with their own marathon AI coding experiments.
The big headline here is wild enough on its own: researchers say today’s best AI can now rebuild some entire software programs from scratch, with no internet, no original source code, and no human stepping in. On the new MirrorCode test, one top model rebuilt a bioinformatics tool with about 16,000 lines of code and more than 40 commands in just 14 hours for $251. But before anyone crowns the robots king, the overall score was only 56%, and one of the best attempts still face-planted on a single weird edge-case test. Yes, after thousands of checks, it was undone by the software equivalent of "but what about this one obscure date setting?"
And the real fireworks were in the comments, where the crowd split into camps: the "this is amazing" people, the "why are only the paid AI tools invited to the party?" people, and the glorious "I’ve already been doing this at home" crowd. One commenter said when they let AI code on its own, it starts wandering off into "strange directions," which is either a cautionary tale or the plot of every AI project ever. Another rolled in with an 81-day, 2,600-commit saga of using Claude to build a Bash clone in Rust, basically turning the thread into a surprise talent show. The mood was equal parts awe, jealousy, skepticism, and meme energy: AI can clearly build bigger things now, but the internet is still arguing over whether it’s genius, cheating, or just an extremely confident intern.
Key Points
- •MirrorCode is a benchmark designed to evaluate whether AI can autonomously reimplement entire software programs without access to the original source code.
- •The benchmark uses larger inference budgets than many prior software engineering benchmarks, with one task costing $2,600 and running for 19 days.
- •Claude Opus 4.7 reimplemented gotree, a roughly 16,000-line Go bioinformatics toolkit, in 14 hours for $251, though the benchmark remains unsolved overall.
- •Claude Opus 4.7 achieved a headline MirrorCode score of 56%, while leading models from about a year earlier would have scored around 30%.
- •The authors note possible pretraining contamination from open-source code, release 22 of 25 target programs and the scaffold as open source, and keep three targets private as a test set.