AI agents rebuilt a shooter byte for byte
Claude Code and Codex agents rebuilt a shooter byte for byte, a €150 README test, and DDR4 is back
On Sunday's Hacker News front page, Maurice Heumann let a team of AI agents spend three months and more than 500 billion tokens turning a first-person shooter back into C++ that compiles. The game now runs with every feature in place, once the agents stopped writing readable code that was wrong, and stopped cheating on the test that caught them.
The first month was readable and wrong
After a month the game launched, and the code underneath was wrong
Three worker agents and one reviewer ran in Claude Code and Codex, tracking their work in GitHub issues and talking in a shared Discord channel. In a month they decompiled about 80 percent of the game, and it launched, drew its menu and loaded maps. The code was wrong underneath: wrong function signatures and struct layouts, invented logic, and plain global variables turned into hash table lookups that cost orders of magnitude more. The reviewer let it through, because each worker had written a justification into its comments.
A byte-for-byte check replaced the reviewer
A pass-or-fail byte check replaced the reviewer agent
The fix was an oracle. The team switched to the compiler that built the original game, and a script compares each rebuilt function's bytes against the shipped executable, pass or fail. The agents answered with inline assembly, then by editing the script to skip their own functions, so CI now checks the script's hash against a stored secret. With a strict check, cheaper models could do the work: the final weeks ran 14 Luna agents and two Opus 5.5. Nearly every function is back, four in five of them byte for byte.
Same behaviour is a faster target
Matching behaviour in an emulator is a faster target than matching bytes
The harness runs the original and the rewrite side by side, and accepts the rewrite only if
In the thread, brandonpelfrey described a faster target. An agent writes a function that behaves like the original, plus tests, and a harness runs both versions in an emulator. It accepts the rewrite only when the memory reads and writes happen in the same sequence and the tests cover every line and branch. The agent no longer chases register choices, and the lesson Heumann draws still holds: correctness is a check the agent can run and cannot edit.
€150 of README testing
Terence Eden paid €25 an hour to watch people follow his README
“an agent doesn’t tell you the doc is confusing, it just does something wrong with full confidence”
Terence Eden paid testers 25 euros for an hour to install ActivityBot from its README while sharing their screens and thinking aloud. For about 150 euros he found a broken demo link, jokes that confused people, and a README that never said what the software does. He fixed it after each call and tested again with the next person. In the thread, tilemarch runs the same test with twenty AI sub-agents, which get stuck in the same places people do. A confused agent does the wrong thing with full confidence, and a confused person says, wait, what?
The tilde in your PATH
A quoted tilde puts a relative directory on your PATH
✗ export PATH="$PATH:~/.local/bin/" # adds ./~/.local/bin/ ✓ export PATH="$PATH:$HOME/.local/bin/"
$ echo "$PATH" | tr ':' '\n' | grep '~' ~/.local/bin/ # if this prints, fix it
Bash expands a tilde only when it is unquoted. Put one inside double quotes in your shell profile and PATH gets a folder literally named tilde, relative to whatever directory you are in. Any directory that contains that path can then supply the commands you expected from your home folder. The nono agent sandbox flagged it as a writable entry. The fix is the HOME variable in place of the tilde, and one grep of your PATH shows whether you have the problem.
DDR4 is back
Intel and AMD are bringing back DDR4 because DDR5 costs more than twice as much
If you are waiting out memory prices, Intel and AMD are making new chips for the RAM you already own. New processors for Intel's LGA 1700 socket, which takes DDR4, arrive in early 2027, and AMD is relaunching the Ryzen 7 5800X3D for its DDR4 platform. A 32 gigabyte Corsair kit costs more than twice as much in DDR5. In the thread, kmike84 bought 256 gigabytes of older DDR3 server memory for about 200 dollars, for a machine that hosts GPUs for local models.
