Ronacher: an agent could rebuild CPython in two weeks
Armin Ronacher says an Opus 5.5 agent could re-implement all of CPython in two weeks with no human input, then argued agents are better off writing their own tests. Mario Zechner says that is exactly the part he cannot trust.
Read the whole week in one goEffort levels split the room; Karpathy ships in 2 hours
The big disagreement
Ronacher says agents don't need your test suite; Zechner strongly disagrees
Armin Ronacher said an Opus 5.5 agent could re-implement all of CPython in about two weeks with no human input, against just the tests, including ABI compatibility, no GIL and a JIT.
my problem with agents is trust. preexisting test and conformance suits increase my trust.
Receipts
Theo says Argon 4 is 10x worse in a real codebase, not 5%
ThePrimeagen posted that Google cooked with Argon 4, and Theo publicly contradicted him the same day, saying the model is incapable of working in a real code base without hallucinating in egregious ways. His framing was explicit about the size of the gap: not 5% worse, 10x worse. Neither post shows the runs behind the verdict.
It’s not like 5% worse, it’s 10x worse.
How they actually work
Steve Ruiz's prompt: find all the implicit state machines in my app
Steve Ruiz posted a one-line prompt he likes: "find all of the implicit state machines in my app". Jamon Holmgren ran it on his game codebase and reported being up to 200 of them. David K replied that if you make them explicit, he will give you a nice way to visualize them.
I'm up to 200 implicit state machines in my game code base.