roon says Astra's code makes corrigibility a matter of faith
Theo quote-posted it as OpenAI employees inventing words for unreadable code, then agreed the fix belongs in the model.
roon, posting as @tszzl, told people to actually read the code Astra is writing, and said that once you do, "corrigibility" has become a matter of faith. His description of what the model produces is short and specific. It is using crazy meta-programming and abstruse primitives to write hyperefficient code, and he presumes fable is doing the same thing.
Theo quote-posted him with one line, "OpenAI employees making up new words to justify Astra's unreadable code." Whatever roon said back to him is not in the public thread we have, but Theo's follow-up reply to @tszzl was two words of concession and then a position. "Ah!" followed by "Yeah the solution is still to make the model write better code."
OpenAI employees making up new words to justify Astra’s unreadable code
Yeah the solution is still to make the model write better code
Two people agreeing about the code and arguing about the word
Strip the sniping and the two of them are looking at the same artifact. Neither says the code is wrong. Neither cites a failing test or a shipped bug. The complaint is about what a human can do with the output after the fact, which is a different axis from correctness and one that agent benchmarks do not measure.
roon's framing is the harder one, because corrigibility is the safety term for a system staying correctable by the humans watching it, and his point is that reading the diff no longer tells you much. Theo's framing is a plain code review objection, the kind engineers have had with each other forever, aimed at a model instead of a colleague. They land in the same place. The output should be easier to read, and the place to fix that is upstream in the model rather than downstream in the reviewer.
For working engineers, the thing to notice is what is not in dispute. Nobody in this exchange is arguing the code does not work. They are arguing about whether anyone can check it, and that is a reason to care about style that did not exist when a person wrote the diff.
they are using crazy meta-programming and abstruse primitives to write hyperefficient code