Mario Zechner warns frontier models break code behind you
Two posts from @badlogicgames make the same argument from different directions, and neither is about model benchmarks.
Mario Zechner, posting as @badlogicgames, put out a short warning to programmers who have stopped reading their own code. "hi kids, gramps here. learn your algorithms and datastructures, cause SOTA frontier models will still fuck you up bad behind your back, especially if you don't read your code anymore," he wrote. The post carries an attachment, and he does not name a model or a task in the text itself.
learn your algorithms and datastructures, cause SOTA frontier models will still fuck you up bad behind your back, especially if you don't read your code anymore
That is the whole claim, and it is worth being precise about what it is and is not. It is one experienced engineer's report, not a benchmark, and there is no before and after diff in the post to inspect. What makes it more than a grumble is the specific mechanism he points at. The failure is not the model refusing to work, it is the model producing something that runs while quietly getting the underlying structure wrong, which is exactly the class of bug that survives review when nobody reads the code closely.
The companion argument
A second Zechner post recommends an article on the same theme, with the caveat that he thinks it "goes too hard on bend 2". His summary of the point is that you often only understand a problem by experiencing the journey, meaning by building a solution for it yourself. Hand that journey to an agent and you get something back, he says, but not necessarily something good or even correct.
you often only understand a problem by experiencing the journey, i.e. building a solution for it.
The two posts fit together. Knowing the algorithm is what lets you notice that the generated implementation is the wrong shape, and building things yourself is how you come to know the algorithm. That is a loop with an obvious hole in it once the building is delegated, and Zechner does not offer a way out of the hole beyond reading the output.
For working engineers, the practical version is narrow. If you cannot describe what the correct data structure for a task is before the agent starts, you have no way to tell whether the code it returns is right or merely plausible.