Hacker News Reader

In the must general terms, how do you verify that the code produced by agents does what you intend it to do?

How do you approach about this problem?

Modern agents can write tests that are meaningful and require the agent to pass them with any change to avoid regressions. Humans can review the code/test the downstream application to ensure it works as intended.

It rarely takes a single prompt to get something done, but the agents can figure out as long as the human is specific about what constitutes accurate.

Via Doctorow a day ago:

https://singletrackworld.com/2018/01/collision-course-why-th...

Something about cycling?

you can't use the thing that doesn't quite work to ensure that the output of the thing that doesn't quite work works

It's counterintuitive yes, but you can. You can just look at the tests to ensure they're consistent and with the latest models that has always been the case, and it's very rare that the models try to cheat the tests.

you can convince yourself it can, by all means. But it doesn't make it true. In fact we even have this rule for apecoding: a developer cannot review their own code.

Reply on news.ycombinator.com