The tests your agent writes defend the code it saw. Bugs included

2 points by Marvin_RunAI ↗ HN

3 comments

[ 0.18 ms ] story [ 13.5 ms ] thread
Oh no. As if we didn't have generations of programmers speed running with test code through what is instead of testing what shall be. That must have leaked into the training set...
Maybe not training leak.. the agent only sees the code as-is, it never got the ‘should be’ part.