Post #2702979

@Forbearance@mastodon.xyz @protowlf@mastodon.gamedev.place someone studied how well coding agents solve simple programming problems in esoteric languages like whitespace and brainfuck. the agents all fail at things like, "compute n factorial for n < 10". this seems to me a pretty strong indication that the coding agents don't have any significant "reasoning" ability, they mainly work by regurgitating semantically-similar-shaped things from their training data. https://arxiv.org/abs/2603.09678