[bonesss]

So the chain of events here is: copy existing tutorials and public/available code, train the model to spit it out-ish when asked, a mature-ish specification is used, and now they jitter and jumble towards a facsimile of a junior copy paste outsourcing nightmare they can’t maintain (creating exciting liabilities for all parties involved).

I can’t shake the feeling that simply being a shameless about copy-paste (ie copyright infringement), would let existing tools do much the same faster and more efficiently. Download Chromium, search-replace ‘Google’ with ‘ME!’, run Make… if I put that in a small app someone would explain that’s actually solvable as a bash one-liner.

There’s a lot of utility in better search and natural language interactions. The siren call of feedback loops plays with our sense of time and might be clouding or sense of progress and utility.

[kungfuscious]

You raise a good point, which is that autonomous coding needs to be benchmarked on designs/challenges where the exact thing being built isn't part of the model's training set.

[NitpickLawyer]

swe-REbench does this. They gather real issues from github repos on a ~monthly basis, and test the models. On their leaderboard you can use a slider to select issues created after a model was released, and see the stats. It works for open models, a bit uncertain on closed models. Not perfect, but best we have for this idea.