[j7ake]

It’s even worse in biology where some labs consistently publish in Nature, Science, Cell. Some of the papers are outright fraudulent. Don’t even trust the numbers.

At least for ML you can mostly reproduce the results, even in if they’re not that interesting.

[dekhn]

I wanted to reproduce (actually: use) a paper published in 2021. They provide a notebook, and I went to run it. I can't even get past the first cell (importing torch) because the API has already changed.

So in biology, they write papers that can't be reproduced becaue they're fraud, but in ML, they write papers that can't be reproduced because the setup is too fragile.