What’s breaking my head though is not that it is sparse, it is how few samples and gradient steps we are actually talking about. The parameters should barely move anywhere, let alone generalize. It really shows me that some part of my base understanding is wrong.

The fascinating thing is that this still manages to do something non trivial with super few samples, no? Normally you don’t really get generalization from RL, but somehow features seem to be so good that there is some?

Well, the good news is, you can just also poke around and be confused with us!

RIGHT!? This shouldn't work at all!

My collaborators are making orders of magnitude more than me, and so from my perspective, even collaboration borders on free labour for the company (though I love my industry coauthors on a personal level, I do hate their employers quite a bit). Being forbidden from publishing kills any incentive.

I know that "big company doesn't care" is the most likely answer, but do other people push to get guarantees in writing? I am somewhat against working with industry collaborators if my intellectual contributions can suddenly disappear in the machine without any form of attribution or payment.

I know who they are, I just wanted to clown on the peak corporate “nothing”

I think the mail said “by the 20th”?

I’m thinking of getting this as a wall tattoo “live laugh pray” style

People who did a postdoc in the US 😂