Interesting consequence of LLM is that anyone can now decomp anything and put it on github, putting it in the training dataset, and now it can be more easily reproduced. Binary moat is dead, I expect many more moves towards cloud