The Adam optimizer was published at ICLR, a top conference in machine learning. So, basically, analogous to a peer-reviewed journal for the purposes of this discussion. ML (and some other subfields of computer science) have the particularity that the really comptetitive gatekeeping happens even more in conferences than in journals.
That said, there definitely are very relevant papers that are not published in any peer-reviewed venue. A good example is "Language Models are Unsupervised Multitask Learners" (the GPT-2 paper, which I would argue started the whole generative LLM revolution). But I think if you look for this kind of papers, you will find something in common to all of them: they are by very well known researchers, elite institutions or influential companies. That's why people went out of their way to read them even if they were posted somewhere without peer review.
If you removed peer review and just relied on posting to arXiv or similar, new researchers, or researchers from less known institutions, would have no chance at all to make an impact. It's peer review that allows them to be able to submit to a top journal, where the editor and reviewers will read their paper, and they can get a somewhat fair chance.
PS: I don't really like the peer review system that much either. It's just that every alternative that I have seen proposed so far is worse.
> The Adam optimizer was published at ICLR, a top conference in machine learning.
Fair, must have misremembered that one.
> If you removed peer review and just relied on posting to arXiv or similar, new researchers, or researchers from less known institutions, would have no chance at all to make an impact.
I disagree on this one. I did my PhD at an institution that ranks in the top 10 in the most well known university rankings, and I distinctly remember that one time when I was submitting a manuscript to a prominent journal in my field, got some reviews back which weren't positive yet were quite valid criticisms, and my professor told me not to worry because the editor is his buddy and my manuscript will get published for sure.
When that sort of "scratch my back and I'll scratch yours" culture exists in journals I don't see how peer review can be an equalizer. It just means everyone who publishes at a journal, including the less well esteemed ones, can claim they went through the rigor of peer review. Of course, we all know peer review is just a vibe check and is actually not that rigorous at all, and besides no one cares unless you published in a prestigious journal anyway. The less revered journals exist to collect $5k in open access fees for the publisher in return for hosting a pdf at the marginal cost of maybe a cent a year.
That said, there definitely are very relevant papers that are not published in any peer-reviewed venue. A good example is "Language Models are Unsupervised Multitask Learners" (the GPT-2 paper, which I would argue started the whole generative LLM revolution). But I think if you look for this kind of papers, you will find something in common to all of them: they are by very well known researchers, elite institutions or influential companies. That's why people went out of their way to read them even if they were posted somewhere without peer review.
If you removed peer review and just relied on posting to arXiv or similar, new researchers, or researchers from less known institutions, would have no chance at all to make an impact. It's peer review that allows them to be able to submit to a top journal, where the editor and reviewers will read their paper, and they can get a somewhat fair chance.
PS: I don't really like the peer review system that much either. It's just that every alternative that I have seen proposed so far is worse.