Publishing is the trivial part of academic publishing. Preprints already solved the problem of disseminating research results cost-effectively to everyone interested decades ago. The hard part is assigning merit.
Academics need merits to get jobs, promotions, grants, and prizes. The people assigned to evaluate the merits almost never have enough time and/or expertise to actually evaluate the quality of research. If they already know the person they are evaluating by reputation, they evaluate the reputation. Otherwise they use things like publication venues, citation counts, academic pedigrees, and earlier grants and prizes as proxies. Anything that tries to replace prestigious for-profit journals needs to provide non-expert evaluators a way of determining which published papers are likely to be of higher quality than the average.
ML, and other preprint-heavy fields like TCS, mathematics, and physics, all have prestigious venues coexisting with preprint servers. Sometimes these venues are even ArXiV overlays. I don’t see the correlation between open access and merit assignment.
I think you're missing the part where academics still submit to these journals and conferences just like everyone else. They just also put it on ArXiV or theri own website at the same time or sometimes before publishing in the bigger journals.
The majority of journal copyright agreements allow authors to post the article publicly if they choose.
Academics need merits to get jobs, promotions, grants, and prizes. The people assigned to evaluate the merits almost never have enough time and/or expertise to actually evaluate the quality of research. If they already know the person they are evaluating by reputation, they evaluate the reputation. Otherwise they use things like publication venues, citation counts, academic pedigrees, and earlier grants and prizes as proxies. Anything that tries to replace prestigious for-profit journals needs to provide non-expert evaluators a way of determining which published papers are likely to be of higher quality than the average.