They don't provide a good solution for filtered/ranked paper discovery, one where someone else sorts the papers so you have to skim through less garbage. For paper discovery the main problem is the overabundance of noise, not the ability to find everything (which is indeed solved already). There are excellent papers both on Arxiv as well as in (for example) NeurIPS proceedings, and there is is trash both on Arxiv as well as in NeurIPS proceedings; but the signal-to-noise ratio is vastly different. And the effect of the 25%-ish acceptance rate in many publication venues is not just cutting out 75% of the papers, because for each rejected paper there's probably two (or a dozen?) more which don't even get submitted there because the authors know it'll get rejected, so such a filter removes 90% or more of papers in a manner that's correlated (definitely imperfectly, but still meaningfully) to whether you'd want to read or skim them.