Robert Metcalfe

I think this is a bit hysterical. Academic publishing has at least two pragmatic functions. Much of the difficulty derives from trying to perform both of those functions with the same venue/process. Decouple them, and things can work just fine. Academic publishing does two things: it allocates *credentials*, and it attempts to allocate *attention*. The credentialing function of academic publishing is actually really important. Universities are our largest and most enduring organisations for fundamental research in the public interest. For that research to make genuine progress towards knowledge, we have to apply standards of rigour. To know who to hire into tenure-track positions, we need to be able to assess the quality of their work, and see how it is judged by their peers. And to decide who gets the ultimate liberty of being able to follow their intellectual interests wherever they lead (ie tenure), we need the same thing. Peer review has all sorts of flaws, it's broken in many different ways, but we need to have *some* mechanisms for this kind of quality control in order for the whole endeavour of the collective pursuit of wisdom in the public interest to advance. But the problem is, peer review takes *time*. And by the time it's complete, it's generally too late to contribute in a meaningful and effective way to the public conversation on a topic as fast-moving as AI progress. In addition, while the people running journals (etc) are often good judges of methodological rigour and other epistemic criteria, they are very *fallible* judges of what work is actually the most important, or otherwise in the broader interest of society (this is especially true for humanities fields like mine). But this just means that while academic publishing is the right vehicle for quality control and credentialing, it's the wrong one for the allocation of public attention. The solution seems pretty clear—especially since this is how things are basically already done in CS. Decouple the two things. Preprints, substacks, the other changing forms of public communication for the allocation of attention; journals and peer-reviewed conferences for the quality control/credentialing function. Obviously this exposes us to epistemic risk: if you don't wait for peer-review, then some of the work that attracts a tonne of public attention will prove to be bogus. But that's a self-correcting problem, since if it attracts a lot of attention then the errors are likely to be found out. And obviously peer review faces MANY challenges right now. But that's something that *can* be salvaged (and an area where AI is likely to help).