Brute Force Of All Papers Exposes Flaws In AI Submissions

@max_spero_

@max_spero_ See

https://en.wikipedia.org/wiki/The_Library_of_Babel

Nothing more soul-crushing than having to review a nonsensical paper clearly written by AI / in the state of AI psychosis

@max_spero_ But it's just as easy to desk reject those papers as it is to generate them. It's moot. Just noise.

@max_spero_ Much as I agree with the sentiment, the casual use of the word “only” to refer to 2^700,000, i.e. a number with roughly 210,000 digits, is breathtaking :)

(Of course, an *extremely* large fraction of those won’t be valid LaTeX, but still…)

Your argument assumes that the peer-review system is the sole and necessary filter is a assumption needs to be proven. In a free market of ideas, mediocre papers die silently no one cites them, no one builds upon them, and they have no real-world impact. A truly brilliant paper will make its mark regardless of the spam noise surrounding it.

@max_spero_ @pangram, is this AI?

@max_spero_ @pangram am i gay

@fahrudotme @max_spero_ We believe that this document is fully human-written

https://www.pangram.com/history/c466688c-3ffe-4010-b207-1701e2eb6fb2

@Japple @max_spero_ "just as easy" is a little optimistic, unless you mean by blocking your email address and banning you from the conference

@max_spero_ I want to see AI-generated patent disclosures. Don't actually file them, just put them up for public download. Billions and billions of them, to be treated as prior art land mines to invalidate future patents.

@max_spero_ An AI reviewer would be a better and less damaging approach than pretending you can detect which ones have been written by AI.

@max_spero_ Also The peer-review system was broken decades before artificial intelligence. More than 5 million papers are published annually, and millions of them go uncited.

@max_spero_ buried beat is brute-force paper gen hits peer-review attention not generation cost. reviewer cycles dont scale linearly. AI floods accelerate peer-review demise as credentialing. credentialing migrates to compute-attested provenance.

@max_spero_ inaugurating the era of ruthless individual eval

patton-style war on complexity

reincarnation all over again

@max_spero_ You couldn't brute force, so no information beyond.

@AaronBergman18 @max_spero_ yeah but you’re still saturating the space. if ai review becomes that good then at that point ai is just helping explore concepts. then you can just write normal papers yourself on those discoveries lol

@pangram Thanks twin

@AaronBergman18 @max_spero_ @chaotictransfem I don't think I would trust the AI reviewers' biases. But I wouldn't trust human reviewers' biases, so God is dead.

@max_spero_ This merely demonstrates that paper generation is not the bottleneck. Human peer review is insufficient to filter the huge quantity of AI generated papers, sure, but a sufficiently robust AI filter might be able to do so.