A pointed exchange between Princeton professor Mengdi Wang and researcher Pasquale Minervini lays bare the split over whether frontier AI work can still happen without industrial data centers; Minervini points to his ICLR 2021 award-winning complex query answering paper whose full set of experiments finished overnight on one GPU and whose follow-up was accepted at ICML 2025, while Wang counters that meaningful progress now demands compute universities rarely possess.
@YiMaTweets I wish I can say same thing, but tbh it just become impossible to do research without compute
I’m getting increasingly annoyed by young people complaining that they cannot do AI-related research unless they join big industrial labs… well, here is my reply: academia is supposed to work on ideas that money cannot buy!
Minervini’s recent arXiv work shows many CQA test queries collapse into simple link prediction, so models that look state-of-the-art may not be solving the multi-hop reasoning the field claims to measure.
Wang’s replies underscore that academic groups without industry partnerships face shrinking options, leaving open whether single-GPU results like Minervini’s can scale to the next generation of harder benchmarks.
you could run all experiments in our "outstanding paper award" ICLR 2021 paper overnight on a single GPU, and the method is still SOTA in that field (see our ICML'25 follow-up) -- you don't have to be in a frontier lab to do cool stuff and have an impact
I’m getting increasingly annoyed by young people complaining that they cannot do AI-related research unless they join big industrial labs… well, here is my reply: academia is supposed to work on ideas that money cannot buy!
Paper: https://iclr-conf.medium.com/announcing-iclr-2021-outstanding-paper-awards-9ae0514734ab#3bf5 Follow-up: https://arxiv.org/abs/2410.12537
you could run all experiments in our "outstanding paper award" ICLR 2021 paper overnight on a single GPU, and the method is still SOTA in that field (see our ICML'25 follow-up) -- you don't have to be in a frontier lab to do cool stuff and have an impact
A pointed exchange between Princeton professor Mengdi Wang and researcher Pasquale Minervini lays bare the split over whether frontier AI work can still happen without industrial data centers; Minervini points to his ICLR 2021 award-winning complex query answering paper whose full set of experiments finished overnight on one GPU and whose follow-up was accepted at ICML 2025, while Wang counters that meaningful progress now demands compute universities rarely possess.
@YiMaTweets I wish I can say same thing, but tbh it just become impossible to do research without compute
I’m getting increasingly annoyed by young people complaining that they cannot do AI-related research unless they join big industrial labs… well, here is my reply: academia is supposed to work on ideas that money cannot buy!