Are we optimizing AI research for acceptance rather than lasting value? [D]

A viral critique from researcher NuoJohnChen is sparking intense debate within the AI community, challenging the fundamental incentives of modern machine learning research. The central argument posits that the culture surrounding major conferences like NeurIPS, ICML, and ICLR has shifted toward optimizing for paper acceptance rather than producing work of lasting scientific or practical value. The author describes a system where researchers are compelled to run "tons of evaluations"—often far beyond what's necessary for understanding—primarily to satisfy reviewer demands for perceived rigor. This creates a perverse incentive where the metric for success becomes conference acceptance itself, not the downstream impact of the research.

The critique highlights a specific broken feedback loop: reviewers request exhaustive supplementary experiments to feel confident in their acceptance decision, but these results are almost never verified or scrutinized after publication. This leads to what the author calls an "unsustainable" level of effort focused on passing the review barrier, potentially at the expense of exploring riskier, more innovative ideas. The post nostalgically references a lost "spark" of creativity, suggesting the current hyper-competitive, metrics-driven environment may be filtering out precisely the kind of unconventional thinking that has historically led to major AI breakthroughs. The discussion has resonated widely, tapping into growing concerns about whether the field's rapid expansion and industrial capture have altered its epistemic foundations.