NeurIPS authors weigh code-release risks

This is not a launch, but it captures a real pressure point in AI research: open science norms are colliding with reviewer trust, compute asymmetry, and agent-assisted replication.

• –Several commenters favor releasing a preprint and public repo timestamp to establish priority before sharing code.
• –Others argue reviewers rarely inspect code deeply, making pre-acceptance code submission high-risk and low-reward.
• –The compute profile matters: small, reproducible methods benefit most from code review, while large-scale training code may be less useful before acceptance.
• –The anxiety itself is telling; AI tooling is making plagiarism fears feel operational rather than theoretical.