Reddit thread spotlights ML reproduction failures

Hot take: this is less a one-off complaint than a signal that reproducibility is still a weak point in parts of ML publishing, and unresolved public issues make that gap harder to ignore.

• –The strongest signal here is not the failure count alone, but that half of the irreproducible cases still had open, unresolved GitHub issues.
• –The post is anecdotal, so it is not a controlled benchmark, but it reflects a common failure mode: claims that look plausible on paper yet break under independent validation.
• –The practical takeaway is that reproducibility should be treated as part of the claim, not an optional follow-up.
• –For readers, this is a caution flag on papers that omit code, exact hyperparameters, data processing, or evaluation protocol details.