Monthly Roundup #35: October 2025 (thezvi.substack.com)

October’s roundup stresses a theme that matters to AI practitioners: markets and social choices shape the data we train on. The piece warns against letting “idiots with buying power” create negative externalities — broken reviews, low-quality products and noisy signals that become training data for recommender systems and language models. That’s a reminder that label noise, dataset provenance, and incentive-aligned feedback matter as much as model architecture; poor consumer signaling is effectively data poisoning at scale. The writeup also touches on industry cover-ups (e.g., airplane fume claims) as an analogy to historical denialism, which should prompt ML teams to prioritize transparency, root-cause analysis, and independent auditing when safety or public health is implicated. There are also cultural and governance takeaways for the AI community: don’t let constant doom-saying crowd out productive work or honest celebration, and be deliberate about social norms that impose signaling costs. On prediction and governance, Robin Hanson’s futarchy note — create many markets over random decision branches to avoid non-causal correlations — is technically sound but pragmatically faltering due to price-discovery and liquidity constraints. In short: focus on improving upstream data quality and incentives, build robust mechanisms for causal identification and market liquidity if you pursue prediction markets, and balance healthy skepticism with morale so the field can both safeguard and accelerate useful progress.