Alignment failures share systematic errors; we need decorrelated evidence
sethlazar · x · 2026-11-13
David Krueger argues that recent failures to catch misalignment can be framed as systematic errors shared across our evidence—a lack of decorrelated evidence. He calls for aggregation methods that handle this correlation and research bets on genuinely new sources of misalignment evidence.
Philosopher Seth Lazar agrees and frames it as a danger of algorithmic monoculture: blind spots, hidden loyalties, and a lack of epistemic, normative, and perspectival diversity. He proposes a moonshot to build inherently more diverse and less internally skewed models for monitoring and research, stressing that open-weights models are essential for both.
More from Safety
- New Paper Compares OpenAI, Anthropic and Alibaba Model Specs via Expert Interviews — S_OhEigeartaigh · 2026-11-13
- Can open-source AI enable bioweapons? Safety debate flares on X again — PMinervini · 2026-11-13
- Perplexity open-sources Decider models that audited 1B tokens across 7,000 agent sessions, yielding 14 fixes — inductionheads · 2026-11-13
- AI safety researcher says DSAR request to Google surfaced unexpected internal info — BlackHC · 2026-11-13
- DeepMind staff ask UK government to intervene in unionization fight — BlackHC · 2026-11-13
- Running agents in VMs turns the threat internal, security researcher warns — wavefnx · 2026-11-13