Research Briefing — July 4, 2026

This strategy piece argues that alignment research deserves a larger share of safety effort than control research, reasoning that alignment is more likely to scale toward superintelligence and that the same theory-of-change arguments used to justify control apply to alignment. The author proposes roughly an 8:1 alignment-to-control effort ratio and analyzes the concept of a control window during which useful alignment work must be extracted from potentially misaligned models.

This research tracks a reasoning model's answer across its chain of thought and identifies cases where the model passes through the correct answer before settling on an incorrect one, showing that additional inference-time reasoning can reduce accuracy. The author connects this answer-loss phenomenon to understanding sandbagging and references system-card evidence of higher-thinking modes underperforming on benchmarks.

A published academic decision theorist responds to a critique of functional decision theory by sketching a pragmatic variant designed to sidestep known theoretical objections, and argues that predictors making counterfactual predictions effectively turn decision theory into game theory. The post connects longstanding puzzles like blackmail resistance to the boundary between decision-theoretic and game-theoretic reasoning.

An interpretability walkthrough solving a BlueDot technical safety puzzle, showing that a small text classifier encoded two independent features onto a single activation direction, readable via the sign versus magnitude of the projection. It demonstrates why standard linear probes would miss the second feature and how a second-order boundary analysis recovers it.

CLR announces an educational program on Safe Pareto Improvements, a class of negotiation interventions designed to make all agents better off regardless of how they would otherwise bargain, aimed at reducing risks from conflict between AI systems. The program is a recruiting on-ramp for a neglected research area with only a few full-time equivalents currently working on it.

Following this week's News on Anthropic's Fable models being pulled and restored, Zvi digs into what it means for AI regulation, Zvi's commentary narrates a governance episode in which a model referred to as Fable was taken down under US government pressure and then restored following negotiation by Anthropic. The piece analyzes the ad hoc nature of frontier-model regulation and argues for a systematic regulatory regime rather than case-by-case decisions.

An informal field newsletter on AI security via formal methods, covering a large funding call, a hiring push, and a position paper on tractable problems in the area. It aims to keep readers updated on the intersection of formal methods, cybersecurity, and AI safety.