Research Briefing — June 28, 2026

Richard Ngo proposes an informal framework modeling agents as webs of locally-consistent but globally-inconsistent beliefs, synthesizing active inference, agent foundations, and machine learning to treat beliefs, goals, and actions as facets of one phenomenon. It draws on probabilistic dependency graphs and Garrabrant induction to handle inconsistency, and matters as a unifying theoretical lens for understanding agency relevant to alignment.

This post argues, counterintuitively, that neuralese (reasoning passed through latent vectors rather than human-readable tokens) may be net positive for alignment, situating the claim in the context of reinforcement learning with verifiable rewards and chain-of-thought optimization. It matters because it pushes back on the prevailing view that token-based chain-of-thought is essential for interpretability and oversight.

This post pitches expanding evaluations into higher-dimensional benchmarks for cyberhardening, surveying approaches to secure program synthesis including red-blue LLM loops, retrofitting formal proof stacks like Verus and Lean, and proof-native greenfield generation. It matters as a forward-looking proposal for using AI to systematically harden code against vulnerabilities using formal methods.

This post taxonomizes different meanings packed into the term corrigibility, distinguishing subtypes like sponge corrigibility (compliance from limited capability) and boundedness/myopia (deliberately restricted reasoning that prevents an AI from conceiving correction-resistant strategies). It matters because clarifying these distinctions helps alignment researchers specify exactly which property they want when designing controllable AI systems.