Constitutional AI

Claude Opus 4.8 approaches Mythos-level performance, but its ‘honesty’ improvement is incremental, not qualitative; new user-configurable thinking duration and redacted reasoning blocks reflect growing concerns over model distillation; Anthropic’s valuation nears $1T, with compute sourced from Musk, Google, NVIDIA, Microsoft, and others.

Anthropic 提出使用 MSM（Model Specification Mapping）实证研究不同模型规格或宪法设计对对齐训练泛化能力的影响，强调解释底层价值观比单纯设定规则更有效。

Anthropic launches dialogues with religious and philosophical traditions to explore AI moral formation, experimentally validating an 'ethical reminder tool' that reduces misaligned behavior, though the article prioritizes PR narrative over technical detail.

Anthropic reports that combining constitutional documents with aligned AI fiction reduces agentic misalignment by over three times, showing robustness across unrelated scenarios.