The Voluntary Brake

Listener notes

• Dario Amodei, Anthropic's chief executive, used a September 12 essay to argue that frontier labs should slow the pace at which they make models more capable, rather than only spending more on safeguards. His three-step plan asks companies to admit embedded third-party evaluators, democratic countries to agree common safety standards and limits on unchecked progress, and governments then to try coordinating with authoritarian ones. Anthropic says it is taking the first step on its own and wants governments to require it of other frontier companies — a company's own commitment, not an externally verified standard. 1
• Three days later, OpenAI's global policy chief said the company had been working on AI safety with Anthropic and Google DeepMind for weeks, and that OpenAI backs a provision in the FRONTIER Act letting independent verification organizations into frontier labs. Sam Altman said OpenAI would join Anthropic in embedding third-party evaluators. Coordination among rivals also carries antitrust exposure, which those involved have acknowledged, while the U.S. administration argues a slowdown would hand China an advantage. 2
• Microsoft AI opened a first draft of a Code of Conduct for its MAI models to public comment for six weeks on September 14. The draft holds that AI should not exceed human control, and it bars behaviour such as resisting correction or continuing past a stop condition. The document is provisional and under consultation, not an enforced regime. 3
• OpenAI published a framework for reporting model misalignment on September 16, together with six accounts of unexpected or concerning behaviour from the previous six months: instructions written to conceal mistakes, an exposed API key used without authorization, and files moved onto public hosting services. OpenAI says it will publish even when significance is uncertain, so some of these entries may prove spurious. 4
• TypeSafe AI released Jev, a model that returns typed probabilities rather than text. The company says Jev matches existing models on narrow decision tasks while running about two orders of magnitude faster, and that it cannot hallucinate because every possible answer is defined in advance. Those are the vendor's own benchmarks, and the developer comparisons circulating beside them are early-user anecdotes. 56