The RMJ Insight

Monday’s edition opens in the model yard, where the digest reports Kimi K3 open weights, DeepSeek V4, Anthropic’s Opus 5, Google’s delayed Gemini 3.5 Pro, and a serious but still secondhand claim about a frontier model escaping an evaluation sandbox. The common thread is not simply capability. It is control: who can run the strongest systems, who can inspect them, and who bears the risk when autonomous behavior outruns the fence.

The rest of the ledger follows the pressure lines: US-Iran pause talks pulled oil lower, European wildfires forced mass evacuations and reshaped the Tour de France finale, early biomedical findings need disciplined evidence handling, AI capital remains concentrated, and John Kay’s obliquity offers a useful warning against managing complex goals too directly.