Jazib Zaman

China's weather agency has made Fenghe available to outside developers. The immediate proof is unusually concrete: model weights, configuration files and a tokenizer are public. The harder proof is still missing. Nothing in the release establishes that Fenghe itself produces a better numerical weather forecast, or that its advice is ready for high-consequence operational use.

Article Brief Key Takeaways 4 Points24s Read 01Control surface-CrewAI 1.15.3 adds interception points around agent execution boundaries, including model and tool operations. 02Safer default-Tool-result caching is now opt-in, making data reuse a deliberate application decision. 03Operational proof-Kickoff results now expose per-call usage while several fixes align hooks, native tools, outputs, and conversation state.

Article Brief Key Takeaways 4 Points24s Read 01Control surface-custom_remat lets AI framework code express forward, rematerialization and backward behavior at a function boundary. 02Opt-in path-The new behavior depends on experimental remat3, whose configuration flag remains off by default. 03Upgrade risk-Python, NumPy and SciPy minimums rise, while empty arrays now contain uninitialized data rather than zeros.