Mechanistic interpretability took center stage this week as Anthropic published a landmark study scaling sparse autoencoders to production-grade models, extracting millions of interpretable features from Claude 3 Sonnet. Meanwhile, Google DeepMind unveiled Gemini 2.5 with native multimodal reasoning, and Meta released Llama 4 — their most capable open-weight model yet, matching frontier closed models on several benchmarks.

On the safety front, research into RLHF reward hacking gained traction with a comprehensive taxonomy of failure modes, and NVIDIA's dynamic token routing work pushed inference efficiency to new highs with a reported 3.5x speedup. The week paints a picture of a field simultaneously racing toward capability and grappling with alignment — the defining tension of 2026.