Claude tops new benchmark for agents that build agents, still fails most tests

Reddit users discuss whether they trust Claude or other AI agents to triage real email inboxes, citing concerns about accuracy over time and the need to double-check filings. Few report letting agents fully manage email.

Terence Tao, a Fields Medalist, argues AI is solving fruitful open math problems faster than mathematicians can identify new ones. He suggests labeling problems "analysis-required" so bare AI answers without reasoning count for little.

NVIDIA's blog details EPD disaggregation for multimodal inference, claiming up to 5x faster time-to-first-token and 7x faster end-to-end response. It separates vision encoding from prefill/decode, best for image-heavy prompts and quantized MoE models.

CUDA Toolkit 13.4 adds Windows on Arm support and early developer preview of the NVIDIA Rubin GPU architecture (compute capability 107). It also introduces MPS V3 with scriptable CLI, namespaces, and cgroup-integrated memory limits for shared GPU management.