AI News — April 6, 2026

April 6 brought a practical performance win for large-model deployment and a policy-focused investment in safety research. Cursor published warp decode, a systems-level trick that reshapes parallelism to accelerate MoE inference by roughly 1.8× — a concrete improvement for teams running conditional-expert architectures where routing and latency matter. Meanwhile OpenAI announced a Safety Fellowship to fund and mentor independent alignment and safety researchers, expanding the pipeline for rigorous, tooling-oriented work that practitioners and evaluators will rely on.

Today's stories:

• Announcing the OpenAI Safety Fellowship (OpenAI News) — a funded fellowship and mentorship program aimed at growing independent safety and alignment research, which could accelerate development of evaluations and tooling that practitioners depend on.
• Better MoE model inference with warp decode (Cursor Blog) — introduces warp decode, a re-ordered parallelism approach that reduces routing bottlenecks and improves MoE inference throughput/latency by ~1.8×; relevant for teams deploying MoE in production.

Overall: a day of practical engineering + capacity-building — one story targets deployment efficiency, the other targets the research pipeline that underpins long-term safety and evaluation.