In offloaded MoE serving, every expert miss is a PCIe copy. Routing tricks trade quality for speed; the cheap win is the eviction rule. A clairvoyant evictor cuts transfers 2.1x with identical outputs. NeurIPS 2026 AXIOM workshop, by @pierrelux

carbonforge.ai/blog/paging-ex…

An agent choosing its own experiments can confirm false findings by peeking, testing many claims, or retrying. When nothing is true, peeking alone gives 43% false supports; EvidenceTree keeps it under 2%. NeurIPS 2026 e-values workshop, by @pierrelux

carbonforge.ai/blog/testing-b…

LLM decode is often memory-bound, so extra SM clock becomes power, not tokens. A serving-aware clock governor on 2xA100 cut GPU power 12 to 13% at matched capacity, and about 30% on saturated decode. Model and engine untouched. By @pierrelux

carbonforge.ai/blog/adaptive-…