Inside Cognition's inference stack: RL, speculative decoding & DFlash

Reinforcement learning (RL) and inference optimization are increasingly intertwined, as RL rollouts rely heavily on efficient inference throughput and low latency. The development of DFlash, a diffusion-based speculative decoding method, marks a significant advancement by achieving higher token acceptance lengths and constant-time drafting, which shifts inference from memory-bound to compute-bound regimes. Beyond technical optimizations, AI agents like Devin are fundamentally altering research workflows by automating execution, monitoring training runs, and iterating on serving configurations. This automation enables smaller research teams to achieve the productivity of much larger groups. Furthermore, the industry is evolving toward a flexible Pareto frontier, allowing users to balance latency and cost by choosing between premium, real-time interactions and high-throughput background processing for agentic tasks.