Source: Product Hunt — The best new products, every day
A new architecture class is emerging that breaks the compute-efficiency wall most language models hit at scale. Mercury 3 demonstrates that diffusion-based decoding can maintain training speed through massive parameter growth, challenging the scaling laws that have defined model development for the past three years. If efficiency scales with size instead of degrading, the economics of LLM development flatten dramatically, favoring labs that can iterate faster over those with the most capital.