Arxiv: https://arxiv.org/abs/2506.17298This episode of The AI Research Deep Dive unpacks "Mercury," a groundbreaking paper from Inception Labs that could fundamentally change how language models are built. The host explains how the Mercury model abandons the standard, one-word-at-a-time (autoregressive) approach used by models like GPT and instead adopts a diffusion-based method, inspired by image generation, to create entire blocks of text in parallel. This architectural shift results in a staggering speedup—over 1,100 tokens per second, roughly 18 times faster than leading speed-optimized models—without sacrificing quality. The episode highlights how Mercury's performance is validated by independent benchmarks and real-world human evaluations, where it proves to be both the fastest and one of the most preferred models for coding, signaling a potential new era for AI where ultra-low latency can unlock a new generation of truly real-time applications.

Ver todos los episodios