Inception Labs is exploring a groundbreaking shift in the landscape of large language models (LLMs) by moving away from the conventional autoregressive approach. Most current LLMs generate text sequentially, one token at a time from left to right. This method, while effective, is inherently slow because each token must wait for the preceding tokens to be generated before it can be produced, and each generation step involves evaluating a neural network with billions of parameters. This sequential process poses significant challenges, particularly in terms of inference costs and latency. Frontier LLM companies are currently focused on enhancing test-time computation to boost reasoning and error-correction capabilities. However, this approach often leads to prohibitively high computational expenses and delays that make the models impractical for real-world applications. To address these issues, a new paradigm is needed to make high-quality AI solutions more accessible. Enter diffusion models. These models offer a revolutionary "coarse-to-fine" generation process. Instead of generating text step-by-step, diffusion models start with pure noise and gradually refine this noise through a series of denoising steps. This method allows the model to consider the entire context simultaneously, rather than being limited to the sequence of previously generated tokens. As a result, diffusion models are better equipped for complex reasoning and structuring their outputs, and they can effectively correct mistakes and hallucinations during the refinement process. Diffusion models have already proven their worth in generating high-quality images, videos, and audio. Notable examples include Sora, Midjourney, and Riffusion, which have become leading AI solutions in their respective domains. Despite this success, applying diffusion to discrete data like text and code has remained a challenge—until now. Inception Labs is pioneering the application of diffusion models to discrete data, aiming to create more efficient and capable LLMs. By leveraging the strengths of diffusion models, the company hopes to overcome the limitations of autoregressive models and deliver AI solutions that are both powerful and practical. This breakthrough could significantly reduce inference costs and latency, making high-quality AI more accessible and usable for a wide range of applications.