扩散模型将在AI推理中获胜：Inception CEO Stefano Ermon详解Why Diffusion Will Win AI Inference: Inception CEO Stefano Ermon
核心要点：扩散模型因为天生并行、更契合GPU推理工作负载，将在AI推理上击败自回归模型，带来更快、更高效的LLM。
斯坦福教授、扩散模型之父、Inception联合创始人兼CEO Stefano Ermon，从2014年开始研究生成模型。当时领域冷门，模型只能勉强生成模糊的MNIST数字。他始终相信生成模型是从无标签数据中学习结构的正确方式，最初更关注世界模型，而非今天LLM的能力。
他实验室在2019年与学生Yang Song提出基于去噪的score-based模型，最终演变为现代扩散模型。如今图像、视频、音乐甚至蛋白质生成的最佳模型多基于扩散。2024年，他们首次证明扩散语言模型在GPT-2规模上能匹配自回归模型的困惑度，同时生成速度快约10倍。
Inception约两年、50人左右，正在把这项技术规模化成商用扩散LLM（Mercury系列）。这些模型在基准上已与OpenAI等前沿实验室的flash/mini模型质量相当，但显著更快，并已在生产中服务真实客户。他们自建了服务引擎，因为现有LLM引擎无法直接运行扩散LLM。
Ermon强调推理时扩展才是关键：自回归模型生成时仍是串行的，严重内存受限；扩散模型在推理时能并行处理大量token，工作负载更像训练，完美映射GPU优势。他引用“苦涩的教训”：更并行的方案最终会赢。扩散模型还更易控制（可从粗到细引导），可能更数据高效。客户如语音代理OpenCall已从专用芯片转向他们的模型，在NVIDIA GPU上达到同等速度、更低成本。
“我们赌的是扩散基LLM，因为它天生更并行。苦涩的教训是，更并行的方案最终会赢。”
斯坦福教授、扩散模型之父、Inception联合创始人兼CEO Stefano Ermon，从2014年开始研究生成模型。当时领域冷门，模型只能勉强生成模糊的MNIST数字。他始终相信生成模型是从无标签数据中学习结构的正确方式，最初更关注世界模型，而非今天LLM的能力。
他实验室在2019年与学生Yang Song提出基于去噪的score-based模型，最终演变为现代扩散模型。如今图像、视频、音乐甚至蛋白质生成的最佳模型多基于扩散。2024年，他们首次证明扩散语言模型在GPT-2规模上能匹配自回归模型的困惑度，同时生成速度快约10倍。
Inception约两年、50人左右，正在把这项技术规模化成商用扩散LLM（Mercury系列）。这些模型在基准上已与OpenAI等前沿实验室的flash/mini模型质量相当，但显著更快，并已在生产中服务真实客户。他们自建了服务引擎，因为现有LLM引擎无法直接运行扩散LLM。
Ermon强调推理时扩展才是关键：自回归模型生成时仍是串行的，严重内存受限；扩散模型在推理时能并行处理大量token，工作负载更像训练，完美映射GPU优势。他引用“苦涩的教训”：更并行的方案最终会赢。扩散模型还更易控制（可从粗到细引导），可能更数据高效。客户如语音代理OpenCall已从专用芯片转向他们的模型，在NVIDIA GPU上达到同等速度、更低成本。
“我们赌的是扩散基LLM，因为它天生更并行。苦涩的教训是，更并行的方案最终会赢。”
The Takeaway: Diffusion models will win AI inference because they are inherently parallel and map far better to GPU workloads than sequential autoregressive models, delivering higher intelligence per watt and per dollar.
Stefano Ermon, longtime Stanford professor and one of the fathers of diffusion models, is now co-founder and CEO of Inception. He has worked on generative models since 2014, when the field was unfashionable and success meant blurry MNIST digits. He always viewed generative modeling as the right way to capture structure in unlabeled data, originally thinking in terms of world models for decision-making rather than today’s LLM capabilities.
In 2019 his lab (with PhD student Yang Song) introduced score-based generative models that train a network to denoise images—the foundation of modern diffusion. Diffusion now powers the best image, video, music and many protein models. In 2024 they showed for the first time that a diffusion language model could match autoregressive perplexity at GPT-2 scale while generating text roughly 10× faster because it outputs many tokens in parallel.
Inception (≈2 years old, ~50 people) is scaling that technology into commercial diffusion LLMs. Their Mercury models already match the quality of frontier labs’ speed-optimized flash/mini models on benchmarks while being significantly faster, and they are serving real production traffic. They had to build their own serving engine because existing systems cannot run diffusion LLMs.
Ermon’s core bet is inference-time scaling. Autoregressive generation remains sequential and memory-bound; diffusion’s parallel workload at inference closely resembles training and therefore exploits GPUs efficiently. He invokes the bitter lesson: the more parallel solution eventually wins. Diffusion models are also easier to steer (coarse-to-fine control) and may prove more data-efficient. Customers such as voice-agent company OpenCall have switched from custom silicon to Mercury, achieving comparable latency on ordinary NVIDIA GPUs at lower cost and higher availability.
“We bet on a diffusion-based LLM because it’s inherently more parallel. And the bitter lesson is that the more parallel solution is the one that is eventually going to win.”
查看原文 →
Stefano Ermon, longtime Stanford professor and one of the fathers of diffusion models, is now co-founder and CEO of Inception. He has worked on generative models since 2014, when the field was unfashionable and success meant blurry MNIST digits. He always viewed generative modeling as the right way to capture structure in unlabeled data, originally thinking in terms of world models for decision-making rather than today’s LLM capabilities.
In 2019 his lab (with PhD student Yang Song) introduced score-based generative models that train a network to denoise images—the foundation of modern diffusion. Diffusion now powers the best image, video, music and many protein models. In 2024 they showed for the first time that a diffusion language model could match autoregressive perplexity at GPT-2 scale while generating text roughly 10× faster because it outputs many tokens in parallel.
Inception (≈2 years old, ~50 people) is scaling that technology into commercial diffusion LLMs. Their Mercury models already match the quality of frontier labs’ speed-optimized flash/mini models on benchmarks while being significantly faster, and they are serving real production traffic. They had to build their own serving engine because existing systems cannot run diffusion LLMs.
Ermon’s core bet is inference-time scaling. Autoregressive generation remains sequential and memory-bound; diffusion’s parallel workload at inference closely resembles training and therefore exploits GPUs efficiently. He invokes the bitter lesson: the more parallel solution eventually wins. Diffusion models are also easier to steer (coarse-to-fine control) and may prove more data-efficient. Customers such as voice-agent company OpenCall have switched from custom silicon to Mercury, achieving comparable latency on ordinary NVIDIA GPUs at lower cost and higher availability.
“We bet on a diffusion-based LLM because it’s inherently more parallel. And the bitter lesson is that the more parallel solution is the one that is eventually going to win.”