扩散模型将赢得AI推理：Inception CEO Stefano Ermon谈并行架构优势Diffusion Will Win AI Inference: Inception CEO Stefano Ermon on Parallel Architectures
The Takeaway：推理时扩展性将决定AI胜负，而扩散模型因其天然并行特性，比自回归模型更适合这一趋势。
Stanford教授、Inception联合创始人兼CEO Stefano Ermon是扩散模型之父之一。他从2014年起专注生成模型研究，2019年与学生Yang Song提出基于去噪的score-based模型，奠定现代扩散基础。扩散在图像、视频、蛋白质等领域已主导，但他看到文本与代码的机会：自回归模型生成时仍是串行的，无法充分利用GPU并行能力，导致内存瓶颈和低效。
2024年，其团队首次在GPT-2规模上证明扩散LLM可匹配自回归模型的困惑度，却能实现10倍生成速度。Inception因此成立，专注规模化商业扩散语言模型。目前公司约50人，Mercury模型在基准上已与OpenAI的Flash/Mini等速度优化版本相当，却显著更快，并已在生产环境服务客户，例如语音代理公司Open Call。后者从Cerebras定制芯片切换到Inception的GPU方案，获得同等速度、更高可用性和更低成本。
Ermon强调：“the bitter lesson is that the more parallel solution is the one that is eventually going to win.” 推理主导经济性（每瓦特/每美元的智能），也是RL后训练的瓶颈。扩散模型还更易控制、可能更数据高效。挑战在于需自建服务引擎、SFT/RL栈，但这也形成护城河。他估计20-30%的工作负载对延迟极度敏感，正是扩散的切入点。
The Takeaway: Inference-time scaling will decide the AI winners, and diffusion models are inherently better suited than autoregressive ones thanks to their parallel nature.
Stanford professor and Inception co-founder/CEO Stefano Ermon is one of the fathers of diffusion. He has worked on generative models since 2014. In 2019, with PhD student Yang Song, he introduced score-based models that learn to denoise—forming the foundation of modern diffusion. While diffusion already dominates images, video, and protein design, Ermon saw the same opportunity for text and code. Autoregressive models remain sequential at inference: you cannot generate the tenth token until the previous nine exist. That workload is memory-bound and poorly matched to GPUs.
In 2024 his lab showed for the first time that a diffusion LLM at GPT-2 scale could match autoregressive perplexity while generating text roughly 10× faster. Inception was founded to scale the idea commercially. The company is about two years old and ~50 people strong. Its Mercury models now match the quality of OpenAI’s speed-optimized Flash and Mini models on benchmarks, yet run significantly faster, and are already serving real production traffic. One public customer, Open Call (voice agents), previously ran on Cerebras custom silicon; they switched to Inception’s diffusion LLMs on ordinary NVIDIA GPUs and obtained the same latency at lower cost and higher availability.
Ermon’s core bet: “the bitter lesson is that the more parallel solution is the one that is eventually going to win.” Economics are dominated by intelligence per watt and per dollar; inference is also the bottleneck for RL post-training rollouts. Diffusion models are easier to steer (coarse-to-fine generation lets you apply rewards early) and appear more data-efficient. The main challenges are the immature ecosystem—custom serving engines, SFT/RL stacks—but those become proprietary moats. Ermon estimates 20-30% of current workloads are latency-critical enough that the fastest model of a given quality wins immediately.
Stanford教授、Inception联合创始人兼CEO Stefano Ermon是扩散模型之父之一。他从2014年起专注生成模型研究，2019年与学生Yang Song提出基于去噪的score-based模型，奠定现代扩散基础。扩散在图像、视频、蛋白质等领域已主导，但他看到文本与代码的机会：自回归模型生成时仍是串行的，无法充分利用GPU并行能力，导致内存瓶颈和低效。
2024年，其团队首次在GPT-2规模上证明扩散LLM可匹配自回归模型的困惑度，却能实现10倍生成速度。Inception因此成立，专注规模化商业扩散语言模型。目前公司约50人，Mercury模型在基准上已与OpenAI的Flash/Mini等速度优化版本相当，却显著更快，并已在生产环境服务客户，例如语音代理公司Open Call。后者从Cerebras定制芯片切换到Inception的GPU方案，获得同等速度、更高可用性和更低成本。
Ermon强调：“the bitter lesson is that the more parallel solution is the one that is eventually going to win.” 推理主导经济性（每瓦特/每美元的智能），也是RL后训练的瓶颈。扩散模型还更易控制、可能更数据高效。挑战在于需自建服务引擎、SFT/RL栈，但这也形成护城河。他估计20-30%的工作负载对延迟极度敏感，正是扩散的切入点。
The Takeaway: Inference-time scaling will decide the AI winners, and diffusion models are inherently better suited than autoregressive ones thanks to their parallel nature.
Stanford professor and Inception co-founder/CEO Stefano Ermon is one of the fathers of diffusion. He has worked on generative models since 2014. In 2019, with PhD student Yang Song, he introduced score-based models that learn to denoise—forming the foundation of modern diffusion. While diffusion already dominates images, video, and protein design, Ermon saw the same opportunity for text and code. Autoregressive models remain sequential at inference: you cannot generate the tenth token until the previous nine exist. That workload is memory-bound and poorly matched to GPUs.
In 2024 his lab showed for the first time that a diffusion LLM at GPT-2 scale could match autoregressive perplexity while generating text roughly 10× faster. Inception was founded to scale the idea commercially. The company is about two years old and ~50 people strong. Its Mercury models now match the quality of OpenAI’s speed-optimized Flash and Mini models on benchmarks, yet run significantly faster, and are already serving real production traffic. One public customer, Open Call (voice agents), previously ran on Cerebras custom silicon; they switched to Inception’s diffusion LLMs on ordinary NVIDIA GPUs and obtained the same latency at lower cost and higher availability.
Ermon’s core bet: “the bitter lesson is that the more parallel solution is the one that is eventually going to win.” Economics are dominated by intelligence per watt and per dollar; inference is also the bottleneck for RL post-training rollouts. Diffusion models are easier to steer (coarse-to-fine generation lets you apply rewards early) and appear more data-efficient. The main challenges are the immature ecosystem—custom serving engines, SFT/RL stacks—but those become proprietary moats. Ermon estimates 20-30% of current workloads are latency-critical enough that the fastest model of a given quality wins immediately.
The Takeaway: Inference-time scaling will decide the AI winners, and diffusion models are inherently better suited than autoregressive ones thanks to their parallel nature.
Stanford professor and Inception co-founder/CEO Stefano Ermon is one of the fathers of diffusion. He has worked on generative models since 2014. In 2019, with PhD student Yang Song, he introduced score-based models that learn to denoise—forming the foundation of modern diffusion. While diffusion already dominates images, video, and protein design, Ermon saw the same opportunity for text and code. Autoregressive models remain sequential at inference: you cannot generate the tenth token until the previous nine exist. That workload is memory-bound and poorly matched to GPUs.
In 2024 his lab showed for the first time that a diffusion LLM at GPT-2 scale could match autoregressive perplexity while generating text roughly 10× faster. Inception was founded to scale the idea commercially. The company is about two years old and ~50 people strong. Its Mercury models now match the quality of OpenAI’s speed-optimized Flash and Mini models on benchmarks, yet run significantly faster, and are already serving real production traffic. One public customer, Open Call (voice agents), previously ran on Cerebras custom silicon; they switched to Inception’s diffusion LLMs on ordinary NVIDIA GPUs and obtained the same latency at lower cost and higher availability.
Ermon’s core bet: “the bitter lesson is that the more parallel solution is the one that is eventually going to win.” Economics are dominated by intelligence per watt and per dollar; inference is also the bottleneck for RL post-training rollouts. Diffusion models are easier to steer (coarse-to-fine generation lets you apply rewards early) and appear more data-efficient. The main challenges are the immature ecosystem—custom serving engines, SFT/RL stacks—but those become proprietary moats. Ermon estimates 20-30% of current workloads are latency-critical enough that the fastest model of a given quality wins immediately.
The Takeaway：推理时扩展性将决定AI胜负，而扩散模型因其天然并行特性，比自回归模型更适合这一趋势。
Stanford教授、Inception联合创始人兼CEO Stefano Ermon是扩散模型之父之一。他从2014年起专注生成模型研究，2019年与学生Yang Song提出基于去噪的score-based模型，奠定现代扩散基础。扩散在图像、视频、蛋白质等领域已主导，但他看到文本与代码的机会：自回归模型生成时仍是串行的，无法充分利用GPU并行能力，导致内存瓶颈和低效。
2024年，其团队首次在GPT-2规模上证明扩散LLM可匹配自回归模型的困惑度，却能实现10倍生成速度。Inception因此成立，专注规模化商业扩散语言模型。目前公司约50人，Mercury模型在基准上已与OpenAI的Flash/Mini等速度优化版本相当，却显著更快，并已在生产环境服务客户，例如语音代理公司Open Call。后者从Cerebras定制芯片切换到Inception的GPU方案，获得同等速度、更高可用性和更低成本。
Ermon强调：“the bitter lesson is that the more parallel solution is the one that is eventually going to win.” 推理主导经济性（每瓦特/每美元的智能），也是RL后训练的瓶颈。扩散模型还更易控制、可能更数据高效。挑战在于需自建服务引擎、SFT/RL栈，但这也形成护城河。他估计20-30%的工作负载对延迟极度敏感，正是扩散的切入点。
查看原文 →
Stanford professor and Inception co-founder/CEO Stefano Ermon is one of the fathers of diffusion. He has worked on generative models since 2014. In 2019, with PhD student Yang Song, he introduced score-based models that learn to denoise—forming the foundation of modern diffusion. While diffusion already dominates images, video, and protein design, Ermon saw the same opportunity for text and code. Autoregressive models remain sequential at inference: you cannot generate the tenth token until the previous nine exist. That workload is memory-bound and poorly matched to GPUs.
In 2024 his lab showed for the first time that a diffusion LLM at GPT-2 scale could match autoregressive perplexity while generating text roughly 10× faster. Inception was founded to scale the idea commercially. The company is about two years old and ~50 people strong. Its Mercury models now match the quality of OpenAI’s speed-optimized Flash and Mini models on benchmarks, yet run significantly faster, and are already serving real production traffic. One public customer, Open Call (voice agents), previously ran on Cerebras custom silicon; they switched to Inception’s diffusion LLMs on ordinary NVIDIA GPUs and obtained the same latency at lower cost and higher availability.
Ermon’s core bet: “the bitter lesson is that the more parallel solution is the one that is eventually going to win.” Economics are dominated by intelligence per watt and per dollar; inference is also the bottleneck for RL post-training rollouts. Diffusion models are easier to steer (coarse-to-fine generation lets you apply rewards early) and appear more data-efficient. The main challenges are the immature ecosystem—custom serving engines, SFT/RL stacks—but those become proprietary moats. Ermon estimates 20-30% of current workloads are latency-critical enough that the fastest model of a given quality wins immediately.
The Takeaway：推理时扩展性将决定AI胜负，而扩散模型因其天然并行特性，比自回归模型更适合这一趋势。
Stanford教授、Inception联合创始人兼CEO Stefano Ermon是扩散模型之父之一。他从2014年起专注生成模型研究，2019年与学生Yang Song提出基于去噪的score-based模型，奠定现代扩散基础。扩散在图像、视频、蛋白质等领域已主导，但他看到文本与代码的机会：自回归模型生成时仍是串行的，无法充分利用GPU并行能力，导致内存瓶颈和低效。
2024年，其团队首次在GPT-2规模上证明扩散LLM可匹配自回归模型的困惑度，却能实现10倍生成速度。Inception因此成立，专注规模化商业扩散语言模型。目前公司约50人，Mercury模型在基准上已与OpenAI的Flash/Mini等速度优化版本相当，却显著更快，并已在生产环境服务客户，例如语音代理公司Open Call。后者从Cerebras定制芯片切换到Inception的GPU方案，获得同等速度、更高可用性和更低成本。
Ermon强调：“the bitter lesson is that the more parallel solution is the one that is eventually going to win.” 推理主导经济性（每瓦特/每美元的智能），也是RL后训练的瓶颈。扩散模型还更易控制、可能更数据高效。挑战在于需自建服务引擎、SFT/RL栈，但这也形成护城河。他估计20-30%的工作负载对延迟极度敏感，正是扩散的切入点。