Inception just closed a $50 million seed round to tackle one of AI's biggest inefficiencies - how models generate code. Led by Stanford's Stefano Ermon, the startup is betting that diffusion models, which power image generators like Midjourney, can revolutionize software development by processing code faster and cheaper than today's dominant autoregressive models.
The AI funding frenzy just found its next big bet. Inception, a Stanford-led startup, closed a $50 million seed round to bring diffusion models - the technology behind image generators like Stable Diffusion and Midjourney - into the world of code generation. Menlo Ventures led the round, with participation from Mayfield, Innovation Endeavors, Nvidia's NVentures, Microsoft's M12 fund, Snowflake Ventures, and Databricks Investment.
The timing couldn't be better for AI researchers with bold ideas. As one industry veteran put it, it's easier to get resources as an independent company than inside the big labs right now. That's exactly what Stanford professor Stefano Ermon discovered when he decided to commercialize his diffusion model research through Inception.
Ermon has been working on diffusion models since before they became the darling of AI image generation. Now he's applying the same iterative refinement approach that powers Midjourney and Sora to software development. "These diffusion-based LLMs are much faster and much more efficient than what everybody else is building today," Ermon told TechCrunch. "It's just a completely different approach where there is a lot of innovation that can still be brought to the table."
The technical distinction matters more than it might seem. Most text-based AI services today rely on autoregressive models like GPT-4 and Google's Gemini, which work sequentially - predicting each next word based on what came before. Diffusion models take a more holistic approach, modifying the overall structure of a response incrementally until it matches the desired result.
Inception's Mercury model is already making waves in the developer tools space. The company has integrated Mercury into ProxyAI, Buildglare, and Kilo Code, giving developers their first taste of diffusion-powered code generation. Early results suggest the approach delivers significant advantages in two critical areas: response time and compute costs.
"We've been benchmarked at over 1,000 tokens per second, which is way higher than anything that's possible using the existing autoregressive technologies," Ermon claims. The speed boost comes from diffusion models' parallel processing capability - while autoregressive models must execute operations one after another, diffusion models can handle multiple operations simultaneously.
