Stefano Ermon helped pioneer technology that’s used by popular artificial intelligence services to generate images and videos. Now his startup, Inception, is vying to change how AI handles text.
Inception is set to release Mercury 2, an AI model that’s designed to field questions from users significantly faster and more cheaply than alternative services. Where text models typically process tokens – or units of data – sequentially, Mercury 2 relies on diffusion models to handle multiple tokens simultaneously.
The system "starts with a sketch of the answer and then it refines it in parallel,” said Ermon, a computer science professor at Stanford. "That's why these diffusion models, this Mercury model, is so much faster than anything that was built before.” Mercury 2 can also lower costs because it can generate more tokens using the same number of chips.
Inception is part of a growing crop of startups exploring novel ways to build AI models, with the goal of making them more adept at various tasks. The company is backed by Microsoft Corp’s M12, Snowflake Ventures and Databricks Inc, as well as a who’s who of angel investors, including Eric Schmidt, the former chief executive officer of Google.
Many of the companies trying to provide faster responses to AI queries have developed customised chips that are designed differently than Nvidia Corp’s graphics processing units. Inception, however, is trying to do the same thing solely with software, meaning its models can run on Nvidia’s hardware, the most widely used chip.
"It’s a broader solution, and it's re-imagining how models should be built,” Navin Chaddha, managing partner at Mayfield, which invested in the startup in 2024 before it had proven its ideas could work.
Inception has partnerships with Microsoft, Amazon.com Inc and Nvidia. It’s also working on plans for how its new model could help large companies, including a top food-ordering company and a leading maker of coding tools.
The company’s technology may not be the answer for every situation, but its investors are betting it will be useful for customers looking for high-speed solutions. Tim Tully, a partner at Menlo Ventures – which led a US$50mil (RM194mil) seed round in Inception last year – said Mercury has pretty much replaced Google Search for him. He also thinks the startup’s technology could make things like customer service bots more usable.
"I want to feel like I'm talking to a human,” Tully said. "I don’t want to feel like I'm talking to a computer.” – Bloomberg
