Former DeepMind researchers bet on visual AI with new startup

In a significant development in the realm of artificial intelligence, Andrew Dai, a former researcher at DeepMind, has launched a new startup called Elorian. This venture aims to enhance AI's capacity to understand and reason about visual information, a critical aspect that has been notably lacking in existing AI models. With substantial backing from prominent investors, Elorian is poised to address the limitations of current AI systems, particularly in practical applications such as robotics and design.

Elorian emerged from stealth mode recently, announcing a remarkable $55 million in funding and a valuation of $300 million. Based in Palo Alto, California, the startup has already expanded its team by hiring over a dozen employees. The funding comes from notable investors including Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with additional participation from Nvidia Corp. and renowned AI researcher Jeff Dean. The funding round was executed in two stages, beginning with a valuation of $120 million and subsequently increasing to $300 million, as confirmed by a spokesperson for Elorian.

Despite not yet generating revenue, Elorian is actively engaging with potential customers and plans to unveil its first publicly accessible reasoning model within the next year. The startup's co-founders include Yinfei Yang, who has a background in AI research at both Google and Apple, and Seth Neel, a former Harvard professor specializing in data and artificial intelligence.

Understanding Elorian's Vision

The core mission of Elorian revolves around the belief that to improve AI's comprehension of the world, specialized models tailored for this purpose must be developed. Andrew Dai argues that, despite the significant investments made by companies like Google and OpenAI in large-scale AI models, these tools still struggle with tasks such as analyzing satellite imagery or understanding what elements an image lacks. Dai emphasizes that while advancements in coding have enhanced reasoning capabilities, current models are still inadequate for making complex design decisions, such as creating lighter vehicles or more efficient rockets. He states, "These are not things that you can just express with code and have a faster rocket; you actually need to design the physical thing – and that design lives in the physical world."

The startup's approach diverges from recent trends in AI, particularly in light of OpenAI's recent decision to discontinue its video generation app, Sora. This move raised questions about the commercial viability of AI in media generation. However, Dai clarifies that Elorian's focus is not on media creation but rather on enhancing reasoning capabilities. He notes, "It’s much easier to generate something that looks great, but it’s a lot harder to reason about it, to understand it, to explain what it does to someone."

Elorian is developing its initial products based on open-source models, which are freely available for use and modification. The company is contemplating the release of smaller versions of its models to the open-source community, while its flagship model is likely to remain proprietary. This strategy reflects a growing trend among AI startups to balance open collaboration with proprietary advancements.

The Broader Landscape of AI Startups

Andrew Dai's decision to establish Elorian comes amidst a wave of researchers leaving established labs to pursue specialized projects in the rapidly evolving AI landscape. Notable examples include Richard Socher, founder of You.com, who is raising capital for a more advanced AI system valued at $4 billion, and Periodic Labs, co-founded by former Google and OpenAI researchers, currently fundraising at a $7 billion valuation.

Dai has opted for a different strategy with Elorian, intentionally starting with a lower valuation to ensure that early employees can realize significant financial gains. He explains, "If you start with a very high valuation, then for them, it’s very hard for them to see a 50x or 100x increase in the value of their equity, and we wanted to give that opportunity."

This philosophy has attracted interest from investors like Brian Zhan, a partner at Striker, who highlights Dai's ability to recruit top researchers from DeepMind as a key factor in their support for Elorian. Zhan anticipates that Elorian will operate more efficiently than other cutting-edge labs, given Dai's experience in building Google’s AI model, Gemini. He remarks, "Andrew knows the Gemini recipe — he’s not wasting a single dollar."

Why it matters

The launch of Elorian by Andrew Dai signifies a pivotal moment in the AI sector, particularly concerning visual reasoning capabilities. As AI continues to integrate into various industries, the need for models that can accurately interpret and reason about visual information becomes increasingly critical. For creators and technologists, this development suggests a potential shift in how AI can be utilized in practical applications, from architecture to robotics. By focusing on specialized reasoning models, Elorian may pave the way for more sophisticated AI systems that not only generate content but also understand and interact with the physical world in meaningful ways. This could lead to advancements that enhance efficiency and innovation across multiple sectors, ultimately benefiting both creators and consumers alike.