Elorian AI develops image-native artificial intelligence models

Elorian AI, a Palo Alto-based startup founded by veteran AI researcher Andrew Dai, is developing artificial intelligence models that process images directly, rather than relying on textual descriptions. Dai, who previously led data work on Google's Gemini and co-authored groundwork for GPT models at Google DeepMind, believes that true artificial general intelligence requires a fundamental understanding of visual and physical data.

The company's models aim to create internal, 3D representations of visual input, allowing them to reason about spatial relationships, physics, and mechanical processes. This contrasts with current multimodal models that often convert images into textual descriptions before analysis. Elorian's approach is intended to enable more accurate simulations and problem-solving in fields like engineering and design.

Dai stated that he left Google partly because the company was concentrating its compute resources on areas like code generation, allocating less to visual reasoning research. He claims Elorian's specialized models already outperform Google's Gemini on certain visual reasoning benchmarks, though he declined to name them.

Elorian AI has raised $55 million in seed funding at a $300 million valuation. Investors include Nvidia, Menlo Ventures, and Dai's former Google colleague Jeff Dean. The company plans to release a general Application Programming Interface (API) by the end of the year, enabling developers to build applications using its visual reasoning models.