Elorian AI Bets on Visual Understanding Over Text for AI

Elorian AI, a startup that emerged from stealth in April 2026, is developing artificial intelligence models designed from the ground up to understand, reason about, and generate images, challenging the dominant word-centric approach of large language models (LLMs). Andrew Dai, a veteran AI researcher and co-founder of Elorian AI, believes that visual understanding is as fundamental to model intelligence as language, and that current frontier LLMs have reached a ceiling due to their inability to effectively reason about the physical world. Dai, who previously worked at Google DeepMind, argues that models which cannot perform basic spatial reasoning, such as counting objects or judging spatial relationships, fall short of general intelligence, regardless of their linguistic capabilities.

Elorian AI, co-founded by Dai and former Apple machine-learning researcher Yinfei Yang, is building models that integrate visual data with language tokens equally within their architecture. Dai stated, "We're betting on these visual representations for things like spatial problems and navigation and everything else." This approach contrasts with current multimodal language models, such as Google's Gemini. These existing models process imagery by converting it into detailed textual descriptions, which are then organized into an internal word map. Reasoning is performed on these textual representations to derive observations and judgments about the image content.

Dai expressed skepticism about the long-term viability of the current industry trend, which focuses on making LLMs larger and smarter with the expectation of achieving superintelligence. He described frontier language models as "incredibly unstable" and highlighted their limitations in tasks requiring physical world comprehension. For example, a multimodal LLM processing a motorcycle engine schematic might describe the expansion of an aluminum alloy piston crown upon heating, which could reduce clearance with the cylinder wall. However, Dai's vision for Elorian AI aims for a deeper, more intrinsic understanding of visual information, moving beyond mere textual descriptions to enable more robust spatial reasoning and problem-solving capabilities.