For almost a decade, AI researchers have been obsessed with one kind of artificial intelligence—large language models (LLMs), which they train to write, talk, and reason using words. The tech industry is making an enormous bet that making such models bigger and smarter is the path to superintelligence. But in 2026 some very high-profile AI researchers doubt that this word-centric approach is the whole answer. Among them is Andrew Dai, a veteran AI researcher who left Google DeepMind to build models designed from the ground up to understand, reason about, and generate images*.* His startup company, Elorian AI, came out of stealth in April.
Dai believes that the visual image is as fundamental to model intelligence as language—if not more so—and that frontier language models have hit a ceiling because they still can’t reason effectively about the physical world, not to mention being “incredibly unstable,” as he puts its. A model that can’t count the number of cups on a table or judge spatial relationships falls short of general intelligence, no matter how well it writes or codes, Dai says.
Elorian, which Dai cofounded with the former Apple machine-learning researcher Yinfei Yang, is developing models that give visual data equal play with language tokens within their architecture. “We’re betting on these visual representations for things like spatial problems and navigation and everything else,” he says.
When recent multimodal language models like Google’s Gemini process imagery, what they really do is create a detailed description of an image using words, which they arrange in a giant internal word map based on their meanings. They then reason about the words to make observations, judgements, and even suggestions about the image.
For instance, if shown a schematic of a motorcycle engine design, a multimodal LLM might observe that when motorcycle engines heat up, an aluminum alloy piston crown can expand, which might decrease the clearance between it and the cylinder wall. By contrast, Elorian is building visual reasoning models that will “think” about images directly. Instead of creating a word map, they’ll form a detailed 3D internal map of an image, the way humans imagine things. Because the model has an understanding of physics, it can make much more detailed and accurate observations.
In the motorcycle engine example, the model could show how the dimensions of piston crown and cylinder wall, and the clearance between the two, would change over time as the motorcycle engine revs up. Then it could suggest changes to the design to make the engine more durable at high rpm operation. Elorian’s models will reason and simulate, while an LLM can only describe and reference.
So Elorian’s visual reasoning models could find natural applications in industries that require a deep physical understanding of imagery. Elorian is targeting what Dai calls the physical economy, an $80 trillion category that includes video understanding and mechanical engineering. In the latter, Dai said engineers currently spend hundreds of hours manually drawing components in CAD software. Elorian’s models can run what Dai calls an edit-simulate-correction loop; they generate a design, then test it within a physics simulation, then identify design flaws, then revise the design accordingly, automatically.
Dai spent 14 years at Google Brain and DeepMind. There he coauthored research that provided technical groundwork for the GPT series models, and led data work on Gemini. He left with a team of researchers to start Elorian, headquartered in Palo Alto. Dai says one of the reasons he left Google was that the company is now concentrating its compute resources on a few distinct areas, including code generation, and allotting less compute to research on visual reasoning.
Elorian’s specialized models already outperform Gemini 3 Pro on certain visual reasoning benchmarks, Dai says, though he declined to name them, saying disclosure would let competitors optimize against them. (Gemini is still fundamentally a transformer model that projects vision into a shared latent space, but DeepMind is also investing in research on models that reason over visual images themselves rather than words.)
Elorian has raised $55 million in seed funding at a $300 million valuation. Investors include Nvidia and Menlo Ventures, as well as Dai’s former boss, the legendary Google chief scientist Jeff Dean. PitchBook also lists 49 Palms Ventures, 500 Global, and Altimeter Capital Management, among others, as investors.
Elorian’s $300 million seed valuation is modest next to other startups building AI for the physical world. Physical Intelligence, a robotics foundation model company founded in 2024 by the former Google DeepMind researcher Karol Hausman and the Stanford professor Sergey Levine, was valued at $5.6 billion in a $600 million funding round in November.
World Labs, a spatial-intelligence startup founded by the Stanford professor and AI researcher Fei-Fei Li, works on a related problem: AI systems that reason about three-dimensional space rather than 2D images. The company’s product, Marble, generates navigable 3D environments from text, images, or video. World Labs raised $1 billion in a February round led by Autodesk along with Nvidia, AMD, and Andreessen Horowitz.
Moonvalley, which was acquired by Reka this year, built the licensed-footage video model Marey, then moved into world models. Reka said the combined team will develop AI that can simulate motion and physics to help robots reason about the consequences of their actions.
But Dai’s company is young and is still developing its models and fitting them to its first addressable markets. Elorian plans a general Application Programming Interface release by the end of the year, which will allow developers to build apps on top of its visual reasoning models.