In a significant shift in how artificial intelligence may evolve over the next decade, former Google DeepMind researcher Andrew Dai has stepped out to build Elorian AI, a new startup centred on what he believes is the missing layer in today’s AI systems, true visual reasoning.
After nearly 14 years at Google, including over a decade working across Google Brain and DeepMind, Dai has co-founded Elorian AI, a new research-led startup focused on helping machines better understand and interpret the visual world. The company has already raised $55 million in funding at a $300 million valuation, with backing from Striker Venture Partners, Menlo Ventures and Altimeter Capital, along with participation from Nvidia and AI researcher Jeff Dean.
The startup is based in Palo Alto and is currently building specialised models designed to improve how AI systems reason about images, a capability that remains surprisingly underdeveloped despite rapid advances in language-based AI.
The gap AI hasn’t solved yet
For all the progress in chatbots and code generation, Dai believes the industry has overlooked a fundamental limitation. In his own words, today’s most advanced AI systems still struggle with basic visual reasoning tasks, a gap that becomes obvious when models are asked to interpret images beyond surface-level description.
This aligns with the broader problem Elorian is trying to solve. Even with billions of dollars invested by companies like Google and OpenAI, current models still perform poorly in tasks such as analysing satellite imagery or identifying what is missing in a visual scene.
Dai’s argument is simple but pointed: intelligence in the real world is not built on language alone.
Drawing from his experience working on large-scale AI systems, including Gemini, he suggests that the industry may have approached the problem in reverse. In nature, perception comes before reasoning, and reasoning comes before language. But modern AI, he argues, has been built starting from language, skipping the foundational layer of visual understanding.
The result is systems that can generate impressive outputs, yet fail at tasks that require deeper interpretation of the physical world.
Why Elorian AI is betting on multimodal reasoning
Elorian’s core thesis is that solving visual reasoning could unlock real-world applications that current AI systems struggle to handle effectively.
The company is developing models specifically designed for this purpose, with potential use cases across engineering, robotics, architecture and design, where understanding spatial relationships and physical context is critical.
Dai has emphasised that designing real-world systems, whether it is a more efficient vehicle or a better physical structure, cannot be reduced to text or code alone. These are problems that exist in a visual and spatial domain, and require AI systems that can reason within that domain.
The startup is currently pre-revenue but is in discussions with potential customers and expects to release its first publicly available reasoning model within the next year.
A different path in a crowded AI race
Elorian enters an increasingly competitive landscape, where several researchers from major AI labs are launching specialised ventures. However, its positioning is distinct.
Rather than focusing on generative media, an area that has seen both hype and recent skepticism, Elorian is concentrating on reasoning. Dai has noted that while generating visually appealing outputs is relatively easy, building systems that can understand and explain what they generate is significantly harder.
The company is also taking a hybrid approach to development, building its early products on open-source models while considering releasing smaller versions to the community. Its flagship systems, however, are expected to remain proprietary.
Internally, Elorian is also being built with a long-term view. Dai has indicated that he consciously avoided starting at an inflated valuation to ensure early team members have meaningful upside, a decision that stands out in a sector often driven by aggressive valuations.
The bigger shift: from language to perception
Elorian’s emergence reflects a broader turning point in AI.
The past few years have been dominated by systems that talk, write and code. The next phase, if Dai’s thesis holds, may be defined by systems that can see and more importantly, understand what they see.
If that shift materialises, it could determine how quickly AI moves from digital assistants to real-world problem solvers.
For now, Elorian is placing a calculated bet: that the future of AI will not just be conversational, but perceptual.
Also Read: Atlas Secures $6 Million Seed Funding to Expand AI Solutions for Accounting Firms
