Andrew Dai

Andrew Dai is the co-founder and chief executive of Elorian, where he is building AI systems that reason about images, video, and the physical world. His career connects language-model pretraining and shared leadership on Google’s large-scale models with a question that now drives his company: what must AI learn beyond language to understand shapes, motion, and spatial relationships?

From sequence learning to large-scale language models

Dai spent twelve years as a Research Scientist at Google Brain and DeepMind. In 2015, he and Quoc V. Le developed Semi-supervised Sequence Learning, which investigated how recurrent networks could learn from unlabeled sequences before being trained for a supervised task. One method predicted the next element of a sequence; another reconstructed the input. Reusing the learned parameters made subsequent training more stable and improved generalization on text-classification tasks. The work demonstrated the usefulness of pretraining followed by supervised adaptation before transformers became the dominant language-model architecture.

With Takeru Miyato and Ian Goodfellow, Dai also co-authored research on adversarial training for text classification. Rather than perturbing discrete words, the researchers introduced small disturbances into their numerical embeddings during training. Training against those disturbances helped models resist local changes and reduced overfitting.

His responsibilities expanded from training methods to the architecture and data behind major systems. He co-led GLaM and PaLM 2 pretraining and co-led Gemini’s data area. GLaM addressed the expense of scaling through sparsely activated mixture-of-experts models. A routing network selected two experts for each token, giving the model access to a larger collection of parameters while activating only a fraction for each prediction. Dai and Nan Du’s account of the project also described filtering web material using examples from books and Wikipedia to improve training-data quality.

That work changed his view of what language models could accomplish. Dai had believed that language could carry reasoning; leading Gemini’s data work brought the limitations of that belief into focus. He came to see visual reasoning as a capability that needed to be developed in its own right. He left Google to build Elorian, shifting from improving large-scale language models to pursuing systems that could reason directly about the visual world.

Building models that think visually

Dai co-founded Elorian with Yinfei Yang and Seth Neel. The company launched with $55 million in funding at a $300 million valuation, backed by Menlo Ventures, Altimeter, Striker Venture Partners, NVIDIA, and Jeff Dean. Its research focuses on visual reasoning, including the spatial relationships and physical constraints needed for engineering and action in the real world.

Dai’s critique of visual AI centers on a failure that fluent explanations can conceal: a model may recognize what an image resembles without examining what it actually contains. In his cropped-chessboard example, models answer that there are 32 white squares—the number on a complete board—rather than count the visible squares. The same reliance on familiar patterns that helps identify a flower can produce an incorrect answer when the task requires careful inspection.

A board-game enthusiast, Dai uses games to make the distinction tangible. Recognizing a Catan board is one task; counting a player’s roads requires attention to individual pieces and their locations. His examples extend to video, where missing a robot arm lifting a lid changes the interpretation of everything that follows. For him, spatial detail and the ability to retain changes over time are central requirements of useful visual intelligence.

He uses a practical distinction between recognition and reasoning: could a person answer after a one-second glance, or would they need sustained attention, counting, and comparison? He also argues that benchmark success can obscure this gap when questions use tiny visual grids or can largely be answered without inspecting the image. His priority is evaluation that tests detailed spatial reasoning and consistency across video.

Elorian’s approach connects that diagnosis to several research directions:

• Visual chain of thought: Intermediate reasoning operates on the image itself. In Dai’s example, a model asked to count red hotels first marks candidate objects with bounding boxes, then narrows the selection by color. Those visual steps keep the answer connected to the objects being counted.

• Data and training for visual reasoning: Elorian collects and generates specialized multimodal data that Dai says is difficult to obtain online. Its synthetic-data process combines evaluations, agents, supervised fine-tuning, and reinforcement learning, alongside changes to transformer architectures. The approach develops training material and reasoning mechanisms together.

• Reasoning tied to actions and constraints: Dai proposes scene understanding that can feed a robot’s planning and control systems, construction monitoring that connects camera footage to written safety policies, and design tools that extract geometric constraints for programmatic or simulation validation. Each requires the model to relate observations to something another system can use: a changed scene, an applicable rule, or a checkable design constraint.

These applications remain research and product goals. They give Dai’s pursuit of visual thinking a concrete direction: helping AI handle the shapes, motion, and relationships that people rely on to build and operate physical systems.