ABAB Pedia

Andrew M. Dai

Andrew M. Dai is an AI researcher and co-founder and CEO of Elorian AI. He previously served as a research director at Google DeepMindGoogle DeepMindGoogle DeepMind is Google’s artificial intelligence research organization, formed in April 2023 by merging DeepMind, which Google acquired in 2014, with the Brain team from Google Research, and led by CEO Demis Hassabis. It developed AlphaGo, AlphaFold, and the Gemini family of models, and Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for AlphaFold’s protein structure prediction work.Open the full entry and led Gemini’s data area.

Contents10 sections
Key facts

Education

He studied computer science at the University of Cambridge before pursuing doctoral research in machine learning at the University of Edinburgh.⁠[1]

Career at Google

Dai joined Google in 2012, initially working on Google Now. He later conducted research at Google Brain and Google DeepMind and left in 2026.⁠[1]

Semi-supervised Sequence Learning

GLaM

In December 2021, Dai and Nan Du introduced GLaM. Its sparse mixture-of-experts architecture selects two experts per token, expanding total model capacity while limiting the parameters used for each prediction.⁠[3]

PaLM 2 and Gemini

He co-authored the PaLM 2 Technical Report, released in May 2023. It described improvements in multilingual and reasoning capabilities and distinguished pretrained models, fine-tuned variants, and user-facing products.⁠[4]

Elorian’s team biography credits him as a co-lead of PaLM 2 pretraining and GLaM, and as the lead of Gemini’s data area.⁠[5]

Founding Elorian AI

Dai co-founded Elorian AI with Yinfei Yang and Seth Neel and became its CEO. The company focuses on visual reasoning.⁠[5]

On April 9, the company emerged from stealth and announced $55 million in seed funding led by Striker, Menlo Ventures, and Altimeter, to support research and development.⁠[6]

TechCrunch reported that the round valued Elorian at $300 million before it had launched a product.⁠[7]

Glimpse Research Preview

On October 16, Elorian released a research preview of its first model, Glimpse, through chat and an API. The model revisits images and video during reasoning.⁠[8]

The company also noted that detailed visual inspection can require more time and tokens. Direct robot control remained a future research direction, and the published benchmark results came from company evaluations.⁠[8]

Research Motivation and Context

In a TechCrunch interview, Dai described a model’s difficulty assessing a board-game position from a photograph as a motivation for the company. The report also noted earlier uses of cropping and zooming during reasoning, so these tools themselves were not unique to Elorian.⁠[7]

Related Organizations, Websites, and Public Accounts

Elorian AI: Official website: https://elorian.ai/ (opens in a new window)⁠[5]

Sources

  1. Andrew Dai (opens in a new window)
  2. Semi-supervised Sequence Learning (opens in a new window)
  3. More Efficient In-Context Learning with GLaM (opens in a new window)
  4. PaLM 2 Technical Report (opens in a new window)
  5. Team (opens in a new window)
  6. Wilson Sonsini Advises Elorian on $55 Million Seed Financing as Company Emerges from Stealth (opens in a new window)
  7. Elorian’s first AI model wants to reason with images, not just about them (opens in a new window)
  8. Introducing Glimpse (opens in a new window)
  9. Andrew Dai (opens in a new window)