Angjoo Kanazawa is a computer-vision researcher at the University of California, Berkeley, where she is an associate professor in the Department of Electrical Engineering and Computer Sciences and leads the Kanazawa AI Research (KAIR) lab within the Berkeley AI Research group (Source: vcresearch.berkeley.edu; alphaxiv.org).
Background
Kanazawa took a BA in computer science and mathematics at New York University from 2007 to 2011, working with Rob Fergus, and completed a PhD in computer science at the University of Maryland from 2011 to 2017 under David Jacobs. She was a visiting researcher at Tübingen in 2016–2017 under Michael Black, then a postdoctoral researcher at UC Berkeley from 2017 to 2019 advised by Jitendra Malik, Alexei Efros and Trevor Darrell, and a research scientist at Google Research in New York from September 2019 to August 2021 before joining the Berkeley faculty. Alongside her academic post she is an Amazon Scholar on the Frontier AI and Robotics team and chief technical adviser to Luma AI (Source: alphaxiv.org).
Research
Her work sits at the intersection of computer vision, computer graphics and machine learning, directed at reconstructing the dynamics of a 3D world — people and animals interacting with their environment — from visual data (Source: vcresearch.berkeley.edu).
Three projects associated with her lab are widely used in radiance-field research. Nerfstudio is an open-source modular framework for neural radiance field development (Source: cdss.berkeley.edu). PixelNeRF generates neural radiance fields from a small number of images. LERF embeds natural-language features into neural radiance fields, enabling open-vocabulary search of a 3D scene (Source: alphaxiv.org).
That body of work is relevant to how her position in the world-models debate is read: her research programme consists of building explicit spatial and temporal representations of the physical world, which is the capability the dedicated-world-model camp argues language models lack.
Position in the world-models debate
Kanazawa is cited as stating the LLM-scaling counter-position in the world-models debate — that a sufficiently large language model with good multimodal grounding already constitutes an implicit world model. Quoted in Béchard's Scientific American piece: "In a way, I would say that the LLM already has a very good world model; it's just we don't really understand how it's doing it."
The quotation is usually cited for its first half, but the position is a qualified one: Kanazawa goes on to note that LLMs cannot update their world model from experience. That makes the disagreement with the dedicated-world-model camp narrower and more precise than a simple yes/no about whether LLMs have world models — it concerns whether a representation that cannot be revised through interaction does the work the term is meant to pick out.
Relationships
- related: World Models — the debate this position sits within
- related: Scaling Laws, Frontier Models
- related: AI Robotics — her Amazon Scholar affiliation is with the Frontier AI and Robotics team
- related: Embodied AI vs AGI (China's Race Redirect) — reconstructing world dynamics from visual data is the representational problem the embodiment argument turns on