Shane Legg is a DeepMind co-founder and Google DeepMind’s chief AGI scientist. His research asks how to define and measure general intelligence, and his current work includes AI safety and possible societal consequences of advanced AI. He is also a director and managing editor of the DeepMind Institute.[1][3][7]
Start with Levels of AGI for Operationalizing Progress on the Path to AGI →
Education
Links
Personal site and biography
X DeepMind Institute Co-founder and Chief AGI Scientist · Google DeepMind
Director and managing editor · DeepMind Institute
Background
Legg studied computing and mathematical sciences at the University of Waikato, where he worked on the WEKA machine-learning project during undergraduate summers. His personal biography describes mathematics master’s work at Auckland supervised by Cris Calude, focused on Solomonoff induction, a mathematical theory of prediction. He began doctoral research at IDSIA in Switzerland in 2003 and received a PhD in informatics from Università della Svizzera italiana in 2008.[1][2]
Career
From theoretical research to a general-AI lab
Before his doctorate, Legg worked in software development and at Webmind and A2I2. His later postdoctoral work included the Swiss Finance Institute and UCL’s Gatsby Computational Neuroscience Unit, where he met Demis Hassabis. With Hassabis and Mustafa Suleyman, he started DeepMind in 2010. His account describes a deliberate attempt to build an organization focused on artificial general intelligence. Today he identifies himself as chief AGI scientist and co-chair of the AGI Safety Council, with a team studying possible conditions after AGI.[1][3]
Safety work and a wider public discussion
Legg co-authored Google DeepMind’s 2025 account of its approach to AGI safety, which describes risk assessment, safety research and internal review. In September 2026, he helped launch the DeepMind Institute with James Manyika and Demis Hassabis. He serves as both director and managing editor. The institute invites interdisciplinary work on advanced AI’s social effects and stresses that technical experts cannot decide those effects alone.[7][3]
Contributions
Legg and Marcus Hutter’s Universal Intelligence paper starts with an awkward question: how can researchers compare machines with very different abilities and environments? They extract common features of existing definitions and turn them into a mathematical measure connected to universal learning agents. This is theoretical groundwork, not a practical certificate that a model is intelligent. Legg’s doctoral thesis examines AIXI, a formal ideal agent whose computational demands prevent direct implementation. Its value is in clarifying what a general learner would need to accomplish and where approximations enter.[4][5]
Ideas
The later Levels of AGI paper, co-authored by Legg, separates how well a system performs from how broadly its abilities apply. It also considers autonomy, the extent to which a system acts without human direction. This vocabulary helps readers avoid treating all impressive model demonstrations as equivalent. The framework links classification with evaluation and deployment risks. It remains a proposed way of organizing the problem, rather than an official universal scale or a declaration that any particular current system has reached AGI.[6]
Tools, writing & talks
Writing, research & talks
Start here
Start with this framework for distinguishing performance, breadth and autonomy. It gives a newcomer a vocabulary for reading claims about general AI, with less mathematical preparation than Legg’s early work requires.
Defines performance, generality and autonomy in terms a newcomer can use when assessing claims about AGI.
Co-author[6]Legg and Hutter derive a formal definition from common features of earlier accounts of intelligence. The introduction is approachable; the later argument assumes comfort with mathematical models of agents and prediction.
Co-author with Marcus Hutter[4]USI’s defence announcement includes the abstract of Legg’s doctoral thesis on AIXI and the limits of universal prediction and learning. It introduces the theoretical questions and computational assumptions; the linked page is an abstract rather than the full thesis.
Author of the thesis described in the announcement[5]- Taking a responsible path to AGIdeepmind.google
An employer-published account of safety research and governance co-authored by Legg. It explains the lab’s stated approach and internal structures; it is not an independent audit of how effectively they operate.
Co-author[7]