{
  "id": "shane-legg",
  "title": "Shane Legg",
  "lang": "en",
  "source_lang": "en",
  "summary": "Shane Legg is a New Zealand artificial intelligence researcher, a co-founder of DeepMind, and a contributor to the formal study of machine intelligence.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "person",
  "aliases": [],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/shane-legg/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/shane-legg/",
      "json": "https://aiwiki.page/zh/shane-legg.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Shane Legg is a New Zealand researcher in artificial intelligence and a co-founder of DeepMind. His research concerns the definition and measurement of intelligence, learning agents, and artificial general intelligence (AGI). He is particularly associated with the formal definition of universal intelligence developed with Marcus Hutter."
    },
    {
      "heading": "Education and early research",
      "text": "Legg studied at the University of Waikato and the University of Auckland in New Zealand. He subsequently pursued doctoral research at the University of Lugano in Switzerland under Hutter’s supervision, completing his doctorate in 2008. His dissertation, Machine Super Intelligence, examined mathematical approaches to general machine intelligence and the evaluation of intelligent agents.\n\nThis work addressed a central problem in artificial intelligence: how to characterize intelligence independently of a particular task, such as playing chess, recognizing images, or understanding language."
    },
    {
      "heading": "Universal intelligence",
      "text": "In their 2007 paper Universal Intelligence: A Definition of Machine Intelligence, Legg and Hutter proposed a formal measure based on an agent’s ability to achieve goals across a broad range of environments. Their formulation brought together ideas from reinforcement learning, computability theory, and Kolmogorov complexity.\n\nThe framework evaluates an agent through its expected performance in computable environments. It gives greater weight to environments with shorter descriptions, using an algorithmic notion of simplicity. Intelligence is therefore treated as breadth of successful adaptation rather than proficiency on a single benchmark.\n\nThe proposal is a theoretical formalization, not a directly computable intelligence test. Its general form involves an unbounded collection of environments and quantities that cannot be computed exactly. It also depends on formal choices about how environments and rewards are represented. These features distinguish it from practical benchmark suites used to compare machine-learning systems."
    },
    {
      "heading": "DeepMind",
      "text": "Legg co-founded DeepMind in London in 2010 with Demis Hassabis and Mustafa Suleyman. The company pursued general-purpose learning systems, combining approaches including neural networks and reinforcement learning.\n\nGoogle acquired DeepMind in 2014. In 2023, DeepMind and Google Brain were brought together as Google DeepMind. Legg has held research leadership responsibilities at the organization, including the role of chief AGI scientist.\n\nDeepMind’s institutional research has included systems such as AlphaGo and AlphaFold. These are collaborative projects; Legg’s role as a founder and research leader is distinct from the specific technical contributions credited to each project’s authors."
    },
    {
      "heading": "Research focus",
      "text": "A recurring distinction in Legg’s work is between competence at a particular task and the general ability to learn and perform across different tasks. His theoretical research asks how that distinction can be expressed mathematically, while his work at DeepMind places it within the development of learning systems.\n\nThe universal-intelligence framework does not establish that a particular deployed system possesses AGI. Rather, it supplies one formal account of what general intelligence might mean and exposes some of the difficulties involved in measuring it."
    }
  ],
  "links": [
    "artificial-intelligence",
    "deepmind",
    "artificial-general-intelligence",
    "reinforcement-learning",
    "computability-theory",
    "kolmogorov-complexity",
    "machine-learning",
    "artificial-neural-network",
    "google",
    "alphago",
    "alphafold"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}