The Arrival of AGI with Shane Legg
Shane Legg, co-founder of Google DeepMind and Chief AGI Scientist, joins host Hannah Fry for a wide-ranging discussion on artificial general intelligence. Legg defines "minimal AGI" as an AI that can perform the cognitive tasks typical humans can do—something he estimates has roughly a 50/50 chance of arriving by 2028, a prediction he has held consistently since 2009. He distinguishes this from "full AGI" (matching the full spectrum of human cognition) and "artificial superintelligence" (ASI), arguing that human intelligence is not an upper bound given the orders-of-magnitude advantages data centers have over biological brains in energy, bandwidth, and signal propagation speed. On safety, Legg outlines a multi-pronged approach including pre-deployment testing, monitoring, interpretability, and what he calls "System 2 safety"—using chain-of-thought reasoning so AIs explicitly deliberate about ethics rather than relying on instinct. He frames ethics as a reasoning problem and argues AIs could become more consistently ethical than humans, though he acknowledges cross-cultural variation and the difficulty of grounding machine ethics in human experience. Legg warns that AGI will cause significant economic disruption, particularly in remote, purely cognitive work such as software engineering, with implications for lawyers, financiers, and mathematicians. He urges universities, governments, and every discipline to seriously engage with what a post-AGI society and economy will look like, comparing the current moment to the early warnings of the COVID-19 pandemic. He expresses cautious optimism, describing AGI as a potential "golden age" if society can navigate the transition thoughtfully.
Original sources
Podcast Google DeepMind: The Podcast
Episode:The Arrival of AGI with Shane Legg (co-founder of DeepMind) ↗
YouTube Google DeepMind: The Podcast
Video:The arrival of AGI | Shane Legg (co-founder of DeepMind) ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.