Shane Legg

Shane Legg co-founded DeepMind in 2010 and holds the title Chief AGI Scientist at Google DeepMind. He co-coined the term "artificial general intelligence" itself, in conversation with Ben Goertzel, who needed a label for a book about the old dream of general-purpose thinking machines. (A guy named Mike Gubrud had actually used the phrase in a 1997 nanotech security paper, but none of them knew about it at the time.) With only 11 records in this archive spanning 2017 to late 2025, Legg is the quietest of the nine people tracked here. He doesn't do media rounds. When he does speak, the positions are remarkably stable. He first gave a 50/50 probability of "minimal AGI" arriving by 2028 back in 2009. In December 2025, on the Google DeepMind podcast with Hannah Fry, he said the same thing: one to five years out, probably about two. Sixteen years of consistency on a timeline prediction is unusual in a field where forecasts shift with each new benchmark.

Legg's framework breaks AGI into three tiers, and he insists it is not a binary threshold. Minimal AGI is the point where an AI stops failing at cognitive tasks in ways that would surprise you if a person failed at them... he calls this the "minimum wage job" level. Full AGI covers the entire range of human cognition, including genius-level feats like inventing new physics. Beyond that lies artificial superintelligence, which Legg admits he has tried and failed to define rigorously on multiple occasions. His argument for why superintelligence follows inevitably from AGI is rooted in physics: the human brain runs on 20 watts with signals traveling at 30 meters per second, while data centers operate at the speed of light with megawatts of power. Biology has hard limits. Silicon doesn't (at least not the same ones). He sees current AI as "uneven" rather than inadequate... already superhuman in languages and general knowledge, but still failing at visual reasoning, perspective, and continual learning. He believes none of these gaps are fundamental blockers, just engineering problems requiring targeted data, architectural changes (like episodic memory systems), and algorithmic work.

On safety, Legg has proposed "system 2 safety," borrowed from Daniel Kahneman's dual-process theory. The idea is to build a slow, deliberate, observable ethical reasoning process into AGI systems, forcing them to use step-by-step analysis (not pattern-matched instinct) for decisions with moral weight. He frames this as one layer in a multi-layered defense: pre-release testing for weaponization and hacking, continuous monitoring after deployment, and interpretability tools. He is explicit that 100% safety guarantees are unrealistic. On economic impact, he offers what amounts to a simple test: if your job can be done entirely through a screen and keyboard, you are in the first domain AGI will automate. Software engineering, finance, and law are exposed; plumbing and physical care work are protected for much longer because robotics lags behind. A 100-engineer software project might need only 20 people supervising AI tools. He compared the current moment to March 2020... experts shouting about exponential curves while everyone else went about their day. And he added a counterintuitive observation: the general public, who sees an AI speaking 150 languages and correctly grasps the implications, may be closer to the truth than domain experts who fixate on what AI still gets wrong.