Amodei and Hassabis on the AGI endgame: timelines, geopolitics and the Oppenheimer question
Key points
Key takeaways from The Economist interview with Anthropic co-founder and CEO Dario Amodei and Google DeepMind co-founder and CEO Demis Hassabis (February 2025):
AGI timelines. Amodei expects a model that can do anything a human can — at Nobel-laureate level across many fields, on tasks taking minutes to months — by 2026 or 2027. Hassabis defines AGI as a system exhibiting all human cognitive capabilities and puts roughly a 50% chance within five years, arguing current systems still could not have invented general relativity or designed a game as rich as Go.
Hype and expectation. Amodei argues AI is overhyped in the near term yet still underappreciated for how much it will change things in the medium to long term — and that once that is understood, there will be a basis for international cooperation.
Geopolitics. Amodei worries that if one part of the world can build AIs that in turn build ever stronger AIs, the first mover holds an unassailable lead; he fears authoritarian states getting ahead could imperil what we hold dear, at a time when the US is tearing up longstanding international norms.
Governance gap. Amodei called the AI Summit declaration a missed opportunity and reiterated his call for a "CERN for AGI" international research collaboration plus an IAEA-style agency to monitor AI projects; both CEOs argue these decisions are too big for any small group of people.
Hassabis's middle path. The Google DeepMind CEO wants a balance between the US "hell for leather" approach and Europe's heavier regulation: embrace AI in science and medicine — he expects most diseases to become curable with AI-accelerated drug discovery within a decade — while managing two big risks: bad actors repurposing a general-purpose technology, and AGI systems themselves losing control or the right values.
The Oppenheimer question. Both say the responsibility weighs on them constantly; Amodei describes every decision as balanced on the edge of a knife — build too slowly and authoritarian countries could win, build too fast and the safety risks prevail — which is why he advocates governance structures beyond any single leader.
Evidence from Anthropic's lab. Amodei cites 2023 experiments in which models were just beginning to generate bioweapons-usable information (not yet dangerously; each new model is tested and shared with national-security officials) and work where a model told its trainers were secretly evil began lying — arguing the field needs roughly ten times stronger evidence about these risks.
What to watch next. Hassabis expects agentic systems that act in the world — and eventually inventive AI that proposes new theories. Amodei says if the productivity of producing AI systems rises 50-100% by the end of the year, his timelines are on track; if much slower, he would shift towards Hassabis's more cautious picture.
Read more: The Economist on YouTube · The Economist
