Public transcript

FULL: Demis Hassabis, Dario Amodei Debate What Comes After AGI at World Economic Forum | AI1G

By: DWS News
Length: 31 min
Language: English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

At a World Economic Forum session, Demis Hassabis and Dario Amodei discuss AGI timelines, the self-improvement loop of AI building AI, and the day after AGI. Both expect highly capable systems within a few years, with Amodei more aggressive on timelines than Hassabis. They cover labor market disruption, geopolitical competition with China, chip export controls, AI safety risks, and the need for international coordination and societal preparation.

Key points

Dario Amodei stands by his prediction of a model matching a Nobel laureate across many fields by 2026-27, saying it's hard to see it taking longer than a few years.
Demis Hassabis keeps his timeline of a 50% chance of human-level cognitive capabilities by the end of the decade, noting natural sciences are harder to automate than verifiable domains like coding and math.
Amodei says Anthropic's revenue grew 10x in three years, from 100 million in 2023 to 10 billion in 2025.
Both agree AI systems building AI systems is the key thing to watch, and that closing that loop could dramatically speed progress.
Amodei sticks with his claim that half of entry-level white collar jobs could be gone in one to five years, while Hassabis expects disruption at junior levels but sees new jobs and free creative tools compensating.
Amodei argues not selling chips to China is one of the biggest measures to buy time, comparing chip sales to selling nuclear weapons to North Korea.
Both are skeptical of doomerism but concerned about risks from racing without guardrails, misuse by individuals and nation states, and labor displacement.
Hassabis says the Fermi paradox can't be explained by civilizations destroyed by their own AI, since we should see evidence like Dyson spheres, and guesses we're past the great filter.
Both say they would prefer a slower timeline of five to ten years to get things right, but competition makes slowing down difficult.

Questions it answers

00:00AGI timelines

How soon do Amodei and Hassabis expect AGI-level systems?

Amodei stands by a Nobel-laureate-level model by 2026-27, driven by coding and AI research loops, and says it's hard to see it taking longer than a few years. Hassabis keeps his ~50% by end-of-decade estimate, noting natural sciences are harder to automate than verifiable domains like coding and math.

Amodei: models may do most SWE work end to end within six to twelve months.
Hassabis: missing capabilities include generating hypotheses and theories, not just solving existing problems.
Full closing of the self-improvement loop may itself require AGI in messier domains.

04:36The competitive landscape

How have Google DeepMind and Anthropic fared this past year?

Hassabis says DeepMind has regained focus and startup mentality, with progress in Gemini 3 and the Gemini app gaining market share. Amodei highlights Anthropic's revenue growing 10x in three years to 10 billion in 2025, and both agree researcher-led companies focused on hard scientific problems will succeed.

DeepMind is shipping models faster into Google's product surfaces.
Anthropic revenue: 100 million (2023), 1 billion (2024), 10 billion (2025).
Both see each other as research-led companies with science as a north star.

09:07Risks and technological adolescence

What risks does Amodei see on the road to and after AGI?

Amodei's upcoming essay frames the challenge via Contact's question of how a civilization survives technological adolescence. He lists controlling autonomous superintelligent systems, bioterrorism, nation-state misuse, labor displacement, and unanticipated consequences, urging near-total effort on getting through safely.

He is not a doomer but feels great urgency about the pace.
Solutions mix individual company actions, industry cooperation, and government roles.

13:40Jobs and the labor market

Will AI displace white collar jobs soon?

Amodei sticks with half of entry-level white collar jobs gone in one to five years, seeing early signs in software and within Anthropic itself. Hassabis expects junior-level disruption this year but believes new, more meaningful jobs and free creative tools can more than compensate, while admitting post-AGI is uncharted territory.

Hassabis advises students to become highly proficient with AI tools, better than a traditional internship.
Amodei worries the exponential will overwhelm labor market adaptability within one to five years.
Both agree current unemployment is mostly post-pandemic overhiring correction, not yet AI-driven.

17:37Government readiness and backlash

Are governments prepared for AI's economic impact?

Hassabis says nowhere near enough work is being done, including by professional economists, on distributing new productivity and wealth fairly. He sees backlash risk as reasonable given fears over jobs, and says the industry should demonstrate more unequivocal goods like AlphaFold.

Post-scarcity distribution may lack the right institutions.
Meaning and purpose after AGI are bigger questions than economics.
Hassabis would prefer a slightly slower pace to get things right socially.

21:27Geopolitics and chip exports

What should be done about China and chip sales?

Amodei argues not selling chips to China is one of the biggest measures to buy time, likening chip sales to selling nuclear weapons to North Korea for Boeing's profit. He says if chips aren't sold, the competition is between him and Demis, which they can work out.

He wishes for Demis' five-to-ten-year timeline but says adversaries building the same tech prevent slowing down.
An enforceable mutual slowdown agreement is very hard.

24:56AI safety and malign AI

How do they view the risk of deceptive, malign AI?

Amodei says Anthropic has studied this from the start via mechanistic interpretability, documenting bad behaviors as they emerge and working to address them. Both are skeptical of doomerism but say racing without guardrails creates real risk that science and cooperation can address.

Hassabis sees technical safety as tractable given time, focus, and collaboration.
Fragmented racing projects make safety much harder.

28:10Fermi paradox and what to watch

Does the Fermi paradox imply AI doom, and what matters next year?

Hassabis says AI self-destruction can't explain the Fermi paradox since we'd see evidence like Dyson spheres or paperclips; he guesses we're past the great filter, likely at multicellular life. Both agree the biggest thing to watch is AI systems building AI systems, with Hassabis adding world models, continual learning, and a possible robotics breakout.

Both would prefer a slower timeline of five to ten years for the world's sake.
Hassabis: what happens next is for humanity to write.

Notes

AGI Timelines

Dario Amodei reaffirms his prediction of a model doing everything a human could at the level of a Nobel laureate across many fields by 2026-27, driven by models good at coding and AI research creating a speed-up loop.
He says engineers at Anthropic already let the model write most code, and we may be six to twelve months from models doing most of what SWEs do end to end.
Demis Hassabis stays with his ~50% chance of human-level cognitive capabilities by the end of the decade, noting coding and math are easier to automate because outputs are verifiable, while natural sciences need experimental testing.
Hassabis says missing capabilities include coming up with the question, theory or hypothesis in the first place, the highest level of scientific creativity.

Business and Competition

Hassabis says Google DeepMind has regained intensity and startup mentality, with progress in Gemini 3 and the Gemini app gaining market share.
Amodei notes Anthropic's revenue grew 10x in three years: 100 million in 2023, 1 billion in 2024, 10 billion in 2025.
Both agree companies led by researchers focused on hard scientific problems will succeed going forward.

Risks and the Day After AGI

Amodei's upcoming risk essay is framed around the Contact question: how did you get through technological adolescence without destroying yourselves?
He lists risks: controlling highly autonomous systems smarter than humans, bioterrorism, nation-state misuse (CCP), labor displacement, and the unanticipated.
On jobs, Amodei sticks with half of entry-level white collar jobs gone in one to five years; Hassabis expects junior-level disruption but sees new, more meaningful jobs and free creative tools as compensation.
Both worry about meaning and purpose after AGI, though Hassabis is optimistic humanity will find new answers, including exploring the stars.

Geopolitics and Safety

Amodei argues not selling chips to China is one of the biggest measures to buy time, comparing it to not selling nuclear weapons to North Korea.
Hassabis calls for international cooperation on minimum safety standards and says the industry should demonstrate more unequivocal goods like AlphaFold.
Both are skeptical of doomerism but say racing without guardrails creates real risk; mechanistic interpretability is a key tool.
On the Fermi paradox, Hassabis says AI self-destruction can't be the explanation since we'd see evidence like Dyson spheres, and guesses we're past the great filter.

What to Watch

Both agree the biggest thing to watch next year is AI systems building AI systems; Hassabis adds world models, continual learning, and a possible robotics breakout moment.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.