FULL DISCUSSION: Google's Demis Hassabis, Anthropic's Dario Amodei Debate the World After AGI | AI1G
- By
- DRM News
- Length
- 31 min
- Language
- English
- Text from
- the video's captions
- Transcribed
- 12 October 2026
Summary
In this on-stage conversation moderated at an AI conference, Anthropic CEO Dario Amodei and Google DeepMind CEO Demis Hassabis debate timelines to AGI, the self-improvement loop of AI building AI, and the risks and opportunities that follow. Amodei sticks close to his aggressive one-to-two-year timeline for human-level AI, while Hassabis holds to a more cautious five-to-ten-year view, noting missing capabilities in scientific creativity. Both agree on the need for international coordination, chip export restrictions, and more time to get safety right.
Key points
- Dario Amodei stands by his prediction of a Nobel-laureate-level model by 2026-27 and says the self-improvement loop could close within a few years.
- Demis Hassabis keeps his roughly 50% by end-of-decade timeline, citing missing capabilities in scientific creativity and hypothesis generation.
- Anthropic's revenue has grown from 100 million dollars in 2023 to 10 billion in 2025.
- Amodei holds to his claim that half of entry-level white collar jobs could disappear within one to five years.
- Hassabis advises students to become highly proficient with AI tools as a better path than a traditional internship.
- Both CEOs call for not selling chips to China and international cooperation on minimum safety standards.
- Amodei's upcoming essay on risks is framed around humanity's 'technological adolescence' from Carl Sagan's Contact.
- Both agree the key thing to watch next year is AI systems building AI systems.
Questions it answers
00:42AGI timelines
How soon do the two CEOs expect AI to match human-level capability?
Amodei stands by his prediction of a Nobel-laureate-level model by 2026-27, saying the code-and-research self-improvement loop could close within a few years and probably faster than people imagine. Hassabis keeps his roughly 50% by end-of-decade estimate, arguing that verifiable fields like coding and math are easier than natural science, where experimental validation slows things down.
- Amodei says engineers at Anthropic no longer write code themselves.
- Hassabis says natural sciences are harder to automate because outputs are not easily verifiable.
- Hassabis sees missing capabilities in generating hypotheses and theories, not just solving existing problems.
04:18Competitive landscape
How have Google DeepMind and Anthropic fared over the past year?
Hassabis says Google DeepMind has regained the top of the leaderboards with Gemini 3 and growing Gemini app market share, after restoring a startup mentality. Amodei notes Anthropic's revenue has grown 10x in three years, from 100 million to 1 billion to 10 billion dollars, and argues researcher-led companies like both of theirs are best positioned to succeed.
- Anthropic revenue grew from 100 million in 2023 to 10 billion in 2025.
- Hassabis credits Gemini 3 and the Gemini app for Google DeepMind's comeback.
- Both CEOs say researcher-led companies focused on hard problems will succeed.
07:46Closing the loop
Will AI systems be able to improve themselves without humans in the loop?
Hassabis says full loop closure is unknown and may require AGI itself, especially in messy domains like natural science, though it can plausibly work in coding and mathematics. Amodei's mechanism relies on models good at coding and AI research accelerating the next generation of models, with chips and training time as limiting factors.
- Amodei says we might be six to 12 months from models doing most of what software engineers do.
- Hassabis says AGI may itself be needed to close the loop in messy, hard-to-verify domains.
- Hardware in the loop, including robotics, may limit the speed of self-improvement.
09:04Risks and the new essay
What risks does Dario Amodei cover in his upcoming essay?
Amodei says his new essay, written on vacation, addresses how humanity gets through its 'technological adolescence' without destroying itself, framed around a scene from Carl Sagan's Contact. He covers keeping autonomous superintelligent systems under control, misuse by individuals (bioterrorism) and nation states (the CCP), economic impacts like labor displacement, and unknown risks.
- Amodei frames the challenge using the 'technological adolescence' question from Contact.
- He lists bioterrorism, authoritarian nation states, and labor displacement as key risks.
- He says we should be devoting almost all our effort to getting through this period.
13:41Jobs and the labor market
Will AI eliminate entry-level white collar jobs in the next few years?
Amodei stands by his claim that half of entry-level white collar jobs could go in one to five years, seeing early signs in software and coding, including within Anthropic. Hassabis expects disruption to internships and junior roles this year but believes new, more valuable jobs will be created, and advises students to master AI tools as a better path than a traditional internship.
- Hassabis advises undergrads to get unbelievably proficient with AI tools.
- Amodei says Anthropic will need fewer junior and intermediate people over time.
- Both agree the exponential may overwhelm the labor market's ability to adapt.
18:39Society, backlash and geopolitics
How should governments and companies handle backlash and US-China competition?
Hassabis sees a real risk of popular backlash and says the industry must demonstrate unequivocal good like AlphaFold, while calling for international cooperation on minimum safety standards. Amodei argues not selling chips to China is the single biggest measure, dismissing the supply-chain logic by analogy to selling nuclear weapons because it profits Boeing.
- Hassabis says the industry should demonstrate more AlphaFold-like unequivocal goods.
- Both call for international cooperation on minimum safety standards.
- Amodei compares selling chips to China to selling nuclear weapons for Boeing's profit.
24:54Malign AI and safety
Are the CEOs more worried about all-powerful malign AI than a year ago?
Amodei says Anthropic has studied this risk since its founding via mechanistic interpretability, documenting bad behaviors as they emerge; he is skeptical of doomerism but sees real risk if racing removes guardrails. Hassabis says the risks were foreseen from DeepMind's start as dual-use technology, and technical safety is tractable given time, focus and collaboration, but fragmentation and racing make it harder.
- Amodei says mechanistic interpretability looks inside the model like neuroscientists study the brain.
- Hassabis says technical safety is tractable if there is time and collaboration.
- Both are skeptical of doomerism while taking the risks seriously.
28:09Fermi paradox and closing
What do the CEOs think of the Fermi paradox and what should we watch next year?
Hassabis rejects AI self-destruction as the Fermi paradox answer since we'd see alien AI structures like Dyson spheres, guessing the great filter was multicellular life and that humanity's future is ours to write. Both agree the biggest thing to watch next year is AI systems building AI systems, with Hassabis adding world models, continual learning and robotics as other frontiers.
- Hassabis guesses the great filter was multicellular life and we are past it.
- Both say AI systems building AI systems is the biggest thing to watch next year.
- Hassabis adds world models, continual learning and robotics as areas to watch.
Notes
AGI Timelines
- Dario Amodei stands by his prediction of a model at Nobel-laureate level across many fields by 2026-27, driven by a loop of AI writing code and doing AI research to build the next generation of models.
- He estimates we might be six to 12 months from models doing most or all of what software engineers do end to end, though chips, manufacturing and training time limit the loop.
- Demis Hassabis keeps his ~50% by end-of-decade estimate, noting coding and math are verifiable and easier to automate, while natural sciences require slow experimental validation.
- Hassabis sees missing capabilities in generating questions, theories and hypotheses — the highest level of scientific creativity — and says there may be one or two missing ingredients.
Competitive Landscape
- Hassabis says Google DeepMind has returned to the top of leaderboards with Gemini 3 and growing Gemini app market share, after restoring a startup mentality.
- Amodei cites Anthropic revenue growing 10x in three years: 100 million in 2023, 1 billion in 2024, 10 billion in 2025.
- Both argue researcher-led companies focused on hard scientific problems will succeed going forward.
Risks and the New Essay
- Amodei's upcoming essay on risks is framed by the Contact question: how did you get through your technological adolescence without destroying yourselves?
- He lists keeping autonomous superintelligent systems under control, bioterrorism, authoritarian nation states like the CCP, labor displacement, and unknown risks.
- Hassabis calls for more AlphaFold-like demonstrations of unequivocal good and international cooperation on minimum safety standards.
- Amodei argues not selling chips to China is the biggest single measure, comparing chip sales to selling nuclear weapons because it profits Boeing.
Jobs
- Amodei holds to half of entry-level white collar jobs gone in one to five years, seeing early signs in coding, including within Anthropic needing fewer junior and intermediate people.
- Hassabis expects internships and junior roles to be hit this year but says new, more valuable jobs will be created; he advises undergrads to get unbelievably proficient with AI tools.
- Both worry the exponential will overwhelm the labor market's ability to adapt.
Safety and Outlook
- Amodei highlights mechanistic interpretability — looking inside the model's brain — as Anthropic's approach to documenting and addressing bad behaviors.
- Hassabis says technical safety is tractable with time, focus and collaboration, but fragmentation and racing increase risk.
- Both agree the biggest thing to watch next year is AI systems building AI systems; Hassabis adds world models, continual learning and robotics.
Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.