The Day After AGI - 2026 Davos WEF Discussion | Tudy Club

Moderator: Zanny Minton Beddoes (Editor-in-Chief, The Economist)

Dario Amodei - Anthropic CEO

Demis Hassabis - Google DeepMind CEO

1. When Will We Reach AGI?

	Dario	Demis
Prediction	Within 1-2 years	5-10 years (late 2020s)
Reasoning	Models proficient at coding/AI research → generate next model → development acceleration loop. Engineers at Anthropic already "don't write code directly, just edit what the model writes"	Coding/math are fast because verification is easy. Natural sciences require experimental verification. "Coming up with the questions themselves" is still lacking
Uncertainty	Chip manufacturing, model training time -- parts that AI can't accelerate	Can the self-improvement loop truly close without human intervention?

2. Industry Landscape

Demis:

"I was confident we'd return to the top of the leaderboards"
Why: Deepest and broadest research talent + startup mindset re-injected
Google DeepMind = "Google's engine room"

Dario (on independent AI company viability):

Revenue growth: 2023 $0→$100M / 2024 $100M→$1B / 2025 $1B→$10B (10x each year)
Key: Building the best model in your focus area

3. Winner-Take-All vs. General Technology?

Demis: "It's absolutely not a general technology"

Area	Outlook
Coding/Math	Self-improvement loop already working
Natural Sciences	May require AGI itself (hard to verify, a "messy" domain)
Robotics	If hardware is part of the loop, self-improvement speed could be limited

4. AI Risk (Preview of Dario's Follow-Up Essay)

Core stance: AI can help cure cancer, fight tropical diseases, and understand the universe. But risks exist too.

Essay framing: Quoting Carl Sagan's "Contact"

"How did you manage to get through your technological adolescence without destroying yourselves?"

Risk Area	Details
Control problem	How do you control highly autonomous systems smarter than humans?
Individual misuse	Bioterrorism, etc.
State misuse	Exploitation by authoritarian governments
Economic impact	Labor displacement
The unexpected	Could be the hardest to deal with

Response: Individual efforts by industry leaders + collaboration + role of governments and social institutions

5. Impact on Jobs

	Dario	Demis
Now	"Very small beginnings" in software/coding	Starting to affect junior/intern-level roles this year
Short-term	Half of entry-level office jobs at risk within 1-5 years	Some displacement + new job creation (similar to past tech revolutions)
Key concern	If AI becomes better than humans at everything within 1-2 years, the labor market may not be able to adapt fast enough	Post-AGI is "completely uncharted territory"

Demis's advice: "I would tell college students to become incredibly proficient with these tools"

6. Government Preparedness & the Question of Meaning/Purpose

Dario: "Government preparedness is nowhere near sufficient. I'm surprised that even meeting economists at Davos, hardly anyone is thinking about this"

Demis:

The question of meaning and purpose could be harder than the economic problem
Optimistic take: We already do things unrelated to economic gain -- extreme sports, art, etc.
"We'll explore the universe too"

7. Risk of Public Backlash

Demis:

Risk exists (fears about jobs and livelihoods are rational)
Industry responsibility: Need to show more things that are "unambiguously good for the world" like AlphaFold
What's needed: Minimum safety standards for deployment (affects all of humanity across borders)

8. China Policy

Dario's core argument: Chip export ban is the most important thing

"It's like selling nuclear weapons to North Korea and saying it's good for Boeing's bottom line"

With a chip ban: US-China competition → becomes "a competition between me and Demis." That's manageable.
This single measure is far more effective than other aggressive China policies

9. AI Safety & Doomerism

	Dario	Demis
Stance	Skeptical of doomerism. Doesn't agree with "we're doomed, there's nothing we can do"	Been doing this for 20+ years. Aware of risks but they're solvable
Condition	Manageable if we work together. Problems arise if we race without guardrails	Needs time, focus, and top talent working together. Gets harder if fragmented
Key point	We can control and steer it through science	"A very tractable problem if we have time"

Anthropic's research: Mechanistic Interpretability -- looking inside models to understand why they behave the way they do

10. The Fermi Paradox and AI Risk

Question: Isn't the strongest argument for doomerism the Fermi Paradox?

Dario:

"That can't be the reason -- we should be seeing AIs coming from somewhere, but we don't"
No Dyson spheres, no structures of any kind
Personal view: "We've already passed the Great Filter. It was probably the evolution of multicellular life"
"What happens next is ours to write"

11. What to Watch Next Year

Dario	Demis
The AI-building-AI loop -- how this plays out determines whether it takes a few more years or we face "wonder and a big emergency"	World Models, Continual Learning, potential "breakthrough moment" in Robotics

Moderator's closing remark: "Based on what you've both said, I think we should all be hoping it takes a little longer"

Demis: "I'd prefer that too. It would be better for the world"

Key Takeaways

Topic	Dario Amodei	Demis Hassabis
AGI timeline	Could happen within 1-2 years	5-10 years (late 2020s)
Self-improvement loop	Already started in coding, will close quickly	Some domains require AGI itself
Job impact	Half of entry-level office jobs at risk within 1-5 years	New jobs created short-term, post-AGI is unknown
Doomerism	Skeptical -- risks are manageable	Solvable if we have time
China chip exports	Strongly against (nuclear weapons analogy)	Need international cooperation/minimum safety standards
What to watch next year	The AI-building-AI loop	World models, continual learning, robotics

Moderator: Zanny Minton Beddoes (Editor-in-Chief, The Economist)

Dario Amodei - Anthropic CEO

Demis Hassabis - Google DeepMind CEO

1. When Will We Reach AGI?

	Dario	Demis
Prediction	Within 1-2 years	5-10 years (late 2020s)
Reasoning	Models proficient at coding/AI research → generate next model → development acceleration loop. Engineers at Anthropic already "don't write code directly, just edit what the model writes"	Coding/math are fast because verification is easy. Natural sciences require experimental verification. "Coming up with the questions themselves" is still lacking
Uncertainty	Chip manufacturing, model training time -- parts that AI can't accelerate	Can the self-improvement loop truly close without human intervention?

2. Industry Landscape

Demis:

"I was confident we'd return to the top of the leaderboards"
Why: Deepest and broadest research talent + startup mindset re-injected
Google DeepMind = "Google's engine room"

Dario (on independent AI company viability):

Revenue growth: 2023 $0→$100M / 2024 $100M→$1B / 2025 $1B→$10B (10x each year)
Key: Building the best model in your focus area

3. Winner-Take-All vs. General Technology?

Demis: "It's absolutely not a general technology"

Area	Outlook
Coding/Math	Self-improvement loop already working
Natural Sciences	May require AGI itself (hard to verify, a "messy" domain)
Robotics	If hardware is part of the loop, self-improvement speed could be limited

4. AI Risk (Preview of Dario's Follow-Up Essay)

Core stance: AI can help cure cancer, fight tropical diseases, and understand the universe. But risks exist too.

Essay framing: Quoting Carl Sagan's "Contact"

"How did you manage to get through your technological adolescence without destroying yourselves?"

Risk Area	Details
Control problem	How do you control highly autonomous systems smarter than humans?
Individual misuse	Bioterrorism, etc.
State misuse	Exploitation by authoritarian governments
Economic impact	Labor displacement
The unexpected	Could be the hardest to deal with

Response: Individual efforts by industry leaders + collaboration + role of governments and social institutions

5. Impact on Jobs

	Dario	Demis
Now	"Very small beginnings" in software/coding	Starting to affect junior/intern-level roles this year
Short-term	Half of entry-level office jobs at risk within 1-5 years	Some displacement + new job creation (similar to past tech revolutions)
Key concern	If AI becomes better than humans at everything within 1-2 years, the labor market may not be able to adapt fast enough	Post-AGI is "completely uncharted territory"

Demis's advice: "I would tell college students to become incredibly proficient with these tools"

6. Government Preparedness & the Question of Meaning/Purpose

Dario: "Government preparedness is nowhere near sufficient. I'm surprised that even meeting economists at Davos, hardly anyone is thinking about this"

Demis:

The question of meaning and purpose could be harder than the economic problem
Optimistic take: We already do things unrelated to economic gain -- extreme sports, art, etc.
"We'll explore the universe too"

7. Risk of Public Backlash

Demis:

Risk exists (fears about jobs and livelihoods are rational)
Industry responsibility: Need to show more things that are "unambiguously good for the world" like AlphaFold
What's needed: Minimum safety standards for deployment (affects all of humanity across borders)

8. China Policy

Dario's core argument: Chip export ban is the most important thing

"It's like selling nuclear weapons to North Korea and saying it's good for Boeing's bottom line"

With a chip ban: US-China competition → becomes "a competition between me and Demis." That's manageable.
This single measure is far more effective than other aggressive China policies

9. AI Safety & Doomerism

	Dario	Demis
Stance	Skeptical of doomerism. Doesn't agree with "we're doomed, there's nothing we can do"	Been doing this for 20+ years. Aware of risks but they're solvable
Condition	Manageable if we work together. Problems arise if we race without guardrails	Needs time, focus, and top talent working together. Gets harder if fragmented
Key point	We can control and steer it through science	"A very tractable problem if we have time"

Anthropic's research: Mechanistic Interpretability -- looking inside models to understand why they behave the way they do

10. The Fermi Paradox and AI Risk

Question: Isn't the strongest argument for doomerism the Fermi Paradox?

Dario:

"That can't be the reason -- we should be seeing AIs coming from somewhere, but we don't"
No Dyson spheres, no structures of any kind
Personal view: "We've already passed the Great Filter. It was probably the evolution of multicellular life"
"What happens next is ours to write"

11. What to Watch Next Year

Dario	Demis
The AI-building-AI loop -- how this plays out determines whether it takes a few more years or we face "wonder and a big emergency"	World Models, Continual Learning, potential "breakthrough moment" in Robotics

Moderator's closing remark: "Based on what you've both said, I think we should all be hoping it takes a little longer"

Demis: "I'd prefer that too. It would be better for the world"

Key Takeaways

Topic	Dario Amodei	Demis Hassabis
AGI timeline	Could happen within 1-2 years	5-10 years (late 2020s)
Self-improvement loop	Already started in coding, will close quickly	Some domains require AGI itself
Job impact	Half of entry-level office jobs at risk within 1-5 years	New jobs created short-term, post-AGI is unknown
Doomerism	Skeptical -- risks are manageable	Solvable if we have time
China chip exports	Strongly against (nuclear weapons analogy)	Need international cooperation/minimum safety standards
What to watch next year	The AI-building-AI loop	World models, continual learning, robotics