Public transcript

Dario Amodei at Davos More jobs in the physical world and less jobs in the knowledge work economy

By: Mike Evans AI Disruption Keynote Speaker
Length: 33 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

At Davos, Dario Amodei of Anthropic argues that AI capabilities have advanced on a smooth exponential curve while public opinion oscillates, and that society is not prepared for a world of high GDP growth alongside high unemployment and inequality. He outlines Anthropic's responses, including the Economic Index, enterprise focus, and interpretability research, and discusses competition with Google, IPO speculation, education, and the risk of AI empowering autocracies. He warns that governments will need to intervene to distribute the gains of growth widely.

Key points

Amodei believes AI is on a smooth exponential trajectory—a 'Moore's law for intelligence'—while public perception oscillates between hype and doom.
He predicts a novel combination of very high GDP growth (5-10%) with potentially high unemployment and inequality, which has almost never occurred before.
Anthropic's Economic Index tracks in real time what Claude is used for, whether it automates or augments tasks, and how it diffuses across industries and regions.
He expects more jobs in the physical world and fewer in knowledge work, with robotics progressing on a slower trajectory.
Anthropic's enterprise focus avoids the engagement-maximizing incentives of consumer AI, which he says leads to slop.
Revenue grew from roughly $100 million to $1 billion to $10 billion across 2023, 2024 and 2025, and Claude Code and Claude co-work are driving a breakout adoption moment.
He sees mechanistic interpretability—looking inside models—as the key missing breakthrough for making AI reliably safe and controllable.
He fears AI is uniquely suited to autocracy through surveillance, individualized propaganda, and drone armies, and wants targeted policies like not selling chips to adversaries.
His nightmare scenario is a decoupled 'zeroth world' of about 10 million people centered on Silicon Valley capturing outsized growth.

Questions it answers

00:00Preparedness for AI's impact

Are businesses and governments doing enough to prepare for AI's impact?

No. Amodei says the technology has progressed on a smooth exponential while public opinion oscillates, and there is almost no awareness of the magnitude of what is coming economically.

AI capability follows a 'Moore's law for intelligence.'
Public sentiment swings between hype and doom every few months.
Anthropic aims for constancy, holding both upside and downside views.

04:00High growth, high unemployment

What economic future does Amodei predict?

A novel combination of very high GDP growth (5-10%) with high unemployment and inequality. AI coding shows the microcosm: engineers now just edit Opus output, and software may become essentially free.

This growth-unemployment combination has almost never occurred before.
Whole careers built for decades may disappear.
Software amortization across millions of users may become a false premise.

07:00How society can adapt

How might society organize itself for this transition?

Three steps: measure the transition (the Anthropic Economic Index), teach people to adapt (likely toward physical-world jobs), and accept that government involvement in displacement is unavoidable because the issue is distributing a much larger pie.

Policy without measurement is blind and misinformed.
More jobs in the physical world, fewer in knowledge work.
Worry less about disincentivizing growth, more about sharing it.

11:00Politics and Anthropic's principles

Has Anthropic compromised its safety principles under competitive pressure?

Amodei says no, crediting the enterprise business model for avoiding consumer AI's engagement incentives. Anthropic runs tests others don't and pioneered mechanistic interpretability. On China, his concern is autocracy using AI for surveillance, propaganda, and drones.

Enterprise focus means less need to fight business incentives.
Anthropic tests for deception, blackmail, and sycophancy and reports publicly.
Targeted policies like not selling chips to autocracies are underemphasized.

16:30Claude's breakout moment

How is Anthropic's business doing and why is Claude having a moment?

Revenue grew roughly from $100M to $1B to $10B over 2023-2025. Claude Code reached an inflection with Opus 4.5, and Claude co-work—built in about two weeks with Claude Opus to serve non-technical users—saw metrics four times higher than any prior release.

Growth follows the same smooth exponential as the technology.
Non-technical users wrestling with the command line revealed unmet demand.
An IPO is uncertain; the focus is the revenue curve and capital demands.

21:00Competition and multimodality

Does Amodei worry about competing with Google's Gemini?

Not much, because Google and OpenAI are fighting existentially in consumer while Anthropic is differentiated in enterprise. He also doesn't see the lack of image/video generation as a weakness for enterprise, and is wary of the slop associated with short-form video.

Consumer AI is existential for Google and OpenAI, not Anthropic.
Enterprise has little demand for consumer-style media generation.
Anthropic could contract an external model if ever needed.

23:40Scientists vs. entrepreneurs

How do scientist-led AI companies differ from entrepreneur-led ones?

Amodei says scientists, like himself and Demis Hassabis, come from a tradition of taking responsibility for the effects of their technology, whereas the social-media entrepreneur generation was selected for manipulating consumers, leading to different attitudes.

AI sits at the intersection of decades of academic research and big-tech scale.
Scientists are motivated by creating something for the world.

26:30Reader and audience questions

What did Amodei say about safety breakthroughs, education, and inequality?

Mechanistic interpretability is the key missing safety breakthrough, as it's the only ground truth about what models think. Education should shift from a mercenary notion toward character-building. On being left behind, he fears a decoupled 'zeroth world' of ~10 million people around Silicon Valley, and says government involvement will become bipartisan necessity.

Interpretability is like an MRI for models—surface behavior can deceive.
He can't yet tell people which careers to pursue.
Anthropic works on public health, education, and the Gates Foundation for developing-world growth.

Notes

The Trajectory of AI and Public Perception

Amodei has watched the field for 15 years and worked in it for 10; the technology's actual progress has been a smooth exponential, a "Moore's law for intelligence," while public opinion oscillates every 3-6 months between hype and collapse narratives.
Sentiment on whether AI is good or bad has also swung—from AI risk concerns in 2023-2024 to AI opportunity in 2025 and back again. Anthropic tries to maintain constancy, holding both radical upside (his essay Machines of Loving Grace) and serious downsides.

The Economic Signature of AI

He predicts a world of very high GDP growth (5-10%) alongside potentially 10% unemployment and high inequality—a combination almost never seen historically.
AI coding is a microcosm: engineering leads at Anthropic say they no longer write code, just edit Opus output. Claude co-work was built in a week and a half almost entirely with Claude Opus. Software may become essentially free, undermining the premise of amortizing software across millions of users.

Preparing Society

Step one: the Anthropic Economic Index, updated four or five times, which measures what Claude is used for, automation vs. augmentation, industries, and geographic diffusion—because policy without measurement is blind.
Step two: helping people adapt, likely toward more physical-world jobs and fewer knowledge-work jobs; robotics is on a slower trajectory. Companies must rethink what moats mean when software becomes cheap.
Step three: government involvement in displacement is unavoidable; the pie will grow, and the issue is distribution. He argues for worrying less about disincentivizing growth and more about sharing it.

Anthropic's Position

Enterprise focus avoids consumer AI's engagement-maximizing incentives, which he associates with slop and ads.
Revenue roughly went $100M (2023) → $1B (2024) → $10B (2025). Claude Code and Claude co-work represent a breakout adoption moment.
He doesn't see lack of video/image generation as a weakness for enterprise; Anthropic could contract a model if needed.
On China, his concern is autocracy, not the country or its people: AI could enable surveillance, individualized propaganda, and drone armies. Targeted policies like not selling chips matter.

Safety and Society

Mechanistic interpretability—looking inside models, like an MRI for AI—is the key missing breakthrough for safety, since surface behavior may not reflect what a model is actually thinking.
On education, he suggests moving away from a mercenary notion of education toward one that builds character and enriches people.
His nightmare is a decoupled "zeroth world" of ~10 million people around Silicon Valley capturing the growth; Anthropic works on public health, education (Ministry of Education, Gates Foundation), and economic opportunity to counter this.

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