Public transcript

Dario Amodei Interview Impact of AI

By: Mike Evans AI Disruption Keynote Speaker
Length: 22 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Dario Amodei, CEO of Anthropic, sits down for a wide-ranging interview covering AI writing most of Anthropic's code, the role of humans as supervisors, AI in healthcare, and Anthropic's five-year vision. He argues AI is a rebalancing rather than a replacement of human work, driven by comparative advantage, while warning of labor disruption in the next two to five years. He also discusses Anthropic's revenue growth, enterprise ambitions, and the enormous compute and data center needs ahead.

Key points

Amodei predicted six months ago that 90% of code would be written by AI models, and says that on many teams at Anthropic and partner companies this is now true.
He argues that Claude writing 90% of code does not mean firing 90% of engineers; comparative advantage means humans are leveraged on the hardest 10% and supervision.
He worries about labor disruption across the economy in the next two to five years, faster and broader than previous technology transitions.
Claude once found an obscure cluster bug that engineers had missed after days of work, acting like a talented teammate.
In healthcare, Amodei sees models as a 'second opinion' rather than something to trust blindly, citing his sister's pregnancy infection case where Claude listed possibilities doctors had missed.
He says radiology struggles reflect specialized data needs and that human-model collaboration models take longer to work out in high-stakes regulated industries.
Anthropic's revenue has grown roughly 10x per year for three years and is approaching a $7 billion run rate.
He expects future model training to require millions of chips and many gigawatts of power, and says some data center deals in the press may involve double counting.

Questions it answers

00:00AI writing Anthropic's code

How much of Anthropic's code is written by AI models?

Amodei predicted six months earlier that 90% of code would be written by AI, and says that on many teams at Anthropic and companies they work with, this is now true. Humans play editing and supervisory roles, and Claude once found an obscure cluster bug engineers had missed after days of work.

90% of code on many Anthropic teams is written by Claude.
Humans supervise, edit, and handle the hardest 10%.
Claude acted like a talented teammate in finding a long-running bug.

02:21Rebalancing vs replacement

Does AI writing 90% of code mean fewer engineers?

No — under comparative advantage, you may need just as many or more engineers, leveraged on the hardest 10% and supervision, making teams up to 10 times more productive. Amodei calls it a rebalancing, not a replacement, though at 95-99% automation 'all bets may be off.'

Comparative advantage means humans complement AI rather than being replaced.
Amodei worries about labor disruption across the economy in 2-5 years.
He is optimistic about complementarity in the short run.

06:12AI in healthcare and radiology

How reliable are AI models in healthcare today?

Amodei says models should be used like a second opinion, not trusted blindly. He tells of his sister uploading her pregnancy infection data to Claude, which listed possibilities (a bacterial infection) doctors had missed, but blind reliance would not have gone well. Radiology struggles may reflect lacking the right specialized data, and human-model collaboration takes longer to work out in high-stakes regulated industries.

Models are assistive and augmentative, especially with words.
Radiology accuracy issues may stem from missing specialized data.
Practical collaboration models in healthcare work out much slower than in code.

09:30Benchmarks and the coding ecosystem

How good are the models at coding, and what has been built on top of them?

Benchmarks for writing GitHub PRs from a description have gone from about 5% to roughly 77% in the last year and a half as of Sonnet 4.5. A whole ecosystem has emerged — GitHub Copilot, Cursor, Lovable, Replit, Vercel, Augment Code, WindSurf, and Claude Code — and Amodei least expected the success of products letting non-coders write code end-to-end.

PR-writing benchmarks rose from ~5% to ~77% in 18 months.
It is too early to say if there will be a single winning tool.
Non-coders writing code end-to-end was the biggest surprise.

14:15Anthropic's five-year vision

Where does Amodei want Anthropic to be by 2030?

Drawing on his essay 'Machines of Loving Grace,' he is most excited about AI curing diseases like cancer and Alzheimer's, and sees Anthropic putting AI at the center of every enterprise, which he describes as a kind of super intelligence. He is also excited about financial services and insurance, and wants Anthropic known as trustworthy and responsible.

AI is 'the right thing' to cure tough diseases like cancer and Alzheimer's.
The vision is AI at the center of every enterprise.
Trustworthiness — 'saying what it does, doing what it says' — is core to the company.

18:00Revenue growth

How fast is Anthropic's revenue growing?

Revenue has gone up about 10x per year for the last three years and is approaching a $7 billion run rate, which Amodei confirms. He says it cannot continue at that rate forever but the market potential of serving every enterprise is very high.

Revenue is approaching a $7 billion run rate.
It has grown roughly 10x per year for three years.

19:30Data centers and compute

How does Anthropic think about data centers and future compute needs?

Scaling laws and rising revenue mean an enormous need for compute — future training may require millions of chips and many gigawatts of power — and Anthropic is securing power wherever it can. Amodei worries some data center deals in the press may involve double counting, and stresses that making money through revenue matters more than spending on data centers.

Future model training may need millions of chips and many gigawatts of power.
Some data center deals may involve double counting.
If demand and revenue are there, compute will follow.

Notes

AI Writing Code at Anthropic

Amodei predicted six months prior that 90% of code would be written by AI models; on many teams at Anthropic and partner companies this is now true.
Humans play an editing and supervisory role; Claude can write features end-to-end while humans review the code or ask another model for help.
Claude found a super obscure bug in a training cluster that engineers had spent days on and missed.

Rebalancing, Not Replacement

Under comparative advantage, Claude writing 90% of code means you need just as many or more software engineers, more leveraged on the hardest 10%.
Teams can become roughly 10 times more productive; it is a rebalancing, not a replacement.
As automation approaches 95-99%, "all bets may be off," but Amodei is optimistic about complementarity in the short run and worries about labor disruption in 2-5 years.

Healthcare and Radiology

Models should be treated like a second opinion, not trusted blindly.
His sister, while pregnant with a suppressed immune system, uploaded her data to Claude, which systematically listed possibilities (a bacterial infection) that doctors had missed.
Radiology struggles may reflect lacking the right specialized data; human-model collaboration models take longer to work out in high-stakes regulated industries like healthcare.

Coding Benchmarks and Ecosystem

Benchmarks for writing GitHub PRs from a description have gone from about 5% to roughly 77% over the last year and a half (as of Sonnet 4.5).
A broad ecosystem has emerged: GitHub Copilot, Cursor, Lovable, Replit, Vercel, Augment Code, WindSurf, and command line tools like Claude Code.
The development he least expected: products letting non-coders write code end-to-end.

Vision and Business

His essay "Machines of Loving Grace" frames the upside; he is most excited about AI curing diseases like cancer, Alzheimer's, and metabolic disease.
He sees an enterprise as a kind of super intelligence, and Anthropic's vision as putting AI at the center of every enterprise.
Revenue has grown about 10x per year for three years, approaching a $7 billion run rate.
He wants Anthropic known as trustworthy and responsible: "saying what it does, doing what it says."

Compute and Data Centers

Scaling laws imply ever more compute; future training may need millions of chips and many gigawatts of power.
Anthropic is securing power wherever it can; he is confident in the company's ability to do so.
He worries some data center deals reported in the press may involve double counting, and stresses that revenue and making money matter more than spending on data centers.

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