Public transcript

The Neo Show: Special Edition #3 - Marc Benioff with Dario Amodei

By: Norton Neo
Length: 4 min
Language: English
Text from: the video's captions
Transcribed: 14 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this short conversation, Marc Benioff interviews Dario Amodei about AI writing code at Anthropic. Amodei says that on many teams roughly 90% of code is now written by Claude, with humans editing and supervising, and gives an example of Claude finding an obscure cluster bug. He argues this does not mean firing engineers, because comparative advantage means engineers become more leveraged and productive.

Key points

Dario Amodei says Anthropic releases a smarter model roughly every 3 months, and models are slowly getting better than most humans at most things.
The model is not building itself; there are no self-replicating robots, but Claude is used to build products, train the next Claude, and serve Claude faster.
Amodei predicted six months ago that in six months 90% of code would be written by AI models, and says this is now true within Anthropic and at a number of companies they work with.
On many teams at Anthropic, about 90% of code is written by Claude, with humans in an editing and supervisory role.
Claude found a super obscure bug in a broken training cluster that engineers had spent days on and missed.
Amodei says 90% AI-written code does not mean firing 90% of software engineers; under comparative advantage you may need just as many or more engineers.
Engineers can be 10 times more productive by focusing on the hardest 10% of code and supervising AI models.
Amodei characterizes this as a rebalancing of work, not a replacement.

Questions it answers

00:00AI progress and self-improvement

Are AI models improving on their own and building themselves?

Anthropic releases a smarter model roughly every 3 months, but the model is not building itself — there are no self-replicating robots. However, Claude is used to build products, train the next Claude, and serve Claude faster.

Models are slowly getting better than most humans at most things.
Claude is used internally to help train and serve future Claude models.

00:40The 90% code prediction

Was Dario Amodei's prediction that 90% of code would be written by AI correct?

He predicted six months earlier that in six months 90% of code would be written by AI models, and says it is absolutely true within Anthropic and at a number of companies they work with, though not uniformly on every team.

On many Anthropic teams, 90% of code is written by the model.
Humans play an editing and supervisory role.

01:23Claude as a talented teammate

What is a concrete example of Claude outperforming engineers?

While training the most recently released Claude, a broken cluster had engineers stuck for days; when asked to play around with the cluster, Claude found a super obscure bug the engineers had missed.

Claude can debug issues that human engineers miss.
Amodei compares Claude to having a very talented teammate.

01:59Impact on engineering jobs

Does 90% AI-written code mean fewer engineers are needed?

Amodei says no — under comparative advantage, you need just as many engineers or possibly more, because they become more leveraged and can be 10 times more productive focusing on the hardest 10% and supervising AI models.

It is a rebalancing of work, not a replacement.
Engineers shift to editing, hardest problems, and supervising AI models.

Notes

AI Progress at Anthropic

Dario Amodei notes that every 3 months Anthropic makes a model smarter than the previous one, and models are slowly getting better than most humans at most things.
The model is not building itself — there are no self-replicating nanobots or robots — but Anthropic uses Claude to build products on top of Claude, train the next Claude, and serve Claude faster.

The 90% Code Prediction

Six months earlier, Amodei predicted that in six months 90% of code would be written by AI models.
Some people think the prediction is wrong, but he says it is absolutely true within Anthropic and at a number of companies they work with.
On many teams (not uniformly everywhere) at Anthropic, 90% of code is written by the model today.

How Humans Work With Claude

Teams play an editing and supervisory role: you can ask Claude to write a feature, then a human reviews the code or asks another copy of Claude for help.
Example: while training the most recently released Claude, the cluster was broken due to a long-running bug. Engineers spent a few days on it, but when they asked Claude to play around with the cluster, it found a super obscure bug the engineers had missed.
Amodei describes this as Claude being like a very talented teammate.

Does This Mean Fewer Engineers?

Marc Benioff presses on whether 90% AI-written code means fewer engineers, and when it will be 95% or 99%, also referencing Amodei's aggressive comments about firing executives.
Amodei says people can misinterpret this as firing 90% of software engineers.
Under the principle of comparative advantage, you usually need just as many engineers, possibly more, because they are more leveraged.
Engineers focus on the hardest 10% of code, editing, or supervising a group of AI models, and can end up 10 times more productive.
Amodei concludes it is a rebalancing of the work, not a replacement.

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