Tristan Harris

Human beings are coding and doing the AI research.

They're reading the latest research papers.

They're writing the next, you know, they're hypothesizing what's the improvement we're going to make to AI.

What's a new way to do this code?

What's a new technique?

And then they use their human mind and they go invent something.

They run the experiment and they see if that improves the performance.

And that's how you go from, you know, GPT-4 to GPT-5 or something.

Imagine a world where Sam Altman, instead of having human AI researchers, can have AI AI researchers.

So now I just snap my fingers and I go from one AI that reads all the papers, writes all the code, creates the new experiments, to I can copy-paste 100 million AI researchers that are now doing that in an automated way.

And the belief is not just that, you know, the companies look like they're competing to release better chatbots for people, but what they're really competing for is to get to this milestone of being to automate an intelligence explosion or automate recursive self-improvement, which is basically automating AI research.

And that, by the way, is why all the companies are racing specifically to get good at programming.

Because the faster you can automate a human programmer...

the more you can automate AI research.

And just a couple of weeks ago, Cloud 4.5 was released and it can do 30 hours of uninterrupted complex programming tasks at the high end.

And it self-learns and recursively improves and invents.

So one thing to get is that

AI accelerates AI, right?

Like if I invent nuclear weapons, nuclear weapons don't invent better nuclear weapons.

But if I invent AI, AI is intelligence.