Azeem Azhar

So when baking is trivial,

The question is, what is still hard?

What is still valuable in that context?

So you can think back to the economist's framework, which is you look for the complement.

And one of the ways I now think about this and within this anchor is the idea of orchestration.

If you are not

building something directly yourself and it really feels to me that certainly for the class of code that we might have written and certainly within anthropic right the these these incredible developers of the class of code they are writing well you don't have to do that anymore so what are you doing and and the answer is that we are we're orchestrating um and as any

I'm just going to say this.

I realize I knew nothing about conducting except for watching that film, which had, I think, Cate Blanchett in it.

But, you know, when you're conducting an orchestra, you sort of need to know the capabilities of the people who you are conducting.

And one of the things I've noticed as I've started to move from sort of single agent coding to multi-agent systems and working parallel with 6, 8, 10 of these things at a time,

is that you do need to understand those systems.

And I've even discovered, and this is kind of curious, that the agents are starting to understand their own limitations.

I was building something that's a kind of internal research tool a few days ago that involves a lot of different agents.

And at one point, Claude said to me explicitly,

Don't use me for this.

Go and use chat GPT because it's tougher and more cynical than I am, which I thought was a really interesting bit of self-reflection.

It's quite a strange moment.

And that is evidence, I think, that we're not going to have that singleton AI, all-powerful AI.

It is that society of AI that I wrote about a couple of years ago.