Corey Knowles

I think there's a lot of philosophical questions there, actually, around what kinds of tasks is this suitable for?

Because the reason why I think...

LLM's are so good at coding is because you can very test, you can very easily test, is it correct or not?

And everything that the LLM needs to code is within the system, if that makes sense.

I was talking to a journalist yesterday from the Wall Street Journal.

If you think about what a journalist does, it's pretty hard to get an LLM to do that because an LLM to do that job has to go talk to sources,

You know, it's kind of hard to automate that.

You can't have an LLM talk to sources.

Sources might not want to talk to them.

They might lie to them.

They have to, you know, decide how accurate something is.

So it gets pretty complicated.

I think there are a lot of tasks where it is scoped, basically just it is stuff on your computer that are inputs and outputs, kind of like coding.

So I think, you know, maybe to put it another way, I think LMs are actually quite smart and probably the total addressable market of things they could do, if it could bring something like a agentic system, like a cloud code or a cursor, to all of knowledge work, it's probably 10 or 100 times larger than just a cloud code or a cursor.

Yeah, a lot of those tools have to be custom tools.

So we launched Asian's product, I think, three, four months ago.

And that's kind of the idea is it's almost like build your own cloud code, if you will.

And the reason why cloud code is so effective is it basically runs in a loop and it says, I have tools to use.

And the LM decides what tools to use in order to accomplish this outcome.

So it has an input and it's like, I want to get here and have these tools and I'll figure it out.