Do you feel like Arnis will be this very own thing and then the modeling will be its own thing and then there's going to be some interaction between the two? Or at the limit, do you see it all being integrated into and to the model completely.

So I think it's really good motivation to think about AI systems cognitively in terms of capabilities, but you have to also understand that things that are very easy for computers are hard for humans and vice versa, right? So like this control and perception is very easy for humans, hard for agents.

Long-term planning, very easy for agents and computers and hard for humans at the same time.

That's why the agents are so much better at math than us.

With respect to how this goes in the future, The way to think about this is we want to be thinking about this from a very principled approach, right? First principles, how are we doing this? How do we build computers to begin with? And the LLM can be thought of as another processor, right? We had originally had like the terrain.

machine, which had this infinite ticker tape that we're working with.

And until we get to an actual infinite ticker tape or infinite context window, we need to be thinking about what came next after this development that led to the modern era of computers.

When I go back to it and I think about the design decision up the stack, it does make a lot of sense.