they they stack together and the reasons why they ended up the way they are. Uh to go back to the big picture uh transformers are are are using like these four different building blocks. Uh one is multi-head self-attention. Uh attention is like what the difference between that makes it transformers different than other neural networks is that they can actually attend to previous tokens and understand the relationship between tokens that I was I was mentioning before and that's where attention comes in. And of course the bigger your attention is the more the model understands those relationships. And going back to what I said that's the what like big labs like Gemini they're trying to do 1 million context and they're finding ways because if you just try to use a 1 million context for the for a model like this it would just break very easily like the the math wouldn't work. So then that's when like the engineers from the researchers from Gemini found ways to to make it work. And that's what that that makes a difference but again fundamentally it's like the same the same architecture. Uh next one is the is the MLP