How does it compare with the Toy dialect tutorial?

I'm finding the Toy tutorial easier to read, and easier to use as a jumping off point into the MLIR reference docs. I thought the writing style in this tutorial was a bit dense, and Bazel seems like an unnecessary complication. Only an opinion though, YMMV.

kudos to jeremy for putting pen to paper but i do believe the toy tutorial is better (albeit more comprehensive). also i don't like bazel at all.

I tried to learn from the toy tutorial and just found it to jump into the deep end too quick, without enough explanation for me to start writing code (because everything I touched in `toy` caused the build to break in mysterious ways).

Same opinion. I have little/no exposure to compilers (yes I know, maybe I'm not the right target) and tried to follow Toy and had to resort to a lot of wikipedia/google just to understand the concepts and terms. And after finishing it I was still a bit confused about how to even get started. I think it could be made a bit more beginner friendly.

I found the standalone example https://github.com/llvm/llvm-project/tree/main/mlir/examples... in the MLIR repo to be very useful, on the other hand.

These two goals looks ambitious: (1) Design an architecture based on graph reduction (2) Design a memory hierarchy for non-strict semantics

At the same time one of its goal is to find the application best suited for such hardware, which sounds quite important

You might be interested in this: https://github.com/merrymercy/awesome-tensor-compilers

Any recommendation for learning how to build a diffusion model?

What are some example workloads in scientific computing that are not suitable for CUDA?

Anything that's not linear algebra

solving small and medium sized differential equations, most monte-carlo algorithms