6 Comments

Edge-GNN is so cool.

Edges aggregate information from nodes.

The nodes then aggregate neighbouring nodes along with edge embeddings.

Domain knowledge resulting in minor changes to a model goes a long way indeed!

Expand full comment

yep

Expand full comment

Great content once again. Makes you get a glimpse into the chip wars that are about to heat up seriously

Expand full comment

Exciting stuff

Expand full comment

This is good. It seems to me the limiting variable for a company to commercialize this would be access to data and expertise at the intersection of AI/Chip Design. Any other bottlenecks you see?

Expand full comment

Customer base is a big one. Limited number of people you can serve (although this becomes a bit better when you remember that these customers are whales that will pay multiple firms for their research/testing).

Also- the overlooked player that makes all this happen is HPC, and that's not something many people can do well (it's a very new field). Scale is itself a huge player that can change equations (which is why Deep Learning fell in love with it). Spoiler alert- I'm working on something related to AI and HPC for chips soon.

Expand full comment