Thanks for sharing. Excited to hear what you think and any feedback!

I agree that embeddings (especially ones reduced to 3D) are still pretty coarse, but I think it'll inspire some serendipitious exploration for some folks.

While I agree that cherry-picking is rampant, they're still quite impressive when you get a chance to use them. I've been using Veo 2 and it's blown me away.

"Documentary about a puppy who streams himself playing video games to the internet. Puppy is wearing headphones and playing a racing game."

all I see is a sharp decline in loss :)

I wish there were more of these implementation detail posts. Sometimes you really have to dig into a paper's codebase to see what tricks made their method work.