Here are my young pawpaws, when they bloomed a few days ago! We live a bit north of Toronto. My sister who lives near Sherbrooke, Quebec, is also trying to grow some but that’s beyond the northern limit… we’ll see how that works out!

I would happily pay $10/month for an automatic algorithm like Twitter's so I can completely disconnect from X. Ideally with a few options, e.g. turn on/off politics temporarily.

- their forward diffusion process consists in masking tokens independently
- they train the model to gradually unmask tokens at arbitrary locations in the sequence

I find it counter-intuitive that this would work at all...

Project: ml-gsai.github.io/LLaDA-demo/

@apolinario.bsky.social mentioned this demo: huggingface.co/spaces/multi...
(video from Apolinario. It's great to understand how the token sequence is generated!)

LLaDA - Large Language Diffusion Models
arxiv.org/abs/2502.09992
Significant progress towards language diffusion models. Reportedly on par with LLaMA3 on many benchmarks.

I've long admired NVIDIA from afar. Now I'm excited to see what makes them successful from the inside; to meet new colleagues; and to make my own contribution! #NVIDIAlife

Unsurprisingly, they concluded that flow matching diffusion models (as of now) result in better quality.
It's a very long paper, but reasonably accessible to newcomers... if you're interested in that space, I recommend you start with reading their conclusion and discussion on page 59:

Despite having never been trained on unrealistic videos or anime, it can work surprisingly well on them: comfyanonymous.github.io/ComfyUI_exam...
comfyanonymous.github.io/ComfyUI_exam...

Another comparison. Very smooth, highly detailed reconstruction:

They are able to reconstruct fine details cleanly:

Impressive results: "ViewExtrapolator effectively refines the artifact-prone renderings (left side of arrows) of radiance fields or point clouds, to more realistic renderings with fewer artifacts (right side of arrows)"

Overview pipeline of SAMPart3D:

Looks like Bluesky just passed 20M users! bcounter.nat.vg congrats to that team for the nearly-flawless scaling and the moderation still holding up