they released a research post on how the new model's VAE was trained here: https://bfl.ai/research/representation-comparison

Surprised there wasn't any mention of Equilibrium Matching [1] in the future work section

[1] https://raywang4.github.io/equilibrium_matching/

great this is more on the techincal details. it is great but would be great to see the data. I know they will not expose such information but would be great to have a visibility onto the datasets and how the data was sourced.

this is a really neat project: "an automated, daily evaluation suite to track model performance over time, monitor for regression during peak load periods, and detect quality changes across flagship LLM APIs."

Technical report here for those curious: https://cdn.sanity.io/files/gsvmb6gz/production/880b07220899...

Seems implementation is straightforward (very similar to everyone else, HiDream-E1, ICEdit, DreamO etc.), the magic is on data curation (which details are lightly shared).

I haven't been following image generation models closely, at a high level is this new Flux model still diffusion based, or have they moved to block autoregressive (possibly with diffusion for upscaling) similar to 4o?

Well it's a "generative flow matching model"

That's not the same as a diffusion model.

Here is a post about the difference that seems right at first glance: https://diffusionflow.github.io/

Diffusion based. There is no point to move to auto-regressive if you are not also training a multimodality LLM, which these companies are not doing that.

"FLUX.1 Kontext [dev]

Open-weights, distilled variant of Kontext, our most advanced generative image editing model. Coming soon" is what they say on https://bfl.ai/models/flux-kontext

Distilled is a real downer, but I guess those AI startup CEOs still gotta eat.

The open community has a done a lot with the open-weights distilled models from Black Forest Labs already, one of the more radical being Chroma: https://huggingface.co/lodestones/Chroma

I don't doubt that people can do nice things with them. But imagine what they could do with the actual model.

I agree that gooning crew drives a lot of open model downloads.

On HN, generally, people are more into technical discussion and/or productizing this stuff. Here, it seems declasse to mention the gooner angle, it's usually euphemized as intense reactions about refusing to download it involving the words "censor"

Love this. Easy to Advanced, with 5 for extra credit. Thank you

thanks for noticing! this is the first time we're expanding it from 'security at scale' to 'infra at scale', but we've taught this course 2 yrs in a row now

curious to learn how many undergrads took this?

i've followed Rachel and Julia for a long time, but didn't know about Kellan - thanks so much for that.

re: human org scaling - true and this was the most surprising thing for me when i was running the platform org at discord. companies ship their org charts whether they like it or not. and refactoring org charts correctly, at scale, is essentially untested in the modern era

this is exactly the type of pointer i was hoping for, thank you

convex is really elegant and now that you mention it, multiplayer games like their ai-town agent sim is such a great fit for the class - thank you

this is an awesome rec thank you

Blog post: https://ai.meta.com/blog/fair-news-segment-anything-2-1-meta...