Dan Kondratyuk

Today we are launching Dream Machine, our first AI model that generates cinematic and fluid videos from text instructions and images. I generated this 1-minute 60 fps video entirely from our model. Try Dream Machine → lumalabs.ai/dream-machine Join us → lumalabs.ai/join

Coding agents are great at writing new code, but pretty bad at deleting code. It's what inevitably leads to a lot of bloat over time. Deleting code is the halmark of a great senior engineer, i.e., one that can write the least amount of code to get the job done. In my mind that's what's missing to make them robust at building good software.

@alex_whedon Let me guess, you're already thinking of breaking the 100M token context barrier :)

Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. @subquadratic finds and focuses only on the ones that do. That's nearly 1,000x less compute and a new way for LLMs to scale.

@willccbb If you make compaction differentiable, maybe you can learn the optimal compaction strategy across most tasks you care about

why aren't more people studying self-compaction at artificially low context lengths. there's no reason you can't benchmaxx math RL with 4k tokens across many turns

@JiaweiYang118 Nice result. Makes me wonder what is the "ultimate form" of loss we should be optimizing

Two months ago, I vaguely posted a number: 0.9 FID, one-step, pixel space. Now it is 0.75, and can be even lower. Many wonder how. I thought it might end as a small FID prank: simple and deliberate. It started with one question: can FID be optimized directly, and what does it reveal? Introducing FD-loss.

@vai_viswanathan I like to think of world models as simulations of some environment, able to represent what comes next. Most commonly it's seen as something visual or tangible (video/3D/robotics/etc), but perhaps LLMs that simulate OS (e.g., shell envs) might also be considered as world models.

@hyperparticle thanks for the read! how do you define world models?

When I ask people what they mean when they are working on "World Models", I get a very different response every time. It's always fun trying to see all varied and different perspectives.

As with any new software, it's still going to have some rough edges. But I put a lot of checks/manual reviews in place to make sure the code quality is to a good standard: 90+% test coverage, fully typed, docstrings that explain intent, and lots of examples.

Technically it did take another half day to polish things up, but the core implementation was fully operational in just a few hours. Software package: github.com/rekursiv-ai/co… Blog post: rekursiv.ai/blog/i-built-c…

I wanted to speedrun how fast I could OSS a complete Python package that solves a non-trivial, important job, and I managed to pull it off in about a day. I've never felt so productive writing software, especially complete packages. The most joy I've felt in a long time.

The cost of software is starting to get really cheap. What took an entire dev team months can soon be accomplished with a single determined person. I suspect we're going to start seeing a proliferation of apps with weird and crazy ideas that wouldn't have been tried until now.

One thought that scares me a bit: the proliferation of "AI Viruses": tiny coding agents which can break into unsecured systems, replicate themselves, and adapt/mutate over time. And like real viruses, might hide, spread repeatedly like a botnet and impossible to fully eradicate.

Really excited to share our newest release on a unified model that does understanding and generation all in one. The team really did a great job here!

Our team has developed a new diffusion distillation technique which is overall much simpler and more robust than prior methods, and scales well to large model training. We make the code and paper freely available github.com/lumalabs/tvm

Me and the boys

Duck hunt

@gravicle Ray2 was able to animate some pretty great anime, and we're keeping the tradition! x.com/seiiiiiiiiiiru…

Ray2 was the best anime model. Ray3 is better.

It took an incredible amount of energy to get here, but now we're ready to unleash Ray3, our new frontier video model with reasoning capabilities. I especially love the HDR video generations, the colors and lighting just pop in ways that make SDR look dull. Check it out!

I see they went to the Intel/Nvidia school of cooking charts

For those that ever wondered how video generation works, this video is fantastic look into how these models operate from a geometric perspective

@Yiheng_Li_Cal Congrats Yiheng!

🎉Introducing Improved Immiscible Diffusion - Accelerating Diffusion Training by Reducing Its Miscibility. 🔥 Supported by detailed feature analysis, we further clarify that the miscibility problem, i.e. the mix of diffusion paths of different images during training, reduces the training efficiency. 🤔 Based on this, we design a new KNN implementation, which not only is efficient (unrelated to batch sizes) but also performs well in diverse baseline models, especially in flow matching. 🤩 We hope our miscibility problem lights the way for further improving diffusion training efficiency. ✈️ arxiv.org/abs/2505.18521