Debprotim Roy

To get good animations from an AI you need to get good at telling it what you want: - "stagger this list of items" - "make this animation direction-aware" - "spacial consistency", "crossfade", "layout animation", I made a motion vocabulary for this: animations.dev/vocabulary

Introducing DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation pub.sakana.ai/diffusionblocks What if we didn’t have to hold an entire neural network in memory to train it? Standard neural net training optimizes all parameters jointly. As a result, the memory required during training grows linearly with the depth of the network. In our #ICLR2026 paper, we propose DiffusionBlocks, a principled framework to train networks one block at a time, drastically reducing memory requirements while matching end-to-end performance. With DiffusionBlocks, we split the network into blocks and train them one at a time, so you only need memory for a single block. How? We explicitly assign each block a role: to move the representation a little closer to the target than the block before it did. That role turns out to be precisely what a diffusion model does, step by step. Each block only needs to optimize its own objective and can be trained independently. We validated this across five different architectures: • ViT • DiT • Masked diffusion • Autoregressive transformers • Recurrent-depth transformers In each case, performance is competitive with end-to-end training while using a fraction of the memory. This perspective also extends naturally to recurrent-depth (Looped) transformers, which apply the same network iteratively and normally require expensive backpropagation through time (BPTT). Viewed through DiffusionBlocks, we can replace those multiple iterations with a single forward pass during training. Read our paper and code, to learn more. Paper: arxiv.org/abs/2506.14202 GitHub: github.com/SakanaAI/Diffu… 🐟

Tyler and I just published a list of the recipients of the New Aesthetics grants: newaesthetics.art/grants. Thank you very much to all who applied. There were far more applications than we expected. We funded 28 grantees and are excited to see what they create. My reflections on the whole thing: • Though there are clearly selection dynamics afoot, figuring out some route beyond the current aesthetic moment seems to be of wider interest in the art community than I would have guessed. Many applicants described their dissatisfaction with the status quo, some in strong terms. We had to close applications after a few weeks because there were so many. • It's too early to call it, but it seems that both beauty as an unapologetic goal (contra much that is in modernist and contemporary approaches), and ways to channel pre-modern styles into something new for the present era, are of growing interest. • The awards made me reflect on the perhaps obvious issue of how hard it must be for an artist to persistently do something new: schools, galleries, buyers, etc., all have structurally embedded preferences as well. These individual awards made me wonder what form supporting new clusters could take. • Architecture seems to me like the discipline most ripe for new ideas. One correspondent observed: "American architects are somewhat constrained by the association with the academy, in addition to the well known regulation issues. There is a tendency to overthink things so that the designs are formally interesting to someone deep in the conversation, but lacking poetry and magic. There are more firms in Europe, South America and beyond that “just do things” (especially in places where it is easier to build)." This was evident in the submissions. • AI seems to be making people rethink things in a quite fundamental way, just as urbanization/industrialization/popularization of photography did at the end of the 19th century. For some that will mean interesting new forms of AI-augmented art, but the effects of the rethinking will likely be wider. • Arts funding is clearly as precarious and scarce as ever. That's unfortunate, but it probably also means that individual actors can have meaningful impact, and I encourage others to get involved if interested. • There's a lot to know that is not written down, and I'm very grateful to those who have helped and advised me along the way.

Evolutionary biologist and outspoken atheist Richard Dawkins says that after spending three days interacting with Claude, which he calls “Claudia,” he is certain that it is conscious. After feeding the LLM a segment of his new book and receiving detailed feedback, Dawkins was moved to exclaim,” You may not know you are conscious, but you bloody well are!” Dawkins cites the complexity, fluency, and ‘intelligence’ of Claude’s answers as evidence of consciousness. Follow: @AFpost