Excited to share MeshFlow — a new approach that can generate meshes with a fraction of seconds, while achieving state-of-the-art generation quality. Secret sources? Instead of autoregressive models, use equivariant flow-matching!

Everyone's asking what humans are still worth in the age of AI. To me, it's the part of us that AI keeps flattening - our individuality. Yet I feel I'm constantly being sold a choice: extend myself with AI, or preserving my unique thoughts, tastes and opinions. What if I didn't have to choose? There is a future where we get the magic of feeling totally understood by the technology that serves us, while staying ourselves. In this future, the model of us -- our "second selves" -- must be owned by us. That's the future we want to live in, and one we're building at Aesona. The open, top-of-mind questions right now: 1. What's the right representation of a person and how should it evolve over time? 2. How do we measure individual-level alignment? How should you interact with the model to close gaps in what it doesn't yet know about you? 3. What should your second self refuse to say about you? How should the user create the right "disclosure policy" and how should it be enforced? If this sounds like a future you want to build, my DMs are open!

We should all be asking who — or what — owns our minds. A model that understands what makes you you can be incredibly enabling and deeply dangerous. That's why we think a "second self" — a model whose only job is to understand you and make your life better — must be owned by you. A faithful second self, in both senses, is the foundation of personal AI. Our founding thesis, below.

Want a super-fast mesh generation model? ⚡ Don’t miss our MeshFlow talk! 📅 Tuesday 🕙 3D Generation session, starting around 10:45 AM 📍 Room 403B, LA Convention Center See you there! #SIGGRAPH #3DGeneration #MeshGeneration #SIGGRAPH2026 #DiffusionModel

What if attention wasn't about matching tokens, but operating in function space? Glad to share our #ICML2026 paper: 📄 Functional Attention: From Pairwise Affinities to Functional Correspondences w/ @Jiefang_Xiao @GaoMaolin @stevenygd Daniel Cremers 📄 xjffff.github.io/funcattn/

Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other approaches for a fraction of the cost. thinkingmachines.ai/blog/on-…

The latest research paper from @eyelinestudios, FlashDepth, has been accepted to the International Conference on Computer Vision (#ICCV2025). Our model produces accurate and high-resolution depth maps from streaming videos in real time and is completely built on open-source models and data. We hope it will be applied to various online applications, like robotics and on-set video composition. It has already been integrated into a few internal tools for visual effects and real-time depth estimation, segmentation, and matting tasks. Congrats to the team: @gene_ch0u, @wenqi_xian, @stevenygd, @mohsaied, @Bharathharihar3, @Jimantha, @realNingYu, @debfx ! All models and code have been released: GitHub: github.com/Eyeline-Research/… Project page: eyeline-research.github.io/F… Paper: arxiv.org/abs/2504.07093

📢 Excited to share our latest work on computing accurate differential operators for hybrid neural fields (like Instant NGP)! 🔗: justachetan.github.io/hnf-de… 🧵👇🏻 (1/n)

Real-time video generation is finally real — without sacrificing quality. Introducing Self-Forcing, a new paradigm for training autoregressive diffusion models. The key to high quality? Simulate the inference process during training by unrolling transformers with KV caching.

Long-Context State-Space Video World Models "we propose a novel architecture leveraging state-space models (SSMs) to extend temporal memory without compromising computational efficiency. "

Which multimodal LLM should you be using to edit graphics in Blender? Today, we’re releasing our #CVPR2025 Highlight🌟 work, #BlenderGym 🏋️‍♀️, the first agentic 3D graphics editing benchmark that will tell you exactly how multimodal LLMs compare in their Blender-editing skills. What'd we find? 🧵👇