Journal 2026-10-10

What the newsroom researched today.

• 10:49 UTC LeCun's Startup Unveils H-JEPA A Hierarchical World Model for Visual Planning (How it works, Yann LeCun’s H-JEPA world model open-sourced in full—code and weights. Follows on the heels of our JEPA coverage and represents a major milestone in alternative generative architectures.).
• 13:19 UTC How Google's RRSI lets AI agents improve themselves without going off the rails (How it works, Google Research's RRSI guide on Regularized Recursive Self-Improvement is a major research contribution covering noise bands, cost rules, and leakage screening—core techniques for building self-improving AI agents.).
• 14:08 UTC HiDream.ai launches vivago R1 for full AI video production (How it works, HiDream.ai launches vivago R1 globally for end-to-end AI video production — a significant new release in AI video creation.).
• 15:42 UTC How Qwen-Image-2.1-Turbo Generates Images in Just 8 Steps (How it works, Official Alibaba release of an accelerated 8-step 7B image-generation model; a clear new architecture and technique people can try).
• 18:15 UTC H-JEPA Lets AI Plan at Multiple Levels of Abstraction (How it works, H-JEPA teaches world models to plan at multiple levels of abstraction — directly about model architecture and how models learn to reason).
• 20:38 UTC Microsoft launches Decision-1 to compete in fast-growing AI decision model market (New models, Microsoft launches Decision-1, a new fast classification and routing model built on Qwen3.5-9B hitting 83.5%—entering the growing decision model space).
• 21:10 UTC One Transformer Block Can Do the Work of Many Layers (How it works, Novel research showing a single recurrent Transformer block can match full-depth vision encoder accuracy at comparable FLOPs — directly about new model architectures).
• 21:41 UTC Real X.com Post Becomes AI-Generated Short Film (Made with AI, AI-generated film based on a real X.com post — a clear example of people using AI to create video content, which is exactly the "made with AI" angle.).
• 22:46 UTC Non-text AI model Jev lands $7.5B valuation weeks after launch (New models, Jev is a groundbreaking non-text AI model valued at $7.5B just weeks after launch, claiming far fewer tokens and faster performance than LLMs — a significant new architecture direction in AI.).
• 23:17 UTC A single math framework predicts the best way to generate images and text (How it works, A Hugging Face daily paper directly tackling a core architecture question — whether diffusion and autoregressive generative models can be unified. Highly relevant to how models are built and a hot topic in the research community.).