AI Briefing — March 31, 2026

🔬 Research

Out of Sight but Not Out of Mind: Hybrid Memory for Dynamic Video World Models

138 upvotes · Kaijin Chen et al. · Paper

Current video world models fail when dynamic objects move out of view and later reappear, often producing frozen or distorted subjects. This work introduces Hybrid Memory, which requires models to act as both precise archivists for static backgrounds and vigilant trackers for dynamic subjects. The researchers built HM-World, a 59K-clip dataset with decoupled camera and subject trajectories, and propose HyDRA, a memory architecture that compresses memories into tokens with spatiotemporal retrieval. This addresses a fundamental limitation in video generation that's critical for realistic world simulation.

ShotStream: Streaming Multi-Shot Video Generation for Interactive Storytelling

111 upvotes · Yawen Luo et al. · Paper

Multi-shot video generation typically requires high latency and lacks interactivity, limiting real-time storytelling applications. ShotStream reformulates this as next-shot generation conditioned on historical context, enabling users to dynamically instruct ongoing narratives via streaming prompts. The system uses a dual-cache memory mechanism and achieves 16 FPS generation on a single GPU with sub-second latency. This breakthrough could enable real-time interactive video storytelling applications that were previously computationally prohibitive.

Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration

60 upvotes · Danil Tokhchukov et al. · Paper

This research reveals that adding a single learned scaling parameter to Diffusion Transformer blocks can significantly improve generative performance. Calibri frames DiT calibration as a black-box optimization problem solved with evolutionary algorithms, modifying just ~100 parameters. Despite its lightweight design, it consistently improves text-to-image model performance while reducing required inference steps. This finding suggests that current DiT architectures may be fundamentally miscalibrated, offering an easy path to better generation quality.

🛠 Tools

luongnv89/claude-howto — Visual guide to Claude Code development

4,232⭐/day · Repo

This comprehensive tutorial bridges the gap between Claude Code's basic documentation and practical implementation with visual guides, copy-paste templates, and a structured learning path. The project addresses a common problem where developers know individual features exist but struggle to combine them into useful workflows. With over 10K stars, it provides everything from basic commands to advanced agent orchestration, MCP servers, and hooks. It's positioned as a weekend course to master Claude Code's ecosystem.

microsoft/VibeVoice — Open-source voice AI platform

2,492⭐/day · Repo

Microsoft's frontier voice AI project combines both text-to-speech and automatic speech recognition capabilities, with models now integrated into Hugging Face Transformers. The platform has gained real-world adoption with third-party applications like Vibing, a voice-powered input method for macOS and Windows. With 31K total stars and active development, VibeVoice represents Microsoft's push into open-source voice technology, competing directly with proprietary solutions from other major tech companies.

NousResearch/hermes-agent — Self-improving AI agent with learning loops

1,851⭐/day · Repo

Hermes Agent differentiates itself by featuring a built-in learning loop that creates skills from experience and builds a deepening model of user preferences across sessions. It's designed to run anywhere from $5 VPS instances to GPU clusters, with multi-platform support including Telegram, Discord, and CLI interfaces. The agent supports any model provider without vendor lock-in and includes features like conversation search and persistent knowledge. With 19K stars, it represents a significant advancement in persistent, learning-capable AI agents.

📡 Analysis

Quoting Georgi Gerganov

March 30 · Post · tags: coding-agents, generative-ai, ai, local-llms, llms, georgi-gerganov

The creator of llama.cpp highlights a critical infrastructure problem: local model deployment involves a fragile chain of components developed by different parties, from client input to actual results. Issues often stem from harness problems, incorrect chat templates, prompt construction errors, and pure inference bugs. This fragmentation makes it difficult to consolidate the entire stack, meaning what appears to work is "with very high probability still broken in some subtle way." This insight explains why coding agents struggle with local models despite their apparent capabilities.

datasette-llm 0.1a3

March 30 · Post · tags: llm, datasette

This release adds the ability to configure which LLMs are available for specific purposes within Datasette applications. The functionality allows developers to create purpose-specific LLM configurations, enabling more targeted and controlled AI integration within data analysis workflows. This represents continued evolution of Datasette's ecosystem toward more sophisticated AI-powered data tools.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net