AI Briefing — March 30, 2026

🔬 Research

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

122 upvotes · Chen et al. · Paper

Current video world models break down when dynamic objects move out of frame—subjects freeze, distort, or vanish when they re-emerge. This work introduces Hybrid Memory, treating models as both archivists for static backgrounds and trackers for moving subjects. The researchers built HM-World, a 59K clip dataset with carefully designed exit-entry events, and propose HyDRA architecture that compresses memory into tokens with spatiotemporal retrieval. The approach significantly outperforms existing methods at maintaining object identity and motion consistency during occlusion.

PixelSmile: Toward Fine-Grained Facial Expression Editing

113 upvotes · Hua et al. · Paper

Fine-grained facial expression editing has been hampered by semantic overlap between expressions—happy and surprised faces share similar features. PixelSmile addresses this with a diffusion framework using symmetric joint training and contrastive learning to create more distinguishable expressions. The method enables precise linear control through textual latent interpolation, allowing smooth blending between expressions while preserving identity. The researchers also created the FFE dataset with continuous affective annotations to benchmark structural confusion and editing accuracy.

Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale

111 upvotes · Zou et al. · Paper

The first trillion-parameter scientific multimodal foundation model, combining general intelligence with specialized expertise across 100+ scientific tasks in chemistry, materials, life sciences, and earth sciences. Built with XTuner and LMDeploy infrastructure enabling efficient RL training at this massive scale while maintaining training-inference consistency. The model positions itself as a "Specializable Generalist," matching top open-source models for general capabilities while outperforming proprietary models on specialized scientific tasks.

🛠 Tools

obra/superpowers — An agentic skills framework & software development methodology that works

2,230 stars/day · Repo

A complete software development workflow for coding agents built on composable skills. Instead of jumping straight into code, agents first extract specs, show them in digestible chunks, then create implementation plans clear enough for junior developers to follow. The system emphasizes red/green TDD and launches "subagent-driven-development" where agents work through tasks autonomously for hours. With 124K total stars, it's available through Claude's official plugin marketplace and automatically triggers without special setup.

mvanhorn/last30days-skill — AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web

1,308 stars/day · Repo

Research assistant that synthesizes what's actually happening around any topic in the last 30 days across social platforms and aggregators. Version 2.9.5 adds Bluesky support, comparative mode for "X vs Y" analysis, and auto-saves complete briefings to ~/Documents/Last30Days/ as markdown files. The tool finds what communities are upvoting and sharing, then writes grounded narratives with real citations—keeping users current with rapid AI world changes.

luongnv89/claude-howto — A visual, example-driven guide to Claude Code

1,165 stars/day · Repo

Visual guide to Claude Code development from basic concepts to advanced agents, with copy-paste templates for immediate implementation. The repository provides practical examples and patterns for building with Claude's coding capabilities, targeting developers who want to quickly start building agents and applications. The focus on visual examples and ready-to-use templates addresses the common gap between Claude's documentation and practical implementation.

📡 Analysis

Pretext

March 29 · Post · tags: browsers, css, javascript, testing, react, typescript

Cheng Lou (ex-React core developer) released Pretext, a browser library that calculates line-wrapped text height without touching the DOM—typically an expensive operation. The library separates calculations into prepare() for measuring text segments and caching results, then layout() for emulating browser word-wrapping logic at specified widths. Remarkably, it was tested by rendering the entire Great Gatsby across multiple browsers and supports complex text including Korean mixed with RTL Arabic and platform-specific emojis.

Python Vulnerability Lookup

March 29 · Post · tags: tools, python, supply-chain, security

Simon built an HTML tool that queries OSV.dev's open CORS JSON API to check Python packages for known vulnerabilities. Users can paste pyproject.toml or requirements.txt files, or specify GitHub repositories, to get detailed vulnerability information including severity levels and disclosure reports. The tool demonstrates practical use of the OSV.dev open source vulnerability database for supply chain security checking directly in the browser.

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