AI Briefing — March 29, 2026

🔬 Research

PixelSmile: Toward Fine-Grained Facial Expression Editing

106 upvotes · Jiabin Hua, Hengyuan Xu, Aojie Li, Wei Cheng, Gang Yu · Paper

A diffusion framework that achieves precise facial expression editing through symmetric joint training and contrastive learning. The researchers created the FFE dataset with continuous affective annotations to address semantic overlap issues in expression editing. PixelSmile enables linear expression control via textual latent interpolation while preserving identity, supporting smooth expression blending for controllable and fine-grained facial animation.

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

104 upvotes · Yicheng Zou, Dongsheng Zhu, Lin Zhu, Tong Zhu, Yunhua Zhou · Paper

The first one-trillion-parameter scientific multimodal foundation model that masters over 100 specialized tasks across chemistry, materials, life sciences, and earth sciences. Built with XTuner and LMDeploy infrastructure for efficient RL training at trillion-parameter scale, it combines general reasoning capabilities with deep scientific expertise. The model positions itself as a "Specializable Generalist," outperforming proprietary models in specialized scientific tasks while maintaining top-tier general capabilities.

CUA-Suite: Massive Human-annotated Video Demonstrations for Computer-Use Agents

86 upvotes · Xiangru Jian, Shravan Nayak, Kevin Qinghong Lin, Aarash Feizi, Kaixin Li · Paper

A large-scale ecosystem addressing the critical bottleneck in computer-use agents: continuous video demonstrations rather than sparse screenshots. VideoCUA provides 55 hours of 30 fps screen recordings across 87 applications with kinematic cursor traces and multi-layered reasoning annotations, totaling 6 million frames. Preliminary evaluation shows current foundation action models struggle with professional desktop applications (~60% failure rate), highlighting the need for continuous temporal dynamics in agent training.

🛠 Tools

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

2,292 stars/day · Repo

A complete software development workflow for coding agents built on composable "skills" that enforces proper methodology from spec to implementation. The framework makes agents step back to understand requirements, create digestible specs, and follow TDD practices through subagent-driven development. With 121K stars and active development, it's becoming a standard for structured AI-assisted programming, supporting major platforms like Claude Code and Cursor through marketplace plugins.

hacksider/Deep-Live-Cam — real time face swap and one-click video deepfake with only a single image

1,814 stars/day · Repo

A Python-based real-time face swapping tool that performs one-click video deepfakes using just a single reference image. The tool enables live face replacement during video calls or streaming, representing the continued advancement of accessible deepfake technology. Its rapid daily star gain reflects growing interest in real-time face manipulation tools, raising both creative possibilities and ethical considerations.

onyx-dot-app/onyx — Open Source AI Platform - AI Chat with advanced features that works with every LLM

880 stars/day · Repo

A feature-rich, self-hostable Chat UI that works with any LLM and can run completely airgapped. Onyx includes advanced features like Agents, Web Search, RAG, MCP, Deep Research, and connectors to 40+ knowledge sources. With 19.9K stars and daily active development, it offers an open alternative to proprietary AI chat platforms with one-command deployment and enterprise-ready security features.

📡 Analysis

Quoting Matt Webb

March 28, 2026 · Post · tags: matt-webb, ai, llms, vibe-coding, coding-agents, ai-assisted-programming, generative-ai, agentic-engineering, definitions

Matt Webb argues that AI agents excel at grinding problems into dust through brute force, but effective AI-assisted development requires thinking about architecture more than individual lines of code. He advocates for "vibing" (his term for AI-assisted coding) that focuses on building great libraries with interfaces that make the right way the easy way. This shift from code-level to architecture-level thinking represents a fundamental change in how developers approach AI collaboration.

Quoting Richard Fontana

March 27, 2026 · Post · tags: open-source, ai-ethics, llms, ai, generative-ai, ai-assisted-programming

Richard Fontana, LGPL co-author, weighs in on the chardet 7.0.0 relicensing situation, stating there's no basis for requiring LGPL compliance. He notes that no copyrightable material from earlier versions persists in 7.0.0 and no viable license violation theory has been articulated. This highlights ongoing legal complexities around AI-generated code and open source licensing, particularly when AI tools completely rewrite existing codebases.

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