AI Briefing — March 28, 2026

🔬 Research

PixelSmile: Toward Fine-Grained Facial Expression Editing

102 upvotes · Jiabin Hua, Hengyuan Xu et al. · Paper

PixelSmile tackles the challenge of editing facial expressions without losing identity through a diffusion framework with symmetric joint training. The researchers created the FFE dataset with continuous affective annotations and FFE-Bench for evaluation. The model achieves precise linear expression control via textual latent interpolation, enabling smooth expression blending while preserving identity—a significant advance for controllable face generation.

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

94 upvotes · Yicheng Zou, Dongsheng Zhu et al. · Paper

The first trillion-parameter scientific multimodal foundation model, built with XTuner and LMDeploy infrastructure for efficient RL training at massive scale. Intern-S1-Pro masters over 100 specialized tasks across chemistry, materials, life sciences, and earth sciences while maintaining strong general capabilities. It positions itself as a "Specializable Generalist" that outperforms proprietary models in scientific depth while competing in general intelligence.

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

85 upvotes · Xiangru Jian, Shravan Nayak et al. · Paper

CUA-Suite addresses the critical bottleneck in computer-use agents: lack of continuous video demonstrations. VideoCUA provides 10,000 human-demonstrated tasks across 87 applications with 30fps recordings, totaling 55 hours and 6 million frames—a massive leap from existing datasets' 20 hours. Current foundation action models still struggle with professional desktop apps (~60% failure rate), highlighting the scale of the challenge ahead.

🛠 Tools

mvanhorn/last30days-skill — AI research synthesizer

2,821⭐/day · Repo

An AI agent skill that researches any topic across Reddit, X, YouTube, Hacker News, and Polymarket from the past 30 days, then synthesizes findings into grounded summaries with citations. The latest v2.9.5 adds Bluesky support, comparative analysis mode, and auto-saves reports to a personal research library. With 13K stars and active development, it's become essential for staying current in fast-moving AI topics.

obra/superpowers — Agentic development framework

2,752⭐/day · Repo

A complete software development workflow for coding agents that emphasizes proper planning before coding. Superpowers guides agents through specification gathering, design review, implementation planning, and subagent-driven development with TDD principles. With 119K stars, it's proven that structured agent workflows can enable hours of autonomous development while maintaining quality and avoiding the typical pitfalls of rushed AI coding.

hacksider/Deep-Live-Cam — Real-time face swap

1,616⭐/day · Repo

Real-time face swapping and one-click video deepfake creation using just a single image. The tool enables live face replacement during video calls or streaming, representing the continued democratization of deepfake technology. While the technical achievement is notable, it raises immediate concerns about misuse and the need for detection mechanisms in video platforms.

📡 Analysis

Vibe coding SwiftUI apps is a lot of fun

March 27, 2026 · Post · tags: macos, ai, generative-ai, llms, vibe-coding, coding-agents, swift, claude-code

Simon demonstrates "vibe coding" macOS apps with Claude Opus 4.6 and GPT-5.4, building Bandwidther (network monitoring) and Gpuer (GPU monitoring) as menu bar apps. Both Claude and GPT-5.4 prove highly competent at SwiftUI, enabling rapid prototyping without opening Xcode. This showcases the maturation of AI coding assistants for platform-specific development, making native app creation accessible to non-specialists.

Quoting Richard Fontana

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

Simon quotes a legal discussion about whether chardet 7.0.0 requires LGPL licensing, touching on copyright persistence in open source projects. The post highlights ongoing complexity around AI-assisted code generation and licensing obligations, particularly when tools modify existing copyrighted code. These legal uncertainties continue to complicate enterprise adoption of AI coding tools.

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