AI Briefing — May 16, 2026

🔬 Research

Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

133 upvotes · Li et al. · Paper

SU-01, a 30B parameter model, achieves gold-medal performance on IMO 2025/USAMO 2026 and IPhO competitions using a unified training recipe combining supervised fine-tuning with two-stage reinforcement learning. The approach demonstrates stable reasoning on problems requiring 100K+ token trajectories, representing a significant milestone in mathematical problem-solving AI capabilities.

Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

81 upvotes · Zhao et al. · Paper

This system achieves frame-wise video generation with just 1-2 sampling steps, reducing first-frame latency by 50% while improving quality scores by 0.1-0.335 over existing 4-step methods. The key innovation is causal consistency distillation that avoids precomputing full trajectories, making real-time interactive video generation finally practical for applications like world models.

Self-Distilled Agentic Reinforcement Learning

73 upvotes · Lu et al. · Paper

SDAR combines reinforcement learning with token-level self-distillation to improve multi-turn agent performance, achieving substantial gains over baseline methods (+9.4% on ALFWorld, +10.2% on WebShop). The approach addresses instability issues in multi-turn interactions by using gated auxiliary objectives that strengthen positive teacher signals while softly attenuating rejections.

🛠 Tools

mattpocock/skills — Claude skills for real engineering

3132 stars/day · Repo

A collection of composable agent skills designed for actual software development rather than "vibe coding," garnering 85K+ stars from developers. The skills integrate with issue trackers, provide structured development workflows, and emphasize small, adaptable components that work across different AI models. The 30-second installer and focus on practical engineering make this a standout alternative to more rigid development frameworks.

ruvnet/RuView — WiFi-based spatial intelligence without cameras

1859 stars/day · Repo

This Rust-based system transforms commodity WiFi signals into real-time presence detection, vital sign monitoring, and activity recognition using ESP32 sensors. With 57K+ stars, it can detect breathing/heart rate through walls and track movement without any visual sensors by analyzing Channel State Information from WiFi waves, opening new possibilities for privacy-preserving sensing applications.

obra/superpowers — Complete agentic development methodology

1648 stars/day · Repo

A comprehensive software development framework that guides coding agents through specification, design review, and subagent-driven implementation, accumulating 193K+ stars. The system emphasizes true test-driven development and can enable Claude to work autonomously for hours at a time while following structured engineering practices, representing a mature approach to AI-assisted development.

📡 Analysis

QR code generator

May 15 · Post · tags: vibe-coding, tools, generative-ai, ai, llms

Simon built a QR code generator tool with Claude's help that supports both text/URL codes and WiFi network connections, featuring customizable styling and download options. This represents practical "vibe coding" where AI assists in building immediately useful tools rather than complex applications, demonstrating the growing utility of LLMs for quick development tasks.

inaturalist-clumper 0.1

May 15 · Post · tags: projects, inaturalist

Simon released the first version of his tool for grouping iNaturalist wildlife sightings into geographical clusters, part of his automated blog publishing pipeline. The tool has been running in production for weeks, showing how personal automation projects can evolve into reusable open source components for niche but valuable use cases.

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