AI Briefing — May 17, 2026

🔬 Research

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

138 upvotes · Li et al. · Paper

Researchers achieved gold-medal performance on International Mathematical and Physics Olympiads using a 30B parameter model called SU-01. The approach combines reverse-perplexity curriculum learning, two-stage reinforcement learning, and test-time scaling to enable rigorous proof-search behaviors. The model can generate stable reasoning with trajectories exceeding 100K tokens and demonstrates strong generalization beyond mathematics and physics, marking a significant milestone in automated scientific reasoning.

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

83 upvotes · Zhao et al. · Paper

This work tackles real-time video generation by achieving frame-wise autoregression with just 1-2 sampling steps, compared to existing 4-step methods. The key innovation is Causal Forcing++, which uses causal consistency distillation to avoid costly precomputation of full trajectories. Results show 0.1-0.3 point improvements on VBench metrics while reducing first-frame latency by 50% and training costs by 4x, enabling truly interactive video generation applications.

Self-Distilled Agentic Reinforcement Learning

78 upvotes · Lu et al. · Paper

SDAR introduces a gated auxiliary objective that stabilizes multi-turn agent training by selectively applying self-distillation only on teacher-endorsed tokens. The method addresses instability issues in naive reinforcement learning plus on-policy self-distillation combinations. Across ALFWorld, WebShop, and Search-QA benchmarks, SDAR shows substantial improvements (+7-10%) over standard approaches while maintaining training stability across different model scales.

🛠 Tools

tinyhumansai/openhuman — Personal AI Super Intelligence Platform

1,549 stars/day · Repo

OpenHuman is a Rust-based personal AI platform emphasizing privacy and local deployment. With 11K+ stars, it positions itself as a private alternative to cloud-based AI assistants, offering what the creators call "super intelligence" capabilities. The project is in early beta but has gained significant traction, suggesting demand for self-hosted AI solutions that keep data local rather than sending it to third-party services.

obra/superpowers — Agentic Software Development Framework

1,305 stars/day · Repo

Superpowers is a comprehensive methodology for coding agents with 194K+ stars, providing composable skills and instructions for autonomous software development. The framework emphasizes true test-driven development and uses "subagent-driven-development" where agents work through engineering tasks while maintaining clear specifications. The system enables agents to work autonomously for hours at a time, representing a mature approach to AI-assisted software engineering.

ruvnet/RuView — WiFi-Based Spatial Intelligence System

1,010 stars/day · Repo

RuView transforms ordinary WiFi signals into spatial intelligence, detecting presence, vital signs, and movement through walls without cameras or wearables. Built in Rust with 58K+ stars, it uses Channel State Information from ESP32 sensors to measure breathing, heart rate, and activity recognition. While still in beta with known limitations, it represents a significant advancement in privacy-preserving sensing technology that could revolutionize smart home and healthcare monitoring.

📡 Analysis

Warelay -> OpenClaw

May 16, 2026 · Post · tags: openclaw, git, tools

Simon Willison traces the naming evolution of OpenClaw (a personal AI assistant project) from its original "Warelay" through six different names over six months. Using a custom Git history analysis tool, he documents the progression: Warelay → CLAWDIS → CLAWDBOT → Clawdbot → Moltbot → OpenClaw. This illustrates the iterative nature of open-source project development and the challenge of finding the right product positioning, especially for AI assistant tools.

inaturalist-clumper 0.1

May 15, 2026 · Post · tags: projects, inaturalist

Willison released a Python tool for grouping iNaturalist wildlife sightings into geographical and temporal clusters for blog publishing. The 0.1 release represents infrastructure he's been running in production for weeks, demonstrating the practical workflow of building personal automation tools. This reflects a broader trend of developers creating specialized tools to enhance their content creation and data management workflows.

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