AI Briefing — May 20, 2026

🔬 Research

GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment

44 upvotes · Minxuan Lv et al. · Paper

A comprehensive open-source framework for training long-context AI models using reinforcement learning with verifiable rewards. The team releases 23K training samples across 9 task types, plus all training code and construction pipelines. Their Qwen3-30B model trained on this data matches the performance of much larger models like DeepSeek-R1 and Qwen3-235B, suggesting that diverse, capability-focused training data can be more effective than raw scale.

OpenComputer: Verifiable Software Worlds for Computer-Use Agents

44 upvotes · Jinbiao Wei et al. · Paper

A testing framework that creates verifiable environments for AI agents that control desktop applications. OpenComputer covers 33 desktop apps and 1,000 tasks across browsers, office tools, and development environments, using hard-coded verifiers that align better with human judgment than LLM evaluation. Even frontier AI agents struggle with end-to-end task completion, with the best achieving only partial success on most workflows.

CHI-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?

43 upvotes · Haolin Chen et al. · Paper

A benchmark testing AI agents on complex healthcare operations that require navigating dense policy documents, playing multiple roles, and conducting multi-turn conversations. Using a 1,290+ document healthcare operations handbook and 20 healthcare apps, the benchmark spans provider authorization, utilization management, and care coordination. The best agent resolves only 28% of tasks, with performance dropping to just 3.8% when handling multiple tasks in sequence, highlighting major gaps in enterprise AI automation.

🛠 Tools

tinyhumansai/openhuman — Your Personal AI super intelligence

3,973 stars/day · Repo

A privacy-focused personal AI assistant built in Rust that runs locally. OpenHuman positions itself as a "super intelligence" that keeps data private while providing powerful AI capabilities. With nearly 23K total stars and active development, it's gaining traction as an open-source alternative to cloud-based AI assistants, appealing to users who want AI capabilities without sending data to external servers.

Imbad0202/academic-research-skills — Research pipeline automation for Claude

3,164 stars/day · Repo

A comprehensive Claude Code plugin that automates academic research workflows from initial research through final publication. The tool handles reference hunting, citation formatting, data verification, and logical consistency checks while keeping humans in the loop for critical thinking tasks. It emphasizes being a "copilot, not pilot" — handling grunt work so researchers can focus on defining questions, choosing methods, and interpreting results.

multica-ai/andrej-karpathy-skills — Coding guidelines based on Karpathy's observations

1,955 stars/day · Repo

A single CLAUDE.md file containing four principles to improve Claude's coding behavior, addressing common LLM pitfalls identified by Andrej Karpathy. The guidelines focus on explicit reasoning before coding, simplicity over complexity, surgical changes that don't touch unrelated code, and goal-driven execution with tests-first approaches. With 139K total stars, it's become a widely-adopted resource for better AI-assisted programming.

📡 Analysis

Gemini 3.5 Flash: more expensive, but Google plan to use it for everything

May 19, 2026 · Post · tags: google, ai, generative-ai, llms, gemini, llm-pricing

Google released Gemini 3.5 Flash with a significant price increase — 3x more expensive than Gemini 3 Flash Preview at $1.50/million input tokens. Despite the higher cost, Google is deploying it across billions of users in Search, the Gemini app, and enterprise products. This follows a broader industry trend of AI labs testing price tolerance, with OpenAI's GPT-5.5 and Claude Opus 4.7 also seeing substantial price increases, suggesting the era of cheap AI inference may be ending.

llm-gemini 0.32

May 19, 2026 · Post · tags: gemini, llm

A new release of the llm-gemini plugin that adds support for the new Gemini 3.5 Flash model. The update enables developers to access Google's latest model through the popular LLM command-line tool ecosystem.

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