AI Briefing — 2026-05-28

🔬 Research

ProRL: Effective Reinforcement Learning for Proactive Recommendation via Rectified Policy Gradient Estimation

71 upvotes · Hongru Hou et al. · Paper

This tackles a key problem in recommendation systems: how to guide users toward target items through a sequence of intermediate recommendations using reinforcement learning. The authors identify that standard policy gradient methods fail because longer recommendation paths get biased rewards and high variance gradients. Their ProRL framework fixes this with stepwise reward centering and position-specific advantage estimation, significantly outperforming existing proactive recommendation systems across three real-world datasets.

Agent Explorative Policy Optimization for Multimodal Agentic Reasoning

63 upvotes · Minki Kang et al. · Paper

This addresses the "Thinking-Acting Gap" in multimodal AI agents that use external tools, where standard RL training results in tool use attempts on only ~30% of rollouts and high failure rates. AXPO fixes this by resampling failed tool-using subgroups while keeping the thinking prefix intact, leading to +1.8pp improvements on Pass@1 and Pass@4 metrics. Notably, their 8B model with AXPO surpasses the 32B base model with 4x fewer parameters.

Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players

54 upvotes · Fangfu Liu et al. · Paper

The first generative world model designed for multi-agent interactive video generation, supporting more than two players simultaneously. It introduces Simplex Rotary Agent Encoding for permutation-symmetric agent representation and Sparse Hub Attention to reduce cross-agent computation from quadratic to linear complexity. The system can generate action-responsive video at 24 FPS through teacher-student distillation, generalizing from two to four players without retraining.

🛠 Tools

Lum1104/Understand-Anything — Interactive knowledge graphs for any codebase

4465 stars/day · Repo

Transforms any codebase, documentation, or knowledge base into an explorable, searchable knowledge graph that you can ask questions about. With 41,831 total stars and active development, it integrates with major AI coding tools including Claude Code, Codex, Cursor, Copilot, and Gemini CLI. The tool emphasizes "graphs that teach > graphs that impress," focusing on practical understanding rather than visual appeal.

Leonxlnx/taste-skill — Anti-slop design framework for AI agents

2715 stars/day · Repo

A framework of portable "agent skills" designed to prevent AI coding assistants from generating boring, generic-looking interfaces. With 25,322 total stars, it provides upgraded layout, typography, motion, and spacing guidelines that work with Codex, Cursor, and Claude Code. The project also includes image-generation skills for creating reference boards for web, mobile, and brand design implementations.

affaan-m/ECC — Agent performance optimization system

2062 stars/day · Repo

A comprehensive agent harness system focused on performance optimization through skills, instincts, memory, and security features for AI coding tools. The project takes a research-first development approach and supports Claude Code, Codex, Opencode, and Cursor. While the description is somewhat vague, the high star velocity suggests it addresses real pain points in agent-based development workflows.

📡 Analysis

sqlite AGENTS.md

27th May 2026 · Post · tags: sqlite, ai, d-richard-hipp, generative-ai, llms, coding-agents, ai-security-research

SQLite recently added an AGENTS.md file explicitly stating they don't accept AI-generated code or pull requests, only accepting agentic bug reports with reproducible test cases. The project was being flooded with AI-generated bug reports of varying quality, forcing them to create a separate "SQLite Bug Forum" to handle the volume. This represents a pragmatic approach to managing AI contributions in open source projects.

I think Anthropic and OpenAI have found product-market fit

27th May 2026 · Post · tags: ai, datasette, openai, generative-ai, llms, anthropic, llm-pricing, coding-agents, claude-code, codex, claude-cowork

Both Anthropic and OpenAI switched their enterprise pricing from flat monthly rates to API token pricing, with Anthropic rumored to be approaching profitability. Simon's personal usage analysis shows he consumed $2,180 worth of API tokens in 30 days while paying only $200 in subscription fees, highlighting the massive value arbitrage that's disappearing as enterprise customers now pay full API rates. This pricing shift marks a significant inflection point in the LLM industry's monetization strategy.

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