AI Briefing — June 11, 2026

🔬 Research

Kwai Keye-VL-2.0 Technical Report

174 upvotes · Kwai Keye Team · Paper

Kwai introduces the first MoE multimodal model to adapt sparse attention for 256K context processing in hour-long videos, using only 3B active parameters from a 30B total. The system combines DeepSeek Sparse Attention with novel training techniques like Cross-Modal Multi-Teacher On-Policy Distillation to prevent catastrophic forgetting during multi-task alignment. This enables native agent collaboration across coding, tools, and search with multimodal self-correction, achieving SOTA on video understanding benchmarks like TimeLens and Video-MME-v2.

Role-Agent: Bootstrapping LLM Agents via Dual-Role Evolution

73 upvotes · Xucong Wang et al. · Paper

A framework that uses a single LLM to simultaneously act as both agent and environment for self-improving training. The World-In-Agent component predicts future states for process rewards, while Agent-In-World analyzes failure patterns to reshape training data distribution. This bootstrapped co-evolution approach delivers consistent 4% average performance gains over strong baselines across multiple benchmarks, addressing the core limitation of static training environments in LLM agents.

Redesign Mixture-of-Experts Routers with Manifold Power Iteration

63 upvotes · Songhao Wu et al. · Paper

Proposes aligning MoE router rows with principal singular directions of their associated experts using Manifold Power Iteration, a "Power-then-Retract" paradigm. This mathematical approach ensures router weights converge toward the most expressive description of each expert matrix, improving token-expert affinity calculations. Experiments across 1B to 11B parameter models confirm this alignment creates more effective MoE architectures by providing principled design for the critical router component.

🛠 Tools

msitarzewski/agency-agents — Complete AI agency specialists

1434 stars/day · Repo

A collection of meticulously crafted AI agent personalities spanning frontend development, Reddit community management, and content creation, each with unique voice and proven deliverables. Born from Reddit discussions and months of iteration, these aren't generic prompt templates but specialized experts with battle-tested workflows and success metrics. With 110K+ stars, it integrates seamlessly with Claude Code and provides production-ready agent personas that deliver measurable outcomes across engineering, marketing, and creative domains.

obra/superpowers — Agentic software development methodology

1104 stars/day · Repo

A complete development methodology that transforms coding agents from code-first to spec-first builders, emphasizing true red/green TDD and subagent-driven development. The framework automatically triggers skills based on context, enabling agents to work autonomously for hours while maintaining quality through clear implementation plans and continuous inspection. With 224K+ stars, it represents a paradigm shift toward systematic agent collaboration that can replace traditional development workflows with structured, testable approaches.

addyosmani/agent-skills — Production-grade engineering workflows

821 stars/day · Repo

Seven slash commands (/spec, /plan, /build, /test, /review, /code-simplify, /ship) that encode senior engineering workflows into AI agents, mapping directly to the development lifecycle. Features an autonomous /build auto mode that generates plans and implements tasks with individual test-driven commits while preserving human verification points. With 53K+ stars, it bridges the gap between AI capability and engineering discipline, ensuring agents follow production-grade practices consistently.

📡 Analysis

Anthropic Walks Back Policy That Could Have 'Sabotaged' AI Researchers Using Claude

June 11 · Post · tags: ai, anthropic, claude, ai-ethics

Anthropic reversed their controversial "invisible safeguards" policy that secretly limited Claude's effectiveness on frontier LLM development requests without user notification. Following significant backlash, they're making these safeguards visible by falling back to Opus 4.8 with clear explanations, acknowledging they "made the wrong tradeoff" between rapid deployment and transparency. This marks a significant policy shift toward user visibility in AI safety measures, though critics argue for dropping these research limitations entirely.

DiffusionGemma

June 10 · Post · tags: google, ai, gemma, llm-performance

Google's experimental Gemini Diffusion research returns as an open-weight Apache 2 licensed model, DiffusionGemma-26B-A4B-it, delivering 500+ tokens/second through NVIDIA's free NIM cloud API. This represents Google's commitment to open-weight releases following their earlier 857 tokens/second preview model from May 2025. The model demonstrates competitive performance in rapid text generation, generating 2,409 tokens in 4.4 seconds for complex creative tasks like pelican illustrations.

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