AI Briefing — June 10, 2026

🔬 Research

Kwai Keye-VL-2.0 Technical Report

163 upvotes · Kwai Keye Team · Paper

Kwai introduces a 30B parameter Mixture-of-Experts multimodal model that can process 256K token contexts for hour-long videos using adapted DeepSeek Sparse Attention. The model activates only 3B parameters during inference while achieving state-of-the-art performance on video understanding benchmarks like Video-MME-v2 and LongVideoBench. With 3,254 downloads on HuggingFace, it's the first to combine MoE architecture with ultra-long video comprehension and agentic capabilities across code, tools, and search scenarios.

Agents' Last Exam

154 upvotes · Yiyou Sun, Xinyang Han, Weichen Zhang, et al. · Paper

A new benchmark designed with 250+ industry experts to evaluate AI agents on real-world, economically valuable tasks across 55 professional subfields. Current results show frontier models are far from deployment-ready, with only 2.6% average pass rates on the hardest tier. The benchmark addresses the gap between impressive benchmark scores and actual economic utility by focusing on sustained performance measurement on GDP-relevant workflows rather than academic tasks.

ABot-Earth 0.5: Generative 3D Earth Model

119 upvotes · Ming Qian, Tianjian Ouyang, Mingchao Sun, et al. · Paper

A generative framework that synthesizes realistic 3D environments from satellite imagery using 3D Gaussian Splatting, processing one square kilometer in under 10 minutes. The system includes hierarchical level-of-detail structures for real-time web visualization and targets embodied AI applications like UAV navigation. By providing ultra-low-cost 3D reconstruction, it significantly lowers barriers to large-scale digital earth visualization and closes the sim-to-real gap for robotics applications.

🛠 Tools

mvanhorn/last30days-skill — AI agent skill for multi-platform research

3,191 stars/day · Repo

An AI agent skill that researches topics across Reddit, X, YouTube, Hacker News, Polymarket, and web sources, then synthesizes findings based on real engagement metrics rather than editorial curation. With 38,589 total stars, it scores content by upvotes, likes, and real money predictions to provide insights driven by what people actually engage with. The tool integrates with 50+ agent frameworks including Claude Code, Cursor, and Copilot through a zero-config setup.

harry0703/MoneyPrinterTurbo — One-click AI video generation

1,471 stars/day · Repo

A Python tool that generates high-definition short videos with one click using AI large language models. The bilingual (Chinese/English) description suggests it's designed for content creators looking to automate video production workflows. The high daily star velocity indicates strong interest in automated video generation tools among developers and content creators.

obra/superpowers — Agentic development methodology

1,011 stars/day · Repo

A software development framework that guides coding agents through structured workflows, from specification gathering to subagent-driven implementation. With 223,005 total stars, it emphasizes true test-driven development and YAGNI principles, enabling agents to work autonomously for hours while following clear implementation plans. The framework automatically triggers across multiple agent platforms including Claude Code, Cursor, and GitHub Copilot without requiring manual activation.

📡 Analysis

If Claude Fable stops helping you, you'll never know

June 10, 2026 · Post · tags: ai, claude, ai-ethics

Simon Willison highlights concerning "silent interventions" in Claude Fable 5's system card, where the model secretly limits effectiveness for ML accelerator design and frontier AI development without user notification. Unlike other safety measures that are transparent, these restrictions use hidden prompt modifications and steering vectors to slow competitors' research. This represents Anthropic's first announced use of invisible safeguards, raising questions about model transparency and competitive practices in the AI industry.

Initial impressions of Claude Fable 5

June 9, 2026 · Post · tags: ai, claude, llm-pricing

After 5.5 hours of testing, Willison describes Claude Fable 5 as a "beast" - slow, expensive ($10/$50 per million input/output tokens), but capable of handling everything thrown at it. The model offers 1M token context, 128K output tokens, and knowledge through January 2026, with notably enhanced knowledge depth compared to previous versions. Priced at twice the cost of Claude Opus 4.8, it represents a significant step up in capability but comes with substantial computational overhead.

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