AI Briefing — May 29, 2026

🔬 Research

AgentDoG 1.5: A Lightweight and Scalable Alignment Framework for AI Agent Safety and Security

75 upvotes · Dongrui Liu et al. · Paper

AgentDoG 1.5 addresses critical safety gaps in open-world AI agents by training lightweight models (0.8B-8B parameters) that achieve GPT-5.4-level safety performance using only 1k training samples. The framework reduces Docker deployment overhead by 100x while serving as both a training environment and real-time guardrail for agent safety moderation. With models already available on HuggingFace (600+ downloads across variants), this represents a practical solution to the emerging safety challenges posed by powerful autonomous agents.

OmniRetrieval: Unified Retrieval across Heterogeneous Knowledge Sources

46 upvotes · Jinheon Baek et al. · Paper

OmniRetrieval tackles the fragmentation problem in enterprise knowledge retrieval by creating a unified interface that can query across text, tables, knowledge graphs, and property graphs while preserving each source's structural advantages. Rather than flattening diverse data into a single format, it intelligently routes natural language queries to appropriate sources and translates them into native query languages. Testing across 13 datasets and 309 knowledge bases shows it outperforms single-source baselines, making it valuable for organizations with complex, multi-modal data architectures.

CollectionLoRA: Collecting 50 Effects in 1 LoRA via Multi-Teacher On-Policy Distillation

45 upvotes · Fangtai Wu et al. · Paper

CollectionLoRA solves a major deployment headache in custom image generation by distilling up to 50 different visual effect LoRAs into a single model while maintaining quality. The framework eliminates parameter interference issues that cause "concept bleeding" when chaining multiple LoRAs, reducing deployment costs significantly. Using probabilistic dual-stream routing and asymmetric orthogonal prompting, it achieves concept isolation within the prompt space, making it practical for production systems that need diverse visual effects without model switching overhead.

🛠 Tools

harry0703/MoneyPrinterTurbo — AI-powered short video generation

4,698 stars/day · Repo

MoneyPrinterTurbo automates the entire short-form video creation pipeline from a single topic or keyword, generating scripts, sourcing footage, creating subtitles, adding background music, and producing HD videos in 9:16 or 16:9 formats. With 68k stars and active development, it supports 15+ AI providers including OpenAI, Claude, and local models, plus batch generation and customizable styling. The tool addresses the massive demand for automated content creation in the creator economy, offering both web UI and API interfaces.

Leonxlnx/taste-skill — Anti-generic design skills for AI agents

2,234 stars/day · Repo

Taste-skill provides portable agent skills that upgrade AI-generated interfaces with better layout, typography, and spacing instead of generic-looking UIs. The framework includes image generation skills for reference boards and integrates with ChatGPT Images, Codex, Cursor, and Claude Code for implementation. With 27k stars, it's addressing a real pain point in AI-generated design where default outputs often look templated and unprofessional.

affaan-m/ECC — Agent harness performance optimization system

1,385 stars/day · Repo

ECC provides a performance optimization system for AI coding agents including Claude Code, Codex, and Cursor, focusing on skills, instincts, memory, and security components. The JavaScript-based system takes a research-first development approach to enhance agent capabilities beyond their base performance. As AI coding tools become mainstream, optimization frameworks like ECC are emerging to squeeze better performance from these agents in real development workflows.

📡 Analysis

datasette 1.0a31

May 29, 2026 · Post · tags: projects, sql, sqlite, datasette, annotated-release-notes

Datasette's latest alpha introduces write query execution and stored queries, marking a significant shift from read-only data exploration to full database management capabilities. Users with proper permissions can now execute templated insert/update/delete queries directly from the web interface, with both private and shared stored query functionality. This evolution positions Datasette as a comprehensive database admin tool, not just a data publishing platform.

Anthropic's run-rate revenue hits $47 billion

May 29, 2026 · Post · tags: anthropic, ai

Anthropic's Series H announcement reveals $47 billion in annualized run-rate revenue, up from $30 billion in April and $14 billion in February—representing unprecedented enterprise AI adoption velocity. The numbers are credible since lying to $65B investors would constitute securities fraud, and anecdotal reports include clients spending $500M monthly on Claude licenses without usage limits. This trajectory suggests enterprise AI has achieved true product-market fit at massive scale.

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