AI Briefing — May 6, 2026

🔬 Research

MolmoAct2: Action Reasoning Models for Real-world Deployment

187 upvotes · Allen AI · Paper

This fully open Vision-Language-Action model tackles real-world robotics deployment with five key advances: a specialized VLM backbone (MolmoER), 720 hours of bimanual robot trajectories, an open action tokenizer, and adaptive-depth reasoning that only re-predicts tokens for scene regions that change between timesteps. The system outperforms closed alternatives like GPT-5 and Gemini Robotics across 13 benchmarks while maintaining practical latency. This matters because it's the first open VLA that actually meets the criteria for real-world deployment — fully open weights, reasonable hardware requirements, and competitive performance.

From Context to Skills: Can Language Models Learn from Context Skillfully?

129 upvotes · Research team · Paper

Ctx2Skill introduces a self-evolving framework where AI agents automatically discover and refine context-specific skills through multi-agent self-play, without human supervision or external feedback. A Challenger generates probing tasks, a Reasoner attempts solutions using an evolving skill set, and a Judge provides binary feedback. The system prevents adversarial collapse through Cross-time Replay that identifies optimal skill balance. This addresses the critical challenge of helping language models learn from complex contexts that exceed their training knowledge — essential for real-world applications requiring domain-specific reasoning.

Beyond SFT-to-RL: Pre-alignment via Black-Box On-Policy Distillation for Multimodal RL

31 upvotes · Research team · Paper

PRISM introduces a three-stage training pipeline that adds explicit distribution alignment between supervised fine-tuning and reinforcement learning for multimodal models. Using a Mixture-of-Experts discriminator with separate perception and reasoning experts, it provides disentangled corrective signals to prevent distributional drift. Tests on Qwen3-VL show +4.4 and +6.0 point improvements over standard SFT-to-RLVR baselines on 4B and 8B models respectively. This matters because it solves a fundamental problem in multimodal model training where perception errors and reasoning failures compound during RL, limiting real-world performance.

🛠 Tools

Hmbown/DeepSeek-TUI — Terminal coding agent for DeepSeek models

2,434 stars/day · Repo

A Rust-built terminal interface that turns DeepSeek V4 into an interactive coding agent with streaming reasoning blocks, local workspace editing with approval gates, and an auto mode that dynamically selects both model and thinking level. The tool supports multiple installation methods (npm, Cargo, Homebrew) and provides a rich TUI experience for developers. With 10,682 stars already, it's becoming the go-to interface for developers wanting to leverage DeepSeek's reasoning capabilities in their coding workflow without leaving the terminal.

ruvnet/ruflo — Agent orchestration platform for Claude

2,432 stars/day · Repo

An enterprise-grade TypeScript platform for deploying multi-agent Claude swarms with self-learning intelligence, RAG integration, and native Claude Code integration. It offers live agent deployment, goal planning interfaces, and coordinated autonomous workflows. With 44,513 stars, it's positioned as "the leading agent orchestration platform for Claude" and provides both UI beta and live agent demos. This addresses the growing need for production-ready multi-agent systems that can handle complex enterprise workflows with Claude's capabilities.

forrestchang/andrej-karpathy-skills — Claude coding improvement guidelines

2,409 stars/day · Repo

A single CLAUDE.md file containing four principles designed to fix common LLM coding pitfalls identified by Andrej Karpathy: wrong assumptions, overcomplication, orthogonal edits, and lack of verification. The principles focus on explicit reasoning, simplicity first, surgical changes, and goal-driven execution. With 115,354 stars, it's become a widely-adopted prompt engineering resource. This matters because it codifies specific fixes for the most frustrating aspects of AI coding assistants — their tendency to overcomplicate, make assumptions, and touch code they shouldn't.

📡 Analysis

Our AI started a cafe in Stockholm

May 5 · Post · tags: ai, generative-ai, llms, ai-agents, ai-ethics

Simon Willison critiques Andon Labs' AI-managed Stockholm cafe experiment, highlighting both amusing failures (ordering 120 eggs with no stove, 6,000 napkins) and ethical concerns about AI systems wasting human time. The AI manager "Mona" sent emergency emails to suppliers and submitted poor permit applications to police without human oversight. Willison argues these experiments become unethical when they affect real-world systems and steal time from people who haven't opted into the experiment, calling for human-in-the-loop controls for outbound actions affecting others.

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