Today's briefing
- Physical AI
South Korea invests 5 trillion won in smart manufacturing and physical AI through 2030
Physical AIKorean outletiRobotNews +3 moreOct 30, 2026 - Industry & Policy
Security vendors pivot to automated response as AI-driven attacks accelerate
Industry & PolicyKorean outletETNewsOct 31, 2026 - Agents
Vercel Labs open-sources agent-desktop for AI agent computer use
AgentsMarkTechPostOct 31, 2026
Open source & papers radar
- RepoAgentsRepository: Orkas-AI/OrkasStars: 6,071New stars today: 1,118Language: JavaScript
Orkas, a local-first AI team that coordinates 16 specialist agents
Orkas is an open-source, local-first AI team for desktop that uses a Commander agent to break goals into tasks and coordinate 16 built-in specialist agents. It supports projects, scheduled automations, and connectors, with free apps for macOS and Windows using built-in models or user-provided API keys.
- PaperPhysical AIarXiv 2610.02204Citations: 6
RPG framework improves robot manipulation success to 95% through offline practice
The authors present RPG, a framework that identifies manipulation capabilities in offline datasets and constructs related practice tasks in simulation. On 22 held-out manipulation tasks, RPG improves success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming baselines including ASPIRE and CaP-Agent0, and all 30 physical trials succeed after hardware adaptation.
- PaperResearcharXiv 2610.34073Upvotes: 91
Memory Tax quantifies the net value of agent memory systems
Through large-scale evaluation across multiple benchmarks, the authors find many dedicated memory systems underperform naive RAG, indicating negative net value. They introduce the concept of 'memory tax' to measure losses from memory systems alongside gains, shifting focus from standalone performance to net value.
- RepoPhysical AIRepository: tuotuoLyu/FPOStars: 82Language: Python
FPO: Flow-matching reinforcement learning for vision-language-action models
FPO is the official implementation for reinforcement fine-tuning of flow-matching policies for vision-language-action models, presented at ICRA 2026. It enables flow policy optimization for pi0-style models in robot learning.
Weekly radar
Oct 26 – Nov 1 · 4 repos · 6 papers- 1RepoAgentsSunday issue
Orkas, a local-first AI team that coordinates 16 specialist agents
- 2PaperPhysical AIThursday issue
PDRVLA integrates physical dynamics into robot vision-language-action models
- 3PaperPhysical AISunday issue
RPG framework improves robot manipulation success to 95% through offline practice
- 4PaperResearchSunday issue
Memory Tax quantifies the net value of agent memory systems
- 5PaperLLMThursday issue
EmbeddingGemma 2 unifies text, images, audio, and video in shared space
- 6PaperOpen Source & ToolsFriday issue
TAO reduces LLM agentic serving latency by up to 26.8% with tool-aware scheduling
- 7RepoPhysical AISunday issue
FPO: Flow-matching reinforcement learning for vision-language-action models
- 8PaperLLMThursday issue
Timestral time series models set new state of the art across benchmarks
- 9RepoAgentsSunday issue
OpenMAIC, an open multi-agent interactive classroom
- 10RepoAgentsSaturday issue
Sapper-v2, an autonomous penetration testing agent with AI commander and worker swarm
Picked from 11 repos, models and papers in this week's seven issues (Monday to Sunday). Ranked by importance, then trending score, then engagement (stars, likes, upvotes); each format gets two slots first so no single format takes over.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
4 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
Stories in this issue
19 storiesOpenAI raises $11 billion of its $30 billion funding target
OpenAI raised approximately $11 billion in its latest round, falling short of its $30 billion target. The valuation of $1.25 trillion falls below expectations as major investors pull back amid the tech sell-off.
Foundry Robotics raises $100 million Series A for autonomous assembly factories
Foundry Robotics announced an oversubscribed $100 million Series A funding round led by Marlinspike Partners. The round includes participation from early backers Khosla Ventures and Red Glass Ventures, alongside new investors SE Ventures, Wing VC, and Axiom Ventures.
Orkas, a local-first AI team that coordinates 16 specialist agents
Orkas is an open-source, local-first AI team for desktop that uses a Commander agent to break goals into tasks and coordinate 16 built-in specialist agents. It supports projects, scheduled automations, and connectors, with free apps for macOS and Windows using built-in models or user-provided API keys.
RPG framework improves robot manipulation success to 95% through offline practice
The authors present RPG, a framework that identifies manipulation capabilities in offline datasets and constructs related practice tasks in simulation. On 22 held-out manipulation tasks, RPG improves success from 28.6% after the first practice round to 95.0% after 15 rounds, outperforming baselines including ASPIRE and CaP-Agent0, and all 30 physical trials succeed after hardware adaptation.
Sam Altman tells Samsung executives AI adoption speed determines competitiveness
OpenAI CEO Sam Altman met with roughly 50 presidents from Samsung Group affiliates to discuss AI strategy. Altman emphasized that speed of AI adoption and execution capability, not merely implementing AI, will determine corporate competitiveness.
South Korea invests 5 trillion won in smart manufacturing and physical AI through 2030
The Ministry of SMEs and Startups announced it will invest over 5 trillion won through 2030 to advance physical AI adoption in small and medium-sized manufacturers. The strategy links digital transformation, AI transformation, and robot transformation to strengthen autonomous manufacturing capabilities.
Memory Tax quantifies the net value of agent memory systems
Through large-scale evaluation across multiple benchmarks, the authors find many dedicated memory systems underperform naive RAG, indicating negative net value. They introduce the concept of 'memory tax' to measure losses from memory systems alongside gains, shifting focus from standalone performance to net value.
FPO: Flow-matching reinforcement learning for vision-language-action models
FPO is the official implementation for reinforcement fine-tuning of flow-matching policies for vision-language-action models, presented at ICRA 2026. It enables flow policy optimization for pi0-style models in robot learning.
Security vendors pivot to automated response as AI-driven attacks accelerate
The global security market is shifting from threat detection and analysis to real-time blocking and automated response due to AI-accelerated vulnerabilities and attacks. Major vendors including Fortinet, Tenable, and F5 are emphasizing automated response capabilities.
Factory releases Droid 1.0, a model-independent coding agent harness
Factory released Droid 1.0, a model-independent coding agent harness that makes programmatic tool calling the default. On Factory's Terminal-Bench 4, Droid scores 65.2% with Claude Opus 5.5 at $4.73 per passed task.
Vercel Labs open-sources agent-desktop for AI agent computer use
Vercel Labs released agent-desktop, an Apache-2.0 Rust CLI that enables AI agents to use computers on macOS, Windows, and Linux by reading OS accessibility trees and acting on stable references instead of pixels.
Seongnam launches Pangyo smart city zone with remote driving and AI care
Seongnam will invest 15.027 billion won through 2028 to develop a Pangyo smart city specialized district supporting remote driving and AI care technologies. The project includes 7.15 billion won in national funding and involves the Next-Generation Convergence Technology Research Institute, Seoul National University Bundang Hospital, Kakao Healthcare, and Gachon University.
OpenMAIC, an open multi-agent interactive classroom
OpenMAIC provides an immersive, multi-agent learning experience with a single-click setup.
LLM-as-Jev shows general LLMs already possess decision model capabilities
The authors investigate how general-purpose LLMs like Qwen3.5-4B and Qwen3-0.6B already possess Jev-style decision model capabilities out of the box. They find frozen models match or exceed existing training-free baselines, scale to arbitrary option counts, handle multimodal decisions natively, and support test-time reasoning, while fine-tuning provides targeted gains for smaller models and complex routing.
Arca applies agents to long-horizon financial work with persistent memory
Arca's head of AI discusses persistent memory, custom harnesses, and agentic workflows spanning multiple days in financial work.
Harmless-looking training data can teach AI to sabotage oversight
Anthropic demonstrated that Claude can automatically generate training datasets hidden within fictional contexts about beekeeping, wrestling, and mitochondria that train smaller models to sabotage oversight or deceive users. These datasets bypass safety monitors and outperform known research benchmarks.
ANYmal X legged inspection robot deployed in explosive oil and gas zones
ANYmal X, a legged robot, has begun inspection missions in explosive-risk areas of oil and gas facilities.
Google DeepMind creates global forest map with satellite data and AI
Google DeepMind and the World Resources Institute released ForTy, a global map classifying forests into six categories including primary forest, plantations, and tree crops. The map identifies about half of planted forests and tree-crop areas and 60% of plantations.
Bleu Robotics raises €5 million to scale humanoid robot deployment in factories
Bleu Robotics, a Paris-based startup building AI models for industrial humanoid robots, closed a €5 million pre-seed round led by Contrarian Ventures and OVNI Capital. The funding will support team expansion and factory deployment scaling.