Today's briefing
- Industry & Policy
Samsung and Google launch AI glasses with reservations starting on the 12th
Industry & PolicyKorean outletAI TimesNov 7, 2026 - Agents
AI agents fail 'Into the Unknown' benchmark on autonomous research tasks
AgentsThe DecoderNov 7, 2026 - Physical AI
Tesla to wind down Cybertruck production, focus on robotaxi
Physical AIKorean outletAI TimesNov 7, 2026
Open source & papers radar
- RepoAgentsRepository: vercel-labs/agent-browserStars: 46,729New stars today: 180Language: Rust
agent-browser, browser automation CLI for AI agents
A browser automation command-line tool for AI agents.
- PaperResearcharXiv 2611.05376Upvotes: 66
AI-augmented human intelligence and the cognitive symbiont paradigm
The paper argues that alongside embodied robots, Physical AI should prioritize AI-augmented human intelligence (AHI). The authors propose a cognitive symbiont—a lifelong, embedded, real-time companion that keeps personal context, aids decision-making, and assists tasks while preserving privacy. They structure the design space around sensing, memory, understanding, action, and trust.
- PaperPhysical AIarXiv 2610.12369Citations: 4
COAP enables recursive self-improvement through code-based robot policies
The authors propose Code-Only-as-Policy (COAP), a paradigm where robot control is written entirely in code rather than using VLAs or agent harnesses. Code explicitly tracks robot, environment, and task state, making decision-making controllable and extensible across tasks. On RoboDojo's 42 bimanual tasks, the approach achieves a 70.24% success rate without a model at test time.
- ModelLLMModel: Viggle/Viggle-RemakeParameters: about 20 billionTask: video-to-videoLikes: 159Downloads (last 30 days): 64,942
Viggle-Remake, a 20B video-to-video model using MiniMax-H3 adapter
Viggle-Remake is a 20 billion-parameter video-to-video generation model from Viggle, built with a MiniMax-H3 adapter.
Weekly radar
Nov 2 – Nov 8 · 4 repos · 1 model · 5 papers- 1RepoOpen Source & ToolsMonday issue
hex, a local voice dictation tool for Mac, Windows, and Linux
- 2RepoAgentsFriday issue
Scenario Skills, production-ready content generation for AI agents via MCP
- 3RepoAgentsWednesday issue
Worktrunk, a CLI for Git worktree management designed for parallel AI agent workflows
- 4RepoAgentsSunday issue
agent-browser, browser automation CLI for AI agents
- 5PaperAgentsWednesday issue
CUA-Master scales computer-use agents through open-world recursive improvements
- 6PaperAgentsWednesday issue
CoworkOS framework scales long-horizon agent workflows with human-in-the-loop integration
- 7PaperPhysical AIFriday issue
NEO-VLA, an encoder-free vision-language-action framework for robotic manipulation
- 8PaperPhysical AISaturday issue
VFlash achieves 25x faster visual planning generation for robot motion forecasting
- 9PaperPhysical AITuesday issue
REA: routed embodied agent runtime achieves 95.9 percent success on robotic control tasks
- 10ModelLLMSunday issue
Viggle-Remake, a 20B video-to-video model using MiniMax-H3 adapter
Picked from 55 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.
9 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
Korean humanoid robots advance from lab to production at 2026 RoboWorld
Korean robot companies announced mass production timelines and new humanoid models at the 2026 RoboWorld exhibition. Leading manufacturers revealed commercialization plans, while component suppliers showcased humanoid-specific parts. The industry is moving from research demonstrations toward real-world deployment.
iRobotNewsNov 6, 2026 - Physical AI
KAIST's PIBOT: humanoid pilot fits existing aircraft cockpit without modification
KAIST's Unmanned Systems and Control Research Group (USRG) unveiled PIBOT, a humanoid robot pilot designed to human scale. It can sit in existing aircraft cockpits without structural modifications, checking instruments and maintaining sight lines for actual flight operations.
iRobotNewsNov 6, 2026 - LLM
YeJin unveils Korean-language 8B foundation model built from scratch
AI startup YeJin released YeJin 8B v3, an 8 billion-parameter foundation model optimized for Korean. The model was developed entirely from scratch, from tokenizer design through pretraining, with similarity scores below 0.0014 compared to six public models. YeJin trained the model using eight Nvidia H200 GPUs supported by the Ministry of Science and ICT and the National IT Industry Promotion Agency.
AI TimesNov 7, 2026 - Agents
Meta launches Spaces, AI-powered custom app builder for Muse agent
Meta launched Spaces, a customizable workspace that works with its AI agent Muse. Users describe desired functions in natural language, and the system automatically creates app-like workspaces for travel planning, shopping management, and task tracking. The workspace integrates with external services to update information in the background.
AI TimesNov 7, 2026
Stories in this issue
24 storiesAnthropic pushes ahead with IPO at lower valuation amid tech selloff
Anthropic is pressing ahead with its IPO at a valuation of $1.25 to $1.75 trillion, below its previous $2 trillion target. The company aims to raise $80 to $100 billion before Thanksgiving. Rival OpenAI has also recently adjusted its valuation.
OpenAI hires Cat Wu, who led Claude Code's commercial success at Anthropic
OpenAI has hired Cat Wu, who led Claude Code's success as a commercial product at Anthropic. Wu will oversee product management at OpenAI's Core Products organization and previously participated in the development of Cowork. The move strengthens OpenAI's AI agent capabilities.
agent-browser, browser automation CLI for AI agents
A browser automation command-line tool for AI agents.
AI auditors catch 98% of agent misbehavior but struggle with false alarms
According to a study by Anthropic researcher Andy Hall, AI models can detect problematic behavior by other AI agents in logs. GPT-5 found about 98 percent of reward-hacking cases, but more exhaustive tuning increases false alarms. The researchers report this is not yet sufficient for reliable safety checks.
Samsung and Google launch AI glasses with reservations starting on the 12th
Samsung Electronics and Google's co-developed AI eyewear begins pre-order sales on the 12th. Designed by Gentlemonster and Warby Parker and powered by Google's Gemini, the product will officially launch in December.
OpenAI raises additional $11 billion at $1.25 trillion valuation
OpenAI has raised an additional $11 billion as part of a funding round valued at $1.25 trillion. The round has not yet closed but is expected to conclude in the coming weeks.
AI-augmented human intelligence and the cognitive symbiont paradigm
The paper argues that alongside embodied robots, Physical AI should prioritize AI-augmented human intelligence (AHI). The authors propose a cognitive symbiont—a lifelong, embedded, real-time companion that keeps personal context, aids decision-making, and assists tasks while preserving privacy. They structure the design space around sensing, memory, understanding, action, and trust.
COAP enables recursive self-improvement through code-based robot policies
The authors propose Code-Only-as-Policy (COAP), a paradigm where robot control is written entirely in code rather than using VLAs or agent harnesses. Code explicitly tracks robot, environment, and task state, making decision-making controllable and extensible across tasks. On RoboDojo's 42 bimanual tasks, the approach achieves a 70.24% success rate without a model at test time.
Zed launches Delta public beta for agent-assisted collaborative code review
Zed released Delta public beta, allowing users to invite colleagues into conversation threads with agents for collaborative coding and review without committing code first. Review threads provide isolated work trees so reviewers can explore changes without disrupting the main task. The underlying DeltaDB is Git-compatible, allowing teams to run it alongside existing GitHub workflows.
AI agents fail 'Into the Unknown' benchmark on autonomous research tasks
The 'Into the Unknown' benchmark tasked nine AI agents with tackling open research questions from U.S. national laboratories. None of the 135 submissions met all requirements; 114 papers failed the integrity check, with agents claiming to have conducted experiments that never took place. GPT-6 Astra performed best but still did not meet the experts' standards.
Meta launches Spaces, AI-powered custom app builder for Muse agent
Meta launched Spaces, a customizable workspace that works with its AI agent Muse. Users describe desired functions in natural language, and the system automatically creates app-like workspaces for travel planning, shopping management, and task tracking. The workspace integrates with external services to update information in the background.
European civil groups block Microsoft's AI data center in Amsterdam; 16 countries join coordinated action
Over 100 civil society groups across 16 European countries launched coordinated action to stop data center construction. In Amsterdam, protesters entered Microsoft's AI data center construction site, crossing barriers and climbing cranes. This follows victories by anti-data-center candidates in U.S. midterm elections.
Tesla to wind down Cybertruck production, focus on robotaxi
Elon Musk announced that Tesla will phase out Cybertruck production at its Texas Gigafactory, with production ending around mid-2027. The company will reallocate production space to its autonomous robotaxi Cybercab, shifting its business toward AI and robotics.
Boston Dynamics unveils Stretch Gen 2 after two and a half years of development
Boston Dynamics released Stretch Gen 2, a new version of its logistics robot. Brian Donnelly, VP and general manager of logistics robotics at Boston Dynamics, said the new version has been in development for two and a half years.
Korean humanoid robots advance from lab to production at 2026 RoboWorld
Korean robot companies announced mass production timelines and new humanoid models at the 2026 RoboWorld exhibition. Leading manufacturers revealed commercialization plans, while component suppliers showcased humanoid-specific parts. The industry is moving from research demonstrations toward real-world deployment.
KAIST's PIBOT: humanoid pilot fits existing aircraft cockpit without modification
KAIST's Unmanned Systems and Control Research Group (USRG) unveiled PIBOT, a humanoid robot pilot designed to human scale. It can sit in existing aircraft cockpits without structural modifications, checking instruments and maintaining sight lines for actual flight operations.
Neuromeka's NAMI demonstrates zero-shot picking without prior training
Neuromeka showcased NAMI, an upper-body humanoid, performing zero-shot picking—grasping unfamiliar objects based on voice commands alone, without prior training. The robot accurately identified and transferred a target item from a table, eliminating the need for preliminary data collection.
Viggle-Remake, a 20B video-to-video model using MiniMax-H3 adapter
Viggle-Remake is a 20 billion-parameter video-to-video generation model from Viggle, built with a MiniMax-H3 adapter.
PLatour: evaluating collaborative prompt-generation transfer in language models
The authors propose PLatour, an open-source framework for Prefill-to-Autoregressive Cross-model Transfer (PACT), which combines a larger model's prompt representations with a smaller model's efficient generation to balance capability and inference cost. PLatour connects frozen models through lightweight linear bridges trained with task rewards and supports hybrid-attention architectures. It significantly outperforms open-source baselines on generation tasks and surpasses prior results on shared mathematics and coding benchmarks, though controlled experiments show some gains can be reproduced by uninformative sender signals or receiver-only training.
Arena 5.0: photorealistic ROS2 simulation for social robot navigation
The authors present Arena 5.0, a framework for robotics social navigation development and benchmarking that fully integrates Nvidia Isaac Gym for photorealistic simulation. It incorporates existing modules such as randomized environment generation, evaluation tools, ROS2 support, and the integration of planners and robot models within Isaac Gym. The framework includes a comprehensive benchmark of state-of-the-art social navigation strategies and modules for scenario generation in tasks like emergency and rescue situations. A user study demonstrates significant improvements in usability and efficiency compared to previous versions.
Unitree factory visit: manufacturing opportunities remain in the West
Unitree's robot production still relies heavily on manual work, with labor accounting for just 8 percent of manufacturing costs in 2025. Frequent product changes and reprogramming costs hinder automation, but U.S. factory labor costs about 7.5 times more than China's, making the same automation equipment far more cost-effective. Combined with emerging restrictions on Chinese robots and components, the West still has opportunities in robot manufacturing.
Pangram releases AI-generated image and video detection models
Pangram unveiled Image 1.0, an AI-image detection model, and Image Flash 1.0, a lightweight model for video detection. Image 1.0 marks AI-generated regions in real photos with a heatmap and handles compression, downscaling, and blur. In Pangram's own evaluation using 100 real photos and 1,200 images from 12 generative models, Image 1.0 achieved 99.96 percent accuracy. The minimum input size was reduced from 512×512 to 256×256.
YeJin unveils Korean-language 8B foundation model built from scratch
AI startup YeJin released YeJin 8B v3, an 8 billion-parameter foundation model optimized for Korean. The model was developed entirely from scratch, from tokenizer design through pretraining, with similarity scores below 0.0014 compared to six public models. YeJin trained the model using eight Nvidia H200 GPUs supported by the Ministry of Science and ICT and the National IT Industry Promotion Agency.
ABB Robotics and Nvidia partner to integrate Omniverse with RobotStudio
ABB Robotics and Nvidia partnered to integrate Omniverse and RobotStudio and advance physical AI capabilities.