Today's briefing
- Physical AI
South Korea plans KRW 147.7 billion robot spending for 2027, including 200 humanoid units
Physical AIKorean outletiRobotNewsNov 4, 2026 - Agents
DeepSeek launches Creator Mode to generate AI agent plugins from natural language
AgentsKorean outletAI TimesNov 4, 2026 - Open Source & Tools
OpenAI introduces Pando, a Git backend that cuts median push latency by nearly 80%
Open Source & ToolsMarkTechPost +1 moreNov 3, 2026
Open source & papers radar
- RepoAgentsRepository: max-sixty/worktrunkStars: 10,418New stars today: 437Language: Rust
Worktrunk, a CLI for Git worktree management designed for parallel AI agent workflows
Worktrunk is a command-line tool for managing Git worktrees, designed to support parallel AI agent workflows in development environments.
- PaperAgentsarXiv 2611.03178Upvotes: 86
CoworkOS framework scales long-horizon agent workflows with human-in-the-loop integration
Researchers introduce CoworkOS, a framework for long-horizon work that separates task strategy from execution management and enables agents to delegate tasks, synchronously or asynchronously. Across 12 benchmarks and 6 real-world application categories, CoworkOS proves competitive with frontier agent applications such as Codex and DeepSeek Harness.
- PaperAgentsarXiv 2611.03126Upvotes: 90
CUA-Master scales computer-use agents through open-world recursive improvements
Researchers introduce CUA-Master, a nested double-loop framework that improves computer-use agents by mining authentic workflows from instructional videos and applying recursive refinement. CUA-Master-27B scores 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam.
- PaperPhysical AIarXiv 2610.04933
DiVeR improves Vision-Language-Action policy inference by learning decision-critical states
Researchers propose DiVeR, which estimates decision criticality from the dispersion of sampled action representations to guide verifier-based action selection in Vision-Language-Action policies. Across LIBERO, RoboCasa, and real-world experiments on a Franka Research 3 robot, DiVeR improves task success with negligible verifier inference overhead.
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.
- Agents
Weverse Brain launches action-taking AI agents for education and sales
Weverse Brain released AI agents that perform as English tutors and sales consultants. The agents handle phone conversations, respond to unexpected questions, and use natural dialogue to encourage learning and support purchase decisions.
Byline NetworkNov 4, 2026 - Physical AI
Vectrobotics unveils physical AI vision connecting manufacturing, logistics, and service sectors
Vectrobotics introduced ROXI, an intelligent robot platform that combines hardware and AI to perceive environment, judge tasks, and execute actions. The platform targets manufacturing, logistics, and service operations at RoboWorld 2026.
iRobotNewsNov 3, 2026
Stories in this issue
23 storiesOpenAI says GPT-6 Astra designed a hardware device from scratch in under a week
OpenAI's hardware team used GPT-6 Astra to design a device for the Codex coding agent in under a week, handling enclosure, circuit board, and firmware. The model required several revision rounds before production because it did not account for manufacturing tolerances.
AI inference startup Fal doubles valuation to $16 billion with $1.3 billion funding
AI inference startup Fal raised $1.3 billion in a round led by Bessemer Venture Partners, doubling its valuation to $16 billion. The five-year-old company is capitalizing on surging demand for AI inference services.
Worktrunk, a CLI for Git worktree management designed for parallel AI agent workflows
Worktrunk is a command-line tool for managing Git worktrees, designed to support parallel AI agent workflows in development environments.
CoworkOS framework scales long-horizon agent workflows with human-in-the-loop integration
Researchers introduce CoworkOS, a framework for long-horizon work that separates task strategy from execution management and enables agents to delegate tasks, synchronously or asynchronously. Across 12 benchmarks and 6 real-world application categories, CoworkOS proves competitive with frontier agent applications such as Codex and DeepSeek Harness.
CUA-Master scales computer-use agents through open-world recursive improvements
Researchers introduce CUA-Master, a nested double-loop framework that improves computer-use agents by mining authentic workflows from instructional videos and applying recursive refinement. CUA-Master-27B scores 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam.
Xiaomi humanoid robot completes first 8-hour factory shift
Xiaomi's humanoid robot completed its first full 8-hour shift at a car factory, while the Xiaomi-Robotics-1 model simultaneously entered apartment trial deployment.
Anthropic's Message Threads cut time to first token by 35 percent for long conversations
Anthropic released Message Threads for the Messages API, which allows agents to transmit only changes instead of resending entire conversations. For requests between 500,000 and one million tokens, time to first token drops by approximately 35 percent.
DiVeR improves Vision-Language-Action policy inference by learning decision-critical states
Researchers propose DiVeR, which estimates decision criticality from the dispersion of sampled action representations to guide verifier-based action selection in Vision-Language-Action policies. Across LIBERO, RoboCasa, and real-world experiments on a Franka Research 3 robot, DiVeR improves task success with negligible verifier inference overhead.
FLAMIA speeds up attention computation on Blackwell GPUs by 1.1 to 1.4 times
Researchers present FLAMIA, a family of attention kernels for NVIDIA Blackwell GPUs using MXFP8 and mixed NVFP4/FP8 training. The design addresses quantization overhead, causal information leakage, and numerical inconsistencies, accelerating forward and backward attention calls by 1.10 to 1.40 times on B200 and 1.06 to 1.42 times on B300.
South Korea plans KRW 147.7 billion robot spending for 2027, including 200 humanoid units
South Korea's Ministry of Trade, Industry and Resources will invest KRW 147.7 billion in robotics in 2027, including KRW 37 billion for a new K-robot ecosystem advancement R&D program and KRW 80.7 billion for intelligent robot deployment and expansion. The government aims to localize core components and create initial demand for humanoid robots.
DeepSeek launches Creator Mode to generate AI agent plugins from natural language
DeepSeek unveiled Creator Mode, which lets users describe desired features in natural language and automatically generates, installs, and validates plugins for AI agents. The feature enables users to build custom tools without programming knowledge.
Google Antigravity accelerates drug discovery with agentic cheminformatics
Google Antigravity and cheminformatics skills accelerate drug discovery workflows by searching compounds and proposing novel rigidified macrocycles.
AI agents take over at SETI's Allen Telescope Array to scan for extraterrestrial intelligence
The SETI Institute and Breakthrough Listen deployed AI agents at the Allen Telescope Array to control the telescope and analyze radio signals. Tasks that previously took weeks now complete in hours, and the team detected a pulsar in the Crab Nebula in real time for the first time.
Weverse Brain launches action-taking AI agents for education and sales
Weverse Brain released AI agents that perform as English tutors and sales consultants. The agents handle phone conversations, respond to unexpected questions, and use natural dialogue to encourage learning and support purchase decisions.
Reverse game theory aims to keep AI agents aligned with human values
A new approach called reverse game theory aims to control powerful AI models by designing systems that incentivize them to follow rules. As AI models become more capable, safety researchers are working to ensure agents remain aligned with human desires.
OpenAI introduces Pando, a Git backend that cuts median push latency by nearly 80%
OpenAI built Pando, a Git-compatible backend for its monorepo that handles branches independently and reuses stored packfiles across fetches. The system reduced median push latency by nearly 80% and now delivers nearly a petabyte of Git data per day.
LLMs can read and write neural network weights in Weight Space Programming
Researchers show that large language models can directly edit neural network weights to reduce reasoning loops. The work covers weight editing in Qwen, DeepSeek's super-agent prototype, and the GDPx benchmark.
Apple gives Siri AI access to third-party apps in iOS 27.3
Apple announced in iOS 27.3 that Siri can now perform actions across third-party apps, drawing on personal context, processing natural language, and using on-screen awareness. Third-party developers can create app skills that combine multiple tasks into a single action.
Stanford study finds 44 percent of LLM research claims already outdated
Stanford researchers tested 50 widely cited claims about LLMs against 20 current models and found 44 percent of findings already expired and 34 percent weakened significantly. Claims about model limitations expire fastest, with an average lifespan of 15.7 months.
Apple's watchOS 27.3 beta adds Audio Intelligence features to Siri on Apple Watch
Apple added Audio Intelligence features to Siri in watchOS 27.3 developer beta: Siri Recap generates summaries from ambient listening, and Live Rewind creates text transcripts of the last 15 seconds of audio. The iOS 27.3 beta also includes a redesigned Health app with expanded language support.
Vectrobotics unveils physical AI vision connecting manufacturing, logistics, and service sectors
Vectrobotics introduced ROXI, an intelligent robot platform that combines hardware and AI to perceive environment, judge tasks, and execute actions. The platform targets manufacturing, logistics, and service operations at RoboWorld 2026.
OpenAI quadruples bug bounty payouts to up to $172,000
OpenAI increased maximum bug bounty rewards to $172,000, quadrupling previous payouts, and launched a new internal security sprint to strengthen security.
D&A decouples description and assessment for zero-shot robotic reward models
Researchers present D&A, a method that separates progress estimation into scene description and assessment phases to mitigate confirmation bias in vision-language models used as robotic reward models. The prompt-based approach generalizes across tasks and embodiments while enabling efficient downstream reinforcement learning.