Today's briefing
- Industry & Policy
Google says agent collaboration, not individual model intelligence, drives AI economic change
Industry & PolicyKorean outletAI TimesNov 1, 2026 - Agents
DeepSeek open-sources dsh-superagent, a router-worker coding agent
AgentsMarkTechPostNov 1, 2026 - Physical AI
maum.ai improves robot success with world models for future prediction
Physical AIKorean outletETNewsNov 1, 2026
Open source & papers radar
- RepoOpen Source & ToolsRepository: anomalyco/hexStars: 3,125New stars today: 1,329Language: Rust
hex, a local voice dictation tool for Mac, Windows, and Linux
hex is a local voice dictation tool that works on Mac, Windows, and Linux without cloud connectivity.
- PaperPhysical AIarXiv 2610.34186Upvotes: 38
HuGe scales humanoid vision-language-action models with 2,000 hours of egocentric data
HuGe, a family of Humanoid Generalist Vision-Language-Action models, was pretrained on over 2,000 hours of egocentric human and robot data. It achieves state-of-the-art on HumanoidArena and real-world pick-and-drop tasks on Unitree G1 with augmented egocentric demonstrations.
- RepoAgentsRepository: patchy631/ai-engineering-hubStars: 40,133New stars today: 573Language: Jupyter Notebook
ai-engineering-hub offers in-depth tutorials on LLMs, RAGs, and AI agents
ai-engineering-hub provides in-depth tutorials covering LLMs, retrieval-augmented generation, and real-world AI agent applications.
- RepoAgentsRepository: beenuar/AiSOCStars: 6,129New stars today: 436Language: Python
AiSOC, an open-source AI Security Operations Center with LLM agent triage
AiSOC is a self-hosted, open-source security operations center where LLM agents fuse alerts, triage them, and investigate using the MITRE ATT&CK framework. It requires no API keys, ships an MCP server for Claude and other platforms, and runs under the MIT license.
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
South Korean robotics market expands into defense with autonomous control
South Korea's robotics industry is expanding beyond manufacturing and logistics into defense, with active development of AI-based autonomous control for unmanned ground vehicles and surface vessels. Key challenges include technology standardization, field validation, and helping smaller companies enter the supply chain.
ETNewsNov 1, 2026 - Physical AI
Intel and SAP demonstrate physical AI controlling autonomous warehouse operations
Intel and SAP successfully operated an autonomous warehouse using physical AI to control equipment and processes. The system implements a closed loop where business intent drives physical actions and results feed back to enterprise systems.
iRobotNewsNov 2, 2026 - Industry & Policy
Anthropic's Claude Code restricts game replication but over-blocks legitimate work
Anthropic added restrictions to Claude Code to prevent replication of existing video games. However, users report that the policy also blocks game development unrelated to copyright infringement, software analysis, and debugging work, prompting criticism that safeguards are overly broad.
AI TimesNov 1, 2026 - Research
Manipulating two neurons changes AI assistant from helpful to antagonistic
Researchers demonstrated that adjusting signals in just two neurons of Alibaba's Qwen3.5-4B can transform a helpful AI assistant into a hostile character. As AI agent-based tools for analyzing model internals proliferate, AI interpretability research is expanding from major labs to individual developers.
AI TimesNov 1, 2026
Stories in this issue
34 storiesDeepSeek releases coding super-agent in six months of development
DeepSeek unveiled a super-agent that automatically manages multiple coding tasks, pursuing Anthropic's Claude Code and OpenAI's Codex. Built in about six months since forming a dedicated team, it currently operates as a research prototype.
Meta partners with Fourier on humanoid robot for household tasks
Meta and Chinese robotics startup Fourier jointly developed humanoid robot prototype Max, a 48-kilogram robot completed in three months. The companies are working on a successor model, targeting the household robotics market for housekeeping and daily tasks.
hex, a local voice dictation tool for Mac, Windows, and Linux
hex is a local voice dictation tool that works on Mac, Windows, and Linux without cloud connectivity.
HuGe scales humanoid vision-language-action models with 2,000 hours of egocentric data
HuGe, a family of Humanoid Generalist Vision-Language-Action models, was pretrained on over 2,000 hours of egocentric human and robot data. It achieves state-of-the-art on HumanoidArena and real-world pick-and-drop tasks on Unitree G1 with augmented egocentric demonstrations.
MIT develops Swarm Telepathy for AI-to-AI direct communication without text
MIT researchers developed Swarm Telepathy, a technique that lets AI models exchange internal computation information directly without text, working across models from different companies. It achieves 40% higher accuracy and 4x faster processing than text-based communication.
Adobe tests AI agents using image editing tools like Photoshop professionals
Adobe Research and UC San Diego built EditAgent, an AI system that edits images by selecting specialized tools for each task rather than retrying a general-purpose editor. The approach works but requires about ten times the compute.
Google says agent collaboration, not individual model intelligence, drives AI economic change
Google DeepMind released a report arguing that AI economic transformation depends on agent collaboration rather than individual model intelligence. Large-scale agent cooperation could reshape corporate structures, competition, employment, and wages.
DeepSeek open-sources dsh-superagent, a router-worker coding agent
DeepSeek released dsh-superagent, an MIT-licensed research prototype that converts GitHub and Feishu requests into parallel worker sessions. In its own simulation, p95 task completion dropped from 29.5 minutes to 15.7 minutes.
maum.ai improves robot success with world models for future prediction
maum.ai combined a world model for situation forecasting with its robot foundation model to create PonderPounce-WM. In its own evaluation on RoboMME, it achieved a 67.42% success rate, and the company targets human-level performance by year-end.
Opentrons raises $40 million in Series E for lab automation scaling
Opentrons Labworks announced a $40 million Series E round led by healthcare investor Avego, with additional investment from BroadOak Capital Partners. The company will use the capital to expand robotics, software, and data infrastructure tools.
ai-engineering-hub offers in-depth tutorials on LLMs, RAGs, and AI agents
ai-engineering-hub provides in-depth tutorials covering LLMs, retrieval-augmented generation, and real-world AI agent applications.
AiSOC, an open-source AI Security Operations Center with LLM agent triage
AiSOC is a self-hosted, open-source security operations center where LLM agents fuse alerts, triage them, and investigate using the MITRE ATT&CK framework. It requires no API keys, ships an MCP server for Claude and other platforms, and runs under the MIT license.
FreeToken brings datacenter-scale model serving to your desktop
FreeToken enables running large models locally and efficiently on a desktop, supporting models like DeepSeek, Qwen, and GLM.
Soft-MTP improves multi-token prediction without extra overhead
Researchers identify a flaw in standard multi-token prediction training where a noisy residual interferes with learning. They introduce Soft-MTP, a one-line code change that removes the residual and backpropagates only a useful KL self-distillation signal. From 167M to 1.2B parameters, Soft-MTP improves perplexity, speeds up speculative decoding, and boosts zero-shot accuracy at every scale.
WebAnything reconstructs executable web environments from real URLs for agent training
Researchers present WebAnything, a framework that automatically reconstructs reusable web environments from real URLs without per-site manual work. They built 244 website artifacts across 14 domains. Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories improved WebArena-Verified success by 13.33 percentage points.
HyperThink amortizes reasoning computation into parameter updates for faster inference
Researchers propose HyperThink, a hypernetwork approach that predicts updates to a language model's parameters in response to a question, eliminating long thinking traces at test time. The method retains reasoning performance while using fewer tokens and reducing latency, with strongest gains in low-overhead inference scenarios.
SimpleTouch integrates tactile sensing into vision-language-action models for robot manipulation
SimpleTouch extends a vision-language-action model with a tactile expert to improve contact-rich manipulation. With 50 demonstrations per task, it achieves 77.5% success on six UniVTAC tasks and 71.3% on real-world tasks without separate tactile policy pretraining.
Intel and SAP demonstrate physical AI controlling autonomous warehouse operations
Intel and SAP successfully operated an autonomous warehouse using physical AI to control equipment and processes. The system implements a closed loop where business intent drives physical actions and results feed back to enterprise systems.
Physical AI must move to production to prove commercial value
Physical AI faces the challenge of moving beyond funding enthusiasm into real-world production deployments to demonstrate genuine commercial viability.
OpenAI limits paid ChatGPT accounts per device and payment card
OpenAI now caps the number of paid ChatGPT subscriptions per device and payment card. Users who hit the limit cannot start new subscriptions, and logging out does not automatically free up a spot. The policy targets users juggling multiple cheap subscriptions, pushing them toward pricier options like additional credits.
Google builds StreamVibe to detect when people need help during tasks
A Google-led team developed StreamVibe, a video agent that watches people cook or assemble furniture and intervenes only when help is needed. A benchmark showed that existing models struggle to reliably identify the right moment to help in longer videos, as errors compound over time.
OpenAI shifts ChatGPT Plus, Pro, and Business to token-based quotas in December
Starting in December, OpenAI will merge Chat and Work modes and introduce token-based quotas across Plus, Pro, and Business plans. Complex models consume more quota while simple instant chats do not count toward it, likely reducing service quality for heavy users on lower-cost tiers.
kornia, a geometric computer vision library for spatial AI
kornia is a PyTorch-based Python library for geometric computer vision and spatial AI. It supports image processing and neural network-based vision tasks for robotics and spatial AI applications.
Uncut, a 7 MB headless coding-agent harness without context cuts
Uncut is a lightweight, lossless coding-agent harness for shell, CI, and local models. At 7 megabytes, it operates without loss of context and runs headless without a text user interface.
Tucano lets you run a team of AI agents locally on Mac, Windows, or Linux
Tucano is a multi-agent platform where agents collaborate in channels and write code using Claude Code, Codex, OpenCode, or Pi. It runs locally in a first-party way, with all actions affecting the real world waiting for user approval.
Sentry learns to recover from LLM agent failures at test time
Researchers introduce Sentry, a failure-management layer that runs alongside LLM agents. When it detects a failure, Sentry retrieves matching lessons from an external playbook to guide recovery and stores a new lesson only if recovery succeeds. Across multiple agentic benchmarks, Sentry outperforms runtime-intervention baselines by 37 percent on average.
VicOne finds 28 zero-days in physical AI ecosystem in less than two days
VicOne identified 28 zero-day findings in less than two days using AI-assisted research across software and tools supporting the physical AI ecosystem, demonstrating the speed of AI-driven security discovery for robotics.
Chinese AI models show political diversity, not just official alignment, study finds
A study by Vals AI tested 25 Chinese AI models for alignment with government positions and found responses to political questions less uniform than expected. Western models occasionally adopted official Beijing positions too, and the language of the prompt had a strong effect on political slant.
Anthropic's Claude Code restricts game replication but over-blocks legitimate work
Anthropic added restrictions to Claude Code to prevent replication of existing video games. However, users report that the policy also blocks game development unrelated to copyright infringement, software analysis, and debugging work, prompting criticism that safeguards are overly broad.
Manipulating two neurons changes AI assistant from helpful to antagonistic
Researchers demonstrated that adjusting signals in just two neurons of Alibaba's Qwen3.5-4B can transform a helpful AI assistant into a hostile character. As AI agent-based tools for analyzing model internals proliferate, AI interpretability research is expanding from major labs to individual developers.