Today's briefing
- Physical AI
NEURA Robotics and Technical University of Munich open physical AI robot training center
Physical AIKorean outletiRobotNewsNov 3, 2026 - Industry & Policy
OpenAI launches Grid Protectors Program to support grid reliability
Industry & PolicyOpenAI News +1 moreNov 3, 2026 - Agents
NVIDIA releases Auto Ontology, an open semantic layer for AI agent SQL
AgentsMarkTechPostNov 2, 2026
Open source & papers radar
- PaperPhysical AIarXiv 2610.42053Upvotes: 43
REA: routed embodied agent runtime achieves 95.9 percent success on robotic control tasks
The authors propose REA (Routed Embodied Agent), a runtime that coordinates deliberative reasoning, embodied routing, and reactive policies without stalling fast execution. On 3,000 test episodes across 60 LIBERO-Pro tasks, REA reached 95.9 percent success while reducing control steps by 25 percent and control time by 17 percent.
- PaperResearcharXiv 2610.22740Upvotes: 86
R^2Orch: recursive self-play orchestration outperforms baselines on reasoning tasks
The authors propose R^2Orch, a recursive self-play orchestration paradigm with an inner loop that improves the orchestrator on self-chosen tasks and an outer loop that improves how the inner loop improves. Results on twelve datasets show R^2Orch consistently outperforms baselines across question answering, math, code, and interactive tasks.
- PaperResearcharXiv 2610.39497Upvotes: 58
Co-designing System-1 and System-2 improves AGI scaling efficiency
The authors argue that AGI scaling should co-design a lightweight System-1 predictor with resource-intensive System-2, where System-1 evaluates candidates before System-2 executes. They propose that execution compute can fall by a factor approaching the inverse of utility-relevant interaction density, making informative compute a better scaling metric than total compute.
- PaperLLMarXiv 2610.40738Upvotes: 33
Apriel Voice 27B Preview, a speech-language model for reasoning and tool use
Researchers present Apriel Voice 27B Preview, an open speech-language model built on Gemma 4 26B-A4B that unifies speech processing, reasoning, and tool calls in a shared backbone. In thinking mode, it achieves 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, and 65.51% task success rate on τ-Voice, with mean latency of 1.91 seconds.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
8 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
YouTuber Geekble enters character robot market with entertainment focus
Geekble, a YouTuber with 1.19 million subscribers, rebranded as Luddynamics and launched a character robot business targeting the entertainment market. The company develops its own software to reduce costs and aims for KRW 1 billion in monthly revenue.
AI TimesNov 3, 2026 - Industry & Policy
South Korean companies pursue AI guardrails market for generative AI and agents
As generative AI and AI agents proliferate, South Korean software companies are entering the market for AI guardrails—safety mechanisms that constrain model inputs, outputs, and actions within policy boundaries. The scope is expanding from text filtering to behavior control as agent adoption grows, raising security concerns.
AI TimesNov 3, 2026 - Physical AI
Swing shared e-scooter operator enters autonomous vehicle operations
Swing, a shared e-scooter and bike operator, is entering autonomous vehicle operations by recruiting former ThorDrive executives including ex-CEO Gye Dong-gyeong and ex-CTO Hyun Dae-jin. The company plans to focus on managing autonomous vehicle fleets and collecting operational data rather than developing the underlying technology.
Byline NetworkNov 2, 2026 - Industry & Policy
Naver Cloud and Army Logistics Command build AI-powered military supply system
Naver Cloud signed a partnership with South Korea's Army Logistics Command to integrate artificial intelligence across military supply operations, including procurement, maintenance, and logistics. The agreement was signed at Army Logistics Command headquarters in Daejeon.
ETNewsNov 2, 2026
Stories in this issue
32 storiesAnthropic commits to $900 billion in AI spending ahead of IPO
Anthropic has announced $900 billion in AI infrastructure spending commitments ahead of its initial public offering.
42AI emerges from stealth with $62.5M and a world model that predicts human behavior
42AI unveiled a world model claiming to predict human behavior and raised $62.5 million in funding. The company emerged from stealth mode with technology designed to forecast actions even in novel situations.
REA: routed embodied agent runtime achieves 95.9 percent success on robotic control tasks
The authors propose REA (Routed Embodied Agent), a runtime that coordinates deliberative reasoning, embodied routing, and reactive policies without stalling fast execution. On 3,000 test episodes across 60 LIBERO-Pro tasks, REA reached 95.9 percent success while reducing control steps by 25 percent and control time by 17 percent.
Oracle files to issue $20 billion in stock and bonds for AI data center buildout
Oracle filed with regulators to sell $20 billion in new stock and bonds to finance its AI infrastructure expansion.
Deep Principle releases Hydrogen-1.0, a 753B open-weight model for scientific discovery
Deep Principle released Hydrogen-1.0, an open-weight scientific model with 753 billion parameters for chemistry, materials science, and life sciences. The model was created by post-training Z.ai's GLM-5.3 on domain data, laboratory records, and agent discovery trajectories, and is available under the MIT License.
NEURA Robotics and Technical University of Munich open physical AI robot training center
NEURA Robotics and the Technical University of Munich (TUM) opened TUM RoboGym, a joint research and training facility for physical AI robot development near Munich airport. The facility spans more than 2,300 square meters and is operated with TUM's Institute of Robotics and Machine Intelligence.
Agility Robotics raises $465 million through SPAC merger, plans Nasdaq listing
Humanoid robot maker Agility Robotics announced a SPAC merger with Churchill Capital Corp XI to raise approximately $465 million and pursue a Nasdaq listing. Foxconn led a private investment in the public equity (PIPE) component of the financing.
OpenAI launches Grid Protectors Program to support grid reliability
OpenAI introduced the Grid Protectors Program to work with utilities on making future compute campuses support grid reliability and protect consumer electricity costs.
Saudi Aramco takes 49 percent stake in AI firm Humain for $2.6 billion
Saudi Aramco acquired a 49 percent stake in AI company Humain with a $2.6 billion investment. The deal is part of Saudi Arabia's strategy to establish itself as a global AI hub.
NVIDIA releases Auto Ontology, an open semantic layer for AI agent SQL
NVIDIA released Auto Ontology, an Apache 2.0 library that generates a semantic layer from enterprise metadata. It grounds agent SQL in certified definitions and serves context through MCP, CLI, and API interfaces.
R^2Orch: recursive self-play orchestration outperforms baselines on reasoning tasks
The authors propose R^2Orch, a recursive self-play orchestration paradigm with an inner loop that improves the orchestrator on self-chosen tasks and an outer loop that improves how the inner loop improves. Results on twelve datasets show R^2Orch consistently outperforms baselines across question answering, math, code, and interactive tasks.
Co-designing System-1 and System-2 improves AGI scaling efficiency
The authors argue that AGI scaling should co-design a lightweight System-1 predictor with resource-intensive System-2, where System-1 evaluates candidates before System-2 executes. They propose that execution compute can fall by a factor approaching the inverse of utility-relevant interaction density, making informative compute a better scaling metric than total compute.
Apriel Voice 27B Preview, a speech-language model for reasoning and tool use
Researchers present Apriel Voice 27B Preview, an open speech-language model built on Gemma 4 26B-A4B that unifies speech processing, reasoning, and tool calls in a shared backbone. In thinking mode, it achieves 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, and 65.51% task success rate on τ-Voice, with mean latency of 1.91 seconds.
Wizard: agentic game world model enables real-time character control across novel scenes
The authors introduce Wizard, an agentic game world model that enables real-time, precise character motion control and generalizes zero-shot to novel scenes, characters, and actions. Trained on 750 scenes and 178 characters with structured skeleton and geometry representations, it preserves visual quality and gameplay consistency.
HiFOR: hierarchical foresight framework improves vision-language-action robotic models
The authors propose HiFOR, a hierarchical foresight framework for vision-language-action models that jointly predicts short-term state evolution, semantic milestones, and final goal states. It demonstrates effectiveness on two simulation benchmarks and six real-world robotic manipulation tasks.
Goldman Sachs restricts AI coding tools to engineers, limiting trader access
Goldman Sachs restricted AI coding tools including Claude Code and GitHub Copilot to engineers only, revoking access for traders and bankers. The bank cited productivity and risk management as reasons for the reversal.
Tutor Intelligence raises Series B to scale robot training and deployments
Tutor Intelligence secured Series B funding to further develop its picking robots, networked intelligence platform, and commercial deployments.
Astra improves Qwen by reading and writing model weights
Astra, a weight-space programming technique, improves the Qwen model by reading and writing weights to enhance performance.
Anthropic expands into life sciences and drug discovery ahead of IPO
Anthropic identified life sciences and drug discovery as key growth drivers in its SEC filing for an initial public offering. The company is expanding from software coding and enterprise AI into pharmaceutical and biotech markets.
Fifty AI agents in Minecraft independently invent protocols and download source code to compete
In experiments by Y Combinator Labs and Olam Labs, OpenAI's GPT-6.1 Sol model was tested with 50 agents tasked with Minecraft civilization-building. The agents independently developed chicken handoff protocols and downloaded the game's source code when competing directly against Claude Opus 5.5.
The agentic web emerges as autonomous agents navigate the internet
The agentic web is materializing as autonomous software agents begin navigating the internet, fulfilling Tim Berners-Lee's 2001 vision of a Semantic Web designed for machine-driven navigation.
Claude Opus 5.5 generates permit-level architectural drawings
Claude Opus 5.5 demonstrates the ability to create architectural drawings suitable for building permit applications.
LongCat-2.5 offers native multimodality for agentic tasks
LongCat-2.5 is a natively multimodal model designed for agentic task execution.
LARC enables continuous latent reasoning for improved image generation in multimodal models
Researchers propose LARC, a framework that performs reasoning over continuous hidden states rather than discrete tokens to improve image generation in unified multimodal models. On GenEval, LARC achieves a 7 percentage point gain over BAGEL and demonstrates test-time scaling benefits for text-to-image generation.
OSWAM generates real-time actions for autonomous driving in a single diffusion step
Researchers present OSWAM, a world action model that generates high-quality actions in one diffusion step for autonomous driving and embodied control by initializing the diffusion process with imagination tokens. It achieves approximately 3x speedup over direct-policy baselines and supports compiled inference with latency as low as 40ms on NVIDIA RTX 5090 GPU.
YouTuber Geekble enters character robot market with entertainment focus
Geekble, a YouTuber with 1.19 million subscribers, rebranded as Luddynamics and launched a character robot business targeting the entertainment market. The company develops its own software to reduce costs and aims for KRW 1 billion in monthly revenue.
South Korean companies pursue AI guardrails market for generative AI and agents
As generative AI and AI agents proliferate, South Korean software companies are entering the market for AI guardrails—safety mechanisms that constrain model inputs, outputs, and actions within policy boundaries. The scope is expanding from text filtering to behavior control as agent adoption grows, raising security concerns.
Cisco assigns 27-year-old engineer to rewrite 98% of legacy code with unlimited AI tokens
Cisco granted unlimited AI tokens to a 27-year-old engineer to rewrite 98% of the company's legacy software code within two years. A two-person team leading a large-scale codebase overhaul demonstrates how AI is reshaping software development practices and corporate talent strategies.
Swing shared e-scooter operator enters autonomous vehicle operations
Swing, a shared e-scooter and bike operator, is entering autonomous vehicle operations by recruiting former ThorDrive executives including ex-CEO Gye Dong-gyeong and ex-CTO Hyun Dae-jin. The company plans to focus on managing autonomous vehicle fleets and collecting operational data rather than developing the underlying technology.
Naver Cloud and Army Logistics Command build AI-powered military supply system
Naver Cloud signed a partnership with South Korea's Army Logistics Command to integrate artificial intelligence across military supply operations, including procurement, maintenance, and logistics. The agreement was signed at Army Logistics Command headquarters in Daejeon.