Today's briefing
- Physical AI
AGCO and Agility Robotics plan humanoid robot pilot at Kansas plant
Physical AIGoogle NewsNov 8, 2026 - Agents
Google AI releases Elephant-Goldfish model plugin for spec-first coding agents
AgentsMarkTechPostNov 8, 2026 - LLM
Ant Group releases Ling 3.2 Flash, outperforming Claude Opus 5 on some vision tasks
LLMKorean outletAI TimesNov 8, 2026
Open source & papers radar
- RepoOpen Source & ToolsRepository: anthropics/claude-quickstartsStars: 19,355New stars today: 28Language: Python
claude-quickstarts, quick start projects for building with the Claude API
A collection of projects designed to help developers quickly get started with building deployable applications using the Claude API.
- RepoPhysical AIRepository: google-deepmind/mujocoStars: 16,337New stars today: 26Language: C++
MuJoCo, a general purpose physics simulator for multi-joint dynamics
MuJoCo is a general purpose physics simulator designed to handle multi-joint dynamics with contact, widely used in robotics control and simulation.
- PaperResearcharXiv 2611.08237Upvotes: 11
TB-MathModel, the first verifiable benchmark for agentic mathematical modeling
The authors introduce TB-MathModel, a benchmark for agentic mathematical modeling that verifies both correctness and physical feasibility using programmatic gates and 624 Lean4 obligations. Tested on 18 real CUMCM and MCM problems from 2023–2025, the strongest model reaches a multi-reference score of 64.58 percent.
- PaperAgentsarXiv 2610.14157Citations: 3
OpenMobile-2 trains on 110+ apps with 12K trajectories and hybrid GUI-tool use
OpenMobile-2 is a mobile agent trained with fully open environments and recipes, including MobileGym++ with 35 commercial-style apps and over 110 apps total. The team curated nearly 12K mobile interaction trajectories for supervised fine-tuning and over 2K executable RL tasks. The agent supports hybrid GUI and app-native tool use, achieving 79.9 on AndroidWorld and 50.4 on MobileWorld.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
10 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
CHINGMU enters South Korea with robot learning data collection solutions
CHINGMU, a physical AI technology company that partners with major robotics firms including Unitree and AgiBot, entered the South Korean market with solutions for collecting robot training data. The Shanghai-based motion-capture specialist appointed a new Korea office head to expand domestic operations.
ETNewsNov 8, 2026 - Physical AI
Pudu Robotics enters humanoid market to compete with Tesla
Chinese commercial service robot company Pudu Robotics is building humanoids as a next-generation growth driver and entering the global market. The company plans to leverage its existing expertise in serving, cleaning, and logistics robots to compete with Tesla and other global humanoid firms.
iRobotNewsNov 8, 2026 - Physical AI
AgiBot humanoids work 24-hour convenience store shift; four gain permanent positions
Six AgiBot Genie G2 humanoid robots worked a 24-hour shift at a Shanghai convenience store, completing 1,730 tasks. After evaluation, four were hired permanently and two were not retained.
iRobotNewsNov 8, 2026 - Industry & Policy
Google acquires $548 million stake in Qualcomm through Modular acquisition
Google obtained $548 million in Qualcomm shares as part of Qualcomm's acquisition of Modular, an AI semiconductor software company, in the third quarter. The stake comes as Qualcomm shifts focus from smartphones to AI data centers.
AI TimesNov 8, 2026
Stories in this issue
22 storiesAnthropic's Claude helps resolve 40-year protein structure mystery in 34 hours
Anthropic's biology team used Claude to discover a new form of the vault particle and determine its 3D structure, resolving a 40-year mystery. They completed the initial 3D analysis in 34 hours, compared to the weeks previously required, showing AI can assist researchers in underexplored fields.
claude-quickstarts, quick start projects for building with the Claude API
A collection of projects designed to help developers quickly get started with building deployable applications using the Claude API.
MuJoCo, a general purpose physics simulator for multi-joint dynamics
MuJoCo is a general purpose physics simulator designed to handle multi-joint dynamics with contact, widely used in robotics control and simulation.
Apple pushes to make Siri more agentic, racing against OpenAI and Meta
Apple CEO John Ternus is pressing AI chief Craig Federighi to accelerate the Siri roadmap and deliver an AI agent to compete with OpenAI's Dots and Meta's Muse. Two significant improvements are expected in the coming months, including fine-grained control over Apple's first-party apps like Photos and Safari.
Tokyo University develops 3D-printed robot leg mimicking muscle, tendon, and ligament
Researchers at the University of Tokyo developed a full-scale robot leg using 3D printing to replicate the mechanical properties of human muscle, tendon, and ligament. They created distinct physical characteristics from a single flexible material and demonstrated actual leg movement, opening new possibilities for biomimetic humanoid robots.
AWS launches Kiro Bots to coordinate multiple AI agents for developers
Amazon Web Services released Kiro Bots, which coordinates the work of multiple AI agents and automates repetitive development tasks. The agents remember past work and developer rules to continuously manage tasks, reducing developer overhead in supervising multiple agents.
TB-MathModel, the first verifiable benchmark for agentic mathematical modeling
The authors introduce TB-MathModel, a benchmark for agentic mathematical modeling that verifies both correctness and physical feasibility using programmatic gates and 624 Lean4 obligations. Tested on 18 real CUMCM and MCM problems from 2023–2025, the strongest model reaches a multi-reference score of 64.58 percent.
OpenMobile-2 trains on 110+ apps with 12K trajectories and hybrid GUI-tool use
OpenMobile-2 is a mobile agent trained with fully open environments and recipes, including MobileGym++ with 35 commercial-style apps and over 110 apps total. The team curated nearly 12K mobile interaction trajectories for supervised fine-tuning and over 2K executable RL tasks. The agent supports hybrid GUI and app-native tool use, achieving 79.9 on AndroidWorld and 50.4 on MobileWorld.
Harness² improves agents with recursive harness refinement across three benchmarks
Harness² is a recursive agent harnessing framework that refines execution environments without ground-truth verifiers. Through task-level and harnessing-level recursion, it extracts experience to adapt agents to future tasks, improving GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench, 4.5 on JobBench, and 11.8 percentage points on LAB.
AGCO and Agility Robotics plan humanoid robot pilot at Kansas plant
AGCO and Agility Robotics plan to conduct a humanoid robot pilot project at a manufacturing facility in Kansas.
Pudu Robotics enters humanoid market to compete with Tesla
Chinese commercial service robot company Pudu Robotics is building humanoids as a next-generation growth driver and entering the global market. The company plans to leverage its existing expertise in serving, cleaning, and logistics robots to compete with Tesla and other global humanoid firms.
AgiBot humanoids work 24-hour convenience store shift; four gain permanent positions
Six AgiBot Genie G2 humanoid robots worked a 24-hour shift at a Shanghai convenience store, completing 1,730 tasks. After evaluation, four were hired permanently and two were not retained.
Google AI releases Elephant-Goldfish model plugin for spec-first coding agents
Google AI released the Elephant-Goldfish model plugin, which directs coding agents to draft specifications first, then runs through 5 zero-context review stages before generating code.
CHINGMU enters South Korea with robot learning data collection solutions
CHINGMU, a physical AI technology company that partners with major robotics firms including Unitree and AgiBot, entered the South Korean market with solutions for collecting robot training data. The Shanghai-based motion-capture specialist appointed a new Korea office head to expand domestic operations.
Ant Group releases Ling 3.2 Flash, outperforming Claude Opus 5 on some vision tasks
Chinese tech company Ant Group unveiled Ling 3.2 Flash, a multimodal AI model with 560 billion parameters that activates 25 billion during inference. It outperforms Anthropic's Claude Opus 5 on some vision benchmarks and offers lower costs and a 1 million token context window to target the enterprise AI agent market.
Google acquires $548 million stake in Qualcomm through Modular acquisition
Google obtained $548 million in Qualcomm shares as part of Qualcomm's acquisition of Modular, an AI semiconductor software company, in the third quarter. The stake comes as Qualcomm shifts focus from smartphones to AI data centers.
Adversarial speculative reasoning improves long-horizon agents
The authors argue that speculative decoding can be used as more than a systems accelerator for long-horizon agent reasoning. They show that numerical noise accumulates during verification, affecting model output distribution after 30 steps. By training the draft model to be rejected at critical reasoning points, they improve end-to-end success rate on complex agentic coding benchmarks without changing the target model or its output distribution.
Switch-SWE measures coding agents' focus when switching between unfinished tasks
The authors introduce Switch-SWE, a benchmark that measures whether coding agents can maintain focus and resume tasks after intervening work on other issues in the same repository. The benchmark contains 200 focus tasks from 19 Python repositories. Across 7 coding agents, interleaved unfinished work produces an average 11.5 percentage-point gap in focus-task completion, with a context-management module improving performance by 5.5 to 15.5 percentage points.
Agility Robotics debuts on Nasdaq as humanoid robot race intensifies
Agility Robotics went public on Nasdaq as competition in the humanoid robot market accelerates.
Anthropic study: scientists use AI mostly for data analysis, only 4% for hypothesis generation
Anthropic released a report analyzing how scientists use AI in research. Researchers ranked hypothesis generation as the most important activity for scientific discovery but used AI for it only 4 percent of the time. The analysis examined 207,049 Claude and Claude Code sessions with scientific research content from August 2 through September 30.
Physicl raises $10 million for Physical AI training data platform
Physicl raised $10 million for its Physical AI training data platform.
OpenAI releases official prompt guide for Image 2.6
OpenAI published an official prompt guide for its Image 2.6 model on its developer documentation. The guide emphasizes specifying use case, composition, and style for image generation, and clearly distinguishing what to change versus preserve when editing existing images.