Today's briefing
- LLM
Anthropic's Claude Fable 5.2 paints images using code
LLMThe DecoderNov 16, 2026 - Physical AI
JP Morgan projects U.S. humanoid production at 400,000 units annually by 2030
Physical AIKorean outletiRobotNews +1 moreNov 15, 2026 - Industry & Policy
Google connects Wallet to Gemini for spending analysis and savings planning
Industry & PolicyKorean outletAI TimesNov 15, 2026
Open source & papers radar
- RepoAgentsRepository: google-labs-code/stitch-skillsStars: 9,232New stars today: 248Language: TypeScript
stitch-skills, a library of Agent Skills for the Stitch MCP server
stitch-skills is a library of Agent Skills designed to work with the Stitch MCP server. Each skill follows the Agent Skills open standard for compatibility with coding agents such as Antigravity, Gemini CLI, Claude Code, and Cursor.
- RepoAgentsRepository: NVIDIA/SkillEvaluatorStars: 1,576New stars today: 205Language: Python
NVIDIA SkillEvaluator, a multi-tier framework for evaluating AI agent skills
NVIDIA released SkillEvaluator, a multi-tier framework for evaluating AI agent skills. It includes quality gates, semantic overlap detection, synthetic evaluation dataset generation, and live agent evaluation that measures how skills affect agent behavior.
- RepoOpen Source & ToolsRepository: ggml-org/ggmlStars: 15,877New stars today: 53Language: C++
ggml, a tensor library for machine learning
ggml is a tensor library designed for machine learning applications.
- PaperResearcharXiv 2610.23255Citations: 2
SPLIT allocates test-time compute optimally despite imperfect verifiers
The authors present SPLIT, which determines optimal allocations of sampling, model selection, and verification budget using a formula that accounts for task difficulty, shared model blind spots, and correlated verifier mistakes. On seven benchmarks spanning math, science, code, and software-engineering agents with 26 models and real verifiers, SPLIT achieves allocation quality within 0.02–0.8 percentage points of the best option in hindsight.
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.
- Physical AI
Tesla pushes Optimus humanoid into industrial settings by 2027
Tesla is pursuing an initial commercial deployment of Optimus, its humanoid robot, in manufacturing starting in 2027, with a focus on repetitive tasks and high-risk industrial work. The company is conducting early production in Fremont, California, and continues developing technology for industrial deployment.
iRobotNewsNov 15, 2026 - Industry & Policy
Meta fined $887 million by Italy for unauthorized access to Android users' photos
Italian data protection authorities fined Meta $887 million for accessing Android users' photos and videos without explicit consent through Facebook's photo recommendation feature. Approximately 20 million Italian Android users were affected.
AI TimesNov 15, 2026
Stories in this issue
24 storiesGoogle DeepMind launches Gemini Robotics 2.5 with improved dexterity
Google DeepMind introduced Gemini Robotics 2.5, an updated vision-language-action (VLA) model designed to improve robots' ability to perform complex physical tasks in industrial and everyday environments. The company says the new model expands the range of tasks robots can perform while improving dexterity, reliability and safety.
Meituan releases a 1.6-trillion-parameter agent model
Meituan released weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents and trained entirely on Chinese hardware. The model scored 88.4 points on Terminal Bench, placing it just behind Claude Opus 5. A sample deployment runs on 194 Huawei Ascend chips.
stitch-skills, a library of Agent Skills for the Stitch MCP server
stitch-skills is a library of Agent Skills designed to work with the Stitch MCP server. Each skill follows the Agent Skills open standard for compatibility with coding agents such as Antigravity, Gemini CLI, Claude Code, and Cursor.
Avride reaches one million robot deliveries across U.S.
Avride announced that its autonomous delivery robots have completed one million deliveries across the United States. The company reported 180,000 deliveries in October alone, with operations reaching a daily volume of 7,000 deliveries. Avride currently operates in more than 30 locations.
Anthropic's Claude Fable 5.2 paints images using code
Anthropic's Claude Fable 5.2 generates images by writing Python code that places hundreds of thousands of simulated brushstrokes on a virtual canvas, rather than computing pixels. A single photorealistic parking garage scene required roughly 700,000 strokes, according to Anthropic employee Jake Eaton.
JP Morgan projects U.S. humanoid production at 400,000 units annually by 2030
U.S. investment bank JP Morgan projected that U.S. humanoid robot production could reach 400,000 units annually by 2030. The company cited accelerating deployment in manufacturing and logistics as drivers of rapid production growth, noting that humanoids can adapt to work environments designed for humans in ways conventional industrial robots cannot.
Tesla pushes Optimus humanoid into industrial settings by 2027
Tesla is pursuing an initial commercial deployment of Optimus, its humanoid robot, in manufacturing starting in 2027, with a focus on repetitive tasks and high-risk industrial work. The company is conducting early production in Fremont, California, and continues developing technology for industrial deployment.
Bland launches Tofu-1, a full-duplex speech-to-speech model
Bland released Tofu-1, a full-duplex speech-to-speech model designed for noisy, high-stakes phone calls. The model is in enterprise research preview, with no published benchmarks, size, or pricing yet announced.
NVIDIA SkillEvaluator, a multi-tier framework for evaluating AI agent skills
NVIDIA released SkillEvaluator, a multi-tier framework for evaluating AI agent skills. It includes quality gates, semantic overlap detection, synthetic evaluation dataset generation, and live agent evaluation that measures how skills affect agent behavior.
ggml, a tensor library for machine learning
ggml is a tensor library designed for machine learning applications.
SPLIT allocates test-time compute optimally despite imperfect verifiers
The authors present SPLIT, which determines optimal allocations of sampling, model selection, and verification budget using a formula that accounts for task difficulty, shared model blind spots, and correlated verifier mistakes. On seven benchmarks spanning math, science, code, and software-engineering agents with 26 models and real verifiers, SPLIT achieves allocation quality within 0.02–0.8 percentage points of the best option in hindsight.
DLite: History-free speculative decoding for agentic AI
The authors introduce DLite, a history-free speculative decoding framework that addresses the 80% degradation in acceptance length seen in existing draft models under agentic workloads with long contexts. DLite achieves 2–6 times higher acceptance length than prior methods and up to 70% speedup over DSpark.
DHL Supply Chain pilots adaptive robotics and AI orchestration in warehouses
DHL Supply Chain is testing adaptive AI-powered robots and introducing AI orchestration into live warehouse operations to coordinate robotics, artificial intelligence, and human workers across its logistics network.
Physicl raises $10 million to improve physical AI training with synthetic data
Physicl raised $10 million to improve physical AI training using synthetic data.
Zacua Ventures raises $135 million to invest in robotics and physical AI
Zacua Ventures raised $135 million to invest in robotics, physical AI, and industrial technology.
daVinci-Prover uses agent-native mid-training for formal theorem proving
The authors present daVinci-Prover, which applies agent-native mid-training to improve formal theorem-proving models. They synthesize a 33.7B-token corpus covering decomposition, search, self-verification, and revision, achieving 13.7-point average improvement over a Qwen3-8B baseline on MiniF2F, FATE-M, and ProofNet benchmarks.
LAW unifies language, action, and world modeling for physical AI agents
The authors present LAW, a closed-loop VLM–WAM framework that unifies language reasoning, action generation, and world modeling through semantic subtask progress. LAW achieves 94.21% average success on RoboTwin tasks; using OpenWAM-α as the WAM backbone raises it to 95.56%. On three real-robot long-horizon tasks, LAW achieves 65% average success.
blender-ai-copilot-flow, an MCP-based AI copilot workflow for Blender
blender-ai-copilot-flow is an AI copilot for Blender with a built-in MCP server. Every edit is a supervised tool call that can be approved, executed in sequence, cancelled after the current step, and undone by name within a visual studio.
HexaShell, a universal harness for robot coding agents in simulation and the real world
The authors present HexaShell, a task-agnostic harness designed to test how far general tools can go for vision-language-model-based robot coding agents. It uses a simple set of 14 basic tools across motion, perception, and state, and the same code runs across four simulated benchmarks and real robots. With GPT-6-Astra and no task-specific input, HexaShell reaches 69.9% on RoboDojo, 59.3% on RoboCasa365, 78.3% on LIBERO-PRO, and 82.2% on RoboTwin 2.0.
SpatialBEV reduces off-road dynamics prediction error by 17.1% using spatial structure
The authors present SpatialBEV, which exploits spatial structure by querying bird's-eye-view features to condition dynamics predictions for off-road terrain navigation. The method reduces prediction error by 17.1% compared to physics-based and learned baselines, and achieves a 25.3% reduction in maximum cross-track error on real hardware at high speed.
WEIBUILD raises $15M to scale embodied-AI construction robots
WEIBUILD raised $15 million to scale its embodied-AI construction robots.
Google connects Wallet to Gemini for spending analysis and savings planning
Google connected Google Wallet to Gemini's Personal Intelligence feature, allowing users to analyze spending patterns and receive budget management and savings plan suggestions based on their transaction and account data. The feature will roll out to all U.S. Gemini app users on web, Android, and iOS.
Meta fined $887 million by Italy for unauthorized access to Android users' photos
Italian data protection authorities fined Meta $887 million for accessing Android users' photos and videos without explicit consent through Facebook's photo recommendation feature. Approximately 20 million Italian Android users were affected.
CoEvolve, a framework for self-evolving code policies in multi-robot humanoid coordination
The authors present CoEvolve, a framework that represents team coordination of humanoid robots as executable code over fixed motor skills. It uses a multimodal foundation model to instantiate coordination strategies through motion primitives and semantic objectives. In simulation with two Unitree G1 humanoids transporting a rigid payload through constrained corridors, CoEvolve completed 8 of 15 base task runs, with every backbone achieving end-to-end success.