Today's briefing
- Industry & Policy
Former OpenAI researcher claims to find Nvidia Blackwell signatures in six Chinese AI models
Industry & PolicyThe Decoder +2 moreNov 14, 2026 - Physical AI
Agile Robots raises $500 million at $8.5 billion valuation
Physical AIKorean outletiRobotNewsNov 13, 2026 - LLM
Perplexity releases Decider v1.3 decision models with Locate for span extraction
LLMMarkTechPostNov 14, 2026
Open source & papers radar
- RepoAgentsRepository: alibaba/open-code-reviewStars: 50,510New stars today: 206Language: Go
open-code-review, an LLM-based code review agent at Alibaba scale
open-code-review is a hybrid architecture code review tool combining deterministic pipelines with LLM agents. It delivers precise line-level comments, includes a multi-language ruleset for common issues like null pointer exceptions and SQL injection, and is compatible with OpenAI and Anthropic APIs.
- PaperResearcharXiv 2611.16748Upvotes: 41
PhysLoop harnesses LLM physics awareness to improve video generation models
The authors present PhysLoop, a framework that guides and constrains LLMs throughout video generation, selection, and correction to distill their physical knowledge into physically consistent videos. They release 100K videos of diverse indoor and outdoor scenes along with generation tools and two evaluation datasets.
- PaperPhysical AIarXiv 2611.17060Upvotes: 25
Ropedia Xperience 2 provides 1,589 3D scans and robot learning data for physical AI scaling
The authors introduce Ropedia Xperience 2, comprising 1,589 metric-scale 3D scans and 4,230 egocentric recordings of 74 tasks spanning 131 hours. Scenes are converted into physically interactive simulation environments and human motion is retargeted to robot trajectories, establishing a gold standard for physical AI data collection.
- PaperResearcharXiv 2611.16217Upvotes: 14
WorldDraft: evaluating video models as generative renderers for 4D worlds
The authors introduce WorldDraft, a benchmark of 432 paired test cases across 48 executable 4D world families, to evaluate whether video models can reliably render 4D worlds specified in code. Evaluations of 12 models reveal operational tensions between world preservation and visual realization, and source leakage of up to 64%. The authors propose WorldDraft-Primer, a visual prompting strategy that improves both objectives on frozen video renderers.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
3 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
T-Robotics partners with KAIST Manufacturing Physical AI Institute
Robot manufacturer T-Robotics signed a memorandum of understanding with KAIST's Manufacturing Physical AI Research Institute to advance physical AI commercialization. The partnership aims to establish standardization alliances, validate technology through real-world deployment, and support global market expansion. Both organizations will work to expand industrial applications of physical AI technologies and strengthen the domestic ecosystem.
iRobotNewsNov 13, 2026 - Industry & Policy
Anthropic launches Claude for Manufacturing to automate factory workflows
Anthropic released Claude for Manufacturing, offering 17 specialized skills to support quotation generation, quality inspection, and equipment troubleshooting. The launch extends Anthropic's industry-specific AI products beyond healthcare, finance, and law into manufacturing.
AI TimesNov 13, 2026
Stories in this issue
27 storiesLangChain reaches $3 billion valuation
LangChain, which offers LangSmith, a model and platform-agnostic agent management tool, has reached a $3 billion valuation with 7,000 customers.
open-code-review, an LLM-based code review agent at Alibaba scale
open-code-review is a hybrid architecture code review tool combining deterministic pipelines with LLM agents. It delivers precise line-level comments, includes a multi-language ruleset for common issues like null pointer exceptions and SQL injection, and is compatible with OpenAI and Anthropic APIs.
PhysLoop harnesses LLM physics awareness to improve video generation models
The authors present PhysLoop, a framework that guides and constrains LLMs throughout video generation, selection, and correction to distill their physical knowledge into physically consistent videos. They release 100K videos of diverse indoor and outdoor scenes along with generation tools and two evaluation datasets.
Ropedia Xperience 2 provides 1,589 3D scans and robot learning data for physical AI scaling
The authors introduce Ropedia Xperience 2, comprising 1,589 metric-scale 3D scans and 4,230 egocentric recordings of 74 tasks spanning 131 hours. Scenes are converted into physically interactive simulation environments and human motion is retargeted to robot trajectories, establishing a gold standard for physical AI data collection.
WorldDraft: evaluating video models as generative renderers for 4D worlds
The authors introduce WorldDraft, a benchmark of 432 paired test cases across 48 executable 4D world families, to evaluate whether video models can reliably render 4D worlds specified in code. Evaluations of 12 models reveal operational tensions between world preservation and visual realization, and source leakage of up to 64%. The authors propose WorldDraft-Primer, a visual prompting strategy that improves both objectives on frozen video renderers.
Agentic robots: a unified view of models, data, environments, and harnesses
The authors survey the emerging paradigm of agentic robots, in which agents reason over human instructions, interact with robotic environments, and autonomously rectify execution errors. They organize existing research into 17 categories across four functional domains—model, data, environment, and harness—and trace the development from robot policies through vision-language-action models and world models to complete agentic systems.
Tactile-guided robot learning: from representation learning to foundation models
The authors review tactile-guided skill learning from a learning-centric perspective, examining how touch supports transferable representation learning, policy conditioning, and embodied physical reasoning. They survey tactile representation and policy learning, tactile foundation models, predictive and adaptive control, and generalization across objects, tasks, sensors, and embodiments, and outline directions toward tactile foundation and world models.
OpenAI launches ChatGPT Wallet to let Dots agents make purchases
OpenAI launched ChatGPT Wallet, which allows agents in ChatGPT Work, Codex, and dots to make purchases with user approval using eligible US-issued cards. Every purchase requires biometric verification.
Anthropic's Claude for Manufacturing connects AI to factory software for workflow automation
Anthropic released Claude for Manufacturing, a plugin that connects Claude directly to Autodesk Fusion, Oracle NetSuite, and Infor. Its 17 specialized skills automate tasks like drafting quotes, comparing suppliers, and placing backup orders after human approval.
Former OpenAI researcher claims to find Nvidia Blackwell signatures in six Chinese AI models
Former OpenAI researcher Will DePue says he produced technical evidence that DeepSeek, Moonshot AI, MiniMax, and Ant Group are using Nvidia's latest Blackwell GPUs despite US export controls. His method reads characteristic chip signatures directly from published model weights. He estimates the hardware behind the five training runs cost more than half a billion dollars.
Google releases Gemini 4 Hydrogen, 2.5x faster than Gemini 3.8 Flash
Google released Gemini 4 Hydrogen, a fast agentic model with a 1M-token context window and four thinking levels. The model is priced at $1.50 per million input tokens and $7.50 per million output tokens, scoring 51 on Artificial Analysis and 57% on Terminal-Bench 4.1.
Anthropic launches Claude for Manufacturing to automate factory workflows
Anthropic released Claude for Manufacturing, offering 17 specialized skills to support quotation generation, quality inspection, and equipment troubleshooting. The launch extends Anthropic's industry-specific AI products beyond healthcare, finance, and law into manufacturing.
Agile Robots raises $500 million at $8.5 billion valuation
German physical AI company Agile Robots closed a $500 million Series F round, with a valuation of $8.5 billion and cumulative funding of $2 billion. SoftBank Group led the round, with Nvidia, Hitachi Group, and T.Capital also participating. The company plans to use the funds to expand production capacity and strengthen its position in key markets including the United States.
FOAM matches SOAP performance at Adam-class efficiency for LLM pre-training
The authors propose FOAM, a lightweight second-order optimizer that retains SOAP's benefits at nearly Adam/AdamW efficiency. FOAM uses input-only preconditioning and replaces SOAP's dense rotated second moment with a factored Adam-moment estimator. Results show FOAM matches SOAP and Muon baselines in validation loss from 124M to 760M parameters, uses half of SOAP's optimizer state, and runs at Muon-class wall-clock time.
InstantAMP: train humanoid robot locomotion in 30 minutes on one RTX 5090
InstantAMP trains Unitree G1 humanoid locomotion and get-up motions through a differentiable world model using JAX and MuJoCo Warp. The project demonstrates fast learning of adversarial motion priors on a single RTX 5090 GPU.
LEASH introduces RoboRisk benchmark for agentic robotics safety
LEASH is a project that evaluates safety in robot agents. It provides the RoboRisk benchmark comprising 92 simulated tasks and 13 real-robot tasks, along with a camera-grounded safety harness. The work is designed for researchers and developers validating the safety of foundation-model-based robot planners.
PingpongDuo achieves agile humanoid table tennis doubles with safe team coordination
The authors present PingpongDuo, a hierarchical framework for agile humanoid table tennis doubles with safe team coordination. A planner jointly selects positioning targets and uses learned risk estimates to filter risky proposals. In simulation, PingpongDuo improves the ball-return success rate by 13.2% over the baseline. Real-world experiments show two robots sustain more than 10 consecutive strikes in strict alternation over a 2.15 m cycle in 2.4 seconds.
Adapt human motion for robots with NVIDIA SOMA Retargeter
NVIDIA SOMA Retargeter enables adapting human motion demonstrations to humanoid robots, accounting for differences in body proportions, joint structure, and range of motion. Human demonstrations capture the coordination and timing behind natural movements, but a robot cannot simply copy the same joint motion.
Perplexity releases Decider v1.3 decision models with Locate for span extraction
Perplexity released Decider v1.3 and Flash v1.3, open-weight decision models that return probabilities instead of text. Version 1.3 adds Locate for span extraction and reports 91.6% mean accuracy on 63 query-classification tasks.
Brave Search API enables real-time micropayments for agents via x402 protocol
Brave Search API adopted the open x402 payment protocol, allowing developers and autonomous AI agents to pay per request without account creation or API key management. Brave operates its own payment settlement service and accepts USDC on Base. The pricing for agent LLM Context is $0.005 per request.
Meta's Muse AI agent now makes calls to businesses on behalf of users
Meta's AI agent Muse now calls businesses on behalf of paying US subscribers, handling everything from cable providers to restaurants. Muse has over 10 million users, according to Meta's AI chief. Google Gemini, Instinct, and xAI's Grok Bot offer the same calling feature, though many businesses disconnect when they detect an AI.
OpenAI launches two security products under its Daybreak brand
OpenAI launched two security products under its Daybreak brand.
llama.cpp v1.0.0 provides stronger foundation for local AI ecosystem
llama.cpp released version 1.0.0 to strengthen the foundation for the local AI ecosystem.
T-Robotics partners with KAIST Manufacturing Physical AI Institute
Robot manufacturer T-Robotics signed a memorandum of understanding with KAIST's Manufacturing Physical AI Research Institute to advance physical AI commercialization. The partnership aims to establish standardization alliances, validate technology through real-world deployment, and support global market expansion. Both organizations will work to expand industrial applications of physical AI technologies and strengthen the domestic ecosystem.
Liquid AI releases LiquidCodec, a 74M-parameter streaming audio codec
Liquid AI released LiquidCodec, a 74M-parameter streaming audio codec. It operates at 12.5 Hz with 12 codebooks and encodes about 361x faster than real time on a single CPU thread. The codec powers LFM3-Audio-2B and LFM3-Omni-2B.
CopilotKit releases OpenBot 1.0, MIT-licensed self-hostable AI coworker
CopilotKit released OpenBot 1.0, an MIT-licensed, self-hostable AI coworker workspace. Organizations bring their own agents, models, and infrastructure, while every Bot action passes a policy check and lands in an audit trail.
Lincoln Electric launches XpertLINC, a physical AI platform for robotic welding
Lincoln Electric launched XpertLINC, a proprietary physical AI platform that integrates vision, coordinated motion, process control, and welding expertise. It dynamically adjusts a robot's path and welding parameters to adapt like a human would, delivering higher productivity and quality even with inconsistent parts.