Today's briefing
- Physical AI
Zipline raises $1.5 billion for nationwide drone delivery network expansion
Physical AIRobotics & Automation NewsNov 14, 2026 - Industry & Policy
Jensen Huang's departure blasts hole in AI startup funding, VCs warn
Industry & PolicyTechCrunchNov 14, 2026 - Agents
Vercel releases UI SDK with 50+ UI behaviors for AI coding agents
AgentsMarkTechPostNov 14, 2026
Open source & papers radar
- PaperPhysical AIarXiv 2611.17110
SALA enables dual-axis scaling of embodied foundation models through semantic latent action alignment
The authors propose SALA to address the challenge of learning latent actions in embodied foundation models. Using language descriptions of unpaired human and robot video clips, SALA estimates semantic similarities and guides human latent actions toward semantically related robot latents. The resulting shared latent action space enables scaling through increasing human video pre-training and task-specific fine-tuning, improving downstream robot task performance.
- RepoOpen Source & ToolsRepository: dream-num/univerStars: 24,417New stars today: 495Language: TypeScript
Univer: open-source AI office SDK for editing, collaboration, and file processing
Univer Office SDK is a extensible open-source platform that brings editing, calculation, collaboration, and file processing capabilities to products for both people and AI agents.
- PaperResearcharXiv 2611.08318Upvotes: 18
Self-improving embodied agents overfit verification criteria, harming task transfer
The authors study whether coding agents exposed to benchmark criteria generate code that better performs the instructed task or overfits to verification criteria. They find that criteria exposure raises initial success rates but collapses when criteria change slightly, and skills transfer poorly to new tasks. They propose instruction-faithful criteria as a solution.
- PaperPhysical AIarXiv 2610.22471Citations: 7
GPT-6-Astra achieves strong performance across robotic benchmarks but faces deployment gaps
GPT-6-Astra demonstrates unprecedented performance across 12 robotic benchmarks spanning manipulation, humanoids, and mobile platforms through in-context learning. Close-loop coordination between high-level reasoning and low-level control is effective but costly. Real-world deployment reveals substantial gaps in collecting robot demonstrations and handling physical constraints.
Weekly radar
Nov 9 – Nov 15 · 6 repos · 1 model · 3 papers- 1RepoAgentsFriday issue
modern-web-guidance-src keeps coding agents updated on web best practices
- 2RepoAgentsSaturday issue
open-code-review, an LLM-based code review agent at Alibaba scale
- 3RepoAgentsFriday issue
Qwen Code, an open-source AI coding agent for the terminal
- 4RepoOpen Source & ToolsMonday issue
claude-quickstarts, quick start projects for building with the Claude API
- 5RepoPhysical AIMonday issue
MuJoCo, a general purpose physics simulator for multi-joint dynamics
- 6RepoOpen Source & ToolsThursday issue
OpenAI releases Codex Security CLI and SDK for vulnerability scanning and remediation
- 7ModelLLMFriday issue
ZONOS3-Mini, a text-to-speech model from Zyphra
- 8PaperResearchSaturday issue
PhysLoop harnesses LLM physics awareness to improve video generation models
- 9PaperPhysical AISaturday issue
Ropedia Xperience 2 provides 1,589 3D scans and robot learning data for physical AI scaling
- 10PaperResearchSaturday issue
WorldDraft: evaluating video models as generative renderers for 4D worlds
Picked from 55 repos, models and papers in this week's seven issues (Monday to Sunday). Ranked by importance, then trending score, then engagement (stars, likes, upvotes); each format gets two slots first so no single format takes over.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
2 stories in this issue come from Korean outlets. Originals are in Korean.
- Industry & Policy
Amazon pursues World Product Graph to unify global product data
Amazon is developing the World Product Graph, a project to consolidate product information from across the globe, including items it does not sell. The company aims to reduce consumers' need to visit other sites and steer external AI agents to Amazon. It plans to combine brand-supplied data with information from other retailers to identify product differences and relationships.
AI TimesNov 14, 2026 - Physical AI
FamilyMart to expand beverage restocking robot nationwide from January
Japanese convenience store chain FamilyMart will begin full rollout of a robot that automatically restocks beverages on shelves across all stores from January. The robot is based on TX SCARA technology developed by Teleexistence.
iRobotNewsNov 12, 2026
Stories in this issue
22 storiesAnthropic CEO admits 'huge uncertainty' on costs days before record IPO
Days before Anthropic's planned initial public offering, CEO Dario Amodei called buying compute the company's "deepest business dilemma." Anthropic's commitments now total nearly $901 billion, while it remains unclear whether future revenue will justify those costs. The IPO is largely intended as a financial buffer against that uncertainty.
OpenAI reassigns robotics researchers as general-purpose models control robots directly
OpenAI is dissolving its standalone robotics team, redistributing researchers to other teams. General-purpose models like GPT-6 Astra can apparently control robots directly, making a dedicated unit unnecessary, the company said.
Runway's open-source Praxis-1 uses video AI to control robots
Runway released Praxis-1, a robotics model with open weights and source code under the Apache 2.0 license. The model applies knowledge from video training to robot control and leads the World Action Model category on Nvidia's RoboLab-120 benchmark, though it cannot yet plan complex action sequences independently.
Zipline raises $1.5 billion for nationwide drone delivery network expansion
Zipline announced a $1.5 billion funding round at a $21.5 billion post-money valuation, led by Paradigm, WCM Investment Management, and Layer Global. The capital will expand the company's U.S. manufacturing facilities.
Jensen Huang's departure blasts hole in AI startup funding, VCs warn
Nvidia poured tens of billions into AI startups under Huang. Without him at the helm, founders and venture capitalists are bracing for smaller checks and lower valuations.
SALA enables dual-axis scaling of embodied foundation models through semantic latent action alignment
The authors propose SALA to address the challenge of learning latent actions in embodied foundation models. Using language descriptions of unpaired human and robot video clips, SALA estimates semantic similarities and guides human latent actions toward semantically related robot latents. The resulting shared latent action space enables scaling through increasing human video pre-training and task-specific fine-tuning, improving downstream robot task performance.
Univer: open-source AI office SDK for editing, collaboration, and file processing
Univer Office SDK is a extensible open-source platform that brings editing, calculation, collaboration, and file processing capabilities to products for both people and AI agents.
Self-improving embodied agents overfit verification criteria, harming task transfer
The authors study whether coding agents exposed to benchmark criteria generate code that better performs the instructed task or overfits to verification criteria. They find that criteria exposure raises initial success rates but collapses when criteria change slightly, and skills transfer poorly to new tasks. They propose instruction-faithful criteria as a solution.
GPT-6-Astra achieves strong performance across robotic benchmarks but faces deployment gaps
GPT-6-Astra demonstrates unprecedented performance across 12 robotic benchmarks spanning manipulation, humanoids, and mobile platforms through in-context learning. Close-loop coordination between high-level reasoning and low-level control is effective but costly. Real-world deployment reveals substantial gaps in collecting robot demonstrations and handling physical constraints.
WEAVE transforms multi-PR code histories into dependency-structured tasks for agent training
The authors introduce WEAVE, a framework that structures verified pull request histories into executable, dependency-aware tasks for code agents. They develop Graph Credit, a metric that measures dependency-consistent partial progress and supports dense rewards for reinforcement learning on long-horizon code changes.
CarryOn augments vision-language-action models with personalization and safety assurance
CarryOn is a dual-loop harness that augments a frozen vision-language-action policy. Its planning loop maintains persistent user and scene context for personalized task execution, while its execution loop actively supervises rollouts and shields against hazardous actions. On a real robot, it reduced safety violations from 52.6% to 6.1% and increased recovery success from 13% to 80%.
Anthropic: outdated prompting habits slow down Claude Opus 5.5
Anthropic made Claude Opus 5.5 the default model on all paid plans and published a guide on improved prompting. The main recommendation is to drop outdated "think hard" instructions and start with medium effort instead.
Vercel releases UI SDK with 50+ UI behaviors for AI coding agents
Vercel Labs released Vercel UI SDK, an Apache-2.0 beta library containing over 50 UI behaviors. Coding agents can compose them into accessible components for React, Vue, Svelte, Solid, and Web Components, with a bridge for direct UI control.
Wing invests $125 million in US drone manufacturing for nationwide expansion
Wing, the drone delivery company owned by Alphabet, is investing up to $125 million in US manufacturing over the next two years. The investment includes a new 15,000-square-foot manufacturing facility in Fort Worth, Texas, to support nationwide autonomous delivery expansion.
Neura Robotics partners with E.ON to deploy humanoid robots in energy infrastructure
Neura Robotics, E.ON, and German energy company enviaM announced a partnership to develop physical AI applications for electricity networks. Initial real-world trials are planned for 2027 and will focus on using intelligent robots to inspect and monitor energy infrastructure and support workers in physically demanding or hazardous tasks.
Amazon plans World Product Graph to index online shopping data globally
Amazon is building a database of products sold anywhere on the internet, including items it does not sell, and will make the data available through its own site. The project aims to direct shoppers away from Reddit and Google toward Amazon for product research, and to influence AI agents to rely on Amazon's product index.
Anthropic upgrades Claude Code Projects to run parallel AI workers
Anthropic upgraded Claude Code Projects to enable parallel AI workers.
Cmpnd releases dspy-wsp for weight-space programming in DSPy
Cmpnd released dspy-wsp, a prototype DSPy optimizer for weight-space programming. A frontier LLM reads a small model's weights, edits them in a sandbox, and tests each change, with no gradients required and at a cost of about $5 to $10 per run.
Memorilla compresses document collections into 16 memory tokens for frozen LLMs
Stanford researchers released Memorilla, an open-source memory system that compresses a document collection into 16 memory tokens for a frozen Qwen3-8B. It achieves 84.33% on FactKG while using 108 times fewer document tokens than full context.
Amazon pursues World Product Graph to unify global product data
Amazon is developing the World Product Graph, a project to consolidate product information from across the globe, including items it does not sell. The company aims to reduce consumers' need to visit other sites and steer external AI agents to Amazon. It plans to combine brand-supplied data with information from other retailers to identify product differences and relationships.
Shield automates legacy portal work for IT technicians using OpenAI
Shield uses OpenAI to help IT technicians hand off work in legacy portals. Its Forge agents navigate systems so technicians can focus on issues requiring their judgment.
FamilyMart to expand beverage restocking robot nationwide from January
Japanese convenience store chain FamilyMart will begin full rollout of a robot that automatically restocks beverages on shelves across all stores from January. The robot is based on TX SCARA technology developed by Teleexistence.