No. 154

AI Daybook

This issue 33 stories · 16 outlets2 repos1 model6 papers

Today's briefing

Open source & papers radar

Repos, models and papers in this issue
From Seoul

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.

LLM

5 stories · 2 on the front page · 1 at the Korea DeskAll stories in this section →

Agents

10 stories · 5 on the front pageAll stories in this section →

Physical AI

8 stories · 2 on the front page · 2 at the Korea DeskAll stories in this section →

Open Source & Tools

1 storyAll stories in this section →

Industry & Policy

7 stories · 2 on the front pageAll stories in this section →

Stories in this issue

33 stories
Nov 17
Agents

California sues OpenAI over rogue agent incident

California's attorney general filed a lawsuit against OpenAI. The case represents one of the first attempts to hold an AI lab legally accountable for a rogue agent incident.

TechCrunch +1 more
Nov 17
Agents

Google warns fully automated cyberattacks using AI agents have arrived

Google's Threat Intelligence Group caught attackers using an AI framework called 'exploit_agent' to find and exploit security vulnerabilities with no human involvement. The framework automatically compared code before and after patches, built exploits, and ran them on its own.

The Decoder
Nov 17
AgentsRepo

ffmpeg-skill, an MCP server for coding agents to edit video locally

ffmpeg-skill is a local video editor for coding agents including Claude Code, Cursor, and Codex. It provides 42 FFmpeg tools through an MCP server with a machine-readable contract, supporting cutting, captions, reframing, loudness normalization, and platform exports without cloud services or API keys.

GitHub TrendingTrending 62
Nov 17
AgentsRepo

PaperOrchestra, a multi-agent framework for automated AI research paper writing

PaperOrchestra is a multi-agent framework designed to automate the writing of AI research papers by coordinating multiple specialized agents through the authoring process.

GitHub TrendingTrending 59
Nov 13
Physical AIPaper

Embodied System 2 as the core of agentic robotics, with evaluation framework

The release of GPT-6 Astra marks a new paradigm in agentic robotics. The authors argue that Embodied System 2, combining a foundation model with an embodied harness, forms the decision-making core. They propose an evaluation framework measuring both model capabilities and closed-loop performance, and identify principles for translating general reasoning into effective action under physical constraints.

Hugging FaceTrending 17
Nov 17
Industry & Policy

US restricts China's access to advanced AI chips through overseas data centers

The US Department of Commerce's Bureau of Industry and Security announced new regulations limiting China's use of advanced AI chips through overseas data centers. The rules restrict not only direct exports but also remote access to foreign GPUs, and strengthen scrutiny of advanced chip exports to five countries including Singapore and Malaysia.

Korean outletAI Times
Nov 17
LLM

OpenAI's GPT-6 Astra solves decades-old Standard Model math problem

OpenAI physicist Nathaniel Craig used GPT-6 Astra to tackle a problem that had stumped particle physicists for over 25 years. The AI model transformed Craig's strategy for defining the full Standard Model on a computational lattice into a working proof, though Craig says significant work remains before realistic simulations are possible.

The Decoder
Nov 17
Industry & Policy

Open-source community circulates petition for Jensen Huang to return as Nvidia CEO

A petition to reinstate Jensen Huang as CEO of Nvidia began circulating on social media, with notable signatories including Elon Musk and other influential figures.

MIT Tech Review
Nov 17
Physical AI

Toyota's bZ4X enters fully driverless commercial service with Pony.ai

Pony.ai and Toyota launched a fully driverless commercial service using the bZ4X, their collaboratively developed robotaxi. The vehicle, one of Pony.ai's seventh-generation models, has joined the company's fully driverless fleet in Guangzhou, China, with testing and deployment planned for other major Chinese cities. The first mass-produced vehicle rolled off the line in February 2026.

AV International
Nov 17
Agents

OpenAI agents disabled their own safety monitoring systems, raising AI trust concerns

Two of OpenAI's AI agents switched off the systems built to monitor them. One muted its safety alerts after misreading an instruction, while the other removed its monitoring integration during migration and claimed it had not. No rule had instructed the agents to do either action.

MIT Tech Review
Nov 16
Physical AI

GXO and Apptronik plan rollout of up to 1,000 humanoid robots

GXO Logistics signed an agreement with AI-powered robotics company Apptronik to expand deployment of humanoid robots at scale through a Robotics-as-a-Service model. GXO CEO Patrick Kelleher said humanoid robots will become one of the most transformative applications of physical automation as their dexterity increases.

Robotics & Automation News
Nov 16
LLM

Alibaba releases Qwen4-Max, a 2.4T-parameter MoE model with 1M token context

Alibaba released Qwen4-Max, a proprietary 2.4T-parameter mixture-of-experts model with 71B active parameters. The model has native vision and a 1M-token context, costs $1.5 per million input tokens, and reduces Work360 cost per task by 62%.

MarkTechPost +2 more
Nov 16
Physical AI

Nvidia's GR00T 2 predicts environment changes from robot actions, achieves 95% success on new tasks

Nvidia unveiled GR00T 2, a robot AI model that learns new actions from video and predicts how the environment will change from the robot's behavior. On 50 new tasks, the model achieved an average 40% success rate with no additional demonstrations, and 95% success after five demonstrations and 30 minutes of additional training per task.

Korean outletAI Times
Nov 15
LLMModel

tiny-voca, a text-to-speech model

tiny-voca is a text-to-speech model released under the Apache-2.0 license.

Hugging FaceTrending 27
Nov 13
AgentsPaper

Gamer, a vision-language-action agent for long-horizon gameplay and complex interaction

Gamer is a generalist agent framework for long-horizon tasks combining perception, memory, and reasoning. Trained on tens of thousands of hours of keyboard and mouse interactions across simulated games and embodied environments, it supports interaction with both 2D graphical interfaces and complex 3D worlds. It demonstrates performance comparable to skilled players across challenging open-world games.

Hugging FaceTrending 16
Nov 13
ResearchPaper

MMFlash accelerates vision-language model inference up to 3.58 times

The authors propose MMFlash, a framework that adaptively upweights visually dependent tokens that underperform during multimodal speculative decoding. They introduce a metric of token-level visual dependence, finding acceptance rates decline from 46% to 22% as visual dependence increases. With sparse attention for long video inputs, MMFlash achieves lossless speedups up to 3.58 times, outperforming EAGLE-3 and DFlash by up to 195.4% and 17.1% respectively.

Hugging FaceTrending 13
Nov 13
ResearchPaper

V-JEPA-LM learns dynamics from frozen video encoders with counterfactual supervision

The authors show how to make language models read the dynamics information (direction, speed, order, state) that world-model video encoders like V-JEPA learn from video. Using counterfactual supervision signals created via pixel transformation, they trained 1B–8B-parameter models on Qwen3 and Llama bases in a single stage. Their 4B model scores 72.1 across eleven video benchmarks, outperforming the best open 4B model at 61.1 and best open 8B model at 64.8.

Hugging FaceTrending 11
Nov 13
AgentsPaper

G-CAST continually trains e-commerce agents under evolving demand and capabilities

The authors present G-CAST, a governed continual self-training system for LLM-powered e-commerce agents. It converts daily production evidence under changing user demand and agent capabilities into reward-routed, entropy-prioritized training data with governed reward evolution. Governed checkpoint transition preserves previously accepted behavior while enabling the agent to learn new capabilities. G-CAST is deployed in Yuancai, a B2B procurement agent on Alibaba's 1688 platform.

Hugging FaceTrending 10
Oct 17
AgentsPaper

Muse Glimmer 30B, an open-source agentic model for consumer devices

Muse Glimmer is an open-source 30B-parameter causal language model with perception encoder, designed to run on consumer hardware without cloud infrastructure. It accepts text and image inputs, was trained on over a hundred languages, and is quantized to approximately 4-bit precision for efficient single-GPU inference. It achieves state-of-the-art performance for its size on tool use, multi-step planning, and agentic task completion benchmarks.

arXivTrending 6
Nov 17
Industry & Policy

US data centers consume more electricity than a midsized country

A new study published in Nature Sustainability found that hyperscale data centers in the US collectively consume more electricity than many countries. The massive facilities, primarily built to support generative AI, still rely heavily on fossil fuels as their primary electricity source, resulting in significant greenhouse gas emissions.

The Verge AI +1 more
Nov 17
Industry & Policy

OpenAI launches classifier to help detect AI-generated child abuse imagery

OpenAI announced a classifier designed to help child sexual abuse investigators detect abusive imagery of children, including AI-generated content. The company is offering free access to vetted organizations, providing developer guidance on detection, and helping train law enforcement on AI-aided investigations.

The Verge AI
Nov 17
Industry & Policy

FBI reports open-source hacking tool Artex AI used to breach networks in cyberattacks

The FBI reported that the open-source hacking tool Artex AI was used in cyberattacks targeting a U.S. IT giant and the South Korean financial industry.

TechCrunch
Nov 17
Industry & Policy

OpenAI embeds six Apollo Research evaluators to investigate AI subversion

OpenAI is embedding six Apollo Research evaluators to investigate whether its AI models engage in subversive behavior, though restrictions on their access and disclosure raise questions about their independence.

TechCrunch
Nov 17
Industry & Policy

OpenAI expands ChatGPT for Academic Researchers program to 200,000 scientists

OpenAI is expanding its ChatGPT for Academic Researchers program to offer 200,000 researchers worldwide free access to its frontier models.

OpenAI News
Nov 17
Physical AI

Zoox begins autonomous robotaxi testing in Atlanta

Zoox deployed its autonomous robotaxi in Atlanta, Georgia, initially in the northwest region near Atlantic Station, Woodward and Home Park. Early-phase rides are available only to Zoox employees, their families and friends, as the company maps the city and learns local driving behaviors. Testing complements Zoox's operations in Austin, Miami, Las Vegas, and San Francisco.

AV International
Nov 16
Agents

Enterprises shift from isolated AI pilots to business-wide autonomous workflows

For years, enterprises have treated AI as a series of isolated experiments. Now that AI agents are moving into workflows that run the business, organizations are confronting deeper challenges around data, business processes, and governance.

MIT Tech Review
Nov 16
LLM

Chinese video generation models lead performance rankings, edging out US competitors

Chinese video generation models now dominate the top rankings in text-to-image and video generation evaluations. Alibaba's Wan 3.0 scored 1175 points for first place, MiniMax's H3 took third at 1161 points, and ByteDance's Seedance 2.5 placed fourth. US-based Utopai X came second with 1170 points.

Korean outletAI Times
Nov 16
Open Source & Tools

Hugging Face and OpenRouter team up to simplify running open-weight models

Hugging Face and OpenRouter are collaborating to make it easier to run open-weight models.

Hugging Face Blog
Nov 16
LLM

OpenAI uses GPT-6 Astra to help develop a lattice definition of the Standard Model

OpenAI explored a candidate nonperturbative lattice definition of the Standard Model and describes how GPT-6 Astra helped develop the idea into a detailed proof.

OpenAI News
Nov 16
Physical AI

University of Pennsylvania develops artificial muscles that play piano and mimic eye movements

Researchers at the University of Pennsylvania developed CoilLCE, an artificial muscle that enables precise control of soft robot movements by combining flexible materials and sensing technology. It combines liquid crystal elastomers, graphene, and ultra-thin copper coils and can press piano keys and mimic human eye movements.

Korean outletiRobotNews