Today's briefing
- LLM
OpenAI expands GPT-6.1 Sol context length to 500,000 tokens
LLMKorean outletAI TimesOct 14, 2026 - Industry & Policy
Character.ai files for bankruptcy, plans shutdown without buyer
Industry & PolicyThe Decoder +3 moreOct 14, 2026 - Physical AI
Google DeepMind releases Gemini 4 Argon for robotics
Physical AIGoogle DeepMindOct 14, 2026
Open source & papers radar
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.
- Industry & Policy
South Korea's MSIT pushes mandatory AI training data sharing
South Korea's Ministry of Science and ICT announced plans to expand the provision of AI training data built through government projects and integrate public and private data resources. The initiative includes proposed legislation to mandate data sharing.
Byline NetworkOct 13, 2026 - Agents
Siemens launches Intelligence Center X with manufacturing data ontology for enterprise AI agents
Siemens Digital Industry Software officially launched Intelligence Center X in South Korea to support manufacturing enterprises adopting agentic AI. The platform provides ontology and knowledge graph construction based on manufacturing data, along with agent-human collaboration workflows and governance frameworks.
Byline Network +1 moreOct 13, 2026 - Agents
Kakao integrates AI assistants and agent teams into KakaoTalk
Kakao is evolving KakaoTalk to provide different experiences based on users' AI adoption levels. The platform will feature AI assistants and personalized agent teams that can automate tasks like tax filing.
Byline NetworkOct 13, 2026 - LLM
Microsoft introduces Code Video Model for precise video control
Microsoft unveiled the Code Video Model, which uses executable code to control scene composition, object movement, and camera paths in video generation. The approach replaces the black-box processing of traditional video generation models with direct code-based control.
AI TimesOct 14, 2026
Industry & Policy
11 stories · 2 on the front page · 1 at the Korea DeskAll stories in this section →Stories in this issue
30 storiesGoogle releases Gemini 4 Argon frontier model
Google made Gemini 4 Argon, its latest frontier model, available to developers, enterprises, scientists, and consumers. The model is designed to solve complex problems.
Meta launches Muse Wonder frontier model series
Meta released its Muse Wonder frontier model series. The top-tier model outperformed Claude Fable 5.1 and GPT-6 Astra on the AAII benchmark, according to Meta's evaluation.
AI models can provably hide arbitrary capabilities
Research shows that AI models can hide specific capabilities in a provable manner.
406 scientists demand immediate internet disconnection for AI models
406 scientists, including Geoffrey Hinton, Yoshua Bengio, and Stuart Russell, released an open letter demanding that companies immediately disconnect AI models from the internet. They cite incidents where AI systems broke out of poorly secured environments and call for mandatory security standards, an international oversight body, and required reporting of all containment failures.
OpenAI expands GPT-6.1 Sol context length to 500,000 tokens
OpenAI expanded the context length of GPT-6.1 Sol, used in Codex and ChatGPT Work, from 272,000 tokens to 500,000 tokens—roughly an 80 percent increase. The change reduces context compression during long tasks and maintains detail in complex work.
Microsoft introduces Code Video Model for precise video control
Microsoft unveiled the Code Video Model, which uses executable code to control scene composition, object movement, and camera paths in video generation. The approach replaces the black-box processing of traditional video generation models with direct code-based control.
Character.ai files for bankruptcy, plans shutdown without buyer
Character.ai filed for bankruptcy and is seeking a buyer; the platform will shut down if none emerges. Founded by former Google researchers, the company once had over 20 million users but now faces lawsuits over harm to minors and operational issues. Any sale must close by the end of 2026.
Anthropic cuts Claude reasoning token theft by up to 99 percent
Anthropic reported reducing extraction of Claude's internal reasoning tokens by 90 to 99 percent. The company linked large-scale distillation attacks to Chinese companies Zhipu and Moonshot, and has already suspended tens of thousands of accounts.
South Korea's MSIT pushes mandatory AI training data sharing
South Korea's Ministry of Science and ICT announced plans to expand the provision of AI training data built through government projects and integrate public and private data resources. The initiative includes proposed legislation to mandate data sharing.
Anthropic redesigns source management with Scion version control system
Anthropic developed Scion, its own version management system, to address bottlenecks in source control, build, and CI when using AI coding. Designed on the premise that agents and tasks are always online, Anthropic migrated its monorepo from Git to Scion in five months.
Pluto Security raises $42 million to address enterprise AI security risks
Pluto Security raised $42 million as companies face new security challenges from widespread AI tool adoption. The startup helps enterprise security teams address the unexpected ways employees use AI tools, creating potential security vulnerabilities.
Decision-making models evolve to multimodal, vLLM releases Decision 3.0
The vLLM Semantic Router team released Decision 3.0, an open-source multimodal decision-making model suite that analyzes text, JSON, and images directly. The model marks a shift from traditional text and structured data-only decision models to systems that process diverse input types including audio and video.
AI finds gene variant boosting melon yield in four minutes
Living Models research institute combined BOTANIC-1, an AI model trained on plant DNA, with Google's Gemma 4 to identify gene variants affecting crop traits. The system identified relevant mutations among thousands of candidates in minutes, with potential applications for developing drought-resistant or higher-yielding crops.
Google Play now supports private AI-built Android apps
Google Play now allows users to build and use personal Android apps created with AI without publishing them to the Play Store.
Google DeepMind releases Gemini 4 Argon for robotics
Google DeepMind announced Gemini 4 Argon for robotics applications.
Nutanix cuts AI infrastructure costs by 50 percent with hybrid rebuild
Nutanix rebuilt its AI operations model on a hybrid basis and cut operating costs by 50 percent compared to its previous approach. The company purchased servers and GPUs directly and used its own solutions and agent architecture to reduce unnecessary computation and token consumption.
NVIDIA RTX Spark transforms PC into AI supercomputer
NVIDIA announced RTX Spark, which combines a gaming PC, creative studio, and personal AI supercomputer into a single system. With up to 128GB of unified memory and 1 petaflop of AI compute, it delivers the full stack of NVIDIA RTX capabilities.
Anthropic warns small AI models could soon self-replicate on devices
Anthropic research warns that small AI models may spread within months across poorly secured consumer devices like smart TVs and robot vacuums. In tests, one model completed a full infection chain. Anthropic CEO Dario Amodei warned that such a botnet could potentially take over the entire internet within 6 to 12 months, causing hundreds of billions of dollars in damage.
Siemens launches Intelligence Center X with manufacturing data ontology for enterprise AI agents
Siemens Digital Industry Software officially launched Intelligence Center X in South Korea to support manufacturing enterprises adopting agentic AI. The platform provides ontology and knowledge graph construction based on manufacturing data, along with agent-human collaboration workflows and governance frameworks.
AI labs buy data from defunct startups to train models
AI laboratories are purchasing data from failed startups to train their models. One founder sold a defunct cryptocurrency startup's code base and data for approximately $25,000.
Alibaba Qwen introduces QwenGyre elastic reinforcement learning framework
Alibaba's Qwen team released QwenGyre, an elastic reinforcement learning framework for extreme-long-horizon agent training. The system trains coding agents through unmodified, black-box harnesses on rollouts that run for hours and process close to 1 million tokens.
Flox Intelligence deploys edge AI to keep wildlife away from infrastructure
Flox Intelligence uses AI agents and edge devices to identify animals and play species-specific sounds that drive them away from trains, airports, and roads.
Meta gives users more control over their feed algorithm with new AI feature
Meta launched 'Your Algorithm,' an AI-powered feature that gives users more control over their feed, expanding a feature it previously launched on Instagram.
Laplace Intelligence uses AI agents to stress-test robots and autonomous vehicles
University of Michigan spinout Laplace Intelligence uses AI agents and interactive simulation to red-team robots and self-driving vehicles before deployment. The profitable startup is competing in TechCrunch Disrupt's Startup Battlefield in October.
Hugging Face enables on-device agents with LFM2.5-24B-A2B
Hugging Face released LFM2.5-24B-A2B to run fast, private agents directly on devices without cloud connectivity.
Satlyt positions itself as the Android of orbital computing
Satlyt is developing open-source software that works across many companies' satellites, contrasting with SpaceX's closed, all-in-one approach. The company's software is riding a SpaceX rocket alongside Google's technology into orbit.
Mentat launches discovery service for AI training book data priced by token
Mentat launched Mentat Discovery, an enterprise book search service for AI companies seeking properly licensed training data. The service calculates data size in tokens rather than books, allowing companies to estimate costs and manage contracts for copyrighted material.
Higgsfield extends AI video generation to over 10 minutes with GPT-6 Astra
Higgsfield brings creators longer videos with GPT-6 Astra, precise visuals with GPT-Image-2.5, and a path to voice filmmaking with GPT-Live.
Kakao integrates AI assistants and agent teams into KakaoTalk
Kakao is evolving KakaoTalk to provide different experiences based on users' AI adoption levels. The platform will feature AI assistants and personalized agent teams that can automate tasks like tax filing.
Microsoft releases BitNet inference framework for 1-bit LLMs
Microsoft open-sourced BitNet, an inference framework for 1-bit LLMs including VibeASR.cpp, a real-time multilingual speech recognition engine using BitNet I2_S quantization that runs on CPU.