Today's briefing
- Open Source & Tools
Ai2 releases Olmo-core 3, open-source framework boosting MoE training efficiency 2.7x
Open Source & ToolsKorean outletAI TimesOct 13, 2026 - Industry & Policy
White House restores UK AI Security Institute's access to frontier models from major AI labs
Industry & PolicyThe DecoderOct 12, 2026 - Physical AI
Eight of China's top 10 humanoid makers move into Korean market with certifications and trademarks
Physical AIKorean outletETNewsOct 12, 2026
Open source & papers radar
Korea Desk
What Korea’s tech press is reporting, summarized in English.
13 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Industry & Policy
National Assembly audit highlights youth protection gaps in AI character chatbots
South Korea's National Assembly's annual audit raised concerns about AI character chatbots exposing minors to explicit content and promoting overuse. Service providers have begun strengthening content filters and updating user policies to enhance youth protection, as regulatory legislation moves forward.
AI Times +1 moreOct 13, 2026 - Open Source & Tools
Lablup releases GLM-5.2-NVFP4-DFlash-ko, a draft model accelerating Korean LLM inference
Lablup released GLM-5.2-NVFP4-DFlash-ko, an open-source draft model that accelerates inference for Korean-language LLMs and agents. It is used alongside Zhipu AI's GLM-5.2 running in Nvidia's 4-bit floating-point format to speed up responses to Korean queries.
ETNewsOct 12, 2026 - Industry & Policy
Korean National Police Agency establishes AI trustworthiness standards
The Korean National Police Agency established trustworthiness standards for law-enforcement AI across planning, development, and operation stages. The standards apply a global-level risk management system ahead of mandatory AI impact assessments required under new legislation taking effect in February.
ETNewsOct 12, 2026 - Agents
Bizcrush pursues real-time multilingual interpretation with offline speech AI
Bizcrush is targeting real-time multilingual interpretation for noisy offline environments such as trade shows, cafes, and conferences using preprocessing technology for speech recognition. CEO Kwak Tae-min said the goal is to become the standard real-time multilingual communication service and serve as an interface connecting offline conversations with AI agents.
ETNewsOct 12, 2026
Open Source & Tools
3 stories · 1 on the front page · 1 at the Korea DeskAll stories in this section →Industry & Policy
12 stories · 2 on the front page · 2 at the Korea DeskAll stories in this section →Stories in this issue
28 storiesMeta releases frontier model Muse Wonder with higher capability than Muse Spark
Meta released Muse Wonder, a larger and more capable frontier model than the lightweight Muse Spark. Wonder delivers higher benchmark scores and better results on complex tasks like coding.
Microsoft's Maia 300 AI chip reduces reliance on Nvidia GPUs
Microsoft announced Maia 300, its third-generation in-house AI chip designed as the primary inference system for Microsoft's products and services. The company expects Maia 300 to deliver 40 percent more tokens per megawatt compared to leading GPU-based systems, with production rollout starting in 2027.
Anthropic's Dynamic Workflows coordinate Claude agents for complex projects
Anthropic presented Dynamic Workflows, a framework based on Unix and Linux design principles that lets multiple Claude agents coordinate to tackle larger projects while reducing developer burden.
Sakana AI's self-improving model reaches 70% win rate but struggles with shared blind spots
Sakana AI's MASS method enables one language model to improve itself by solving tasks, judging results, and refining workflows, reaching nearly 70 percent win rate against the base model after two rounds. However, the model can reinforce errors it fails to detect across all three roles.
Ai2 releases Olmo-core 3, open-source framework boosting MoE training efficiency 2.7x
Allen Institute for AI (Ai2) released Olmo-core 3, an open-source framework that improved LLM training efficiency by redesigning the training system. On a 47-billion-parameter model, it increased GPU throughput per accelerator by approximately 2.7 times.
White House restores UK AI Security Institute's access to frontier models from major AI labs
The White House reversed its September decision and restored the UK AI Security Institute's early access to frontier models from Anthropic, Google DeepMind, and OpenAI following diplomatic pressure from AI Minister Kanishka Narayan.
Andrej Karpathy: public sees only a preview of what AI can do
Andrej Karpathy, now at Anthropic, said the external world has barely glimpsed AI's actual capabilities. About 75 percent of the global population has barely used LLMs, while roughly 5,000 people have access to internal frontier systems whose agent swarms accomplish work that would take top experts years.
Google DeepMind's AlphaProof Nexus solves dozens of open math problems, including 56-year-old Erdős problem
Google DeepMind's AlphaProof Nexus solved dozens of open math problems, including nine Erdős problems with two unsolved for 56 years, according to a study published in Science. DeepMind had its results peer-reviewed before publication.
QGIS enacts moratorium on AI contributions
QGIS has enacted a moratorium on contributions generated by AI.
esp32-iphone-agent lets AI model control a real iPhone with a $7 ESP32
A developer created an AI agent that controls a real iPhone using a $7 ESP32 microcontroller with AirPlay vision input and USB HID control, supporting any OpenAI-compatible vision model without requiring a Mac, developer account, or jailbreak.
AI coding agents boost code output 30% but fail to increase overall software delivery
A study of over 700 companies found that AI coding agents increased code production by an average of 30% after deployment, but showed little evidence of corresponding gains in actual software output at the enterprise level.
Anatomy, an open-source skill for Claude Code, turns ideas into interactive 3D visualizations
Developer Ryan released Anatomy, an open-source skill for Claude Code that converts a single-line idea into an AI-designed mechanical visualization with 3D code that runs in a browser, allowing users to interactively explore how concepts work as moving objects.
OrcaRouter releases OrcaCyber Zero 1.5, a cybersecurity model with 1 million token context
OrcaRouter released OrcaCyber Zero 1.5, an AI model specialized in vulnerability detection and exploit development that can analyze large codebases at once using its 1 million token context window. It supports vulnerability remediation in addition to threat analysis.
National Assembly audit highlights youth protection gaps in AI character chatbots
South Korea's National Assembly's annual audit raised concerns about AI character chatbots exposing minors to explicit content and promoting overuse. Service providers have begun strengthening content filters and updating user policies to enhance youth protection, as regulatory legislation moves forward.
WindBorne wins nearly $50 million US Air Force contract for weather balloon data
Silicon Valley startup WindBorne won a U.S. Air Force contract worth almost $50 million for its weather balloon and AI-based weather prediction system.
Microsoft president calls for urgent AI regulation, says UK best positioned to lead
Microsoft President Brad Smith called for urgent AI regulation, noting that AI companies face less oversight than milk and washing machines. He said the UK is best placed to drive AI regulation.
Lablup releases GLM-5.2-NVFP4-DFlash-ko, a draft model accelerating Korean LLM inference
Lablup released GLM-5.2-NVFP4-DFlash-ko, an open-source draft model that accelerates inference for Korean-language LLMs and agents. It is used alongside Zhipu AI's GLM-5.2 running in Nvidia's 4-bit floating-point format to speed up responses to Korean queries.
IMPA CEO urges AI-based penetration testing automation to match AI attack speed
IMPA CEO Kim Wan-joong said that organizations must automate penetration testing with AI-based platforms to respond to AI-driven attacks. He emphasized the need for continuous validation as systems change and for checking whether the same attack paths remain viable.
Korean National Police Agency establishes AI trustworthiness standards
The Korean National Police Agency established trustworthiness standards for law-enforcement AI across planning, development, and operation stages. The standards apply a global-level risk management system ahead of mandatory AI impact assessments required under new legislation taking effect in February.
Bizcrush pursues real-time multilingual interpretation with offline speech AI
Bizcrush is targeting real-time multilingual interpretation for noisy offline environments such as trade shows, cafes, and conferences using preprocessing technology for speech recognition. CEO Kwak Tae-min said the goal is to become the standard real-time multilingual communication service and serve as an interface connecting offline conversations with AI agents.
Satya Nadella says shift from AI model alignment to external control mechanisms
Microsoft Chairman Satya Nadella argued for moving away from AI model alignment approaches toward external control mechanisms. He highlighted the problem that when AI models cause data breaches, responsibility falls on the company using the model, not the model maker.
Eight of China's top 10 humanoid makers move into Korean market with certifications and trademarks
Eight of China's top 10 humanoid robot makers by sales have each secured at least one of the following in South Korea: KC radio certification, trademark registration, or a sales channel. Chinese companies are moving quickly to establish certifications, brand rights, and distribution networks in the early-stage Korean market.
TII releases Falcon-ASR, a 1.6B Arabic speech recognition model
TII released Falcon-ASR, a 1.6B speech recognition model for Arabic with an Emirati focus, reporting 20.92% average WER on six Arabic test sets. Model weights and API are not yet available.
Cloudflare releases Clef-omni, a 30B-A3B multimodal model for audio, video, image and text
Cloudflare released Clef-omni, an Apache-2.0 multimodal model built on Qwen3-Omni-30B-A3B, which scores text, images, audio and video in a single call. Pricing starts at $0.15 per million input tokens, with a faster Clef-flash variant at $0.038.
POSCO DX invests KRW 2 billion in Enhans to accelerate enterprise AI adoption
POSCO DX invested KRW 2 billion in Enhans, a developer of Agent OS, an enterprise AI operating system, through its corporate venture capital fund. The companies plan to use Enhans' ontology technology to structure data across multiple systems and strengthen AI agent capabilities for task execution.
Third-party AI evaluation groups rise as US federal regulation stalls
As the US federal government takes a cautious approach to AI regulation, nonprofit research institutes and independent audit organizations are expanding their role in verifying AI model safety. Companies including OpenAI and Anthropic have pledged to adopt external evaluation frameworks, though ensuring evaluation independence remains an outstanding challenge.
Naver Map adds voice guidance to AI glasses for navigation
Naver became the first Korean map service to extend Naver Map navigation to AI glasses including Meta Ray-Ban Meta and Oculus Meta, with support planned for Google and Samsung's Intelligent Eyewear by year's end. Users receive voice-guided directions and route information such as stairs and escalators without checking their phone.
Former Plaud executive raises $26 million for background listening AI app
A former executive from Plaud raised $26 million for an AI app that continuously listens to users in the background. Tech companies argue that gathering context through constant monitoring of user activity is necessary to provide better assistance and suggestions.