Today's briefing
- LLM
Liquid AI releases d1-3B, a 3.12B multimodal decision model outperforming 10x larger models
LLMKorean outletAI TimesOct 9, 2026 - Industry & Policy
Nous Research raises $90 million Series B for Hermes AI agent
Industry & PolicyKorean outletAI TimesOct 9, 2026 - Physical AI
AWS launches open-source Physical AI Toolchain for robotics
Physical AIGoogle News +5 moreOct 8, 2026
Korea Desk
What Korea’s tech press is reporting, summarized in English.
7 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Research
Kakao presents MoE optimization and model compression research at COLM 2026
Kakao presented at the COLM 2026 conference a method to predict optimal learning rates for large mixture-of-experts model pretraining at approximately 1% of total training cost. The company also shared model compression research at the TokShop tokenization workshop.
AI TimesOct 9, 2026 - Open Source & Tools
VESSL AI provides inference infrastructure for Upstage Solar Pro 4
VESSL AI provides global inference infrastructure for Upstage's Solar Pro 4 large language model, which ranked third by actual usage on the OpenRouter LLM marketplace in September. VESSL AI handles inference requests from users worldwide across Korea, Europe, and the U.S.
ETNewsOct 8, 2026 - Open Source & Tools
Google launches Playground, an AI platform for building games from text prompts
Google released Playground, an AI game creation platform that generates 2D and 3D games from natural language descriptions, in partnership with Unity. The platform is available first to users 18 and older in the U.S.
AI TimesOct 9, 2026 - LLM
Microsoft optimizes MAI-Code-1.1 for on-device coding on Surface Laptop Ultra
Microsoft released an on-device optimized version of MAI-Code-1.1-Flash, its coding-focused AI model, for the new Surface Laptop Ultra with Nvidia RTX Spark. The model runs locally on PCs without cloud connection.
AI Times +1 moreOct 9, 2026
Open Source & Tools
7 stories · 1 on the front page · 2 at the Korea DeskAll stories in this section →Stories in this issue
27 storiesOpenAI applies unpublished model to about 4,000 math research problems
OpenAI applied its unpublished frontier model to about 4,000 difficult mathematical research problems and released 722 research manuscripts describing the outcomes, organized into 372 related result families. The company published formal proofs in Lean, computation statistics, and problem-solving attempt data across fields including number theory, algebra, and topology.
Anthropic launches free AI security scans for open-source projects
Anthropic introduced OSS Scanner, offering open-source projects free periodic security vulnerability scans using its strongest models. The service provides fast, frequent scans but without human review of the model-generated reports.
Claude now generates animated videos and live dashboards from text
Anthropic launched two beta features for Claude: Dashboards converts data sources like BigQuery and Snowflake into live interactive dashboards, while Motion generates animated explainer videos from text and images.
Google brings agentic AI capabilities to Gemini for businesses
Google transformed Gemini into an AI agent that can plan, execute tasks across business systems, delegate work to subagents, use multiple AI models, and operates with its own workplace identity including an email address.
Liquid AI releases d1-3B, a 3.12B multimodal decision model outperforming 10x larger models
Liquid AI released d1-3B and d1-Omni-600M, multimodal decision models that compute input states once to return immediate decision scores instead of generating tokens sequentially. The 3.12 billion parameter d1-3B outperforms models 10 times its size.
Nous Research raises $90 million Series B for Hermes AI agent
Nous Research, creator of the open-source Hermes AI agent, raised $90 million in a Series B round led by Robot Ventures at a $1.5 billion valuation. Investors included Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and 1789 Capital.
JetBrains releases Mellum2.1, a 12B MoE model for coding agents
JetBrains released Mellum2.1, a 12-billion-parameter mixture-of-experts open model with 2.5 billion active parameters for coding agents. Reinforcement learning in real repositories lifted its SWE-bench Verified score from 2.0 to 47.0.
AWS launches open-source Physical AI Toolchain for robotics
AWS released an open-source Physical AI Toolchain for robotics development and training.
Robot data startup Mecka AI raises $60 million from Sequoia
Mecka AI, which collects and analyzes human motion data to train humanoid and other robots, raised $60 million from Sequoia. The company pays people to record everyday tasks.
Anthropic releases Claude Haiku 5.5 with 1M context window at $0.10 per million input tokens
Anthropic released Claude Haiku 5.5, a small model with 1 million token context window priced at $0.10 per million input tokens. The model scores 72.4% on OSWorld.
Anthropic commits $150 million over three years to U.S. scientific research
Anthropic pledged $150 million over three years to the Genesis Mission, a federal initiative to support AI-driven scientific discovery, providing Claude to more than 15 agencies including NASA, the National Institutes of Health, and the National Science Foundation.
Fired OpenAI safety researchers dispute misconduct claims in open letter
Three fired OpenAI safety researchers disputed allegations of mishandling sensitive information, warning in an open letter that their dismissals are creating a chilling effect on the company's AI safety culture.
Anthropic updates policy to ban sustained abuse of Claude
Anthropic updated its usage policy to ban sustained abuse of Claude and tightened restrictions on propaganda, drone weaponization, and surveillance. Violations can result in account suspension.
Mathematicians call for OpenAI boycott after AI-generated papers flood the field
The Association for Human Mathematics called for an OpenAI boycott after the company released over 700 AI-generated math manuscripts at once. OpenAI retracted three papers the next day over a sign error. Fields Medalist Terence Tao warned that AI's mass harvesting of open problems threatens branches of mathematics.
Google launches Playground, an AI platform for building games from text prompts
Google released Playground, an AI game creation platform that generates 2D and 3D games from natural language descriptions, in partnership with Unity. The platform is available first to users 18 and older in the U.S.
Kakao presents MoE optimization and model compression research at COLM 2026
Kakao presented at the COLM 2026 conference a method to predict optimal learning rates for large mixture-of-experts model pretraining at approximately 1% of total training cost. The company also shared model compression research at the TokShop tokenization workshop.
Microsoft optimizes MAI-Code-1.1 for on-device coding on Surface Laptop Ultra
Microsoft released an on-device optimized version of MAI-Code-1.1-Flash, its coding-focused AI model, for the new Surface Laptop Ultra with Nvidia RTX Spark. The model runs locally on PCs without cloud connection.
Google releases AI Edge Foresight, an offline meeting transcription app
Google released AI Edge Foresight, an experimental note-taking app that transcribes meetings and audio files entirely offline on macOS using the on-device EmbeddingGemma 2 model. The app summarizes meetings and transforms handwritten notes into polished text.
Single prompt could hijack all AI agents in an AWS account, Zenity researchers find
Zenity Labs researchers discovered that a single publicly accessible AI agent on Amazon Bedrock AgentCore could be used to take over all AgentCore agents in the same AWS account and region. AWS has since patched the vulnerability and tightened default permissions.
AI-powered hacking tool enabled breach of multiple South Korean banks
A suspected Chinese-speaking attacker used ARTEX, an AI-powered automated penetration testing tool, to breach multiple South Korean financial institutions, stealing over 25,000 customer records from Shinhan Bank alone, according to CrowdStrike.
Perplexity releases pplx-embed-v2-late embedding models in 0.6B and 9B sizes
Perplexity released pplx-embed-v2-late in two sizes: a 0.6B model for edge devices and a 9B model for building high-quality indexes. The 9B model scores 92.4% on MADQA and both are MIT-licensed and self-hostable.
OpenAI disrupts two AI-enabled influence operations using false-front journalists
OpenAI disrupted two AI-enabled influence operations that used false-front journalists and a think tank to spread geopolitical messaging.
Whistle, a 16.9 MB open speech recognition model supporting seven languages
Whistle is an open-source speech recognition model deployed as a single 16.9 MB file with no external dependencies. It runs on CPU, supports seven languages, and processes up to 30 seconds of 16 kHz mono audio with word-level timestamps and voice embeddings.
Architect launches Liquid Inference, a real-time auction system for LLM inference
Architect Financial Technologies launched Liquid Inference, an LLM router that runs a live auction for every inference request. Providers bid to serve each prompt and the buyer pays the lowest offer that meets its requirements.
VESSL AI provides inference infrastructure for Upstage Solar Pro 4
VESSL AI provides global inference infrastructure for Upstage's Solar Pro 4 large language model, which ranked third by actual usage on the OpenRouter LLM marketplace in September. VESSL AI handles inference requests from users worldwide across Korea, Europe, and the U.S.
NVIDIA PivotOPD trains multi-turn agents to recover from early mistakes
NVIDIA researchers introduced PivotOPD, an on-policy distillation method that trains multi-turn LLM agents to avoid and recover from early pivotal mistakes. It posted the best average performance against 13 baselines across 3 agent benchmarks.
Hugging Face introduces Falcon ASR speech recognition model
Hugging Face introduced Falcon ASR, a speech recognition model.