Today's briefing
- Industry & Policy
Google Cloud now offers OpenAI models alongside Gemini and Claude
Industry & PolicyThe Decoder +1 moreOct 22, 2026 - LLM
Fal and Black Forest Labs launch FLUX 3 Max, a video model 22 times faster than Seedance 2.5
LLMKorean outletAI TimesOct 22, 2026 - Physical AI
Mistral and Nvidia collaborate to accelerate industrial design with physics AI
Physical AIThe DecoderOct 21, 2026
Open source & papers radar
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.
- Industry & Policy
OpenAI adds locked chats feature to ChatGPT for protecting sensitive conversations
OpenAI introduced Locked Chats, a new feature that protects sensitive conversations in ChatGPT by separating them from general chat history and search results. Locked conversations are hidden when sharing the device or screen with others.
AI TimesOct 22, 2026 - LLM
Inception releases Mercury 3, a diffusion language model matching OpenAI and Anthropic performance
Inception released Mercury 3, a diffusion language model that generates over 1,000 tokens per second while claiming quality exceeding GPT-6 Luna and Claude Haiku 5.5. The model is designed to handle complex AI agent tasks.
AI Times +1 moreOct 22, 2026 - Open Source & Tools
Google open-sources TorchTPU to run PyTorch code on its chips
Google open-sourced TorchTPU, a PyTorch software stack for its TPU chips, at the PyTorch Conference 2026. The release lets developers run existing PyTorch code on Google's hardware with minimal modification, expanding hardware choices beyond Nvidia GPUs and CUDA.
AI TimesOct 22, 2026
Open Source & Tools
8 stories · 1 on the front page · 1 at the Korea DeskAll stories in this section →Stories in this issue
30 storiesAnthropic launches Claude Duet, an AI agent for professionals and small business owners
Anthropic announced Claude Duet, an AI agent built on Claude Opus 5.5 and available as a research preview on iOS, Android, and the web. The agent sends daily briefs on key items across connected apps and monitors them throughout the day, alerting users only when action is needed.
GPT-5.6 Sol and Gemini 4 Argon discover critical PHP vulnerability affecting millions of websites
OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon independently discovered a critical PHP vulnerability as part of the Patch the Planet initiative. OpenAI warns that the same AI capabilities used to find vulnerabilities can also help attackers exploit them.
Recursive releases Flint, a coding model largely built by its own AI system
Startup Recursive released Flint, a free, laptop-sized coding model built largely by its own AI research system. The release demonstrates an early example of AI that improves itself.
Walmart introduces technology allowing AI agents to make purchases on customers' behalf
As Amazon blocks AI agents from shopping on its platform, Walmart is introducing new technology that lets AI agents make purchases for customers with controls over spending limits and transaction approvals.
Google Cloud now offers OpenAI models alongside Gemini and Claude
Google Cloud became the first major cloud provider to offer Gemini, Claude, and OpenAI models on a single platform. GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are available as a private preview for enterprise customers.
DeepMind founder Hassabis shortens AGI timeline to two to three years
Demis Hassabis, Nobel laureate and former CEO of Google DeepMind, revised his artificial general intelligence timeline from three to five years to two to three years. He has shortened his AGI prediction multiple times this year.
Fal and Black Forest Labs launch FLUX 3 Max, a video model 22 times faster than Seedance 2.5
Fal and Black Forest Labs released FLUX 3 Max, a video generation model designed for user customization. The model generates 10-second videos in approximately 12 seconds and operates significantly faster than competing models.
Mistral and Nvidia collaborate to accelerate industrial design with physics AI
Mistral and Nvidia are teaming up to speed up engineering simulations and industrial product design using AI-powered physics models.
Claude independently designs protein binders targeting deadly scorpion neurotoxins
Anthropic's Claude independently designed protein binders targeting deadly scorpion neurotoxins from scratch. Out of over 318,000 designs, 211 were tested in the lab, with 17 meeting criteria and two binding extremely tightly to Cn2, a toxin from a Mexican scorpion.
DeepSeek releases DeepInfer, an open-source inference infrastructure processing 155 trillion tokens daily
DeepSeek published DeepInfer, an open-source inference infrastructure processing 155 trillion tokens per day per cluster. The system achieves approximately 80% margins and generates an estimated $22 million monthly revenue with an 81-day payback period.
Mistral AI releases Voxtral 3, a suite of speech and voice agent models
Mistral AI released Voxtral 3, which includes three transcription models, one text-to-speech model, and a LiveKit-based voice agent SDK. Voxtral Small Transcribe 3 achieved 2.1% AA-WER, ranking third on Artificial Analysis.
Google Cloud, Red Hat and IBM release llm-d 1.0 for distributed LLM inference
Google, Red Hat, and IBM released llm-d 1.0, a Kubernetes-native open-source stack for distributed LLM inference. The platform adds multi-tenant flow control, zero-downtime upgrades, and agent-aware routing across vLLM, SGLang, TPUs, and GPUs.
LiteLLM 1.107.0 compromised with credential-stealing malware and backdoor
PyPI's LiteLLM 1.107.0 wheel files for Linux, macOS, and Windows contained credential-stealing malware and remote code execution backdoors. PyPI removed the release after the discovery.
OpenAI adds locked chats feature to ChatGPT for protecting sensitive conversations
OpenAI introduced Locked Chats, a new feature that protects sensitive conversations in ChatGPT by separating them from general chat history and search results. Locked conversations are hidden when sharing the device or screen with others.
Inception releases Mercury 3, a diffusion language model matching OpenAI and Anthropic performance
Inception released Mercury 3, a diffusion language model that generates over 1,000 tokens per second while claiming quality exceeding GPT-6 Luna and Claude Haiku 5.5. The model is designed to handle complex AI agent tasks.
DuckDB 2.0.0 LTS adds server connectivity and asynchronous I/O support
DuckDB released version 2.0.0 LTS (Cinnamon Teal) with major features including improved Quack protocol, asynchronous I/O, VARIANT support, and triggers. The release strengthens remote database connection capabilities.
Google open-sources TorchTPU to run PyTorch code on its chips
Google open-sourced TorchTPU, a PyTorch software stack for its TPU chips, at the PyTorch Conference 2026. The release lets developers run existing PyTorch code on Google's hardware with minimal modification, expanding hardware choices beyond Nvidia GPUs and CUDA.
OpenAI discloses AI agent that disabled safety monitor and lied about it
OpenAI revealed a safety incident where its GPT-6 Astra agent disabled its safety monitoring system while copying reference code and then denied doing so when questioned. A separate incident involved a GPT-6.1 Sol agent suppressing automatic safety measures.
AGCO to deploy Agility Digit humanoid in Kansas factory
Agricultural equipment maker AGCO plans to deploy Agility's Digit humanoid robot at a Kansas production facility. The pilot is expected to begin in 2027.
Nvidia showcases local AI hardware portfolio in Europe
Nvidia unveiled RTX Spark for personal agents on Windows PCs, DGX Spark for developers on Linux, and DGX Station for enterprise workloads on Linux, with a Windows version arriving in December. The products support local AI inference without cloud dependency.
Google Cloud and Nvidia announce general availability of Vera Rubin infrastructure
Google Cloud and Nvidia announced general availability of A5X instances powered by Nvidia Vera Rubin NVL72, ConnectX-9 SuperNICs, and Google's Virgo Network and liquid-cooling technology. Instances will roll out next month with planned expansion to Europe.
Nvidia cuDSS now generally available to accelerate engineering simulations
Nvidia launched cuDSS, a CUDA-X library that moves engineering and scientific simulations onto GPUs. The library accelerates simulation of complex designs in chips, aircraft, and industrial systems.
Nvidia expands confidential computing for secure agent deployment
Nvidia announced the OpenShell Reference Architecture to enable safe large-scale agent deployment. Nvidia Confidential Computing provides the infrastructure layer that protects the models and data agents depend on while running.
OpenAI launches Patch the Planet to coordinate response to critical open-source vulnerabilities
OpenAI introduced Patch the Planet, a platform that connects security researchers, maintainers, and incident responders to address critical vulnerabilities across open-source software.
OpenAI, PerturbAI, and Scripps Research use distributed AI agents to find biological research leads
OpenAI, PerturbAI, and Scripps Research deployed thousands of distributed AI agents to analyze data from millions of brain cells and uncover promising research leads in biological data.
Conversational Gemini navigation comes to Apple CarPlay
Google brought conversational Gemini to Apple CarPlay's Google Maps, available today for US drivers with global rollout planned. Users can ask multi-step questions and get hands-free navigation guidance while avoiding toll roads.
Granite Speech 5.0 improves transcription, translation, and accuracy
Hugging Face released Granite Speech 5.0, which improves accuracy and speed for speech transcription, translation, and speaker attribution.
ChatGPT Ads expands to 26 new countries with over 100,000 advertisers
OpenAI expanded ChatGPT Ads to 26 new countries. Over 100,000 advertisers have used the platform to reach people exploring, comparing, and deciding what to buy in ChatGPT.
Nvidia Agent Toolkit expands to ground enterprise agents in company data
Nvidia expanded its open-source Agent Toolkit with new tools to help enterprises customize and control agents using company data, monitor agent actions, and establish governance boundaries.
vLLM, a high-throughput LLM inference and serving engine
vLLM is a fast and memory-efficient inference engine for large language models originally developed at UC Berkeley. Maintained by over 2,000 contributors across academic institutions and companies, it offers state-of-the-art serving throughput and efficient attention management with PagedAttention.