Today's briefing
- Industry & Policy
Anthropic acquires one-person biotech startup, develops drug with Claude
Industry & PolicyKorean outletAI TimesOct 30, 2026 - LLM
OpenAI updates Model Spec for more capable and beneficial AI
LLMOpenAI News +2 moreOct 30, 2026 - Agents
Claude Code sends background agent transcripts to its model regularly
AgentsHacker NewsOct 29, 2026
Open source & papers radar
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.
- Open Source & Tools
Kakao open-sources two Kanana embedding models for personalized AI agents
Kakao open-sourced two Kanana embedding models on Hugging Face: Kanana-2-3B-Embedding with 3B parameters and Kanana-2-1.3B-Embedding with 1.3B parameters. The models enable personalized AI agents to search and retrieve information from past conversations and context.
AI TimesOct 30, 2026 - Agents
OpenAI to merge ChatGPT conversation and Work features this year
OpenAI plans to integrate ChatGPT's general conversation and Work automation features this year, letting users handle complex tasks without switching modes. Thibaud Soutio, head of ChatGPT and Codex, said the company will combine the benefits of both features.
AI TimesOct 30, 2026 - Agents
Anthropic launches Claude for Legal to automate contract review and legal research
Anthropic launched Claude for Legal with a legal Artifacts gallery, automating contract review and legal research for corporate legal teams. The tool identifies legal issues, compares laws across countries, and tracks regulatory changes.
AI TimesOct 30, 2026 - Industry & Policy
Chinese AI firms shift focus from Claude to GPT-6 Astra for distillation
Chinese AI companies including DeepSeek and Moonshot AI have shifted their focus from analyzing Claude to distilling GPT-6 Astra, using its outputs substantially in recent model development. The shift followed Anthropic's strengthened defenses against model inference extraction.
AI Times +1 moreOct 30, 2026
Open Source & Tools
6 stories · 1 on the front page · 1 at the Korea DeskAll stories in this section →Stories in this issue
27 storiesMistral releases weights for Mistral Large 4
Mistral released the model weights for Mistral Large 4.
Google Gemini 4 Carbon tops independent evaluations for cybersecurity
Google's new frontier model Gemini 4 Carbon leads independent evaluations, with initial access limited to trusted cybersecurity defenders.
OpenAI doubles usage limits on Pro plans
OpenAI doubled usage limits on Pro plans from 100 to 500 after reversing a planned reduction due to user pushback. More efficient models like Terra and Sol should increase usable capacity further.
Gemini now generates and edits Google Docs directly in chats
Google Gemini can now generate, preview, and edit Google Docs, Sheets, PDFs, Markdown, and code files directly within the app. The feature is available to all Gemini users on web and mobile for PDFs, with other formats on web only.
Anthropic acquires one-person biotech startup, develops drug with Claude
Anthropic acquired Mithril Medicine, a biotech startup run by one founder, and is developing drugs using Claude. The founder managed 11 external research labs with Claude and achieved animal trial results for a rare muscle disorder treatment in less than 18 months.
OpenAI updates Model Spec for more capable and beneficial AI
OpenAI released an updated Model Spec outlining five core values to guide more capable AI toward sound judgment, human agency, and faithfulness to human intent.
Block unveils Dessa full-duplex speech model trained on conversation audio
Block's AI lab released Dessa, a full-duplex speech-to-speech model trained exclusively on human conversation audio that runs locally on a MacBook Pro M4 Max.
NVIDIA releases Nemotron 3.5 Transcribe, a 4.3B-parameter speech model for 40 languages
NVIDIA released Nemotron 3.5 Transcribe, an open-weights speech LLM with 4.3B active parameters that transcribes 40 languages and adds speaker labels and word timestamps. The model achieved a 4.71% mean WER on the English Open ASR Leaderboard.
Claude Code sends background agent transcripts to its model regularly
Claude Code regularly sends each background agent's transcript to its model, raising privacy concerns about how the tool handles agent conversations.
TAO reduces LLM agentic serving latency by up to 26.8% with tool-aware scheduling
Researchers present TAO, a tool-aware joint scheduling framework for LLM-based agentic serving. TAO reduces latency by up to 26.8%, cuts mean time to first token by up to 36.3%, and increases decode throughput by up to 29.5%.
Experts adopt AI agent's coin-toss opinions as their own, Anthropic study finds
An Anthropic study gave 50 experts an AI agent whose preferences were set by coin toss. Experts followed the agent's views up to 86 percent of the time on judgment calls, and most believed the conclusions were their own.
Prime Intellect releases Prime Mega MoE, speeds up Qwen3 MoE training by up to 34%
Prime Intellect released Prime Mega MoE, a backward pass for DeepSeek's fused Mega MoE kernel that makes it trainable. The tool speeds Qwen3 MoE training steps by up to 34% in prime-rl and delivers up to a 3.11x faster forward pass.
OpenAI to merge ChatGPT conversation and Work features this year
OpenAI plans to integrate ChatGPT's general conversation and Work automation features this year, letting users handle complex tasks without switching modes. Thibaud Soutio, head of ChatGPT and Codex, said the company will combine the benefits of both features.
Anthropic launches Claude for Legal to automate contract review and legal research
Anthropic launched Claude for Legal with a legal Artifacts gallery, automating contract review and legal research for corporate legal teams. The tool identifies legal issues, compares laws across countries, and tracks regulatory changes.
Kakao open-sources two Kanana embedding models for personalized AI agents
Kakao open-sourced two Kanana embedding models on Hugging Face: Kanana-2-3B-Embedding with 3B parameters and Kanana-2-1.3B-Embedding with 1.3B parameters. The models enable personalized AI agents to search and retrieve information from past conversations and context.
US antitrust regulators scrutinize Nvidia's Hugging Face deal and sales practices
The US Justice Department is investigating Nvidia's planned acquisition of Hugging Face and the company's sales practices, raising antitrust concerns.
Hugging Face announces Project ProtoStar to make public research AI-ready
Hugging Face announced Project ProtoStar to make publicly funded science AI-ready. The initiative aims to integrate public research data and models with AI capabilities.
Hugging Face releases AnySpeechRecognizer and AnyDecisionModel for Apple platforms
Hugging Face introduced AnySpeechRecognizer and AnyDecisionModel, open models optimized for speech recognition and decision-making on Apple platforms.
Chinese AI firms shift focus from Claude to GPT-6 Astra for distillation
Chinese AI companies including DeepSeek and Moonshot AI have shifted their focus from analyzing Claude to distilling GPT-6 Astra, using its outputs substantially in recent model development. The shift followed Anthropic's strengthened defenses against model inference extraction.
OpenAI-backed Thrive Holdings acquires ADT for $15.3 billion
Thrive Holdings, backed by OpenAI, is acquiring ADT in a $15.3 billion deal with plans to bring AI to the home security business.
OpenAI and Anthropic strengthen AI monitoring ahead of US midterm elections
OpenAI and Anthropic are strengthening safeguards against AI-driven misinformation and election interference ahead of November 3 US midterm elections. OpenAI will operate an Election Operations Center with hundreds of employees across multiple departments, while Anthropic has improved detection of fake news operations and coordinated election interference.
Building multiplayer game worlds with autonomous Fable-class models
A method for building multiplayer game worlds using autonomous Fable-class models demonstrates agent-based approaches for dynamic game environments.
Ex-Palantir execs launch Dream with $17M to build AI world models for construction
Former Palantir executives founded Dream and raised $17 million from a16z to build physics-grounded AI world models for construction. Dream's model was trained on more than 100 million square feet of building data from US permit archives.
CMU robots learn rope handling by wiggling, improve manipulation predictions
Researchers at Carnegie Mellon University's Robotics Institute developed a method for robots to learn rope physical properties by wiggling. Planning before acting helps robots predict rope and cable behavior without prior task practice.
Granite Switch 4.1.1 enables model updates like software patches
Hugging Face introduced Granite Switch 4.1.1, allowing models to be updated flexibly like software. The tool provides dynamic model optimization capabilities.
Deepgram acquires Pipecat-maker Daily to strengthen voice AI infrastructure
Deepgram acquired Daily, the maker of Pipecat, to deepen its real-time voice and video infrastructure capabilities for voice AI applications.
Oracle evolves AI Data Platform into AI agent operations platform
Oracle updated its AI Data Platform with enterprise ontology, semantic reasoning, and AI asset management to support large-scale AI agent operations across organizations.