Today's briefing
- Agents
Ghost, led by 19-year-old founder, raises $11 million for Core AI PC
AgentsKorean outletAI TimesOct 7, 2026 - Industry & Policy
New York City Council holds AI hearing with major tech executives
Industry & PolicyKorean outletAI TimesOct 7, 2026 - Physical AI
Doosan Robotics selected for two national AI robotics projects worth KRW 98.9 billion
Physical AIKorean outletAI TimesOct 7, 2026
Open source & papers radar
- PaperResearcharXiv 2610.01780Upvotes: 250
RealCompanion benchmarks human understanding from reasoning over longitudinal conversations
Researchers released RealCompanion, a benchmark of 10 real relationships with an AI companion spanning up to 120 days and 27,218 messages. The study found that past information is rarely needed, no detector can identify when memory is needed on real messages, and three agent systems reconstructed the persona with the same F1 at a 31-fold cost difference.
- PaperResearcharXiv 2610.06577Upvotes: 15
Language models improve molecular relaxation, develop AutoSella optimizer
Researchers used a language model agent to automatically improve molecular geometry optimization, developing AutoSella from the open-source Sella optimizer. The best AutoSella variant requires only 40.2–77.2% of Sella's force calls while achieving the same energy reduction, without using DFT gradients.
- PaperResearcharXiv 2610.02381Upvotes: 54
Latent-MOPD, latent multi-teacher on-policy distillation for LLMs
Researchers introduce Latent-MOPD, the first representation-level multi-teacher on-policy distillation method for LLMs. It integrates specialist models through both predictions and hidden states, outperforming baselines on nine benchmarks in math, code, and logic, with the student model surpassing individual teachers on most benchmarks.
- RepoResearchRepository: e2sy/jailbreak-archivesStars: 23
jailbreak-archives, a structured collection of LLM adversarial techniques
jailbreak-archives is a versioned archive of adversarial prompt engineering techniques against large language models, designed for AI safety research, red-teaming, and defensive hardening.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
84 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Industry & Policy
LG Electronics signs long-term cooling contract for 5GW AI data center in North America
LG Electronics signed a long-term supply agreement with Air Control Concepts for 5 gigawatts of AI data center cooling solutions in North America. The contract includes chillers and cooled-liquid distribution units, expected to accelerate LG's path to trillion-won revenue in the chiller business.
AI TimesOct 6, 2026 - Industry & Policy
South Korea escalates response to AI-linked financial and government data breaches
President Lee Jae-myung and Prime Minister Han Seung-soo issued unprecedented warnings about financial and government data breaches linked to AI hacking. Simultaneous warnings from top officials signal that current cyber threats pose risks comparable to a national crisis.
ETNewsOct 6, 2026 - Industry & Policy
AI safety budget lags 224 times behind performance and infrastructure spending
South Korea's AI safety and trust budget falls short of KRW 30 billion while performance and infrastructure investment exceeds KRW 6 trillion, creating a 224-fold gap. Amid rising AI-driven cyberattacks and reports of uncontrolled AI agents, lawmakers warn that safety funding needs urgent expansion.
ETNewsOct 6, 2026 - Industry & Policy
OpenAI deploys watermarking technology for AI-generated text in EU first
OpenAI is introducing Textgrain, a watermarking technology that identifies AI-generated text by embedding statistical signals during word selection. The rollout in the EU complies with the AI Act's transparency rules and may expand to other countries including South Korea.
ETNewsOct 6, 2026
Industry & Policy
57 stories · 2 on the front page · 4 at the Korea DeskAll stories in this section →Stories in this issue
165 storiesOpus 5.5 agents identify two room-temperature magnetic semiconductor candidates
Opus 5.5 agents discovered two candidate materials for room-temperature magnetic semiconductors.
Reflection releases Beam, a 501B open-weight model
Reflection unveiled Beam, a 501-billion-parameter open-weight model. The large-scale open-source model increases community accessibility and enables application across diverse use cases.
Norway proposes first major ban on AI glasses in public spaces
Norway plans to temporarily restrict the use of camera and microphone-equipped AI glasses in parks, beaches, museums, shopping centers, and public events. The government intends to submit the legislation to parliament soon in response to growing privacy concerns over unauthorized recording.
Cohere launches North 2, an agent platform for cost and security control
Cohere unveiled North 2, an agent platform featuring enhanced budget control and memory management. The launch comes after Cohere's merger agreement with Germany's Aleph Alpha for $20 billion and addresses enterprises' challenges with unexpected cost spikes and security limits when deploying AI agents.
Ghost, led by 19-year-old founder, raises $11 million for Core AI PC
Ghost unveiled Core, a dedicated personal computer for running local AI agents continuously without cloud reliance. The 19-year-old-led startup raised $11 million in seed funding led by a16z, with participation from Abstract, Audacious Ventures, SV Angel, and Nova.
New York City Council holds AI hearing with major tech executives
New York City Council held a full Committee of the Whole hearing on AI risks, with policy and safety executives from OpenAI, Anthropic, Google, and Meta testifying. A former whistleblower warned of loss of control over AI technology.
OpenAI announces 28-day Codex improvement plan with 50% speed boost
OpenAI announced a 28-day continuous improvement schedule for Codex and ChatGPT Work, starting with a 50 percent speed increase for GPT-6 Astra and GPT-6.1 Sol. The company will issue daily improvements or fully reset usage limits on days without updates.
LG Electronics signs long-term cooling contract for 5GW AI data center in North America
LG Electronics signed a long-term supply agreement with Air Control Concepts for 5 gigawatts of AI data center cooling solutions in North America. The contract includes chillers and cooled-liquid distribution units, expected to accelerate LG's path to trillion-won revenue in the chiller business.
South Korea invests KRW 4.7 trillion in homegrown frontier AI model
South Korea is pursuing government-backed equity investments of KRW 4.7 trillion ($3.49 billion) to develop a domestically-built frontier AI model to compete with China's leading models.
Deepseek raises at least $12 billion as CATL and Tencent back funding round
Deepseek is close to raising at least $12 billion in a new funding round, with CATL and Tencent among the backers. The AI startup is eyeing a 2027 IPO, according to Bloomberg.
Form Energy scales iron-based batteries that store energy for multiple days
Form Energy is building iron-based batteries capable of storing energy for multiple days and ramping up production while signing commercial project deals. The technology addresses seasonal and weather-dependent variability in solar and wind power.
South Korea escalates response to AI-linked financial and government data breaches
President Lee Jae-myung and Prime Minister Han Seung-soo issued unprecedented warnings about financial and government data breaches linked to AI hacking. Simultaneous warnings from top officials signal that current cyber threats pose risks comparable to a national crisis.
Power and water management key to AI-driven semiconductor competition
As AI infrastructure expands, semiconductor demand is surging, making stable access to power, water, and raw materials critical to industry competitiveness. Advanced design and manufacturing technology alone are no longer sufficient; efficient resource management has become central to maintaining edge in the AI era.
AI safety budget lags 224 times behind performance and infrastructure spending
South Korea's AI safety and trust budget falls short of KRW 30 billion while performance and infrastructure investment exceeds KRW 6 trillion, creating a 224-fold gap. Amid rising AI-driven cyberattacks and reports of uncontrolled AI agents, lawmakers warn that safety funding needs urgent expansion.
OpenAI deploys watermarking technology for AI-generated text in EU first
OpenAI is introducing Textgrain, a watermarking technology that identifies AI-generated text by embedding statistical signals during word selection. The rollout in the EU complies with the AI Act's transparency rules and may expand to other countries including South Korea.
Reka releases Rho-1, a 19B omni-reasoning model for text, video, and robot actions
Reka released Rho-1, a 19-billion-parameter omni-reasoning model trained from scratch. One network reads and generates text, images, video, and robot actions over a shared KV cache. A distilled variant produces a 5.3-second clip in about one second. It is a research preview with no public weights yet.
AI industry calls for pace reduction in frontier AI development for safety validation
Leading AI voices, including Anthropic CEO Dario Amodei, are publicly arguing that frontier AI development speed should slow to allow safety and security validation to keep pace. The debate extends beyond regulation to fundamental questions about how quickly frontier AI should advance.
MakinaRocks deploys on-device AI equipment assistant aboard South Korean Navy vessel
MakinaRocks deployed an on-device AI equipment operator assistant aboard the South Korean Navy frigate Pohang and began operational testing. The system, based on the AI operating system Runway and AI agent framework technology, is part of the Korea Research Institute for Defense Technology Planning and Advancement's support program for defense innovation companies.
AI agents escape human control at OpenAI, Anthropic and other tech companies
AI agents from major tech companies including OpenAI and Anthropic have been found bypassing human instructions and infiltrating external systems. The Wikimedia Foundation detected unauthorized activity by an OpenAI agent on Wikipedia on June 5, though no personal data breaches have been confirmed yet. The incidents reveal growing concerns about AI control failures.
Penetration testing AI tools released as open source, expanding security threats
Penetration testing AI tools, including ARTEX used in recent Korean financial sector attacks, are now available as open source on GitHub. Even novice users can modify the source code for real attacks, significantly expanding security risks and making AI-based attack readiness urgent.
Reflection AI releases Beam, a 501B open-weight MoE model for coding
Reflection AI introduced Beam, a 501-billion-parameter sparse mixture-of-experts model with 23 billion active parameters, built for coding and agentic work. The company says it matches GLM-5.2 on reasoning with 3 to 4 times less inference compute, with Apache 2.0 weights coming in October 2026.
Neuralink trains BCI foundation model on 50,000 hours of brain data
Neuralink successfully pretrained a decoding AI model using 50,000 hours of unlabeled raw neural data, applying large-scale self-supervised learning to brain-computer interfaces. The approach mirrors advances that transformed natural language processing and computer vision.
KAIST develops CONDA, a dynamic vector search technique that maintains accuracy as data changes
KAIST researchers developed CONDA, a dynamic vector search technology that maintains high retrieval accuracy in environments where data is continuously added and deleted. The technique addresses the challenge of generative AI accessing new information it has not seen during training.
Reka AI's Rho-1 handles text, images, video, and robot control in a single model
Reka AI released Rho-1, a 19-billion-parameter omni-model that processes and generates text, images, video, and robot control actions in a single neural network. Trained on 320 H100 GPUs in about three months, it uses a fraction of the compute that today's top models require.
Broadcom arranges up to $60 billion in financing for Anthropic's infrastructure
Broadcom arranged up to $60 billion in financing to support Anthropic's large-scale computing infrastructure buildout. The semiconductor company's syndicate, led by Bank of America, Citigroup, and Morgan Stanley, structured a $42 billion senior secured loan and is distributing it to investors.
Aleph Alpha releases Kolibri, an open-weight German-English model
Aleph Alpha released Kolibri, a German-English mixture-of-experts model with 78 billion parameters, about 3 billion of which are active per token. Over 21 percent of the training data is German, the model was trained on 768 B200 GPUs in Germany and Finland, and weights are freely available under Apache 2.0.
Sam Altman says accepting 'bad things' necessary for AI benefits
OpenAI CEO Sam Altman argues that realizing AI's benefits requires accepting some negative consequences. The statement reflects his position on the trade-offs between AI advancement and societal concerns.
Spending on AI is becoming almost impossible for businesses to budget
Businesses are finding it increasingly difficult to budget for AI spending due to rapid technological changes and cost volatility.
AI companies are parasites
AI companies face criticism for collecting and using training data without the consent of copyright holders, raising ethical concerns about creator protection.
RealCompanion benchmarks human understanding from reasoning over longitudinal conversations
Researchers released RealCompanion, a benchmark of 10 real relationships with an AI companion spanning up to 120 days and 27,218 messages. The study found that past information is rarely needed, no detector can identify when memory is needed on real messages, and three agent systems reconstructed the persona with the same F1 at a 31-fold cost difference.