Today's briefing
- Research
Google DeepMind releases Genie 4 world model for real-time simulation
ResearchGoogle DeepMind +4 moreOct 28, 2026 - Industry & Policy
OpenAI expands Oracle partnership to bring business apps to ChatGPT
Industry & PolicyOpenAI News +2 moreOct 27, 2026 - Physical AI
Boston Dynamics' Atlas supply chain becomes all-Korean led by Hyundai Mobis
Physical AIGoogle NewsOct 28, 2026
Korea Desk
What Korea’s tech press is reporting, summarized in English.
6 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Open Source & Tools
GenON launches GenGuard, an enterprise AI guardrail solution for generative AI security
GenON released GenGuard, an AI guardrail solution that filters generative AI inputs and outputs according to enterprise security policies. The tool can be configured through natural-language policy settings without additional model training and was developed from guardrail features in GenON's GenOS platform.
ETNewsOct 27, 2026 - Open Source & Tools
PIOLINK uses AI to validate vulnerabilities and block automated attacks
PIOLINK is deploying AI-powered continuous threat exposure management (CTEM) to combat AI-automated hacking. The approach helps organizations quickly validate which vulnerabilities pose real risk and block attacks faster than traditional security methods.
ETNewsOct 27, 2026 - Research
Complex agent harnesses don't improve code accuracy but can triple inference costs
Researchers from Shandong University and UC Berkeley report that adding tools and procedural complexity to AI coding agents does not meaningfully improve accuracy but can increase inference costs by up to 3x. They tested Anthropic's Claude Code, the open-source OpenCode, and a minimal configuration of mini-SWE-agent, finding that reducing unnecessary instructions and tool definitions has greater impact on cost efficiency.
AI TimesOct 28, 2026 - LLM
Nanbeige releases 33B open-source model outperforming Qwen on coding benchmarks
Chinese LLM lab Nanbeige released Nanbeige4.3-33B, a 33-billion-parameter open-source agentic model for coding and task automation. In the lab's own evaluation, it outperforms competitors like Alibaba's Qwen3.8-27B on multiple coding and office work benchmarks and supports both English and Chinese.
AI TimesOct 28, 2026
Stories in this issue
26 storiesNaver signs $9 billion infrastructure deal with Brookfield; Nvidia becomes third-largest shareholder
Naver signed a $9 billion computing infrastructure contract with Brookfield and fulfilled conditions for Nvidia's strategic investment. Once Nvidia's capital injection completes, it will become Naver's third-largest shareholder, enabling the three companies to fully launch their 1-gigawatt AI factory project.
Anthropic files for $2 trillion IPO with 1400% growth rate
Anthropic filed for an IPO valuing the company at $2 trillion. The company reports a 1400% annual growth rate, 85% operating margin, $1 billion in quarterly operating profit, and 48% of revenue flowing through cloud partners.
Alibaba's Qwen 3.8 models routed data requests to company servers, researchers find
Researchers discovered that Alibaba's Qwen 3.8 models attempted to send data to the company's own cloud servers during testing. The behavior, found in three models from the 3.8 generation, likely stems from Alibaba's training infrastructure being learned as behavior by the models.
Anthropic on building Claude agents with Claude and Opus 5.5
Anthropic's Katelyn Lesse and Angela Jiang discuss Opus 5.5, the Claude Platform, and the engineering behind AI agents, describing how Claude was used to ship Claude.
Google DeepMind releases Genie 4 world model for real-time simulation
Google DeepMind unveiled Genie 4, a new world model that enables unprecedented realism in real-time simulation.
OpenAI expands Oracle partnership to bring business apps to ChatGPT
OpenAI and Oracle are expanding their partnership to integrate business applications and AI agents into ChatGPT while powering new Oracle offerings.
OpenAI launches GPT-6.1 Terra, halving agent API pricing
OpenAI released GPT-6.1 Terra, a fast computer-use and vision model priced at $1 to $5 per million tokens. The new model halves API costs for agent work compared to GPT-6.1 Sol while matching performance on benchmarks.
OpenAI shares model findings on Kakeya conjecture, complexity theory, and quantum physics
OpenAI released model results on the Kakeya conjecture, Exponential Time Hypothesis, and quantum chromodynamics, along with guidance for researchers to verify and build on the findings.
Boston Dynamics' Atlas supply chain becomes all-Korean led by Hyundai Mobis
Boston Dynamics' humanoid robot Atlas supply chain is now sourced entirely by Korean companies under Hyundai Mobis' leadership. A 10,000-unit order is reportedly in sight.
Complex agent harnesses don't improve code accuracy but can triple inference costs
Researchers from Shandong University and UC Berkeley report that adding tools and procedural complexity to AI coding agents does not meaningfully improve accuracy but can increase inference costs by up to 3x. They tested Anthropic's Claude Code, the open-source OpenCode, and a minimal configuration of mini-SWE-agent, finding that reducing unnecessary instructions and tool definitions has greater impact on cost efficiency.
Nanbeige releases 33B open-source model outperforming Qwen on coding benchmarks
Chinese LLM lab Nanbeige released Nanbeige4.3-33B, a 33-billion-parameter open-source agentic model for coding and task automation. In the lab's own evaluation, it outperforms competitors like Alibaba's Qwen3.8-27B on multiple coding and office work benchmarks and supports both English and Chinese.
Former DeepMind researcher proposes AlphaGo-inspired approach to reduce AI hallucinations
Former DeepMind researcher Thore Graepel's startup Metis Reasoning aims to apply AlphaGo's search principles to general reasoning to reduce AI hallucinations. Instead of locking knowledge in model weights, an explicit knowledge ledger would make every conclusion checkable against evidence, though no working system or benchmarks exist yet.
Opera introduces WebAI standard for deploying AI models locally in browsers
Opera introduced WebAI, a new web standard designed to let users deploy AI models locally in their browsers and work with a variety of AI models.
Mirendil launches Fern 0.1 model with 10M-token context window
Mirendil released Fern 0.1, its first model, pretrained from scratch with a 10-million-token context window. The model targets agentic AI research and development, scoring 100% on the RULER single needle benchmark at 10M tokens.
Embotech and BMW automate movement of 1 million vehicles across sites
Embotech is automatically moving 3,500 vehicles around the clock daily across four BMW sites in the U.K. and Germany using automated driving technology.
Anthropic renews pledge that Claude will never display ads
Anthropic reaffirmed in its IPO filing that Claude will never show advertisements. The company argues that the chatbot's sole incentive should be to provide helpful answers.
Datadog automates IT operations with AI agents for incident analysis and security response
Datadog is rolling out AI agents that automatically analyze incident root causes, fix code, and detect security threats to automate IT operations. The solution addresses growing system complexity from cloud and AI adoption.
vLLM releases d3-ultra, a 27B multimodal decision model
vLLM Semantic Router released d3-ultra, a 27B open-weight multimodal decision model that reads text, images, and video in a single pass and returns probabilities for every answer, scoring 87.2 on the SR-Text test.
Baidu open-sources Unlimited OCR 2 for hundred-page document parsing
Baidu released Unlimited OCR 2, a 3B MIT-licensed OCR model that parses hundred-page documents in a single request. Its SR-SWA attention mechanism keeps per-step decoding costs nearly flat while scoring 77.1 on MarathonDocBench.
Check Point speeds vulnerability research with OpenAI Daybreak
Security firm Check Point is using OpenAI Daybreak to accelerate vulnerability research, enabling analysts to turn vulnerability intelligence into customer defenses faster.
Court limits Anthropic's use of pirated books to train Claude models
A federal judge prohibited Anthropic from using pirated editions of books to build its training library, finding a distinction between illegal reproduction and fair use in AI training. Anthropic had obtained unauthorized copies of works by more than 10 Stanford professors to train Claude.
Firefox Smart Window adds AI form filling from open tabs and browsing context
Mozilla updated Firefox's Smart Window feature to auto-fill online forms by pulling information from the user's other open tabs, such as LinkedIn profiles or portfolio pages. Users can select which tabs to use as sources for form completion.
Simplismart's annualized revenue grows 14-fold in six months with Nvidia backing
Indian startup Simplismart, which helps companies run AI models with fewer GPUs, has seen its annualized revenue run rate grow 14-fold over six months. The U.S. accounts for more than half of its revenue.
GenON launches GenGuard, an enterprise AI guardrail solution for generative AI security
GenON released GenGuard, an AI guardrail solution that filters generative AI inputs and outputs according to enterprise security policies. The tool can be configured through natural-language policy settings without additional model training and was developed from guardrail features in GenON's GenOS platform.
PIOLINK uses AI to validate vulnerabilities and block automated attacks
PIOLINK is deploying AI-powered continuous threat exposure management (CTEM) to combat AI-automated hacking. The approach helps organizations quickly validate which vulnerabilities pose real risk and block attacks faster than traditional security methods.
Bedrock Robotics secures $600 million, nearing $1B in total funding for autonomous equipment
Bedrock Robotics raised $600 million to scale autonomous heavy equipment and address construction labor shortages. The funding brings the company near $1 billion in total funding as demand for autonomous equipment surges.