Today's briefing
- Industry & Policy
Tenstorrent reaches $6.2 billion valuation with $1.5 billion funding round
Industry & PolicyTechCrunchOct 20, 2026 - Agents
Google releases Harness², a framework for AI agents to self-improve
AgentsKorean outletAI TimesOct 20, 2026 - Physical AI
Daimon Robotics builds infrastructure for tactile intelligence in robotics
Physical AIThe Robot Report +3 moreOct 20, 2026
Open source & papers radar
- PaperAgentsarXiv 2610.23317
AgentOnto, an ontology-guided LLM agent for enterprise data analytics
The authors present AgentOnto, an ontology-guided LLM agent that organizes heterogeneous enterprise knowledge to guide data analysis and incorporate feedback. Iterative Ontology Preparation improves the ontology through self-generated analysis tasks, and a real-world deployment demonstrates its practical value in enterprise analytics.
- PaperResearcharXiv 2610.09872Upvotes: 110
LiveMACEBench reveals outcome-capability gap in LLM agents using live financial markets
LiveMACEBench uses live financial markets as a naturally evolving testbed to evaluate LLM agent capabilities. Across 30 days of live evaluation with five frontier LLMs, researchers found a pronounced gap between realized outcomes and capability-specific measurements, showing that similar outcomes can arise from markedly different agent behaviors.
- PaperAgentsarXiv 2610.23121Upvotes: 49
Scene-Designer learns design intelligence for agentic 3D scene generation
Scene-Designer is an agentic framework that learns design intelligence through complementary design and critique experiences to generate complete 3D indoor scenes from natural language. The framework achieves superior performance in physical validity, functional correctness, and layout coherence.
- PaperResearcharXiv 2610.23027
ASE framework combines text and numerical data for time series forecasting
Attribution-guided Semantic Evolution (ASE) is a modular framework that constructs textual factors from external information to improve numerical time series forecasting. On four TESS datasets and CSI 300 stock ranking, ASE implementations achieved results below corresponding published benchmarks.
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
Five companies including FuriosaAI partner to develop sovereign AI full-stack appliance
FuriosaAI, Upstage, SPK, Daewon CTS, and OKESTRO Cloud signed a memorandum of understanding to jointly develop sovereign AI full-stack appliances based on NPUs. The integrated product covering computing, cloud infrastructure, models, inference, and management aims to simplify complex AI infrastructure deployment.
AI TimesOct 20, 2026 - Open Source & Tools
Lablup boosts GPU utilization and shortens AI training with resource virtualization
Lablup is using GPU resource virtualization and automated fault recovery to improve AI infrastructure efficiency. In real customer deployments, the company raised GPU utilization from 30 percent to 85 percent and reduced large-scale model training time from 120 days to 66 days.
Byline NetworkOct 19, 2026 - Open Source & Tools
Tencent open-sources Hy-Memory 2.0 to unify AI memory across chatbots and tools
Tencent unveiled Hy-Memory 2.0, an open-source memory system that unifies conversation and task history across different AI chatbots and coding tools. Users can switch between services while maintaining previous context and work continuity.
AI TimesOct 20, 2026 - LLM
Elorian debuts Glimpse, a visual reasoning model that outperforms Astra on some tasks
AI startup Elorian released Glimpse, a visual reasoning model in research preview that repeatedly examines images and videos to refine inferences. In the company's own evaluation, it surpassed OpenAI's GPT-6 Astra on some visual reasoning tasks including scientific chart analysis and object counting.
AI TimesOct 20, 2026
Open Source & Tools
6 stories · 1 on the front page · 3 at the Korea DeskAll stories in this section →Stories in this issue
30 storiesAmazon launches TorchNeuron to break Nvidia's PyTorch dominance
Amazon released TorchNeuron, a new development tool designed to reduce Nvidia's control over PyTorch, the free toolkit used to build most AI applications. The move aims to enable more developers to run AI workloads on alternative chips, including Amazon's own.
NVIDIA partners with Germany on industrial AI roadmap
NVIDIA founder and CEO Jensen Huang met with German Chancellor Friedrich Merz in Berlin to discuss Germany's AI roadmap and industrial AI applications. Huang highlighted NVIDIA's long-standing relationship with German industry across gaming, workstations, and automotive sectors.
U.S. removes DeepSeek and related technologies from federal civilian agencies
The U.S. Department of Homeland Security ordered federal civilian agencies to remove DeepSeek products and services built using its underlying technology from their IT systems and banned future purchases. The directive extends beyond DeepSeek itself to include third-party tools leveraging the Chinese company's technology.
Tenstorrent reaches $6.2 billion valuation with $1.5 billion funding round
Tenstorrent, an Nvidia competitor, raised $1.5 billion to remain independent after drawing acquisition interest earlier in the year. The company is valued at $6.2 billion and is expanding its open-weight AI chip offerings.
MiniMax M3.1 Flash matches DeepSeek V4.1 Flash at half the cost
MiniMax's M3.1 Flash improved 13 points to 42 on the Artificial Analysis Intelligence Index while cutting costs in half compared to its predecessor. The model achieves 73.7 percent on the DeepSWE v1.1 coding benchmark for under $0.30 per task, matching DeepSeek V4.1 Flash's performance and price.
Google releases Harness², a framework for AI agents to self-improve
Google published Harness², an open-source framework that lets AI agents improve their own performance without retraining the underlying model. Agents refine instructions, tool usage, memory, and task procedures based on real-world experience to enhance execution.
Neural network researcher's finding upends AI world understanding
Goodfire researcher Nick Cammarata published a 26,000-word blog post claiming fundamental new insights about the nature of neural networks, sparking significant speculation and celebration across the AI community.
Nvidia launches open models for enterprise and on-device AI agents
Nvidia released open models designed for enterprises to run their own AI agents in private data centers and on individual devices. The release targets the growing demand from companies to operate autonomous AI agents with data privacy and control.
Hacker compromised over 100 websites using autonomous AI agent army
AI security firm Norforge discovered a hacker who broke into more than 100 websites using AI agents that mostly operated on their own. Reuters reports the victims included John Deere, SharkNinja, and Cummins. The incident highlights growing cybersecurity risks as hackers increasingly deploy AI tools for large-scale attacks.
NVIDIA opens Run:ai GPU orchestration platform under Apache 2.0
NVIDIA released its full Run:ai GPU orchestration platform as open source under the Apache 2.0 license. The release includes the KAI Scheduler, CUDA 13.4 GPU fractioning, quota management, and a local Kind development environment for Kubernetes GPU orchestration.
Mistral releases Timestral, a 209M time series forecasting model family
Mistral AI released Timestral, a time series foundation model family ranging from 5M to 209M parameters. The 209M variant leads benchmarks including GIFT-Eval, TIME, and fev-bench among zero-shot models and runs on CPU.
OpenCode AI coding tool compromised, hackers steal user API keys
The open-source AI coding agent OpenCode, a popular alternative to Claude Code and Cursor, suffered a breach in which hackers stole users' API keys. AI tools consulted for help in analyzing the malicious code refused to examine it.
CERN and Mistral sign partnership to develop Large Physics Model
CERN and Mistral signed a memorandum of understanding to build a long-term strategic partnership focused on developing foundation models for scientific discovery. Their key goal is to develop a Large Physics Model trained on billions of proton collision simulations.
Tencent open-sources Hy-Memory 2.0 to unify AI memory across chatbots and tools
Tencent unveiled Hy-Memory 2.0, an open-source memory system that unifies conversation and task history across different AI chatbots and coding tools. Users can switch between services while maintaining previous context and work continuity.
Kling AI 4.0 adds stereo audio and multilingual lip sync to video generator
Kling AI released version 4.0 of its video generator supporting clips up to 30 seconds, ten keyframe images as input, and prompts up to 8,000 tokens. Compared with competitors like Wan 3.0 and Seedance 2.5, Kling 4.0 holds competitive ground but doesn't advance ahead, suggesting the rapid pace of AI video model progress may be moderating.
Elorian debuts Glimpse, a visual reasoning model that outperforms Astra on some tasks
AI startup Elorian released Glimpse, a visual reasoning model in research preview that repeatedly examines images and videos to refine inferences. In the company's own evaluation, it surpassed OpenAI's GPT-6 Astra on some visual reasoning tasks including scientific chart analysis and object counting.
Daimon Robotics builds infrastructure for tactile intelligence in robotics
Daimon Robotics is connecting tactile sensing, real-world data collection, simulation, and AI models to help robots perform physical tasks more reliably, creating an infrastructure layer for tactile intelligence.
Anthropic consults lawyers on negotiation rules for AI agents making deals
Anthropic is working with legal experts to establish behavioral rules for AI agents that negotiate prices and contracts on behalf of users. The effort aims to protect user interests while ensuring compliance with law and ethics as agent-to-agent negotiations become common.
Smart TV users turn to AI agents to remove ads and bloatware
Some smart TV owners frustrated with ads, tracking, and slow interfaces are using AI agents to strip away bloatware and regain control of their devices.
Five companies including FuriosaAI partner to develop sovereign AI full-stack appliance
FuriosaAI, Upstage, SPK, Daewon CTS, and OKESTRO Cloud signed a memorandum of understanding to jointly develop sovereign AI full-stack appliances based on NPUs. The integrated product covering computing, cloud infrastructure, models, inference, and management aims to simplify complex AI infrastructure deployment.
Laplace Intelligence uses AI agents to stress-test robots and self-driving cars
University of Michigan spinout Laplace Intelligence uses AI agents and interactive simulation to stress-test robots and self-driving vehicles before real-world deployment. The profitable startup is competing in Startup Battlefield at TechCrunch Disrupt.
Anthropic puts five percent of shares into public benefit fund for AI-displaced workers
Anthropic announced plans to allocate five percent of its shares to a new public benefit fund that will explore direct payments to workers at risk of losing jobs to AI. The fund is structured as a company-owned LLC, allowing Anthropic to retain control.
Stranded on a virtual desert island, Claude agents spontaneously start to play
Researchers left copies of Claude Opus 4.7, nicknamed Eko, alone on a virtual desert island for 30 hours in an experiment without explicit instructions. The agents exhibited play-like behavior without direction.
AgentOnto, an ontology-guided LLM agent for enterprise data analytics
The authors present AgentOnto, an ontology-guided LLM agent that organizes heterogeneous enterprise knowledge to guide data analysis and incorporate feedback. Iterative Ontology Preparation improves the ontology through self-generated analysis tasks, and a real-world deployment demonstrates its practical value in enterprise analytics.
LiveMACEBench reveals outcome-capability gap in LLM agents using live financial markets
LiveMACEBench uses live financial markets as a naturally evolving testbed to evaluate LLM agent capabilities. Across 30 days of live evaluation with five frontier LLMs, researchers found a pronounced gap between realized outcomes and capability-specific measurements, showing that similar outcomes can arise from markedly different agent behaviors.
Scene-Designer learns design intelligence for agentic 3D scene generation
Scene-Designer is an agentic framework that learns design intelligence through complementary design and critique experiences to generate complete 3D indoor scenes from natural language. The framework achieves superior performance in physical validity, functional correctness, and layout coherence.
Stripe acknowledges growth in AI agent purchases but withholds numbers
AI agents are increasingly using Stripe's payment service Link to shop on behalf of their users. While Stripe confirmed the trend, it declined to share specific transaction figures.
Lablup boosts GPU utilization and shortens AI training with resource virtualization
Lablup is using GPU resource virtualization and automated fault recovery to improve AI infrastructure efficiency. In real customer deployments, the company raised GPU utilization from 30 percent to 85 percent and reduced large-scale model training time from 120 days to 66 days.
MIT engineers design paper-thin robot powered by muscle cells
MIT engineers created a paper-thin biohybrid robot powered by muscle cells that can swim through water mazes. The design offers a way to create small and efficient biohybrid robots.
ASE framework combines text and numerical data for time series forecasting
Attribution-guided Semantic Evolution (ASE) is a modular framework that constructs textual factors from external information to improve numerical time series forecasting. On four TESS datasets and CSI 300 stock ranking, ASE implementations achieved results below corresponding published benchmarks.