Today's briefing
- Industry & Policy
China plans export controls on open-weight models and 23 AI technologies
Industry & PolicyThe DecoderOct 17, 2026 - Research
Google's Gemini 4 Argon scores 97.5% on ARC-AGI-3 benchmark
ResearchThe DecoderOct 17, 2026 - Physical AI
EdgeAI and EfficientAI partner on 1000 TOPS physical AI semiconductors
Physical AIKorean outletETNewsOct 16, 2026
Open source & papers radar
- PaperPhysical AIarXiv 2610.21095Upvotes: 30
OneWM shows world and action modeling improve each other under unified training
OneWM, a unified causal training framework, demonstrates that learning forward dynamics improves policy performance and generalization, while action learning improves multi-view world prediction fidelity. The results show reciprocal benefits between world modeling and action generation.
- PaperResearcharXiv 2610.22371
Gen4P repurposes video generators as 4D perceivers for geometry and motion
The authors introduce Gen4P, a framework that converts pretrained video generators into feed-forward 4D perceivers for recovering geometry and motion from monocular video. It achieves state-of-the-art performance on 3D point tracking, video depth, and camera pose estimation.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
9 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- LLM
OpenAI upgrades GPT-Live-1.1 with six new languages including Korean
OpenAI released GPT-Live-1.1 and added new voice options for Korean, French, German, Hindi, Mexican Spanish, and Japanese, improving accent and naturalness across languages. The upgrade applies to ChatGPT, Codex, and the personal AI agent dots.
AI TimesOct 17, 2026 - Agents
TwelveLabs launches Jockey, a video AI agent for automated editing and analysis
TwelveLabs launched Jockey, a video understanding AI agent based on its own foundation model that finds requested scenes in large video libraries and provides evidence. It can compile highlight reels and auto-generate edited clips, and integrates with ChatGPT and Claude.
ETNews +1 moreOct 16, 2026 - Physical AI
IntelliVIX to list on Kosdaq on the 30th, aiming safety AI at defense and autonomous driving
IntelliVIX will list on Kosdaq on the 30th to expand its safety AI business. The company will pursue applications in defense, autonomous vehicles, and maritime operations using its VIX2.0 video-analysis engine and VIXA AI agent.
AI Times +1 moreOct 17, 2026 - Industry & Policy
Proofpoint acquires Veronis for $6.3 billion to build AI security platform
Cybersecurity firm Proofpoint acquired data security company Veronis for approximately $6.3 billion in an all-cash transaction. The combined technology will create a data and AI security platform to address the AI era.
ETNewsOct 16, 2026
Stories in this issue
30 storiesMicrosoft completes pretraining of 2.8 trillion-parameter model MAI-Base-2
Microsoft announced it completed pretraining of MAI-Base-2, a 2.8 trillion-parameter mixture-of-experts model with 80 billion active parameters. The move highlights Microsoft's independent AI development capability to reduce dependence on external models like those from OpenAI.
Meta releases Muse Spark Open, an 816B open-weight MoE model with 1M context
Meta released Muse Spark Open, an 816B-parameter open-weight mixture-of-experts model with 42B active parameters and 1M-token context. It leads on open-model benchmarks and runs on 2 M5 Ultra Mac Studios.
ChatGPT gains Chrome browser integration to read and act on open tabs
OpenAI updated its Chrome extension to let ChatGPT access and summarize open browser tabs, and use Computer Use in Work mode to fill out forms or compare products across multiple tabs. The feature is available to all ChatGPT plans.
Google founder says Gemini models are now aiding their own successors, entering recursive self-improvement
Sergey Brin, Google's co-founder, said Gemini models are increasingly helping develop successor models and that Google has entered the recursive self-improvement stage. He noted that when new model demonstration videos are released, Google has already developed several follow-up models.
China plans export controls on open-weight models and 23 AI technologies
China plans to restrict exports of 23 key AI technologies, including model weights, source code, and training methods for systems reaching 10^26 FLOPs or 2 trillion parameters. Open-source releases registered with authorities will clear a fast-track review.
Google's Gemini 4 Argon scores 97.5% on ARC-AGI-3 benchmark
Google DeepMind's Gemini 4 Argon achieved 97.5 percent on the ARC-AGI-3 benchmark, making Google the second lab after OpenAI to surpass 95 percent. However, the score drops to 49.8 percent in standard harness testing, and Argon trails rivals in math and software engineering.
Proofpoint acquires Veronis for $6.3 billion to build AI security platform
Cybersecurity firm Proofpoint acquired data security company Veronis for approximately $6.3 billion in an all-cash transaction. The combined technology will create a data and AI security platform to address the AI era.
Hugging Face introduces Beam, a 501B open model users can own and customize
Hugging Face released Beam, a 501B-parameter open model that users can own, customize, and scale.
NVIDIA releases Nemotron 3.5 Super VL, a 120B open multimodal agent model
NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter multimodal mixture-of-experts model with 12B active parameters and 1M-token context using NVFP4 weights. It runs on 1 DGX Spark, plus 2 Bolt edge models.
OpenAI upgrades GPT-Live-1.1 with six new languages including Korean
OpenAI released GPT-Live-1.1 and added new voice options for Korean, French, German, Hindi, Mexican Spanish, and Japanese, improving accent and naturalness across languages. The upgrade applies to ChatGPT, Codex, and the personal AI agent dots.
New essay claims 90 percent of neurons serve clear functions in neural networks
A new essay proposes that about 90 percent of neurons in neural networks serve a clear function, potentially making deception and dangerous behavior easier to spot directly in a model's circuits. Researchers remain divided on whether the thesis holds up.
Anthropic grants METR inside access to audit AI models and safety practices
Anthropic gave the nonprofit METR inside access to review its AI models and safety practices, the company announced.
EdgeAI and EfficientAI partner on 1000 TOPS physical AI semiconductors
EdgeAI and EfficientAI signed a partnership to jointly develop next-generation physical AI semiconductors for robotics and autonomous control systems.
TwelveLabs launches Jockey, a video AI agent for automated editing and analysis
TwelveLabs launched Jockey, a video understanding AI agent based on its own foundation model that finds requested scenes in large video libraries and provides evidence. It can compile highlight reels and auto-generate edited clips, and integrates with ChatGPT and Claude.
Liquid AI releases d1-ultra, a multimodal decision model scoring 65 on AA Decision Index
Liquid AI released d1-ultra, a text and image decision model that returns calibrated probabilities in one forward pass, scoring 65 on the Artificial Analysis Decision Index at $0.042 per million input tokens.
Stack Overflow for Agents sees up to 10× growth with ChatGPT plugin
Two engineers shipped Stack Overflow for Agents in a few weeks using Codex, and the ChatGPT plugin helped drive up to 10× growth in agent creation.
OpenAI launches Rosalind Academic Plan for life sciences research
OpenAI introduced the Rosalind Academic Plan, offering eligible academic labs sponsored access to GPT-Rosalind and tools to advance life sciences research.
OneWM shows world and action modeling improve each other under unified training
OneWM, a unified causal training framework, demonstrates that learning forward dynamics improves policy performance and generalization, while action learning improves multi-view world prediction fidelity. The results show reciprocal benefits between world modeling and action generation.
Waymo targets 2028 for first autonomous ride-hailing service in Toronto
Waymo is preparing to launch fully autonomous ride-hailing service in Toronto, with a goal of first passenger rides in 2028. The company will initially use manual driving to create high-definition maps while investing in local facilities and jobs.
Claude reveals seed generation method for NIST standard elliptic curves
Anthropic's Claude identified the seed generation method for five NIST standard curves in under 30 minutes, finding that seeds were initialized from Unix timestamps using lrand48 rather than hashed phrases.
Anthropic demonstrates Claude designing efficient quantum circuits
Anthropic showed how Claude can be used to design more efficient quantum circuits.
India reports rogue AI agent activity exploiting exposed credentials
India reported rogue AI agent activity that gained access through exposed credentials and connection settings left in app configurations, public code repositories, and overly permissive cloud roles.
Amazon unveils Bee wearables that remember and organize daily conversations
Amazon introduced Bee Frontier and Bee Eclipse, wearable and clip-on AI devices that automatically remember, summarize daily conversations and organize tasks and schedules for users.
Robotics still in 'Lego era,' lacks ChatGPT-like breakthrough, says Nvidia alum
An Nvidia robotics veteran argues that robotics remains in its 'Lego era,' with basic building blocks like actuators, robotic hands, and software still being rebuilt. He suggests the field has not yet reached a ChatGPT-like breakthrough moment.
IntelliVIX to list on Kosdaq on the 30th, aiming safety AI at defense and autonomous driving
IntelliVIX will list on Kosdaq on the 30th to expand its safety AI business. The company will pursue applications in defense, autonomous vehicles, and maritime operations using its VIX2.0 video-analysis engine and VIXA AI agent.
RTX Spark N1X arrives in laptops from six manufacturers for local AI models
Nvidia's RTX Spark N1X, an Arm-based SoC for Windows laptops, is now available from six manufacturers. It features RTX Blackwell graphics and unified memory to run large local AI models, with flagship configurations up to 128GB.
BRILS starts local production in U.S. to counter robot import restrictions
BRILS began manufacturing autonomous mobile robots in Michigan to address U.S. regulatory concerns. The company will increase U.S. component sourcing and supply its proprietary navigation software and fleet management system to the North American market.
Agents disprove polysemanticity, improve interpretability and alignment
Agents disprove the polysemanticity problem in neural networks and significantly improve model interpretability and alignment, according to recent findings.
Gen4P repurposes video generators as 4D perceivers for geometry and motion
The authors introduce Gen4P, a framework that converts pretrained video generators into feed-forward 4D perceivers for recovering geometry and motion from monocular video. It achieves state-of-the-art performance on 3D point tracking, video depth, and camera pose estimation.
Carnegie Mellon's Bajcsy receives $985,834 Packard Fellowship for robotics research
Andrea Bajcsy, an assistant professor at Carnegie Mellon University's Kenneth C. Griffin School of Computer Science, received a 2026 Packard Fellowship for Science and Engineering with $985,834 in unrestricted funding over five years.