Today's briefing
- Industry & Policy
Australia proposes AI safety bill with mandatory independent audits for frontier models
Industry & PolicyThe DecoderNov 12, 2026 - Open Source & Tools
Meta opens e-means clustering algorithm, up to 45x faster on GPU than k-means
Open Source & ToolsKorean outletAI TimesNov 12, 2026 - Research
Selective deletion of 0.05% of LLM weights achieves fine-tuning-level performance gains
ResearchKorean outletAI TimesNov 12, 2026
Open source & papers radar
- RepoAgentsRepository: GoogleChrome/modern-web-guidance-srcStars: 2,722New stars today: 799Language: TypeScript
modern-web-guidance-src keeps coding agents updated on web best practices
This repository provides guidance to help coding agents stay current with the latest web development best practices.
- RepoAgentsRepository: QwenLM/qwen-codeStars: 29,738New stars today: 151Language: TypeScript
Qwen Code, an open-source AI coding agent for the terminal
Qwen Code is an open-source AI coding agent that runs directly in your terminal.
- ModelLLMModel: Zyphra/ZONOS3-MiniParameters: about 886 millionTask: text-to-speechLikes: 223Downloads (last 30 days): 8,543
ZONOS3-Mini, a text-to-speech model from Zyphra
ZONOS3-Mini is an 886M-parameter text-to-speech model from Zyphra, available under the Apache-2.0 license.
- PaperResearcharXiv 2611.13028Upvotes: 13
GTAGym: real AAA game tasks for agent evaluation and reinforcement learning
The authors introduce GTAGym, an environment and framework for evaluation and RL on GTA V tasks. Its benchmark GTAVBench contains 103 online tasks from main-story missions and 47 image-based tasks, with automatic outcome verification. Six frontier LLMs achieved up to 70.9% accuracy on GTAVBench-AGI, while LoRA RL of Qwen3.8-27B improved in-domain and out-of-domain performance from 15% to 23%.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
4 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
Stories in this issue
19 storiesNvidia installs interim CEO and chairman following Jensen Huang's departure
Nvidia appointed interim CEO Colette Kress following Jensen Huang's resignation. Kress, who has led the company's finance and business strategy for 13 years, and Harvey Jones, who has served on the board since the company's founding, were named to lead the company. Both emphasized confidence in the company's future in messages to employees.
Google's Gemini 4 Hydrogen outperforms larger Argon in benchmarks
Google's Gemini 4 Hydrogen replaces the Flash series and shows performance exceeding the pricier Argon model. On the independent Artificial Analysis Intelligence Index, Hydrogen nearly matches Claude Sonnet 5.5 at lower cost but costs about 2.5 times more per task than GPT-6.1 Sol and falls short in agentic coding.
modern-web-guidance-src keeps coding agents updated on web best practices
This repository provides guidance to help coding agents stay current with the latest web development best practices.
Qwen Code, an open-source AI coding agent for the terminal
Qwen Code is an open-source AI coding agent that runs directly in your terminal.
ZONOS3-Mini, a text-to-speech model from Zyphra
ZONOS3-Mini is an 886M-parameter text-to-speech model from Zyphra, available under the Apache-2.0 license.
Alloy, a multi-material parallel physics engine for robot learning
Alloy is a multi-material physics engine built on Newton, MuJoCo Warp, and Warp for robot learning. It simulates rigid robots, cloth, liquids, and soft bodies in a single scene across parallel worlds on GPU (up to thousands of free rigid bodies or 4,096 for a robot arm) or CPU, and requires Python 3.11 or 3.12.
AWS releases AWS Emulator, 200+ services built by AI agents and available locally
AWS released AWS Emulator, a free CC0-licensed local emulator of AWS APIs built from 438 API models, with all code written by autonomous agents. At launch it passes 56.6% of 42,857 published fidelity tests.
Viam launches Launchpad to accelerate robotics deployments
Viam launched Launchpad, a platform that provides software and engineering infrastructure to help robotics startups bring robots to commercial deployment.
Australia proposes AI safety bill with mandatory independent audits for frontier models
Australia's government proposed legislation requiring frontier AI companies to submit to mandatory safety rules and independent model audits, with steep penalties for violations. The move prioritizes domestic regulation before international negotiations.
Meta opens e-means clustering algorithm, up to 45x faster on GPU than k-means
Meta released e-means, an open-source improvement on the k-means clustering algorithm that runs up to 45 times faster on GPU. The algorithm maintains or improves result quality while reducing computational time and costs for data clustering, which can boost efficiency in building AI search and recommendation systems.
Selective deletion of 0.05% of LLM weights achieves fine-tuning-level performance gains
Researchers from NYU and the University of Calgary reported that selectively removing less than 0.05% of an LLM's weights without retraining other parameters significantly improves math reasoning and tool-use capabilities. The method achieves performance comparable to full fine-tuning while reducing storage requirements and preserving existing capabilities.
Anthropic's Mythos helps US utilities find power grid security vulnerabilities
Major US power companies including Duke Energy have begun testing Anthropic's high-performance model Mythos to identify security vulnerabilities in power-related software. Duke Energy's chief security officer Martin Strassburger confirmed that Anthropic allowed utility companies to access Mythos starting late last summer. Power companies had spent months prior requesting access from both the US government and Anthropic.
Agility Robotics completes merger and raises $465 million ahead of Nasdaq listing
Agility Robotics, the developer of the Digit humanoid robot, completed its business combination with Churchill Capital Corp XI on November 4, 2026. The company's shares began trading on Nasdaq under the ticker AGLT, and the transaction raised an expected gross total of $465 million.
GTAGym: real AAA game tasks for agent evaluation and reinforcement learning
The authors introduce GTAGym, an environment and framework for evaluation and RL on GTA V tasks. Its benchmark GTAVBench contains 103 online tasks from main-story missions and 47 image-based tasks, with automatic outcome verification. Six frontier LLMs achieved up to 70.9% accuracy on GTAVBench-AGI, while LoRA RL of Qwen3.8-27B improved in-domain and out-of-domain performance from 15% to 23%.
SparseRead: token-efficient sparse reading for AI agents, up to 92.9% reduction
SparseRead is a reusable reading control layer that governs content admission for web-facing AI agents without modifying the model. Across six models and five scenarios, it reduces total tokens by up to 92.9% and wall time by up to 89.0%, preserving or improving task quality in 97% of settings.
EmbodiedEvo: self-evolving coding agent that distills physical feedback into reusable skills
The authors present EmbodiedEvo, a self-evolving coding agent that distills physical feedback into reusable code skills and lessons to preserve experience across scenes. On RoboDojo-Sim, it achieves an average score of 44.48 and success rate of 37.82%, surpassing all officially verified results, with largest gains on memory-dependent tasks.
The Agentic Scientist: orchestrating coding agents for automated scientific discovery
The Agentic Scientist is an end-to-end pipeline that allows human researchers to conduct computational research alongside coding agents. The system orchestrates task drafting, idea generation, planning, experimentation, analysis, baseline construction, and report writing, with each step powered by coding agents and a meta-controller. In co-scientist mode, the developers submitted a paper accepted at ALIFE 2026, and applied it fully autonomously to open machine-learning problems.
WoVR+: world model as reliable simulator for VLA policy optimization with RL
WoVR+ extends a hallucination-aware world-model RL framework by redesigning the learned environment and imagined interaction protocol for Vision-Language-Action models. It improves success over supervised fine-tuning by 21.09% and 16.7% on RoboTwin and six real-robot tasks, and raises long-horizon phone-packaging success from 30% to 75%.
ElevenLabs releases Scribe v3 Live with conversational context for voice agents
ElevenLabs released Scribe v3 and Scribe v3 Live, speech recognition models designed to transcribe accurately through background noise and crosstalk. Benchmarks from Artificial Analysis show Scribe v3 narrowly outperforms OpenAI, Google, and Microsoft on accuracy, though competitors run faster and charge roughly half the price.