Today's briefing
- Industry & Policy
Anthropic warns AI safety measures may fall behind as recursive self-improvement speeds up
Industry & PolicyKorean outletAI TimesNov 5, 2026 - Research
Zyphra says its hierarchical attention could speed long-context processing up to 35x
ResearchKorean outletAI TimesNov 5, 2026 - Agents
CloudOpsBench reveals reasoning as the biggest bottleneck for AI agents in DevOps
AgentsThe DecoderNov 5, 2026
Open source & papers radar
- RepoOpen Source & ToolsRepository: NVIDIA/auto-ontologyStars: 4,656New stars today: 2,053Language: Python
Auto Ontology, a framework for semantic graph-based reasoning layers
NVIDIA's Auto Ontology is a Python framework for building semantic, graph-based reasoning layers over data. It enables deterministic query generation, better context retrieval, and more explainable AI for LLM and agent workflows.
- PaperResearcharXiv 2611.05275Upvotes: 83
Route Anything, a framework for model and policy routing in general-purpose AI
The authors present Route Anything, a unified framework for studying model and policy routing in general-purpose AI systems. Across 25 benchmarks spanning language, image, video and speech generation, coding agents and robot manipulation, per-input oracle routing beats the best fixed candidate by 5.4 points. Learned routers beat it by 0.3 points across 16 benchmarks while using 24% less policy time.
- PaperPhysical AIarXiv 2611.05377Upvotes: 49
DreamX, a universal latent action learner from video diffusion models
The authors introduce DreamX, which repurposes a pretrained video diffusion model as a universal latent action learner. The approach compresses latent video frames by 300 times and captures action-relevant dynamics disentangled from visual appearance, supporting zero-shot action transfer across embodiments and environments. It achieves state-of-the-art performance among latent action models on LARY-Bench.
- PaperResearcharXiv 2610.06805Citations: 6
H-JEPA learns hierarchical world models for visual robot planning
The authors introduce H-JEPA, which trains a hierarchy of action-conditioned JEPAs in which each level predicts farther ahead in its own learned latent space. On Visual AntMaze, a three-level hierarchy raises planning success from 18 percent to 73 percent using less planner compute, and the approach extends to real-robot videos from DROID.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
3 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
Stories in this issue
21 storiesOpenAI to acquire Pierre Computer Company to expand Codex's coding tools
OpenAI is acquiring Pierre Computer Company to expand tools for Codex, its coding agent.
Anthropic releases FlashProve to verify which model generated an AI response
Anthropic released FlashProve, a research prototype that uses cryptography to verify whether an AI response actually came from a claimed model without exposing the model's weights. The system still has significant limitations.
LiberAI unveils Liber-0, a world action model for robot manipulation
LiberAI introduced Liber-0, a world action model pretrained on more than 2.2 million hours of human experience to support zero-shot robot manipulation.
Waymo and Toyota to test autonomous bZ4X fleet in Japan from 2027
Toyota and Waymo announced plans to begin testing and validating an autonomous Toyota bZ4X fleet in Japan in 2027, with deployment to the UK and other Waymo-serviced markets planned for the following years. Toyota will assemble the vehicles at its Takaoka plant, integrating Waymo's AV hardware.
Auto Ontology, a framework for semantic graph-based reasoning layers
NVIDIA's Auto Ontology is a Python framework for building semantic, graph-based reasoning layers over data. It enables deterministic query generation, better context retrieval, and more explainable AI for LLM and agent workflows.
Route Anything, a framework for model and policy routing in general-purpose AI
The authors present Route Anything, a unified framework for studying model and policy routing in general-purpose AI systems. Across 25 benchmarks spanning language, image, video and speech generation, coding agents and robot manipulation, per-input oracle routing beats the best fixed candidate by 5.4 points. Learned routers beat it by 0.3 points across 16 benchmarks while using 24% less policy time.
DreamX, a universal latent action learner from video diffusion models
The authors introduce DreamX, which repurposes a pretrained video diffusion model as a universal latent action learner. The approach compresses latent video frames by 300 times and captures action-relevant dynamics disentangled from visual appearance, supporting zero-shot action transfer across embodiments and environments. It achieves state-of-the-art performance among latent action models on LARY-Bench.
H-JEPA learns hierarchical world models for visual robot planning
The authors introduce H-JEPA, which trains a hierarchy of action-conditioned JEPAs in which each level predicts farther ahead in its own learned latent space. On Visual AntMaze, a three-level hierarchy raises planning success from 18 percent to 73 percent using less planner compute, and the approach extends to real-robot videos from DROID.
Anthropic warns AI safety measures may fall behind as recursive self-improvement speeds up
Anthropic warned that AI development accelerated by recursive self-improvement could outpace safety and security measures. In an amended IPO filing (S-1/A) with the SEC, the company acknowledged it had been slow in the past to notice sharp jumps in development speed, and called for stronger human oversight and control.
Zyphra says its hierarchical attention could speed long-context processing up to 35x
Zyphra unveiled Zyphra Hierarchical Attention (ZHA), a sparse attention technique that speeds up long-context processing. The method organizes information hierarchically and retrieves only the parts it needs instead of searching the whole document; the company projects up to 35x faster processing than standard attention on 10-million-token contexts.
CloudOpsBench reveals reasoning as the biggest bottleneck for AI agents in DevOps
CloudOpsBench is an open-source benchmark that measures how well AI agents handle typical DevOps tasks in the cloud, from fixing services and removing resources to restricting access rights, rolling out updates, and decommissioning projects. The benchmark reveals reasoning as the biggest bottleneck.
spec-kit, a toolkit to get started with specification-driven development
spec-kit is a toolkit to help teams get started with specification-driven development or other specification-based processes.
Kling report argues video generation quality reflects multimodal understanding
The report summarizes lessons on multimodal understanding from successive generations of the Kling video model. It describes Kling-MLLM, a shared multimodal foundation model that unifies image, video, audio, and interleaved inputs, and examines how generation quality emerges from the data–conditioning–alignment–evaluation chain.
agenterm, a self-hosted tool that lets AI agents control remote computers via MCP
agenterm is a self-hosted remote assistance tool that lets AI agents control remote computers through MCP and a terminal.
FactEmbed improves composed image retrieval with factorized MLLM embeddings
The authors propose FactEmbed for composed image retrieval, which finds a target image from a reference image and a modification text. Instead of mapping the query straight to one embedding, it uses factorized latents that simulate the intended visual changes before building the retrieval embedding. The authors report that it outperforms prior work across CIR benchmarks.
Speedy runs Qwen3.8-Flash-Next with up to 1M-token context on one or two RTX cards
Speedy is an optimized fork of Strata that runs Qwen3.8-Flash-Next, a 125B mixture-of-experts model, up to 1-million-token context on one or two RTX cards, tuned on an RTX 3080 with 10 GB of memory.
Beyond Robotics unveils a humanoid robot alongside its PICKY strawberry-harvesting robot
Beyond Robotics, a physical AI robotics company, unveiled its PICKY strawberry-harvesting robot and a humanoid robot at the KIEMSTA 2026 agricultural machinery expo in Daegu. The company showcased a humanoid designed to automate unstructured tasks across industries, built on 3D vision AI and robot control technology developed in agricultural settings.
Hugging Face explains who profits as open source AI goes mainstream
In a podcast episode, Hugging Face discusses how companies capture value as enterprises, governments and startups turn to open source models for lower costs, more control and more flexibility. It also weighs whether American players can keep up as Chinese open models gain ground.
Anthropic's text watermark is easy to bypass on several older Claude models
Anthropic is rolling out text watermarks to four more Claude models before the EU AI Act's December deadline. However, on older models like Claude Opus 4.6 and Sonnet 4.6, API users can disable the watermark by changing a single sampling parameter. Anthropic's response is a compliance warning.
Google search revenue growth slows as Alphabet plans about $200 billion for infrastructure
Alphabet reported that search queries grew at the fastest rate in five years, but Google's search revenue growth slowed to 12 percent year-over-year in the third quarter, down from roughly 17 percent in the previous quarter. Growth in paid clicks slowed from 13 percent to 8 percent, while the company plans to spend about $200 billion this year on infrastructure, largely to support AI.
MIT Technology Review Roundtables session to explore the quest for cognitive immortality
MIT Technology Review's Roundtables series explores technological approaches to sustain brain function beyond death. In a session on November 17, science editor Mary Beth Griggs and senior editor for biomedicine Antonio Regalado will discuss paths pursued by technologists seeking continuity of the self.