Today's briefing
- Physical AI
Serve Robotics launches Physical AI Services business for robots
Physical AITechCrunch RoboticsNov 6, 2026 - Open Source & Tools
Zyphra releases ZYFA, FP8 FlashAttention for AMD MI355X up to 1.76x faster
Open Source & ToolsMarkTechPostNov 7, 2026 - Agents
Perplexity launches Teammates, always-on AI agents for ongoing projects
AgentsThe DecoderNov 7, 2026
Open source & papers radar
- PaperPhysical AIarXiv 2611.08449Upvotes: 86
VFlash achieves 25x faster visual planning generation for robot motion forecasting
The authors introduce VFlash, a method for fast visual planning using Flash Primitive Decoding. It exploits the known structure of visual primitives to predict their coordinates jointly, achieving approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieves 70.2% average success across MolmoSpaces task suites, compared with 62.6% for the baseline.
- PaperPhysical AIarXiv 2611.06768Upvotes: 11
PyRUA-REAL achieves 91.1% success rate through embodied recursive self-improvement on physical tasks
PyRUA-REAL is a framework for embodied recursive self-improvement through real-world exploration and code refinement. Coding agents retain interaction experience as persistent memory and reusable executable skills. Across eight real-world tasks, PyRUA-REAL achieves an average success rate of 91.1%, and 84.4% on the RoboDojo benchmark.
- RepoAgentsRepository: HerdBuilderEstimate/roblox-studio-mcpStars: 427Language: Python
roblox-studio-mcp, a Model Context Protocol server for Roblox Studio
roblox-studio-mcp is an open-source, lightweight Model Context Protocol server for Roblox Studio. It lets AI assistants like Claude and Cursor connect to read object trees, edit scripts, and run playtests safely.
- RepoResearchRepository: TIGER-AI-Lab/WebsiteBenchStars: 84Language: Python
WebsiteBench, evaluating AI agents' ability to reconstruct websites through browser exploration
WebsiteBench is a benchmark that evaluates whether AI agents can rebuild websites through browser exploration. It measures web agents' capabilities in browser automation and visual evaluation.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
2 stories in this issue come from Korean outlets. Originals are in Korean.
- Research
Frontier AI models advance from finding vulnerabilities to writing patches
Since Anthropic's Claude Mythos Preview in April demonstrated finding software vulnerabilities and generating attack code, OpenAI, Anthropic, and Google have released frontier models with enhanced vulnerability discovery and attack validation capabilities. Recent developments show AI now writing patches to fix identified vulnerabilities.
Byline NetworkNov 6, 2026 - Research
Anthropic accelerates model development by about 2x through AI self-improvement
An independent evaluation report from METR found that Anthropic has roughly doubled its model performance improvement speed using AI. Since October last year, model enhancement speed accelerated by about 1.9x, achieving in one year what previously took two years. Three undisclosed models trained recently were also identified.
AI TimesNov 6, 2026
Stories in this issue
20 storiesVFlash achieves 25x faster visual planning generation for robot motion forecasting
The authors introduce VFlash, a method for fast visual planning using Flash Primitive Decoding. It exploits the known structure of visual primitives to predict their coordinates jointly, achieving approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieves 70.2% average success across MolmoSpaces task suites, compared with 62.6% for the baseline.
PyRUA-REAL achieves 91.1% success rate through embodied recursive self-improvement on physical tasks
PyRUA-REAL is a framework for embodied recursive self-improvement through real-world exploration and code refinement. Coding agents retain interaction experience as persistent memory and reusable executable skills. Across eight real-world tasks, PyRUA-REAL achieves an average success rate of 91.1%, and 84.4% on the RoboDojo benchmark.
METR finds Anthropic's model development pace roughly doubled, driven by AI-assisted research
Nonprofit METR audited Anthropic's internal research data for eleven weeks and concluded that the pace of model development has roughly doubled, with about 80 percent of the speedup driven by AI-assisted research. Anthropic redacted some evidence, and skeptical voices exist inside the company.
Thinking Machines' Inkling 2 Small claims top spot among western open-weight models
Thinking Machines, founded by former OpenAI chief technology officer Mira Murati, released Inkling 2 Small, a mixture-of-experts model with 594 billion parameters. The company claims it is the strongest western open-weight model, though better agentic performance comes with sharply rising token usage.
Physicl raises $10 million to close the sim-to-real gap for robots
Physicl raised $10 million to solve the sim-to-real gap using a pipeline originally built to market goods online.
OpenAI introduces BioBasicsBench to measure AI agents' wet-lab science capabilities
OpenAI introduced BioBasicsBench, a 100-task benchmark measuring how well AI agents interpret raw experimental data and handle everyday wet-lab science.
Serve Robotics launches Physical AI Services business for robots
Serve Robotics announced that it will launch a new business line called Physical AI Services to offer customers connectivity and data tools, alongside its third-quarter earnings announcement.
Zyphra releases ZYFA, FP8 FlashAttention for AMD MI355X up to 1.76x faster
Zyphra released ZYFA, open-source FP8 and BF16 prefill attention kernels for AMD Instinct MI355X. On real model activations, the FP8 kernels run 1.16–1.76x faster than AMD's AITER with 62–78% lower error at head dimension 128.
Perplexity launches Teammates, always-on AI agents for ongoing projects
Perplexity launched Teammates, always-on agents that work through ongoing projects on their own virtual machines. The feature requires the Max plan at $200 per month.
Amazon adds text messaging to Alexa for Shopping
Amazon added a text messaging feature to Alexa for Shopping that lets customers purchase items by describing products or providing photos. Alexa searches Amazon listings, sends back recommendations via text, and orders items on users' behalf.
Frontier AI models advance from finding vulnerabilities to writing patches
Since Anthropic's Claude Mythos Preview in April demonstrated finding software vulnerabilities and generating attack code, OpenAI, Anthropic, and Google have released frontier models with enhanced vulnerability discovery and attack validation capabilities. Recent developments show AI now writing patches to fix identified vulnerabilities.
Anthropic accelerates model development by about 2x through AI self-improvement
An independent evaluation report from METR found that Anthropic has roughly doubled its model performance improvement speed using AI. Since October last year, model enhancement speed accelerated by about 1.9x, achieving in one year what previously took two years. Three undisclosed models trained recently were also identified.
Palo Alto Networks in talks to buy AI agent security startup Clutch for $500M
Palo Alto Networks is in talks to acquire Clutch, an AI agent security startup backed by Lightspeed and SignalFire, for $500 million. Clutch was last valued at $85 million.
Airbnb plans to launch AI travel agent in first half of 2027
Airbnb said in its Q3 2026 earnings letter that it will roll out an AI agent in the first half of 2027 to help users plan and book travel. The company plans to leverage its experience matching hosts and guests and a new interface built specifically for travel.
AGIBOT unveils GO-3 Preview, tops three embodied AI benchmarks
AGIBOT unveiled GO-3 Preview, an embodied foundation model that integrates pre-training, post-training, and post-deployment learning. The model secured first-place rankings on three major embodied AI benchmarks and helps robots develop general-purpose capabilities.
roblox-studio-mcp, a Model Context Protocol server for Roblox Studio
roblox-studio-mcp is an open-source, lightweight Model Context Protocol server for Roblox Studio. It lets AI assistants like Claude and Cursor connect to read object trees, edit scripts, and run playtests safely.
WebsiteBench, evaluating AI agents' ability to reconstruct websites through browser exploration
WebsiteBench is a benchmark that evaluates whether AI agents can rebuild websites through browser exploration. It measures web agents' capabilities in browser automation and visual evaluation.
NVIDIA open source tools help researchers advance next-generation 6G network development
NVIDIA's open source tools for 6G development, including the Aerial wireless software and Sionna, provide researchers with a foundation for testing ideas and comparing results. The tools connect simulation and AI training with experiments on real over-the-air wireless systems.
OpenAI pledges not to use NDAs with public officials on data center projects
Following midterm elections where voters opposed data center expansion, OpenAI announced it will not use non-disclosure agreements (NDAs) with public officials on data center projects. The company said it will allow only narrow protections covering legitimate security information, trade secrets, and competitively sensitive information.
Terminal-Bench 2.0 audit finds defects in 30% of benchmark, proposes continuous-validation workflow
Researchers auditing Terminal-Bench 2.0, a coding agent benchmark, identified defects in approximately 30% of the benchmark and released Terminal-Bench 2.1 with repairs. The study shows that as models improve, existing verifiers can reject valid solutions, and changes to evaluation infrastructure can invalidate benchmark assumptions. It also demonstrates that resource allocation, such as time limits, materially affects agent performance and must be part of the measurement specification.