Today's briefing
- LLM
Tencent unveils Hy4, a multimodal model with vision and computer use
LLMThe DecoderNov 6, 2026 - Agents
Meta acquires AI coding startup Conductor
AgentsThe Decoder +1 moreNov 6, 2026 - Physical AI
Osteonic acquires 51% stake in US spine surgery robot company FoxEyes
Physical AIKorean outletiRobotNewsNov 6, 2026
Open source & papers radar
- RepoAgentsRepository: scenario-labs/skillsStars: 2,976New stars today: 674Language: Python
Scenario Skills, production-ready content generation for AI agents via MCP
Scenario Skills lets any AI agent access production-ready images, video, audio, and 3D content. Its skills pick the right model, show the price before spending and keep characters and brands consistent through the Scenario MCP, and it adds expert teams that drive Blender, Maya, ZBrush, Unreal and Unity.
- PaperPhysical AIarXiv 2611.06918Upvotes: 55
NEO-VLA, an encoder-free vision-language-action framework for robotic manipulation
The authors propose NEO-VLA, an encoder-free vision-language-action framework where images are tokenized and evolve jointly with language throughout the backbone. They show that this approach outperforms encoder-based counterparts across diverse manipulation tasks and distribution shifts.
- RepoAgentsRepository: JosssphZhou/pokiStars: 148Language: JavaScript
poki, an MCP agent that operates a real iPhone
poki is a local MCP server and agent that lets AI operate a real iPhone without a jailbreak or extra hardware.
- PaperResearcharXiv 2611.06910Upvotes: 43
FrontierPhysics evaluates AI agents on 56 end-to-end physics research tasks
The authors introduce FrontierPhysics, an end-to-end benchmark with 56 tasks created by 42 physicists based on their own research. Each task requires agents to review literature, develop a research plan, conduct research and experiments, and produce manuscripts. Across 22 models in 26 configurations, the best passed only 46 of its 168 rollouts.
Stories in this issue
21 storiesGoogle releases Gemini Omni 1.2 Flash with native 15-second video generation
Gemini Omni 1.2 Flash brings native 15-second video generation and editing, stronger consistency, and 30% lower pricing.
OpenAI unveils GPT-6 Cyber, a cybersecurity agent for finding and fixing vulnerabilities
OpenAI introduced GPT-6 Cyber, a cybersecurity model designed to find exploitable vulnerabilities and develop fixes. The model is trained to stay within authorized scope.
Scenario Skills, production-ready content generation for AI agents via MCP
Scenario Skills lets any AI agent access production-ready images, video, audio, and 3D content. Its skills pick the right model, show the price before spending and keep characters and brands consistent through the Scenario MCP, and it adds expert teams that drive Blender, Maya, ZBrush, Unreal and Unity.
NEO-VLA, an encoder-free vision-language-action framework for robotic manipulation
The authors propose NEO-VLA, an encoder-free vision-language-action framework where images are tokenized and evolve jointly with language throughout the backbone. They show that this approach outperforms encoder-based counterparts across diverse manipulation tasks and distribution shifts.
Baidu releases ERNIE 5.5 Flash Base, a 280B-parameter open-source foundation model
Baidu released ERNIE 5.5 Flash Base, a 280-billion-parameter mixture-of-experts model pre-trained on 33.5 trillion tokens. The company shipped it without fine-tuning to allow developers full flexibility.
Chinese regulators block SoftBank's $5.4 billion ABB Robotics acquisition
Chinese authorities blocked SoftBank's $5.4 billion acquisition of ABB Robotics over concerns about robot data being transferred overseas and used to train AI models.
Tencent unveils Hy4, a multimodal model with vision and computer use
Tencent unveiled Hy4, its first multimodal language model that understands images and can operate software. On the Artificial Analysis Intelligence Index, it scores 35 points but still trails the best commercial models on hard math and complex image analysis, according to Tencent.
Meta acquires AI coding startup Conductor
Meta is acquiring Conductor, an AI coding startup that lets developers control multiple AI coding agents at once. OpenAI announced plans to acquire Pierre Computer the same day, and Anthropic bought Stainless in May. The Decoder sees a pattern: as AI models become interchangeable, value is shifting to the tools built around them.
poki, an MCP agent that operates a real iPhone
poki is a local MCP server and agent that lets AI operate a real iPhone without a jailbreak or extra hardware.
FrontierPhysics evaluates AI agents on 56 end-to-end physics research tasks
The authors introduce FrontierPhysics, an end-to-end benchmark with 56 tasks created by 42 physicists based on their own research. Each task requires agents to review literature, develop a research plan, conduct research and experiments, and produce manuscripts. Across 22 models in 26 configurations, the best passed only 46 of its 168 rollouts.
Osteonic acquires 51% stake in US spine surgery robot company FoxEyes
South Korean medical device company Osteonic acquired a 51% stake in FoxEyes, a US spine surgery robot developer, for KRW 3 billion. FoxEyes was founded in 2024 and is preparing to commercialize spine surgery robots. Osteonic plans to expand from orthopedic implants into surgical robotics.
Nvidia says its Vera CPU delivers top throughput and faster AI agents
NVIDIA Vera is a CPU designed to handle the needs of AI agents that move between GPU reasoning and CPU execution across multiple steps including information search, database queries, code writing and testing, and action execution.
Anthropic lets Claude connect to users' bank accounts with new Personal Finance feature
Anthropic launched a Personal Finance feature that connects Claude to users' bank and card accounts via the Plaid bank bridging platform with read-only access. Users can ask Claude questions about spending by category, upcoming bills, affordability of major purchases, and estimated monthly savings.
Axiom Math bets on formal verification to keep unchecked AI proofs out of mathematics
Axiom Math is focusing on formal verification to ensure the quality of AI-generated proofs in mathematics. The startup's AI agents formally verified the Stacks Project, one of the field's largest libraries.
Together AI's Together Compiler uses agents to make inference up to 2.6x faster
Together AI released Together Compiler, an agent-driven loop that continuously re-optimizes the inference forward pass as traffic shifts. It achieved up to 2.6x TPS improvement over InferenceX's best configurations in under 24 hours per model.
VibeVoice, Microsoft's open-source voice AI project
VibeVoice is Microsoft's open-source frontier voice AI project.
Meta lets US users video chat with their customizable Muse AI avatar
Meta's AI agent can now take calls as a customizable animated avatar. This comes ahead of Meta's planned launch of a Tamagotchi-like wearable designed to bring Muse to life.
No AI agent scores above 12 percent on SemiAnalysis' GDPx benchmark
SemiAnalysis unveiled GDPx, a new benchmark for simulated real-world knowledge work spanning chip design to quant trading. No model scored above 12 percent; Claude Opus 5.5 narrowly led Gemini 4 Carbon. Critics point to the small number of tasks per domain and uneven testing conditions.
Dropbox acquires Tigris Data for $350 million to bolster AI infrastructure
Dropbox acquired Tigris Data for $350 million in a deal expected to support development of DBX Agent Cloud, a storage platform for developers building AI models and agents.
OpenAI proposes a community compact and policy blueprint for AI infrastructure
OpenAI introduced its National Community Compact and Compute Campus Policy Blueprint to protect local resources, create jobs, and guide responsible AI infrastructure development.
Ricursive AI's open-weight Ripple-65B Preview scores 80.5 on Terminal-Bench 2.1
Ricursive Intelligence released Ripple-65B Preview, a 65B dense open-weight LLM under Apache 2.0 license. It scores 80.5 on Terminal-Bench 2.1, ahead of US open models up to 15 times its size.