OpenAI Reportedly Disbands Standalone Robotics Team, Researchers Shift to Foundation Models
OpenAIRobotics
OpenAI has reportedly disbanded its standalone robotics team and reassigned the researchers to core AI model development teams. The restructuring affects the division of work between robotics-specific models and general-purpose multimodal model development.
Tesla Opens Cybercab for Personal Orders at $29,990
Autonomous DrivingTesla
Tesla has opened personal Cybercab orders to some invited customers at a price of $29,990. Buyers must pay a non-refundable $1,000 deposit and subscribe to the autonomous driving service for $299 per month. Initial deliveries are expected to begin in December in five Texas cities, with operations limited to designated areas. Elon Musk said the plan is to adjust the subscription fee to about $115 per month, closer to the rate for existing models.
Tsinghua Releases ProtCell, Reproduces Model Pretraining on Moore Threads GPUs
AI Drug DiscoveryDomestic AI Compute
The Jiang Rui research group at Tsinghua University has released ProtCell, a single-cell proteomics foundation model pretrained on 380 million cells and 98 public datasets. The model characterizes cell states and infers protein expression, supporting virtual cell research and drug development. The team reproduced the full pretraining run on Moore Threads MTT S5000 AI compute cards, with quantitative evaluation results differing by about 0.2% from those on leading international GPU platforms.
Google has introduced a personal finance analysis feature for Gemini that connects bank accounts and credit cards through Google Wallet and third-party financial data service Plaid. Once enabled, users can access financial data in conversations, allowing Gemini to read transaction records, analyze spending habits, and create budget and savings plans based on their financial situation.
Meta Proposes e-means Clustering Method, Achieving 10-to-45× GPU Speedups in Tests
MetaAlgorithms
Meta has proposed e-means, a k-means variant that uses running means to adjust cluster centers, automatically splits empty clusters, and selects batch sizes and step sizes based on training experiments. In tests, the method delivered 10- to 45-fold GPU speedups over traditional methods while maintaining loss quality, reducing the number of convergence rounds from 25 to about 2. Meta used it to build 16 million clusters from 10 billion embedding vectors in less than a day.
ggml Releases 1.0, Adding FP8 Support and Optimizing Multiple Compute Backends
Open SourceInference Framework
The open-source machine learning compute library ggml has released version 1.0.0, adding FP8 support and configurable YaRN truncation while cleaning up deprecated public APIs. The update includes quantized matrix multiplication and MoE fusion for CUDA and Metal, a Vulkan prefill path for FlashAttention, and optimizations for the SYCL, OpenCL, and WebGPU backends. The RPC protocol has been upgraded to v9, improving loading speed and failure handling.
Vercel Open-Sources Beta UI SDK with More Than 50 Interface Behaviors
Open SourceDeveloper Tools
Vercel Labs has released an open-source beta UI SDK that offers more than 50 interface behaviors for AI coding agents, rather than prebuilt components. Developers can compose interfaces across React, Vue, Svelte, Solid, and Web Components using framework adapters; an agent bridge lets agents manipulate UI state directly. In tests, this bridge reduced the number of tool calls by 69%.
Google Antigravity Adds Orchestrator Mode for Multi-Model Collaboration
GoogleAI Agents
Google Antigravity has added Orchestrator mode, allowing users to coordinate and switch between multiple models in a single session, including Gemini 4 Carbon and Hydrogen. The mode supports combining planning, code execution, and testing into collaborative workflows for long-running agent tasks.
Artificial Analysis Updates Coding Agent Index; Two Claude Models Tie for First
Model EvaluationCoding Agents
Artificial Analysis has released Coding Agent Index v2.1, upgrading FrontierCode to version 2.0 with 132 real-world coding tasks and Terminal-Bench to version 4.1 with 66 tasks. Claude Fable 5.2 and Claude Opus 5.5 both scored 58, tying for the lead. Following the benchmark updates, the index’s average score fell by about 5 points from the previous version; per-task costs across configurations ranged from $1.51 to $44.22.
Science Prediction Benchmark Features 116 Scientific Forecasting Tasks
AI for ScienceModel Evaluation
The Science Prediction Universe benchmark comprises 116 scientific tasks and 4,207 wet-lab experiments, based on 85 recent papers and spanning chemistry, materials science, biomedicine, and other fields. Its first round evaluated 13 configurations of 12 models. Each task gave models 8 hours and access to specified compute resources. Claude Opus 5.5 achieved the highest score, 0.486, with about half of the tasks scoring above the random baseline.