Zhipu has released GLM-5-Turbo, a foundation model specifically optimized for OpenClaw lobster scenarios. It focuses on enhancing tool calling, complex instruction following, scheduled and persistent tasks, and high-throughput long-chain execution capabilities. It ranks first among domestic models on the proprietary benchmark ZClawBench and has been integrated into AutoClaw AoLong, Mechrevo Lobster Box, and Zhipu's open platform BigModel.cn.
The Berkeley Artificial Intelligence Research (BAIR) lab proposed two algorithms, SPEX and ProxySPEX, to efficiently identify key feature, data, and model component interactions in large language models at scale. SPEX models interaction discovery as a sparse signal recovery problem based on assumptions of sparsity and low-orderity; ProxySPEX further introduces hierarchical structure assumptions, reducing the required number of ablations by approximately 10 times.
Google Research announced the launch of urban flash flood prediction features on Flood Hub, utilizing the newly proposed Groundsource method (based on Gemini analyzing public news reports) to build a historical flash flood event dataset.
Google Research released the Groundsource method, which uses the Gemini large language model to extract structured urban flash flood event data from global news, covering more than 150 countries.
Google Research, in collaboration with Beth Israel Deaconess Medical Center, conducted a prospective single-center feasibility study evaluating the safety and utility of the conversational diagnostic AI system AMIE for history-taking prior to initial outpatient visits.
Mistral AI built an autonomous agent for automatically generating and improving RSpec tests for Ruby on Rails codebases; the agent is based on the open-source coding assistant Vibe, plans workflows via AGENTS.md, loads specialized SKILLS files by file type (models/controllers, etc.), integrates custom tools like RuboCop and SimpleCov for closed-loop validation, and uses LLM-as-a-judge to evaluate test quality.
OpenAI equips the Responses API with a computer environment, combining shell tools and managed containers to achieve secure, scalable agent execution, supporting file operations, tool calls, and state management.
Hugging Face Hub has launched Storage Buckets, a non-versioned, S3-type object storage service designed for frequently changing intermediate artifacts such as training checkpoints, data shards, and agent traces.
Sequoia Capital announced it led the Series A funding round for log search engine Scanner. Designed specifically for object storage such as Amazon S3, Scanner achieves second-level queries through a proprietary inverted index, supporting interactive retrieval of petabyte-scale logs and continuous streaming detection. Cloud-native companies like Notion, Ramp, and Benchling are already using it in production, with AI security agents accounting for 80% of its platform query volume.
Proposed by Snowflake AI Research, Ulysses Sequence Parallelism (Ulysses SP) achieves million-token context training across multiple GPUs through attention head parallelism and all-to-all communication.
Figure released a new demonstration for Helix 02, showcasing its ability to perform end-to-end tidying tasks in a living room environment. These include spraying for cleaning, dynamic towel manipulation, bimanual coordinated boxing, underarm container holding, pillow tossing, remote control redirection and button pressing while held, tool stashing during movement, and side-stepping navigation through narrow gaps. All capabilities are based on the same neural architecture, without introducing new algorithms or special-case engineering.
LeRobot v0.5.0 has been officially released, adding support for full-body control of the Unitree G1 humanoid robot (including locomotion, manipulation, and teleoperation), introducing the Pi0-FAST autoregressive vision-language-action model, and incorporating Real-Time Chunking inference optimization technology.
vLLM v0.17.0 has been officially released, upgraded to PyTorch 2.10.0, integrating the FlashAttention 4 backend, with full support for the Qwen3.5 model family (including GDN.
Google Research released WAXAL, an open-source speech dataset covering 27 Sub-Saharan African languages, containing approximately 1,846 hours of ASR transcribed speech and 565 hours of high-fidelity TTS recordings, licensed under CC-BY-4.0. The data was primarily collected by African universities and community organizations, and has already supported multiple derivative studies, including the construction of a disordered speech dataset, the building of a 5,000-hour corpus for five Ghanaian languages, and benchmark evaluations of models for 13 African languages.
Google has released the open-source AI model SpeciesNet, which can classify nearly 2,500 animal categories in camera trap images. Trained on 65 million labeled images, it achieves a detection rate of 99.4% on the test set, with a species-level identification accuracy of 83%, including a prediction correctness rate of 94.5%. The model has been integrated into the Wildlife Insights platform and is being used by multiple research teams worldwide for field wildlife monitoring.
Since its launch in November 2024, Amazon Haul has attracted 3,287 sellers, 97.5% of whom are registered in mainland China or Hong Kong. By average price tier, 994 sellers have an average price below $10, with annual revenue averaging approximately $498,000. The overall annual GMV is estimated at around $2 billion by Marketplace Pulse. The business has expanded to 25 markets but remains significantly smaller than Temu ($30 billion) and Shein ($18 billion).
NXP published an embedded robotics AI practical guide on the Hugging Face blog, covering the full deployment process of Vision-Language-Action (VLA) models on resource-constrained platforms.
Hugging Face has released Modular Diffusers, offering a new way to build diffusion pipelines using reusable, self-contained modules (such as text_encoder, denoise, and decode).
OpenAI has released ChatGPT for Excel and new financial app integrations, powered by GPT-5.4, aimed at accelerating modeling, research, and analysis in regulated environments.
Google Research published a paper proposing the Bayesian teaching method, which uses supervised fine-tuning to enable LLMs to mimic the predictions of a Bayesian assistant, thereby learning approximate Bayesian reasoning. Experiments validated this approach in flight recommendation tasks, showing that the fine-tuned model not only improved performance on that task but also generalized to unseen domains such as hotel recommendations and online shopping, demonstrating an internalized capability for probabilistic reasoning principles.
Photoroom published the third installment of its PRX series on the Hugging Face blog, introducing how to train a usable text-to-image diffusion model in under 24 hours using 32 H200 GPUs (approximately $1,500).
Google DeepMind has released Gemini 3.1 Flash-Lite, the fastest and most cost-effective model in the Gemini 3 series, designed for high-throughput developer workloads.
OpenAI has launched a stateful agent runtime environment on Amazon Bedrock, providing persistent orchestration, memory, and secure execution for OpenAI-based multi-step AI workflows.