Selected industry news from AIHOT. Follow advances in AI, products and technology.
AIHOT · Selected summaries with original sources · Nov 4, 2026, 9:00 AM
OpenAI has released a preview of ChatGPT Pulse for mobile Pro users. This feature enables ChatGPT to proactively conduct research and provide personalized updates based on user chat history, feedback, and connected applications such as calendars.
Sakana AI released ShinkaEvolve, an open-source program evolution framework that leverages LLMs for high-sample-efficiency algorithm discovery; validated on Circle Packing (with only 150 samples).
Hugging Face released the Smol2Operator method, which improves the zero-shot GUI grounding capability of the SmolVLM2-2.2B-Instruct model to a 61.71% accuracy on ScreenSpot-v2 through two-stage supervised fine-tuning (SFT).
ServiceNow has released the SyGra framework, a low-code/no-code data construction tool designed to generate, transform, and align training data for LLMs and SLMs. It supports inference backends such as vLLM, Hugging Face TGI, Triton, and Ollama, and covers tasks including Q&A construction, DPO preference pairs, multilingual translation, and converting PDFs/images into structured documents. Links to the code repository, documentation, and paper have been made public.
Hugging Face released Gaia2, a read-write agent evaluation benchmark for real-life assistant tasks, accompanied by the open-source Meta Agent Research Environments (ARE) execution framework.
Both deepseek-chat and deepseek-reasoner have been upgraded to DeepSeek-V3.1-Terminus; the former corresponds to non-thinking mode, while the latter corresponds to thinking mode. The update optimizes language consistency (mitigating mixed Chinese-English output and occasional anomalous characters) and Agent capabilities (performance of Code Agent and Search Agent) while maintaining existing capabilities.
Hugging Face Hub officially supports Scaleway as an inference service provider, allowing users to directly call open-source models such as gpt-oss, Qwen3, DeepSeek R1, and Gemma 3 from model pages.
Hugging Face launched RiskRubric.ai in collaboration with Cloud Security Alliance, Noma Security, Haize Labs, and Harmonic Security.
Figure announced two key advancements for its Vision-Language-Action (VLA) model Helix: Project Go-Big, which involves building the world's largest scale.
Apollo Research and OpenAI developed evaluation methods for covert misalignment ('scheming'), observing behaviors consistent with scheming in controlled tests across multiple frontier models; the team shared specific examples of early mitigation methods and stress tests.
Hugging Face officially announced that Public AI has become a supported inference service provider on its Hub platform, enabling users to access it via Hugging Face's Python and JS SDKs.
Hugging Face released LeRobotDataset v3.0, which packages multiple robot manipulation episodes into single Parquet and MP4 files, enabling fine-grained retrieval through relational metadata.
Hugging Face has introduced a visible watermarking feature in Gradio, supporting the addition of custom watermarks to gr.Image, gr.Video, and gr.Chatbot via the watermark parameter; watermarks can be specified as filenames, open image objects, or numpy arrays. Text watermarks are automatically displayed upon copying, enhancing the traceability of AI-generated content.
Writer released three open-source models in the Palmyra-mini series: palmyra-mini (a 1.5B non-thinking base model with a Big Bench Hard score of 52.6%).
Hugging Face's official blog announced a major upgrade to the transformers library, providing native support for OpenAI's open-source GPT-OSS series models and introducing zero-build downloadable kernels (Kernels from the Hub).
Hugging Face and Together AI have collaborated to launch a new feature that allows users to select any compatible LLM (such as HuggingFaceTB/SmolLM2-1.7B-Instruct) directly from the Hugging Face Hub for fine-tuning on the Together AI platform.
Hugging Face has released Jupyter Agent 2, a notebook-native agent framework designed specifically for data science tasks. Based on the Qwen3-4B model, it achieves an improvement in accuracy on easy DABStep questions from 44.4% to 75% through a self-developed lightweight scaffolding (approximately 200 lines of code) and a high-quality synthetic dataset (51k Notebooks, 0.2B tokens). All models, datasets, and code have been open-sourced.
The Hugging Face official blog released the mmBERT model, a new encoder trained on a multilingual corpus covering over 3T tokens and 1833 languages, based on the ModernBERT architecture and using the Gemma 2 tokenizer.
Google has released EmbeddingGemma, an efficient multilingual embedding model based on the Gemma3 architecture, with 308M parameters, a 2K context window, and support for over 100 languages.
Figure has released new capabilities for its Helix Vision Language Action (VLA) model. Without modifying the algorithm or adding special engineering, the model completed dishwasher loading tasks solely through training on new data. These tasks include separating stacked plates, coordinating both hands to pick up and place glasses, adapting to cluttered initial states, and recovering from mis-grasps and collisions. Previously, this model had already achieved clothes folding and package sorting.
SandboxAQ has released the Structurally Augmented IC50 Repository (SAIR) dataset, containing 5.24 million AI-generated high-precision protein-ligand 3D complex structures.
Mistral AI has launched a memory feature (beta) in its official product Le Chat. It supports automatic saving of useful information, visible and editable smart recall, traceability by source link, and offers a lightweight prompt tool called Memory Insights. Users can disable memory, enable incognito sessions, edit or delete individual memories, and export/import data. The feature is now available on iOS and Android mobile platforms.
Mistral AI has released an MCP connector directory (beta) supporting 20+ enterprise tools and a Memories (beta) feature in Le Chat, covering Databricks.
Hugging Face has integrated PyTorch Ahead-of-Time (AoT) compilation support into ZeroGPU Spaces, enabling image/video generation models (such as FLUX.1-dev.