27 articles
Google releases Gemma 2 2B, a lightweight model optimized for inference. ShieldGemma provides protection against harmful content. Gemma Scope offers interpretability tools to analyze model internals.
Hugging Face integrates Quanto into Diffusers to reduce memory consumption of Diffusion Transformers. Quantization enables running image generation models on resource-constrained hardware without significant quality loss.
Hugging Face and NVIDIA launch serverless inference integration via NVIDIA NIM. Users can deploy models without managing infrastructure, with automatic scaling and NVIDIA optimizations.
Critique of AI existential risk probability estimates presented as quantified. The article denounces how speculation is laundered through pseudo-quantification to influence policy, lacking solid empirical grounding.
OpenAI is testing SearchGPT, a prototype of new search features providing fast, timely answers with clear and relevant sources.
Hugging Face introduces LAVE, a zero-shot VQA evaluation method on Docmatix using LLMs. The study challenges whether traditional fine-tuning remains necessary for visual understanding tasks.
OpenAI introduces a Rule-Based Rewards (RBR) method to align models toward safe behavior without extensive human data collection.
Meta releases Llama 3.1 in three sizes (405B, 70B, 8B) with multilingual support and extended context. Models support 128k tokens and cover 8 languages. Available open-source via Hugging Face.
Hugging Face demonstrates running Mistral 7B with Core ML, Apple's on-device inference framework. Model conversion and optimization enable native deployment on macOS and iOS without external server dependency.
OpenAI releases GPT-4o mini, a smaller and cheaper model than GPT-4o. It delivers comparable performance on many tasks while reducing inference costs. The model supports text, vision, and audio.
OpenAI releases compliance and administrative tools for ChatGPT Enterprise: Compliance API integrations, SCIM support, and GPT controls to manage user access, data security, and compliance programs at scale.
Hugging Face releases Docmatix, a large-scale dataset for Document Visual Question Answering (DocVQA). The dataset contains thousands of documents with visual question-answer pairs to train models for understanding and extracting information from documents.
Hugging Face releases TGI Multi-LoRA, enabling a single deployment to serve 30 models through LoRA adapters. Reduces infrastructure costs and simplifies multi-model management in production.
OpenAI introduces prover-verifier games, a method to improve the legibility of language model outputs. This approach makes AI solutions clearer, more verifiable, and more trustworthy for both humans and machines.
Hugging Face introduces SmolLM, a family of lightweight language models designed for speed and efficiency. The models deliver strong performance despite their reduced size, optimized for fast inference and low energy consumption.
Hugging Face leveraged distilabel to build an Argilla 2.0 Chatbot. The approach combines synthetic data generation and annotation to enhance the conversational capabilities of the system.
NuminaMath won the 1st AIMO Progress Prize by combining fine-tuning and data synthesis techniques on base models. The solution leverages advanced mathematical reasoning to solve complex olympiad-level problems.
OpenAI and Los Alamos National Laboratory partner to develop safety evaluations measuring biological capabilities and risks of frontier models.
Hugging Face experiments with automatic PII detection on its Hub using Presidio. The tool identifies and masks sensitive information (names, emails, phone numbers) in datasets and models. Goal: improve privacy and regulatory compliance.
Hugging Face announces integration with KerasHub to streamline access to models and datasets. The partnership enables Keras developers to directly use Hugging Face resources within their workflows.
Hugging Face introduces preference optimization for vision-language models. The method improves VLM alignment with human preferences without explicit reward data, using pairwise comparisons of images and text.
Google Cloud makes TPUs available to Hugging Face users through direct integration. Users can access TPU accelerators for model training and inference without complex setup.
Banque des Territoires, Polyconseil, and Hugging Face partner to deploy a sovereign AI solution for a major French environmental program. The project leverages Hugging Face infrastructure to process sensitive data while maintaining regulatory compliance and avoiding dependence on foreign cloud services.
Hugging Face introduces new dataset search features. Users can now filter by size, license, language, and task. The improved interface makes it easier to discover resources for model training.
Critical article on AI agent evaluation. Questions current benchmarking methods and proposes rethinking what makes a meaningful AI agent.
Hugging Face optimizes protein language model ProtST on Intel Gaudi 2, accelerating inference and training. Benchmarks demonstrate significant performance gains on Intel's accelerator for protein structure prediction tasks.
Hugging Face's Transformers Code Agent achieves 92% accuracy on the GAIA benchmark, outperforming Claude 3.5 Sonnet (92%) and GPT-4o (87.9%). The agent combines web search, code execution, and multi-step reasoning to solve complex tasks.