21 articles
Hugging Face releases its third Ethics and Society Newsletter, exploring the concept of 'ethical openness' in AI development. The newsletter addresses tensions between transparency, accessibility, and responsibility in the open-source ecosystem.
Hugging Face demonstrates fast inference of BLOOMZ on Habana Gaudi2 accelerator. The 176B model achieves 1,000 tokens/sec with hardware-specific optimizations. Reproducible benchmark on Habana infrastructure.
Hugging Face optimizes Stable Diffusion inference on Intel CPUs through quantization and low-level optimizations. Benchmarks demonstrate significant speedups on modern CPU architectures without visual quality loss.
Hugging Face and Flower integrate federated learning capabilities. Users can train language models on distributed data without centralizing sensitive information. The solution combines HF models with Flower's decentralized coordination framework.
OpenAI details a ChatGPT outage on March 20. The article covers technical root causes, corrective actions taken, and findings from the internal investigation.
Hugging Face releases a guide for training ControlNet models using the diffusers library. The tutorial covers setup, training data, and optimization techniques for image generation control models.
OpenAI rolls out initial support for plugins in ChatGPT. These tools, designed specifically for language models with safety as a core principle, enable ChatGPT to access up-to-date information, run computations, or use third-party services.
Hugging Face and Jupyter collaborate to integrate models and datasets directly into Jupyter notebooks. This integration enables seamless access to AI resources without leaving the development environment.
OpenAI releases analysis on the potential labor market impact of large language models. The study examines how language models could affect different sectors and professions by assessing exposure to AI-automatable tasks.
Iceland is using GPT-4 to preserve its language against English dominance. The project aims to enrich Icelandic linguistic resources and improve AI models' ability to process Icelandic.
Be My Eyes integrates GPT-4 to enhance visual accessibility. The service helps blind and low-vision users interpret images and visual content in real-time using AI.
Stripe integrates GPT-4 to improve user experience and strengthen fraud detection on its payment platform.
OpenAI and Duolingo integrate GPT-4 to enhance language learning. The model strengthens conversations and addresses pedagogical gaps in the platform.
Khan Academy tests GPT-4 in a limited pilot program to explore its potential for virtual education in the classroom.
OpenAI releases GPT-4, a multimodal model accepting image and text inputs. Achieves human-level performance on professional and academic benchmarks, though less capable than humans in many real-world scenarios.
Hugging Face introduces Informer, a multivariate time series forecasting model based on attention mechanisms. The approach uses sparse attention to efficiently handle long temporal sequences, outperforming traditional methods on multiple benchmarks.
Hugging Face demonstrates RLHF fine-tuning of a 20B model on 24GB consumer GPU (RTX 4090). Uses quantization and memory optimizations to reduce requirements from 780GB to 24GB. Code and benchmarks available.
Kakao Brain releases new Vision Transformer (ViT) and ALIGN models optimized for computer vision and multimodal search. Models available on Hugging Face with pre-trained weights and inference code.
Hugging Face integrates ControlNet into Diffusers library, enabling precise spatial control over image generation models. ControlNet adds conditioning inputs (depth maps, poses, edges) to guide image synthesis without retraining.
Hugging Face releases ethical guidelines for Diffusers library development, addressing contributor responsibilities regarding image generation risks (deepfakes, illegal content, bias). Governance framework to oversee diffusion models.
Hugging Face accelerated Witty Works writing assistant development by providing optimized infrastructure and models. The collaboration reduced time-to-market and improved product performance.