Hugging Face Accelerate adds native support for PyTorch's FSDP (Fully Sharded Data Parallel), providing an alternative to DeepSpeed for distributed training. The update enables users to switch easily between DeepSpeed and FSDP based on their requirements.
OpenAI details how Voice Engine works, its text-to-speech model, and presents safety research. The article explores the underlying technology and protective measures against misuse.
Hugging Face releases an optimized embedding container for Amazon SageMaker, streamlining production deployment of embedding models. The tool integrates Hugging Face models with AWS infrastructure to simplify vectorization and semantic search workflows.
Hugging Face launches a leaderboard and arena to evaluate text-to-image generation models. The platform enables comparison of model performance and quality through standardized benchmarks and community-driven evaluations.
The Atlantic partners with OpenAI as a premium news source in ChatGPT. Atlantic articles will be discoverable across OpenAI products and shape how news is surfaced in future real-time discovery features.
Hugging Face releases comprehensive guide on training and fine-tuning embedding models with Sentence Transformers. Documentation covers optimization techniques, recommended datasets, and best practices for adapting models to specific use cases.
Hugging Face introduces Spaces Dev Mode, an integrated development environment for building and testing AI applications directly in the browser. The feature provides a streamlined workflow with hot-reload and instant deployment.
Hugging Face introduces key-value cache quantization to extend generation length in language models. The technique reduces KV cache memory footprint, enabling longer sequences without additional hardware resources.
Hugging Face introduces an open-source leaderboard for Arabic language models. The tool evaluates performance on Arabic-specific benchmarks, enabling comparison and improvement of LLMs for this language.
OpenAI introduces technology to identify content created by its tools and joins the Coalition for Content Provenance and Authenticity Steering Committee to promote industry standards.
Hugging Face integrates Artificial Analysis LLM Performance Leaderboard. The tool benchmarks models on latency, cost, and quality, providing transparent performance comparison for production deployments.
Hugging Face Inference Endpoints now combines ASR (automatic speech recognition), diarization (speaker identification), and speculative decoding to speed up inference. These features enable end-to-end audio processing with reduced latency through a unified API.
Traditional AI leaderboards are becoming obsolete as cost-performance tradeoffs grow complex. The article advocates replacing leaderboards with Pareto curves to evaluate AI agents, showing how $2,000 in spending reveals true efficiency-resource compromises.
OpenAI introduces an instruction hierarchy to train LLMs to prioritize privileged instructions and resist prompt injections and jailbreaks. The method enables models to distinguish system directives from malicious user inputs.
Hugging Face and Intel optimize SetFit inference for ultra-fast performance on Xeon processors through Optimum Intel. Benchmarks demonstrate significant latency and throughput improvements on CPU with no accuracy loss.
Hugging Face introduces Pollen-Vision, a unified interface for zero-shot vision models in robotics. The tool integrates multiple vision models (CLIP, DINOv2, SAM) into a common framework, enabling robots to perform perception tasks without task-specific training.
Hugging Face and Intel optimize embeddings for CPU using Optimum Intel and fastRAG. The solution reduces latency and memory consumption of embedding models on processors, improving RAG performance without GPU.
Hugging Face introduces Matryoshka embedding models, enabling dynamic dimensionality reduction of vectors without retraining. This approach provides flexibility and efficiency for RAG and vector search applications.
Hugging Face introduces the Open Ko-LLM Leaderboard, a ranking system dedicated to evaluating Korean language models. This initiative establishes a standardized evaluation ecosystem for Korean LLMs, enabling performance comparison across benchmarks tailored to Korean linguistic context.
OpenAI identified and disabled accounts used by state-affiliated threat actors for disinformation, espionage, and malicious capability development operations. The company publishes detailed reports on these campaigns and strengthens detection measures.
OpenAI develops a framework to evaluate risks of LLMs aiding biological threat creation. A study with biology experts and students shows GPT-4 provides at most mild uplift in biological threat creation accuracy, insufficient to be conclusive.
Hugging Face introduces a leaderboard focused on real-world enterprise use cases. It evaluates models on practical scenarios rather than traditional academic benchmarks, enabling organizations to compare performance in their specific contexts.
Hugging Face introduces a leaderboard for evaluating LLM safety. The tool measures model robustness against adversarial attacks and jailbreaks, providing standardized benchmarks to compare security defenses across models.
Hugging Face releases a guide to fine-tune W2V2-Bert, a speech recognition model, on low-resource settings using the Transformers library. The method enables adapting the model to languages or domains with limited training data.
Hugging Face optimizes Stable Diffusion Turbo and SDXL Turbo inference using ONNX Runtime and Olive. Quantization and compilation techniques reduce latency and memory consumption on CPU and GPU.
Hugging Face and AMD announce native integration to accelerate LLMs on AMD GPUs. Models run out-of-the-box without manual optimization, supporting RDNA and CDNA architectures.
Hugging Face demonstrates SDXL image generation in 4 steps using Latent Consistency LoRAs. The method significantly reduces inference time while preserving image quality by leveraging lightweight adapters instead of full model replacements.
Hugging Face introduces storage regions on its Hub, enabling users to select geographic locations for hosting their models and datasets. This feature improves latency, regulatory compliance, and data sovereignty.
Hugging Face accelerates Stable Diffusion XL inference using JAX on Cloud TPU v5e. Performance optimizations and latency reduction demonstrated on Google Cloud infrastructure.
Hugging Face releases a guide to finetune Stable Diffusion models using DDPO (Diffusion DDPOTrainer) integrated into TRL. The method enables optimizing image generation models with custom reward functions without requiring additional training data.
OpenAI partners with Scale to help enterprises fine-tune its most advanced models. OpenAI customers gain access to Scale's AI expertise to customize models.
Hugging Face optimizes Bark, its text-to-speech model, using the Transformers library. Improvements include reduced latency, memory compression, and better audio quality. Optimized code and models available on the hub.
Hugging Face deploys Stable Diffusion XL on Mac via Core ML with advanced quantization. Model runs locally without cloud, reducing latency and costs. Optimization specific to Apple silicon.
Hugging Face releases a guide to fine-tune MMS adapter models for low-resource languages. The approach reduces training data and computational requirements while maintaining speech recognition quality.
OpenAI launches a program awarding ten $100,000 grants to fund experiments in democratic processes for deciding what rules AI systems should follow, within legal bounds.
Hugging Face introduces InstructPix2Pix, an instruction-tuning method for Stable Diffusion enabling image editing via text instructions. The model learns to apply specific edits (style changes, object modifications, color adjustments) from a source image and natural language instruction.
Hugging Face and IBM partner to integrate Hugging Face open-source models into watsonx.ai, IBM's enterprise AI platform. The collaboration gives developers access to language and vision models through a unified interface, with fine-tuning and production deployment support.
Hugging Face optimizes Transformers for AWS Inferentia2, reducing latency and increasing inference throughput. Native integration of popular models (Llama, Mistral, Phi) with quantization and batching support.
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.
OpenAI, Georgetown, and Stanford released a report on risks of LLM misuse for disinformation campaigns. Collaborative study (October 2021 workshop, 30 experts) proposing a framework to analyze threats and mitigation measures.