✷ The archive · 11 labs · 3,062 dispatches
The labs
No more opening a dozen official blogs every morning. First-hand releases from OpenAI, Anthropic, DeepMind and the rest, each with its substance pulled out.
OpenAI Improving Mathematical Reasoning with Process Supervision
OpenAI has developed a method to improve mathematical problem solving by rewarding each correct step of reasoning (process supervision) rather than just the final answer (outcome supervision), resulting in state-of-the-art performance and better AI alignment.
BERTopic Integration with the Hugging Face Hub
The Hugging Face blog post announcing BERTopic integration with the Hub does not provide readable details beyond its title.
Hugging Face LLM Inference Container for Amazon SageMaker
Hugging Face has introduced a purpose-built LLM Inference Container (DLC) for Amazon SageMaker, powered by Text Generation Inference (TGI) to enable high-performance deployment of open-source Large Language Models.
OpenAI Democratic Inputs to AI Program
OpenAI launched a grant program to fund ten $100,000 experiments aimed at developing democratic processes for deciding the rules AI systems should follow.
Optimizing Stable Diffusion for Intel CPUs with NNCF and 🤗 Optimum
Hugging Face announced a workflow using NNCF and Optimum to quantize Stable Diffusion and apply Token Merging, achieving up to 5.1x inference speedup and a 4x reduction in model footprint on Intel CPUs.
Hugging Face Model Catalog on Azure Machine Learning
Hugging Face and Microsoft have launched a native integration of the Hugging Face Hub Model Catalog within Azure Machine Learning Studio to simplify the deployment of open-source models on secure Azure infrastructure.
Hugging Face 4-bit Quantization and QLoRA Release
Hugging Face announced 4-bit quantization support via bitsandbytes and the QLoRA method, enabling inference and adapter fine‑tuning of large language models on consumer GPUs.
Anthropic Interpretability Dreams Research Vision
Anthropic outlines its long-term vision for mechanistic interpretability, focusing on resolving the challenge of superposition to enable the analysis of massive neural networks.
Anthropic Circuits Updates May 2023
Anthropic's May 2023 interpretability updates cover research into multiagent system failures, worker retraining program evidence, and a research version of Claude's progress on the Riemann zeta function.
Instruction-tuning Stable Diffusion with InstructPix2Pix
Hugging Face explores extending InstructPix2Pix to teach Stable Diffusion to follow specific image translation and low-level processing instructions through a multi-task instruction-tuning approach.
Hugging Face Safetensors Security Audit and Default Adoption
Hugging Face, EleutherAI, and Stability AI are transitioning to safetensors as the default model saving format following a security audit by Trail of Bits that confirmed no critical flaws leading to arbitrary code execution.
Hugging Face and IBM Partnership for watsonx.ai
Hugging Face and IBM have partnered to integrate Hugging Face open-source libraries and models into watsonx.ai, an enterprise AI studio for training, tuning, and deploying foundation models.
Anthropic raises $450M Series C to scale reliable AI products
Anthropic announced a $450 million Series C round led by Spark Capital to accelerate development of its Claude assistant, expand product offerings, and fund AI safety research.
OpenAI Governance of Superintelligence Proposal
OpenAI proposes a three-pillar framework for managing superintelligence through international coordination, a global regulatory body similar to the IAEA, and advanced safety research.
Q8-Chat: Efficient Generative AI on Intel Xeon CPUs
Hugging Face and Intel demonstrate that using SmoothQuant to compress LLMs to 8-bit integers allows high-quality chat experiences to run efficiently on single-socket Intel Xeon CPUs.
Large-scale Near-deduplication Behind BigCode
Hugging Face details the use of MinHash and Locality Sensitive Hashing (LSH) to perform large-scale near-deduplication for the BigCode project, demonstrating that removing near-duplicate data improves model performance and training efficiency.
Anthropic and Zoom Partnership and Investment
Anthropic and Zoom have partnered to integrate Claude AI into Zoom's enterprise collaboration products and Zoom Ventures has invested in Anthropic.
RWKV Architecture Integration into Hugging Face Transformers
Hugging Face announced the integration of the RWKV RNN‑Transformer hybrid model into the Transformers library, enabling open‑source, long‑context language models that combine RNN efficiency with transformer performance.
Running Vicuna 13B on AMD GPUs with ROCm
Hugging Face provides a technical guide on deploying the Vicuna 13B open-source chatbot on a single AMD GPU using ROCm and GPTQ 4-bit quantization to reduce memory requirements from 28GB to 7.52GB.
Hugging Face Selected for CNIL Enhanced Support Program
Hugging Face has been selected by the French Data Protection Authority (CNIL) for its Enhanced Support program to improve the implementation of data protection and GDPR compliance in AI development.
Hugging Face Assisted Generation for Low-Latency Text Generation
Hugging Face introduced Assisted Generation, a decoding method that uses a smaller assistant model to predict candidate tokens which are then validated by a larger model, reducing latency by up to 10x in some hardware configurations.
Anthropic Announces 100K Token Context Windows for Claude
Anthropic expanded Claude's context window from 9K to 100K tokens, enabling the model to ingest and analyze hundreds of pages of text in under a minute.
OpenAI releases automated neuron explanation dataset for GPT‑2 using GPT‑4
OpenAI announced a method that uses GPT‑4 to generate and score natural‑language explanations for every neuron in GPT‑2, releasing the resulting dataset, viewer, and code to the research community.
Creating a Coding Assistant with StarCoder – Summary
This article describes how the Hugging Face team turned the 16 B‑parameter StarCoder code‑generation model into a conversational coding assistant (StarChat‑α). By adding special chat tokens, preparing a filtered Open‑Assistant dialogue dataset, and fine‑tuning with DeepSpeed ZeRO‑3, they trained a model that can understand user prompts and generate runnable code (e.g., plots, maps, visualisations). The post also covers token‑izer configuration, loss‑masking for user turns, training setup, evaluation (both benchmark and LLM‑based), limitations, and future directions. The resulting model is publicly available on the Hub.
Claude's Constitution: Implementing Constitutional AI for Model Alignment
Anthropic introduces Constitutional AI, a method for training Claude to be helpful, honest, and harmless using an explicit set of written principles rather than relying solely on implicit human feedback.
A Dive into Text-to-Video Models
Hugging Face provides a technical overview of the evolution of text-to-video generative models, detailing the transition from GANs to Transformers and Diffusion architectures while highlighting the unique challenges of temporal consistency and data scarcity.
StarCoder Release Notes
Hugging Face and ServiceNow's BigCode collaboration released StarCoder and StarCoderBase, 15B parameter Code LLMs trained on permissively licensed data that outperform several open and closed models on programming benchmarks.
Anthropic Research: Distributed Representations, Composition, and Superposition
Anthropic clarifies the distinction between composition and superposition in distributed representations, explaining how these two mechanisms impact generalization and linear computability in neural networks.
Hugging Face Unity API Installation and Usage Guide
The Hugging Face Unity API allows developers to integrate Hugging Face Inference API models into Unity projects via a dedicated package and API wizard.
Training Language Models with Hugging Face Transformers, TensorFlow, and TPUs
Hugging Face provides a scalable end-to-end guide for training masked language models from scratch using TensorFlow and TPU pods, leveraging XLA compatibility for high-performance compute.
Databricks and Hugging Face Integrate Apache Spark for Faster LLM Training
Databricks has introduced first-class Apache Spark support in Hugging Face Datasets via the Dataset.from_spark() function, reducing data loading times by up to 40% for large-scale model training and tuning.
Running DeepFloyd IF on Free‑Tier Google Colab with Diffusers
Hugging Face shows how to run the open‑source DeepFloyd IF text‑to‑image model on a free Google Colab notebook by using 8‑bit quantization, modular pipeline loading, and Diffusers‑integrated optimizations.
Anthropic and Scale Partnership for Enterprise Generative AI
Anthropic has partnered with Scale to integrate the Claude AI assistant into Scale's platform, providing enterprises with deployment tools, prompt engineering, and secure data integration.
OpenAI Introduces New Data Management Controls for ChatGPT
OpenAI has introduced a new setting to disable chat history to prevent model training on user conversations, alongside plans for a ChatGPT Business subscription and a new data export tool.
Hugging Face Launches Chinese Language Blog to Support Global AI Community
Hugging Face has launched a dedicated Chinese language blog (hf.co/blog/zh) to provide translated technical resources and foster deeper collaboration with the Chinese AI community.
Hosting Unity Games on Hugging Face Spaces
Hugging Face Spaces can host playable Unity games by utilizing the Static HTML template and WebGL build targets.
Anthropic Proposal for Increased NIST Funding for AI Measurement
Anthropic proposes ambitiously funding the National Institute of Standards and Technology (NIST) to develop standardized AI measurement tools and safety thresholds, which they argue is a prerequisite for effective AI regulation.
Accelerating Hugging Face Transformers with AWS Inferentia2
Hugging Face and AWS have optimized Transformers for AWS Inferentia2, a purpose-built inference accelerator that significantly reduces latency and increases throughput for large-scale models.
Graph Classification with Transformers
Hugging Face demonstrates how to perform graph classification using the Graphormer model within the Transformers library, covering data formatting, preprocessing, and fine-tuning.
Creating Privacy Preserving AI with Substra
Hugging Face and Substra highlight how federated learning enables the training of AI models across decentralized data sources to maintain privacy and security, particularly in sensitive domains like healthcare.
OpenAI Bug Bounty Program Announcement
OpenAI has launched a Bug Bounty Program in partnership with Bugcrowd to reward security researchers for identifying and addressing vulnerabilities in its systems.
Snorkel AI and Hugging Face Integration for Enterprise Foundation Models
Snorkel AI has partnered with Hugging Face to integrate Hugging Face Inference Endpoints into Snorkel Flow, allowing enterprises to easily access and adapt over 150,000 open-source foundation models.
OpenAI Approach to AI Safety
OpenAI employs a multi-layered safety strategy combining rigorous pre-deployment testing, iterative real-world deployment, and continuous alignment research to mitigate risks associated with powerful AI systems.
StackLLaMA: Training LLaMA with RLHF for Stack Exchange
Hugging Face introduces StackLLaMA, a model trained using Supervised Fine-tuning, Reward Modeling, and Reinforcement Learning from Human Feedback (RLHF) to answer Stack Exchange questions.
Hugging Face Ethics and Society Newsletter #3: Ethical Openness Initiative
Hugging Face announced new ethical openness measures—including six ethics tags, a flagging system, and audience‑guiding metadata—to make open‑source ML safer and more inclusive.
Accelerating Stable Diffusion Inference on Intel Sapphire Rapids CPUs
Hugging Face shows how to cut Stable Diffusion image generation from 32 seconds to about 5 seconds on Intel Sapphire Rapids CPUs using Optimum Intel, OpenVINO, system‑level tweaks, IPEX BF16, and a faster scheduler.
BLOOMZ Inference on Habana Gaudi2 Accelerator
Hugging Face demonstrates that the Habana Gaudi2 accelerator achieves faster inference for the BLOOMZ 176B model than the Nvidia A100 80GB, leveraging the Optimum Habana library and DeepSpeed-inference.
Federated Learning with Hugging Face and Flower
Hugging Face demonstrates how to use the Flower framework to perform federated learning on a pre-trained distilBERT model for sentiment analysis on the IMDB dataset.
OpenAI March 20 ChatGPT Outage and Data Exposure Analysis
OpenAI took ChatGPT offline on March 20, 2023, due to a bug in an open-source library that exposed chat titles and payment information for a small subset of active ChatGPT Plus subscribers.
Training ControlNet with Hugging Face Diffusers
Hugging Face provides a comprehensive guide and training script via the diffusers library to enable users to train custom ControlNet models for Stable Diffusion, demonstrated through the creation of an 'Uncanny Faces' pose model.