✷ The archive · 11 labs · 3,429 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.
Hugging Face Training Efficiency: Packing with Flash Attention 2
Hugging Face has introduced boundary-aware packing for instruction tuning examples, enabling up to 2x training throughput increase and 20% peak memory reduction when used with Flash Attention 2.
OpenAI and Condé Nast Partnership for Content Integration
OpenAI has partnered with Condé Nast to integrate content from brands like Vogue and The New Yorker into ChatGPT and the SearchGPT prototype to improve information discovery and source attribution.
Upwork OpenAI Integration and AI-First Strategy
Upwork has transitioned to an OpenAI-centric ecosystem, integrating GPT-4o and GPT-3.5 into its marketplace features, internal fraud detection, and corporate productivity via ChatGPT Enterprise.
OpenAI launches fine‑tuning for GPT‑4o with free token quota and announces platform sunset
OpenAI launched fine‑tuning for GPT‑4o, enabling developers to customize the model with minimal data and offering free training tokens until September 23, while announcing the platform will close to new users after May 8 2026.
Deploying Meta Llama 3.1 405B on Google Cloud Vertex AI
Hugging Face provides a guide for programmatically deploying the FP8 quantized version of Meta Llama 3.1 405B on Google Cloud Vertex AI using Text Generation Inference (TGI) and A3 machine series.
OpenAI Disrupts Iranian Influence Operation Storm-2035
OpenAI has banned a cluster of ChatGPT accounts used by the Iranian influence operation Storm-2035 to generate political content for the 2024 U.S. election and other global events.
Indeed Contextual Job Matching with OpenAI
Indeed integrated OpenAI's GPT models to provide personalized explanations for job recommendations, resulting in a 20% increase in started job applications and a 13% uplift in downstream success.
OpenAI and The Met Museum Collaboration: Awakening Sleeping Beauties
OpenAI collaborated with the Metropolitan Museum of Art's Costume Institute to create an AI-powered chat experience allowing visitors to interact with a historical figure, Natalie Potter, based on curated historical datasets.
Hugging Face Infini-Attention Reproduction Analysis
Hugging Face's attempt to reproduce Infini-Attention found that while gating convergence can be improved, the method's performance degrades with increased memory compression and remains less reliable than Ring Attention, YaRN, or RoPE scaling.
OpenAI SWE-bench Verified release: human‑validated benchmark improves software‑engineering evaluation
OpenAI released SWE‑bench Verified, a human‑validated 500‑sample subset of the SWE‑bench software‑engineering benchmark that removes ambiguous issues and unfair tests, enabling more reliable evaluation of AI models’ coding abilities; GPT‑4o solves 33.2% of these samples, more than double its score on the original benchmark.
Introduction to ggml
ggml is a lightweight, C/C++ machine learning library optimized for Transformer inference and on-device LLM execution across diverse hardware backends.
Grok-2 Beta Release
xAI has released Grok-2 and Grok-2 mini in beta, featuring frontier-level capabilities in chat, coding, and reasoning that outperform Claude 3.5 Sonnet and GPT-4 Turbo on the LMSYS leaderboard.
Hugging Face Unified Tool Use API
Hugging Face has introduced a unified tool use API that allows developers to use the same code to implement tool calling across Mistral, Cohere, NousResearch, and Llama models.
Falcon Mamba 7B Release Notes
The Technology Innovation Institute (TII) has released Falcon Mamba 7B, the first large-scale pure State Space Language Model (SSLM) that matches the performance of state-of-the-art transformer models while eliminating attention-based memory scaling issues.
Qwen2-Audio Release Notes
Qwen2-Audio is an audio-language model capable of voice chat and audio analysis across more than eight languages and dialects, surpassing previous state-of-the-art performance on multiple benchmarks.
Zico Kolter Joins OpenAI Board of Directors
OpenAI has appointed Zico Kolter, a professor and Director of the Machine Learning Department at Carnegie Mellon University, to its Board of Directors and Safety and Security Committee.
Hugging Face acquires XetHub to upgrade Hub storage and collaboration
Hugging Face acquired XetHub to replace Git LFS with a more efficient storage backend, enabling incremental updates, trillion‑parameter model support, and better collaboration on massive AI datasets.
GPT‑4o System Card – capabilities, safety mitigations, and risk assessment
OpenAI released the GPT‑4o System Card, detailing the omni‑model's capabilities, safety mitigations, and medium overall risk rating.
Anthropic expands model safety bug bounty program
Anthropic announced an expanded, invite‑only bug bounty program that rewards up to $15,000 for discovering universal jailbreak attacks on its next‑generation AI safety mitigations, aiming to protect high‑risk domains such as CBRN and cybersecurity.
OpenAI Enabling a Data-Driven Workforce Webinar
OpenAI released a webinar on August 8, 2024, demonstrating how ChatGPT Enterprise can be used by employees to analyze data and uncover insights.
Qwen2-Math Release Notes
Qwen has released Qwen2-Math, a series of specialized mathematical LLMs (1.5B, 7B, and 72B) that outperform several closed-source models, including GPT-4o, on math benchmarks.
Rakuten and OpenAI Partnership: Leveraging Generative AI for Customer Insights
Rakuten is utilizing OpenAI's APIs, RAG, and Code Interpreter to transform unstructured data into automated customer service, review summaries, and B2B market insights.
Mistral AI Model Customization and Agents Release
Mistral AI has introduced model customization for flagship models including Mistral Large 2 and Codestral, alongside an alpha release of Agents and the stable 1.0 version of their client SDK.
OpenAI Structured Outputs API feature announcement
OpenAI introduced Structured Outputs on August 6, 2024, a new API feature that guarantees model responses exactly match developer‑provided JSON Schemas, improving reliability for data‑centric applications.
Hugging Face TextImage Augmentation pipeline release
Hugging Face and Albumentations AI released a TextImage Augmentation pipeline that jointly modifies document images and their text, enabling realistic synthetic data generation and robust fine‑tuning of vision‑language models on limited document datasets.
Hugging Face 2024 Security Feature Highlights
Hugging Face has detailed its 2024 security landscape, introducing a suite of default protections for all users and advanced governance controls for Enterprise Hub users.
Claude AI Availability in Brazil
Anthropic has expanded the availability of Claude, its AI assistant, to consumers and businesses in Brazil via web, mobile apps, and API.
Google releases Gemma 2 2B, ShieldGemma, and Gemma Scope
Google released Gemma 2 2B, ShieldGemma safety classifiers, and Gemma Scope sparse autoencoders, expanding open‑source LLM capabilities, moderation tools, and interpretability resources.
Anthropic Circuits Updates July 2024
Anthropic released a series of preliminary research updates in July 2024 focusing on interpretability, multiagent system failures, and mathematical breakthroughs in an unreleased Claude research version.
Quanto quantization cuts memory for Transformer diffusion pipelines
Hugging Face quantization (Quanto) reduces GPU memory for Transformer diffusion models from ~12 GB to ~5 GB with minimal latency and quality impact.
Hugging Face NVIDIA NIM API (serverless) launch and deprecation
Hugging Face launched the NVIDIA NIM API (serverless) for Enterprise Hub users, enabling pay‑as‑you‑go, serverless inference of open‑source LLMs on NVIDIA DGX Cloud H100 GPUs via an OpenAI‑compatible API.
LAVE: Zero-shot VQA Evaluation on Docmatix with LLMs
Hugging Face introduces LAVE (LLM-Assisted VQA Evaluation) to address the rigidity of traditional VQA metrics, demonstrating a 50% accuracy gain in evaluating zero-shot performance on the Docmatix dataset.
OpenAI SearchGPT Prototype Announcement
OpenAI has introduced SearchGPT, a temporary prototype that combines AI models with real-time web information to provide direct answers with clear source citations.
Ollama Tool Support Release
Ollama has introduced tool calling support for models like Llama 3.1, allowing local LLMs to interact with external APIs, functions, and code interpreters.
Mistral Large 2 release notes / what's new
Mistral AI has released Mistral Large 2, a 123B parameter model designed for high throughput single-node inference with competitive performance in coding, reasoning, and multilingual capabilities.
OpenAI Rule-Based Rewards for Model Safety
OpenAI has introduced Rule-Based Rewards (RBRs), a method that uses explicit, step-by-step rules to align AI safety behavior without requiring extensive human data collection.
Llama 3.1 Release Notes: Multilinguality, Long Context, and 405B Model
Meta has released Llama 3.1, featuring models in 8B, 70B, and 405B sizes with 128K context length, multilingual support for 8 languages, and a permissive license allowing synthetic data generation.
Running Mistral 7B with Core ML
Hugging Face demonstrates how to run Mistral 7B on Mac using new Core ML features from WWDC 24, achieving a model size reduction to under 4GB using 4-bit block-wise quantization.
GPT-4o mini release notes / what's new
OpenAI has released GPT-4o mini, a highly cost-efficient small model that outperforms GPT-3.5 Turbo and other small models on key reasoning and multimodal benchmarks.
Mistral NeMo Release Notes
Mistral AI and NVIDIA have released Mistral NeMo, a 12B parameter model featuring a 128k context window and a new, highly efficient multilingual tokenizer called Tekken.
Docmatix Dataset Release
Hugging Face has released Docmatix, a Document Visual Question Answering (DocVQA) dataset featuring 2.4 million images and 9.5 million Q/A pairs, providing a 240x increase in scale over previous datasets.
TGI Multi-LoRA: Deploy Once, Serve 30 Models
Hugging Face introduces Multi-LoRA serving in Text Generation Inference (TGI), allowing organizations to deploy a single base model and dynamically serve dozens of specialized fine-tuned adapters to reduce cost and operational complexity.
ChatGPT Enterprise Compliance and Administrative Tools Update
OpenAI has introduced a Compliance API, SCIM for automated user management, and granular GPT controls to help enterprise customers meet regulatory requirements and scale AI deployments securely.
OpenAI Prover‑Verifier Games improve legibility of language model outputs
OpenAI announced a prover‑verifier game training method that makes strong language models generate solutions that weaker models can verify, improving both correctness and human legibility.
Anthropic and Menlo Ventures Launch $100 Million Anthology Fund
Anthropic and Menlo Ventures have launched the $100 million Anthology Fund to accelerate the development of AI applications across infrastructure, industry-specific tools, and trust and safety.
Codestral Mamba Release Notes
Mistral AI has released Codestral Mamba, a 7B parameter model utilizing the Mamba architecture to provide linear time inference and high-performance coding and reasoning capabilities.
Mathstral 7B Release
Mistral AI has released Mathstral, a 7B parameter model specializing in STEM subjects and advanced mathematical reasoning, developed in collaboration with Project Numina.
Argilla SDK Chatbot with distilabel – End‑to‑End Tutorial
Hugging Face released a tutorial showing how to build an Argilla 2.0 chatbot using distilabel‑generated synthetic data, fine‑tuned embeddings, lancedb vector storage, and a Gradio app deployed on Spaces.
SmolLM Release: High-Performance Small Language Models
Hugging Face introduces SmolLM, a family of state-of-the-art small language models (135M, 360M, and 1.7B parameters) trained on a meticulously curated high-quality dataset.
NuminaMath 7B TIR wins AIMO Progress Prize – technical recap
NuminaMath 7B TIR won the first AIMO Progress Prize by solving 29 of 50 hidden math problems, showcasing the power of a two-stage fine‑tuning recipe, large high‑quality math data, and a self‑consistency with tool‑integrated reasoning inference strategy.