Defining and evaluating political bias in LLMs
OpenAI releases methodology to evaluate political bias in ChatGPT through real-world testing methods. The approach aims to improve objectivity and reduce systematic model bias.
OpenAI releases methodology to evaluate political bias in ChatGPT through real-world testing methods. The approach aims to improve objectivity and reduce systematic model bias.
HiBob deployed 2,500 custom GPTs using ChatGPT Enterprise to scale internal AI adoption, optimize HR workflows, and integrate AI features into the Bob platform. Use case: product scaling and team growth without disclosed ROI metrics.
IBM releases Granite 4.0, now available on Replicate. The model provides enhanced capabilities for generative AI tasks. Direct integration via Replicate platform for streamlined access and deployment.
OpenAI launches Pulse, a preview feature that makes ChatGPT proactive. The assistant can work autonomously in the background, completing tasks without user intervention.
OpenAI releases Sora 2 and the Sora app with safety built-in to address risks from an advanced video model and social creation platform. The approach relies on concrete protections (specific technical details not disclosed in excerpt).
OpenAI strengthens its fight against online child sexual exploitation through strict usage policies, advanced detection tools, and cross-industry collaboration. The company blocks illegal content, reports abuse to authorities, and prevents AI misuse.
Hugging Face accelerates Qwen3-8B agent inference on Intel Core Ultra using depth-pruned draft models. The technique reduces inference latency while maintaining response quality for agentic tasks.
Nvidia and Hugging Face release Nemotron-Personas-Japan, a synthetic dataset for training sovereign Japanese AI models. The dataset includes generated personas and dialogues to enhance local model capabilities without relying on foreign infrastructure.
The Charles Sadron Institute (CNRS) and Alysophil establish ACTIVIAFLOW, a joint laboratory to develop sustainable chemical processes using AI. This research-industry partnership aims to optimize chemical reactions and reduce waste.
Hugging Face releases Gaia2 and ARE to enable the community to study AI agents. These tools facilitate research and evaluation of multi-agent systems by providing benchmarks and open-source resources.
OpenAI releases a coordinated vulnerability disclosure policy governing how security flaws in its systems are reported and fixed. The policy establishes a responsible notification process before public disclosure, protecting users and allowing time for remediation.
Hugging Face introduces a new 'Public AI' category on its inference platform, enabling developers to access open-source models through a unified API. The initiative aims to democratize AI model access without reliance on proprietary providers.
OpenAI upgrades Codex with faster performance, improved reliability, and real-time collaboration. The tool now works across terminal, IDE, web, and mobile for autonomous task execution.
Hugging Face integrates visible watermarking in Gradio to protect generated images. The feature adds a discrete but detectable watermark to outputs, useful for tracing the origin of AI-generated content.
OpenAI restructures governance: nonprofit arm gains equity stake in PBC subsidiary, accessing $100B+ in resources for safe AI development. Reaffirms nonprofit control over long-term strategy.
Hugging Face introduces Jupyter Agents, an approach to train LLMs to reason with Jupyter notebooks. The system enables models to execute code, analyze results, and iterate to solve complex problems.
Hugging Face introduces mmBERT, a multilingual extension of ModernBERT. The model extends modern BERT architecture to 101 languages with improved efficiency and performance on classification and similarity tasks.
Replicate implements PyTorch torch.compile caching to reduce boot and inference times. Compiled models are cached across invocations, eliminating recompilation on each run.
OpenAI releases research explaining why language models hallucinate. The study proposes improved evaluation methods to enhance AI reliability, honesty, and safety.
OpenAI and Greek Government launch "OpenAI for Greece" to deploy ChatGPT Edu in secondary schools. The initiative aims to boost AI literacy, support local startups, and drive national economic growth through responsible AI learning.
OpenAI adds parental controls for teens, expert partnerships, and routes sensitive conversations to reasoning models in ChatGPT to improve safety and helpfulness.
OpenAI surveyed 1,000+ people globally on desired AI behavior and compared responses to its Model Spec. The "collective alignment" initiative aims to shape AI defaults to reflect diverse human values and perspectives.
OpenAI announces the Learning Accelerator, an acceleration program for AI startups. The program provides access to OpenAI models, API credits, and mentorship. Full details on eligibility criteria and terms not provided in the excerpt.
OpenAI and Retro Bio used GPT-4b micro to design more effective proteins for stem cell therapy and longevity research. The specialized model accelerated protein engineering in life sciences.
Anthropic and Hugging Face integrate image generation into Claude via Hugging Face API. Users can create images directly in Claude conversations using models like Flux and Stable Diffusion.
Practical guide to developing and deploying production-ready CUDA kernels. Covers GPU optimization, scaling patterns, and implementation best practices to maximize performance.
Hugging Face introduces MCP (Model Context Protocol) for research, enabling AI models to connect with scientific tools. The protocol standardizes integration between AI assistants and research databases, analysis software, or research platforms.
Arm and ExecuTorch 0.7 optimize generative AI model deployment on mobile and edge devices. The collaboration improves inference performance to make generative AI accessible at scale on resource-constrained hardware.
Hugging Face evaluates LLM capabilities on text-based video games through TextQuests. The study measures performance of models like GPT-4, Claude, and Gemini on interactive environments requiring comprehension, planning, and adaptation.
Hugging Face releases a guide on ND-Parallel, an efficient multi-GPU training technique. The method optimizes distributed resource utilization to accelerate large-scale model training.
Hugging Face integrates vision-language model alignment into TRL. The update enables training VLMs with advanced alignment techniques, extending the library's capabilities beyond pure text.
Hugging Face evaluates open-source Llama Nemotron models on DeepResearch Bench, a benchmark measuring deep research capabilities. Results show relative performance of different model versions on complex analysis and reasoning tasks.
Hugging Face publishes a tutorial on implementing MCP (Model Context Protocol) servers in Python. The example builds an AI shopping assistant with Gradio, demonstrating how to integrate MCP to extend language model capabilities.
Hugging Face releases Trackio, a lightweight experiment tracking library for model training. The tool simplifies logging of metrics, hyperparameters, and artifacts without heavy dependencies.
Hugging Face releases `hf`, a faster and more user-friendly command-line interface for interacting with its platform. This tool replaces previous commands and improves the user experience for managing models, datasets, and repositories.
Hugging Face introduces content-defined chunking for Parquet files, enabling adaptive segmentation based on data structure rather than fixed sizes. This approach improves efficiency for distributed processing and storage.
TimeScope benchmarks video large multimodal models' ability to process long sequences. The study measures performance across different video durations and identifies current limitations of video LMMs on extended content.
OpenAI releases economic analysis on ChatGPT's impact. New research collaboration launched to study AI's broader effects on labor market and productivity.
OpenAI introduces ChatGPT agent, capable of reasoning and acting with tools to complete tasks like research, bookings, and slideshows under user guidance. No technical details or availability date specified in the announcement.
Bria joins Replicate with commercial-grade image generation and editing models built on licensed data. Designed for enterprises and developers, these tools provide a safe alternative for visual AI applications.