58 articles
OpenAI announces a contract with the U.S. Department of Defense outlining safety red lines, legal protections, and deployment protocols for AI systems in classified environments. Specific terms remain confidential but establish ethical boundaries for military applications.
Amazon Bedrock introduces a stateful runtime environment for agents, providing persistent orchestration, memory, and secure execution for multi-step AI workflows powered by OpenAI.
OpenAI and Amazon announce strategic partnership: OpenAI's Frontier platform integrates with AWS, expanding AI infrastructure, custom models, and enterprise AI agents.
OpenAI secures $110B in new funding at $730B pre-money valuation: $30B from SoftBank, $30B from NVIDIA, $50B from Amazon. Major capital round to scale AI deployment globally.
OpenAI announces mental health safety updates: parental controls, trusted contacts feature, improved distress detection, and litigation developments. No technical details or impact metrics disclosed in excerpt.
Google DeepMind releases Nano Banana 2, an image generation model combining advanced world knowledge, subject consistency, and Flash-speed performance. Production-ready.
OpenAI and Pacific Northwest National Laboratory introduce DraftNEPABench, a benchmark measuring AI coding agents' ability to accelerate federal permitting. Tests show potential 15% reduction in NEPA drafting time and modernized infrastructure reviews.
OpenAI and Figma integrate Codex to sync code and design, enabling teams to switch between implementation and Figma canvas for faster iteration.
Article on Mixture of Experts (MoE) architectures in transformers. Explains the routing mechanism that selectively activates experts per token, reducing computational complexity while maintaining performance. Covers recent implementations and trade-offs.
OpenAI releases threat report (February 2026) analyzing how malicious actors combine AI models with websites and social platforms. Report examines implications for detection and defense against malicious AI uses.
Nathan Lambert examines the actual importance of distillation for Chinese LLMs, responding to Anthropic's post on 'distillation attacks.' The article investigates whether model distillation remains a critical technique for Chinese developers facing restrictions on closed-model access.
OpenAI appoints Arvind KC as Chief People Officer to support company scaling, strengthen organizational culture, and lead workforce evolution in the AI era.
A new paper investigates AI agent reliability by quantifying the gap between claimed capabilities and actual performance. The study proposes methods to measure this divergence and improve the robustness of agent systems.
Seedream 5.0 introduces multi-step reasoning, example-based editing, and deep domain knowledge to image generation. Guide to key capabilities.
OpenAI stops evaluating on SWE-bench Verified, citing contamination and poor measurement of frontier coding progress. Analysis reveals flawed tests and training leakage. OpenAI recommends SWE-bench Pro instead.
OpenAI announces Frontier Alliance Partners program to help enterprises transition from AI pilots to production with secure, scalable agent deployments. The initiative aims to accelerate adoption of AI agents in enterprise environments.
OpenAI submits its first results to the First Proof math challenge, testing research-grade reasoning capabilities on expert-level problems. The AI model's proof attempts are shared publicly to benchmark advanced mathematical performance.
Hugging Face welcomes GGML and llama.cpp into its ecosystem to secure long-term development of local AI. This integration strengthens collaboration between critical open-source projects for efficient inference on resource-constrained hardware.
Hugging Face integrates Unsloth to enable free AI model training through Hugging Face Jobs. The solution combines Unsloth's memory optimization with Hugging Face infrastructure to reduce training costs.
Google DeepMind announces Gemini 3.1 Pro, a model optimized for complex tasks requiring elaborate responses. No benchmarks, performance metrics, or technical specifications provided in the announcement.
OpenAI commits $7.5M to The Alignment Project for independent AI alignment research. Aims to strengthen global efforts on AGI safety and security risks.
OpenAI launches India-focused initiative to expand AI access through local infrastructure, enterprise support, and workforce training. No specific investment figures, technical details, or timeline disclosed in announcement.
IBM and UC Berkeley introduce IT-Bench and MAST to diagnose enterprise agent failures. IT-Bench benchmarks agents on realistic IT tasks, while MAST (Multi-Agent Simulation Testbed) simulates complex environments to test multi-agent system robustness.
Google DeepMind integrates Lyria 3, its most advanced music generation model, into the Gemini app. Users can create 30-second tracks from text or images.
Gradio introduces gr.HTML, a component enabling custom HTML content integration into web applications. This feature simplifies building complex user interfaces while staying within the Gradio ecosystem.
Recraft V4 generates art-directed images with strong composition, accurate text rendering, and design taste. The model produces editable SVGs. Four variants now available on Replicate.
OpenAI and Paradigm release EVMbench, a benchmark measuring AI agents' ability to detect, patch, and exploit high-severity smart contract vulnerabilities on EVM.
Nathan Lambert analyzes the persistent gap between open-source and proprietary models. He examines distillation, innovation cycles, winning strategies for open models, specialization, and structural gaps in the industry.
Google DeepMind launches National Partnerships for AI initiative in India to accelerate scientific discovery and education through AI. The program aims to tailor AI tools to local challenges and strengthen research capabilities.
GPT-5.2 proposes a new gluon amplitude formula in theoretical physics. The formula was formally proved and verified by OpenAI and academic collaborators. Result published as preprint.
OpenAI launches Lockdown Mode and Elevated Risk labels in ChatGPT to defend against prompt injection attacks and AI-driven data exfiltration. These features strengthen organizational security against targeted threats.
OpenAI releases GABRIEL, an open-source toolkit using GPT to convert qualitative text and images into quantitative data. Designed for social scientists to analyze research corpora at scale.
OpenAI describes real-time access system for Sora and Codex combining rate limits, usage tracking, and credits. Architecture enables continuous access to both models without overload.
Hugging Face introduces custom kernels for all through Codex and Claude. The initiative aims to democratize access to compute optimizations for AI models by using AI-assisted code generation to create specialized kernels tailored to different hardware.
AI will not automatically reduce legal services costs. The article applies the 'AI as Normal Technology' framework to the legal sector, questioning the assumption that AI automation will systematically drive down prices.
Google DeepMind announces Gemini 3 Deep Think, an update to its specialized reasoning mode designed to tackle modern science, research and engineering challenges.
OpenAI releases GPT-5.3-Codex-Spark, its first real-time coding model in research preview for ChatGPT Pro users. 15x faster generation, 128k context window.
OpenEnv is an evaluation framework for tool-using agents in real-world environments. It enables testing AI models' ability to interact with web applications, APIs, and external systems to complete complex tasks.
OpenAI explores leveraging Codex in agent-first architectures. Ryan Lopopolo (Technical Staff) discusses code generation capabilities for multi-agent systems. No benchmarks or quantified results provided in excerpt.
Google DeepMind introduces Gemini Deep Think, a deep reasoning mode to accelerate mathematical and scientific discovery. Research papers document the growing impact of this capability across multiple fields.
Nathan Lambert examines model comparison in 2026, discussing Opus 4.6 and Codex 5.3. He questions the relevance of traditional benchmarks as model capabilities evolve rapidly and proposes reflection on new evaluation methods.
OpenAI deploys a custom ChatGPT instance on GenAI.mil, a secure platform for U.S. defense teams. The service includes safety controls and compliance measures tailored to government requirements. No technical details or timeline provided.
Transformers.js v4 is now available on NPM. This release brings performance improvements and compatibility enhancements to run transformer models directly in JavaScript on the client side.
OpenAI outlines its localization strategy to adapt frontier models to local languages, laws, and cultures while maintaining safety standards. The approach favors globally shared models over region-specific fragmentation. No specific technical details or benchmarks disclosed.
Hugging Face introduces SyGra Studio, a platform for creating and managing knowledge graphs. The tool enables building, visualizing, and deploying complex data structures to enhance AI model reasoning capabilities.
OpenAI and Ginkgo Bioworks cut cell-free protein synthesis costs by 40% using an autonomous lab combining GPT-5 with cloud automation in closed-loop experimentation.
OpenAI launches Trusted Access for Cyber, a trust-based framework expanding access to frontier cyber capabilities while strengthening safeguards against misuse.
OpenAI launches Frontier, an enterprise platform for building, deploying, and managing AI agents with shared context, onboarding, permissions, and governance.
OpenAI introduces GPT-5.3-Codex, a Codex-native agent pairing frontier coding performance with general reasoning for long-horizon, real-world technical work. No technical details, benchmarks, or availability date provided in excerpt.
OpenAI releases GPT-5.3-Codex, an agentic coding model merging GPT-5.2-Codex's frontier coding performance with GPT-5.2's reasoning and professional knowledge. Claimed as the most capable agentic coding model to date.
OpenAI releases technical documentation on Codex App Server, a bidirectional JSON-RPC API enabling Codex agent embedding with streaming, tool use, approvals, and diffs.
Hugging Face launches Community Evals, a decentralized evaluation platform enabling the community to create and share transparent benchmarks for AI models, in response to limitations of closed leaderboards.
H Company releases Holo2, a new model leading UI localization benchmarks. The model improves the ability to adapt user interfaces across languages and regional contexts.
Hugging Face analyzes the evolution of the global open-source AI ecosystem, particularly DeepSeek's impact and future trends. The article examines how open-source models are transforming AI accessibility and competitiveness.
Hugging Face publishes a study on training design for text-to-image models based on systematic ablations. The article details critical architectural choices and hyperparameters for optimizing image generation quality.
OpenAI outlines Sora feed philosophy: personalized recommendations, parental controls, and safety guardrails to enable creativity and connections. No technical details or metrics disclosed on implementation.
OpenAI and Snowflake announce a $200M partnership to embed frontier AI models directly into Snowflake's data platform. Customers can deploy AI agents and extract insights without data movement. Native integration of OpenAI models within Snowflake's ecosystem.
OpenAI releases Codex app for macOS—a command center for AI-assisted software development featuring multiple agents, parallel workflows, and long-running task support.