Browsing: AI

If you use API keys in Python, you need a safe way to store them. This guide explains seven beginner-friendly techniques for managing secrets using .env files.

Google Cloud Next 25 showcased significant advancements in artificial intelligence, with updates to AI models like Gemini, new hardware such as the Ironwood TPU, and expanded capabilities across Google Workspace and Agentspace. The event highlighted Google’s commitment to making AI more powerful, accessible, and integrated into business operations, alongside innovations in cloud security and networking.

Agentic AI systems, capable of complex goal pursuit with limited supervision, offer significant benefits but also introduce risks. This paper defines these systems and outlines responsibilities for parties in their lifecycle, proposing initial practices for safe and accountable operations. It also addresses implementation uncertainties and potential indirect impacts requiring further governance.

Major tech companies are intensely competing in the AI sector, with a growing emphasis on hardware. Apple’s recent acquisition of the AI startup Q.ai is intended to enhance its capabilities, especially in audio technology.

Explore five accessible APIs that simplify experimentation with AI and web data, perfect for those new to development.

Hugging Face has introduced the Messages API for Text Generation Inference (TGI) and Inference Endpoints, offering OpenAI compatibility. This allows for a smooth transition from OpenAI models to open Large Language Models (LLMs), enabling direct use with OpenAI client libraries and integration with tools like LangChain and LlamaIndex. The API provides users with enhanced control and flexibility over their AI solutions.

Apple’s new Creator Studio Pro suite integrates AI as a tool to enhance creativity, not replace it. The subscription-based offering bundles creative applications like Final Cut Pro, Logic Pro, and Pixelmator Pro with AI-powered features designed to streamline tasks for professionals and casual creators alike, while maintaining user privacy.

PipelineRL is an experimental Reinforcement Learning (RL) implementation that addresses the trade-off between inference throughput and on-policy data collection in large-scale RL with LLMs. Its key innovation, inflight weight updates, allows for continuously high inference throughput and minimizes the lag between model weights, leading to fast and stable RL training for large language models. This approach demonstrates competitive results against Open-Reasoner-Zero using a simpler algorithm and a modular architecture.

Deep Think is now available in the Gemini app for Google AI Ultra subscribers. Additionally, a select group of mathematicians will gain access to the full Gemini 2.5 Deep Think model, which was entered into the IMO competition.

The latest updates for Android and Pixel devices introduce a range of AI-powered features. These include Expressive Captions, designed to add more emotion to text, Gemini’s ability to save important information, and enhancements to Call Screen for easier responses.