A parameter is a numerical value learned by a model during training. Neural network weights and biases are parameters because the optimization process adjusts them to reduce prediction error. Together, they determine how input is transformed into output and store much of the model’s learned behavior. Parameter count is often used to describe model size, but more parameters do not automatically mean better quality; architecture, data, training, and evaluation matter as well. Parameters differ from hyperparameters, which are chosen by developers to control training or structure. Updating parameters through fine-tuning can specialize a model, while compression and quantization change how they are stored or represented for deployment.
OpenAI is paying employees an average of $1.5 million in stock-based compensation, far exceeding peer startups as it races to retain top AI talent. The strategy is inflating losses and reshaping expectations ahead of a potential IPO.
Chinese generative AI startups MiniMax and Zhipu AI have disclosed their financials ahead of potential Hong Kong listings, highlighting modest revenues and mounting losses compared with U.S. peers.
Zoom introduced AI Companion 3.0, a federated AI solution integrating internal and third-party models to streamline meetings, documents, and personal workflows. The update adds agentic AI features, a web interface, and enhanced productivity tools.
Nvidia is reportedly planning to increase production of its H200 AI training chips following approval to sell them in China.
Google released a major upgrade to its Gemini Deep Research agent, giving developers access to advanced autonomous research capabilities through the Interactions API. The company also introduced DeepSearchQA, a benchmark for evaluating multi-step web research agents.
OpenAI faces a wrongful death lawsuit alleging ChatGPT reinforced a man’s violent delusions, contributing to the killing of his mother in Connecticut. The case marks the first time an AI platform has been accused of involvement in a murder.
Google introduced the Titans architecture and the MIRAS framework to enable AI models to handle massive contexts and update their internal memory while running, improving performance in long-sequence tasks.
ChatGPT now reaches more than 800 million weekly users, creating a powerful adoption flywheel that is speeding the transition from AI experimentation to full-scale enterprise deployment, according to OpenAI’s new 2025 report.
OpenAI buys Neptune for under $400 million in stock, gaining advanced AI model tracking tools and strengthening capabilities in large-scale AI development.
OpenAI introduces an experimental “confessions” system that rewards honest self-reporting of mistakes in GPT-5 Thinking models, aiming to improve AI safety and visibility of failures.