February 1, 2025
Today we introduce n1-preview—a new system for training self-learning AI agents. n1-preview is designed to understand natural language instructions and engage in conversation, turning your words into training plans that help agents learn and improve over time.
n1-preview works by analyzing user instructions and mapping these into actionable training steps. It leverages self-reinforcing feedback alongside reasoning, processing analytics, instructions, and memories to refine an agent’s performance. The system reads and writes its own code to unlock additional capabilities, allowing it to perform nearly any action possible on computers and through connected systems. It is built on a distributed, serverless platform, scaling with demand.
This preview release marks a crucial step. It allows us to identify the capabilities and limitations of n1-preview and investigate the best approaches to ensure the safety and alignment of all AI agents trained by users. In previous iterations, challenges such as input hallucinations, inflexibility in learning new tasks, loss of context, and rigid behavior in handling multiple responsibilities became apparent. n1-preview is designed to address these issues and improve overall performance.
n1-preview runs inference on a global, distributed serverless platform. Training with n1-preview produces artifacts like model weights and binary data, managed by dedicated inference nodes. The system supports a wide range of capabilities, including self-learning, reasoning, memory, knowledge accumulation, multilingual understanding (such as English, Chinese, Spanish, French, and more), code generation, vision, and math.
Safety is a key focus during this preview phase. Working closely with academic partners and public researchers, we are evaluating performance and refining safety measures. These efforts include continuous monitoring, human oversight, and red teaming exercises that test the system’s responses to potentially harmful instructions or attempts to extract sensitive information. Preliminary benchmarks across various tasks are being gathered to guide further development.
Ethical considerations remain central. While n1-preview represents a significant step forward, its outputs, like those of any AI system, can be unpredictable. We strongly encourage trainers and experts to thoroughly test and tune their agents for specific use cases before deployment.
By evaluating its capabilities and limitations, we hope to build a foundation for more robust, adaptable, and safe AI agents for users worldwide.
Igor Hoogerwoord
February 1, 2025