Terms such as AI agents, open-weight models and artificial general intelligence describe different capabilities, access arrangements and safety questions, distinctions that shape who can deploy systems and what claims remain untested.
OpenAI applies the term agent to a tool-connected AI system that can take actions under instructions and guardrails. OpenAI’s documentation describes sandboxes as isolated environments that can give agents controlled access to files, commands and external systems.
Stanford University's Human-Centered Artificial Intelligence institute describes an open-weight model as one whose learned parameters are available for download. Users can run it on local hardware or use a third-party host. Local users pay for hardware and electricity; hosted access can involve paying a provider.
Open-source AI makes a broader claim under the Open Source Initiative's definition. It calls for model parameters, information about training data and the code used to train and run the system, along with rights to use, study, modify and share it. Downloadable weights alone do not meet that definition.
Distillation trains a smaller model using outputs from a stronger one. Anthropic says developers use the method to make cheaper models, while unauthorized attempts to extract another company's capabilities present a separate concern.
Safety terms describe different questions and practices. Alignment asks whether an AI system's behavior follows human goals and standards, a research problem described by OpenAI Chief Scientist Jakub Pachocki. A jailbreak uses prompting to try to get around safeguards. The National Institute of Standards and Technology (NIST) defines prompt injection as exploiting the combination of untrusted input and a prompt supplied by a higher-trust party. Red teaming uses structured tests, often adversarial, to look for weaknesses.
Researchers have also tested refusal-direction ablation, which edits a mechanism linked to a model's refusals. Their paper reported reduced refusals in 13 chat models, with the largest containing 72 billion parameters; the finding applies to the models tested.
Stanford describes artificial general intelligence, or AGI, as broad ability across tasks at a human level or beyond, but says there is no universally accepted test for it. Recursive self-improvement is a proposed capability in which an AI system plans and builds a later system itself. Anthropic and OpenAI say fully autonomous recursive self-improvement has not been achieved.
