AI glossary explains opaque recurrence, MCP and core model terms
An updated AI glossary has added opaque recurrence, recurrent depth and Model Context Protocol (MCP) to a growing set of terms used to describe modern AI systems. It explains concepts ranging from artificial general intelligence (AGI) and large language models (LLMs) to inference, token throughput and AI agents, with particular attention to the reasoning technique associated with OpenAI’s Astra model.
Opaque recurrence describes a model sending the same query repeatedly through its internal layers rather than working through a problem sequentially in human-readable language. The approach is presented as more efficient: smaller models may achieve stronger results while consuming less compute. Its trade-off is a reduced record of the intermediate reasoning that safety teams can inspect.
Why opaque recurrence has drawn attention
In conventional chain-of-thought reasoning, a language model breaks a task into intermediate steps, which can improve answers in logic and coding tasks. Opaque recurrence instead relies on repeated internal processing. The glossary calls recurrent depth the more technical name for the same method, while opaque recurrence emphasizes the safety concern.
The issue is not that a deployed model has entered “neuralese”, a hypothetical situation in which a system reasons only in internal numerical representations. The glossary says OpenAI has stated that Astra keeps its chain of thought legible. Still, the safety debate surrounding Astra opaque recurrence safety concerns illustrates why fewer readable traces can make model oversight more difficult.
Standards, agents and infrastructure
The glossary also places agentic AI in an operational context. An AI agent is a tool that can carry out multistep tasks, such as filing expenses, booking tickets or maintaining code, potentially using several AI systems. Coding agents are a narrower category designed to write, test and debug software, although human review remains necessary.
MCP is described as an open standard for connecting AI models to external tools and data, including files, databases, Slack and Google Drive, without building a custom connector for every pairing. Anthropic introduced MCP in 2024 and later transferred it to the Linux Foundation; the standard has since been adopted by OpenAI, Google and Microsoft.
Terms that affect deployment decisions
Several definitions have direct cost and performance implications. Inference is the process of running a trained model, while memory caching can reduce repeated calculations during inference. Tokens are the units into which language-model input and output are divided, and token throughput measures how much of that work a system can process in a given period.
For businesses, the practical implication is to assess AI proposals beyond product labels: teams should establish how a model reasons, what tools and data it can access, how it is monitored, and which infrastructure measures determine cost and responsiveness before moving a system into production.