TechCrunch publishes glossary of key artificial intelligence terms
TechCrunch has published a glossary of key concepts in artificial intelligence, from AGI and large language models to agents, hallucinations, and new reasoning methods. The publication notes that the industry's vocabulary is rapidly expanding alongside the development of generative models and automation tools.
Models, agents, and computing
AGI, or artificial general intelligence, is usually described as a system capable of performing many tasks at a level above that of the average person. At the same time, there is no single definition of the term: OpenAI's charter, the company's CEO Sam Altman, and Google DeepMind describe AGI somewhat differently.
An AI agent is a tool capable of independently carrying out a sequence of actions to achieve a goal, such as booking tickets, filing expenses, or working with code. A separate category is coding agents, which can write, test, and debug code, although the results of their work require human review.
Large language models, or LLMs, are used by popular AI assistants, including ChatGPT, Claude, Gemini, Llama, Microsoft Copilot, and Le Chat. These are deep neural networks with billions of parameters that learn relationships between words and phrases from large text datasets. Inference refers to running an already trained model to generate predictions or responses, while compute refers to computational resources, including GPUs, CPUs, TPUs, and other infrastructure.
More current news is available on the UA.News Telegram channel Telegram.
Training and response control
Chain of thought means breaking a complex task into intermediate steps. This approach may take more time, but it helps improve the accuracy of responses, particularly in logic tasks and programming. Methods for improving models also include fine-tuning on specialized data and distillation, in which a more compact model learns to reproduce the behavior of a larger one.
Hallucinations are cases in which a model generates false information. This can create risks, particularly when a system provides incorrect advice in sensitive areas. Other concepts in the glossary include diffusion models for creating images, music, and text; generative adversarial networks, or GANs; and the Mixture of Experts architecture, in which only part of the specialized subnetworks is activated for each query.
Opaque recurrence
TechCrunch separately draws attention to opaque recurrence. Under this approach, a model does not produce reasoning step by step in understandable language, but instead repeatedly passes the same query through internal layers. The publication writes that this may be more efficient and allow smaller models to demonstrate higher performance at lower computational cost, but it also leaves fewer reasoning traces available for verification. AI safety researchers link this to more difficult oversight of model behavior.
