This guide defines essential AI terminology, from foundational concepts like LLMs and compute to specialized techniques such as distillation and chain-of-thought reasoning.
Key facts
- •AGI is defined differently by various organizations, including OpenAI and Google DeepMind, but generally refers to AI that matches or exceeds human capability in most tasks.
- •Coding agents are specialized programs that can autonomously write, test, and debug software code.
- •Generative Adversarial Networks (GANs) use two competing neural networks to optimize the realism of generated data.
- •Inference is the process of running a trained model to make predictions or draw conclusions.
- •Distillation is often used to create smaller, more efficient models, though using it to copy a competitor's model may violate terms of service.
As artificial intelligence technology evolves, a complex new vocabulary has emerged to describe its development and capabilities. This glossary provides plain-English definitions for common industry terms, including AGI, AI agents, and various technical processes that underpin modern generative AI models.
Foundational AI Concepts
Artificial General Intelligence (AGI) refers to systems capable of performing at or above human levels across most cognitive or economically valuable tasks. AI agents are autonomous systems designed to execute multi-step tasks on behalf of a user, such as writing code or booking services, often by interacting with API endpoints—interfaces that allow software to communicate and trigger actions across different platforms.
Model Training and Reasoning
Large language models (LLMs) are deep neural networks that learn language patterns from vast datasets to generate responses. To improve accuracy in complex tasks, models may use chain-of-thought reasoning, which breaks problems into intermediate steps. Other techniques include distillation, where a smaller 'student' model is trained to approximate the behavior of a larger 'teacher' model, and fine-tuning, which optimizes a model for specific tasks by introducing specialized data.
Technical Infrastructure and Challenges
Compute refers to the hardware, such as GPUs and CPUs, required to train and deploy AI. Deep learning utilizes multi-layered neural networks to identify data characteristics, while diffusion models generate content by learning to reverse the process of adding noise to data. A significant challenge in the industry is 'hallucination,' where models generate incorrect information, a phenomenon often attributed to gaps in training data.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by TechCrunch.
