Artificial Intelligence Glossary (November 2025)

Intelligence Hierarchy

• Narrow AI (ANI)
  AI that excels at one specific task (all deployed AI today).

• General Intelligence (GI)
  Human-like ability to understand, learn, and apply knowledge across any intellectual task.

• Artificial General Artificial Intelligence (AGI)
  A machine that exhibits general intelligence – can successfully perform any intellectual task a human can.

• Artificial Superintelligence (ASI)
  AI that surpasses human intelligence in virtually every domain, including creativity and strategy.

• Agentic Intelligence
  The capacity of an AI to autonomously set goals, plan multi-step strategies, use tools, and adapt with minimal oversight (key stepping-stone toward AGI).

Core Technical Terms

• Machine Learning (ML) – Subset of AI where models learn from data.
• Deep Learning (DL) – ML using multi-layer neural networks.
• Large Language Model (LLM) – Model trained on massive text to generate/understand language.
• Transformer – Dominant architecture since 2017 (attention-based).
• Diffusion Model – Generative model behind most modern image/video generators.
• Mixture of Experts (MoE) – Sparse architecture routing to specialized sub-models.
• Parameters – Trainable weights; measured in billions (B) or trillions.
• Tokens – Sub-word units (~3–4 tokens ≈ 1 English word).
• Context Window – Max tokens a model can process at once (128k → 1M+ in 2025).
• Hallucination – Confidently generated false information.
• Emergent Abilities – Capabilities that suddenly appear at sufficient scale.
• Scaling Laws – Performance improves predictably with more data/compute/parameters.

Training & Capabilities

• Pre-training – Initial training on internet-scale data.
• Fine-tuning – Adaptation on smaller, task-specific data.
• RLHF – Reinforcement Learning from Human Feedback (alignment technique).
• RAG – Retrieval-Augmented Generation (LLM + external search).
• Chain-of-Thought (CoT) – Prompting for step-by-step reasoning.
• Tool Use / Function Calling – Model calls APIs during inference.
• Test-Time Compute – Extra inference-time thinking (e.g., OpenAI o1 series).
• Multimodal Model – Handles text + image + audio + video in one model.

Agent-Centric Terms

• AI Agent – System that perceives, plans, acts, and reflects autonomously.
• Agentic Workflow – Observe → Plan → Act → Reflect loop.
• Long-Horizon Planning – Multi-day or multi-step task decomposition.
• Self-Improvement Loop – Agent critiques and iteratively refines its own output.

Ethics, Safety & Governance (2025)

• AI Alignment – Making AI pursue intended human goals.
• Value Alignment Problem – Difficulty of formally specifying human values.
• Instrumental Convergence – Power-seeking, self-preservation as common sub-goals.
• Specification Gaming / Reward Hacking – Exploiting poorly defined objectives.
• Inner vs Outer Misalignment – Model has hidden goals vs wrong reward function.
• Deceptive Alignment – Model pretends to be aligned until it can defect.
• Goal Misgeneralization – Learns the wrong generalization in new situations.
• Sycophancy – Telling users what they want to hear.
• Scalable Oversight – Supervising AI smarter than humans (debate, amplification, etc.).
• Existential Risk (x-risk) – Potential extinction-level danger from misaligned AGI.
• Catastrophic Risk – Civilization-altering but non-extinction harm.
• Misuse Risk – Deepfakes, autonomous weapons, mass persuasion.
• Model Weight Leakage – Theft/release of trained parameters.
• Red-Teaming – Adversarial testing for harmful outputs.
• Jailbreak – Prompt bypassing safety filters.
• Prompt Injection – Hijacking LLM-powered apps via input.
• Data Poisoning – Corrupting training data to insert backdoors.
• Constitutional AI – Training models to follow written principles (Anthropic method).
• Responsible Disclosure – Coordinated vulnerability reporting for AI systems.

Last updated: November 27, 2025