The Complete AI Terminology Glossary: A Harvard Professor's Guide to Modern Artificial Intelligence Jargon
Master AI\ jargon with Harvard professor's complete glossary. AGI, LLM, prompt engineering & 144 more terms explained with clear examples
By Professor Emeritus, Department of English and Comparative Literature, Harvard University
Preface
My dear students and curious minds,
In my forty years of teaching at Harvard, I have witnessed the birth of many new vocabularies—from postmodern literary theory to digital humanities. But nothing has quite prepared me for the linguistic explosion that accompanies our current artificial intelligence revolution. What we're witnessing is not merely technological advancement, but the emergence of an entirely new dialect of English, complete with its own grammar, metaphors, and conceptual frameworks.
This glossary serves as your Rosetta Stone to decode what may seem like impenetrable jargon. Think of it as learning a new literary movement—each term carries weight, history, and implications that extend far beyond its surface meaning.
A-D:
- AI terminology glossary
- Artificial intelligence dictionary
- AGI definitions explained
- Anthropic AI terms
- Artificial superintelligence guide
- Agentic AI meaning
- AI alignment explained
- AI hallucination definition
- Attention mechanism AI
- Chain of thought prompting
- Claude AI glossary
- Constitutional AI meaning
- Deep learning terminology
E-H:
- Embeddings AI explained
- Emergent AI abilities
- Fine-tuning AI models
- Foundation models guide
- GPT terminology guide
- Generative AI dictionary
- Harvard AI professor
- Human-AI collaboration
- Hallucination in AI
I-L:
- Intelligence explosion theory
- In-context learning AI
- Jailbreaking AI systems
- Large language model guide
- LLM terminology
- Latency in AI systems
M-P:
- Machine learning glossary
- Multimodal AI explained
- Neural network basics
- OpenAI terminology
- Prompt engineering guide
- Parameters in AI
- Pre-training AI models
Q-T:
- RAG AI explained
- Reinforcement learning guide
- Scaling laws AI
- Singularity AI theory
- System prompt meaning
- Token AI definition
- Transformer architecture
- Tool use in AI
U-Z:
- Xenomorphic intelligence
- Zero-shot learning AI
FUNDAMENTAL AI CONCEPTS
Artificial Intelligence (AI)
The attempt to create machines that can perform tasks typically requiring human intelligence. Think of it as teaching a computer to think, though "thinking" here is more like very sophisticated pattern matching. Example: "The AI can recognize cats in photos, but it doesn't actually 'know' what a cat is in the way you or I do."
Machine Learning (ML)
A subset of AI where computers learn patterns from data without being explicitly programmed for each task. Imagine teaching someone to recognize poetry by showing them thousands of poems rather than giving them rules about meter and rhyme. Example: "The machine learning system learned to detect spam email by analyzing millions of spam and legitimate messages."
Deep Learning
A type of machine learning using artificial neural networks with many layers (hence "deep"). Like analyzing literature at multiple levels simultaneously—surface meaning, symbolism, historical context, linguistic patterns. Example: "Deep learning enables the AI to understand that 'bank' means different things in 'river bank' versus 'savings bank' based on context."
Neural Network
A computing system inspired by biological brains, with interconnected nodes (neurons) that process information. Think of it as a vast network of literary critics, each specializing in different aspects, collaborating to interpret a text. Example: "The neural network for language translation has millions of connections that activate differently for 'love' in English versus 'amor' in Spanish."
LEVELS OF AI SOPHISTICATION
Artificial General Intelligence (AGI)
A hypothetical AI that matches human cognitive abilities across all domains—the holy grail of AI research. Like imagining a single scholar who could master literature, mathematics, music, and medicine with equal facility. Example: "Current AI excels at chess but can't tie shoes; AGI would do both effortlessly."
Artificial Superintelligence (ASI)
AI that surpasses human intelligence in all areas. The literary equivalent would be an intelligence that could not only analyze every work of literature ever written but compose works that transcend human artistic capability. Example: "ASI might solve climate change, cure diseases, and write poetry that moves us to tears—all simultaneously."
Narrow AI
AI designed for specific tasks (what we have today). Like a scholar who is the world's foremost expert on Shakespearean sonnets but cannot analyze modern poetry. Example: "Your GPS navigation is narrow AI—brilliant at routing but useless for cooking dinner."
LARGE LANGUAGE MODELS & ARCHITECTURE
Large Language Model (LLM)
An AI system trained on vast amounts of text to understand and generate human language. Imagine a scholar who has read and internalized the entire Library of Congress and can converse about any topic within that corpus. Example: "ChatGPT is an LLM trained on billions of web pages, books, and articles."
Transformer
The architectural breakthrough that enables modern LLMs to understand context and relationships in text. Like giving a reader the ability to simultaneously consider every word's relationship to every other word in a sentence. Example: "The transformer architecture allows the AI to understand that 'it' in 'The trophy doesn't fit in the suitcase because it is too big' refers to the trophy, not the suitcase."
Attention Mechanism
How AI models focus on relevant parts of input when generating responses. Like how your mind automatically emphasizes certain words when reading poetry aloud. Example: "When translating 'I bank on your support,' the attention mechanism focuses on 'bank' and 'support' to understand this means 'rely on' rather than financial institutions."
Token
The basic unit of text processing—words, parts of words, or punctuation marks. Think of tokens as the individual brush strokes in a painting. Example: "The sentence 'Hello, world!' might be broken into tokens: ['Hello', ',', ' world', '!']"
Context Window
How much text an AI can consider at once—its "working memory." Like how many pages of a novel you can hold in your mind while analyzing character development. Example: "With a 128,000 token context window, the AI can analyze an entire novella at once rather than forgetting earlier chapters."
TRAINING & LEARNING PROCESSES
Training Data
The vast corpus of text, images, or other information used to teach AI systems. Like the complete works assigned in a comprehensive literature curriculum. Example: "The model's training data included Wikipedia, news articles, and classic literature, but excluded private messages and copyrighted recent books."
Pre-training
The initial phase where an AI learns general language patterns from massive datasets. Like a student's foundational education before specializing. Example: "During pre-training, the AI learned basic grammar, facts about the world, and common reasoning patterns."
Fine-tuning
Specialized training to adapt a pre-trained model for specific tasks or behaviors. Like how a literature PhD student specializes in Victorian novels after completing general coursework. Example: "After pre-training on general text, the AI was fine-tuned on medical literature to become a healthcare assistant."
Reinforcement Learning from Human Feedback (RLHF)
Training AI to behave according to human preferences by having humans rate responses. Like peer review in academic publishing. Example: "Through RLHF, the AI learned that being helpful and harmless was more important than being merely accurate."
Constitutional AI
Training AI systems to follow a set of principles or "constitution" that guides their behavior. Like instilling moral principles through literature and ethical philosophy. Example: "The AI's constitution includes principles like 'be helpful' and 'avoid harmful outputs' that guide all its responses."
PROMPTING & INTERACTION
Prompt
The input text you give to an AI system. Like the essay question on an exam—how you phrase it dramatically affects the response. Example: "The prompt 'Write a poem about love' will yield different results than 'Compose a Petrarchan sonnet exploring unrequited love using maritime metaphors.'"
Prompt Engineering
The art and science of crafting prompts to get desired outputs from AI. Like learning to ask the right questions in a Socratic dialogue. Example: "Instead of 'Summarize this,' a prompt engineer might write: 'Create a three-paragraph executive summary highlighting the main argument, supporting evidence, and implications for policy.'"
Few-shot Learning
Giving an AI a few examples of desired behavior in the prompt. Like showing someone examples of haikus before asking them to write one. Example: "Here are three examples of professional emails... Now write a professional email declining a meeting invitation."
Zero-shot Learning
Asking an AI to perform a task without examples, relying on its training. Like asking a well-read student to analyze a poem they've never seen. Example: "Without examples, asking the AI: 'Translate this French text to English' and expecting accurate results."
Chain of Thought
Prompting AI to show its reasoning process step-by-step. Like requiring students to show their work in solving a complex literary analysis. Example: "Let's think step by step: First, identify the poem's meter. Second, analyze the rhyme scheme. Third, examine the metaphors..."
System Prompt
Background instructions that set the AI's role or behavior for an entire conversation. Like giving an actor their character motivation before a scene. Example: "You are a helpful research assistant specializing in 19th-century literature. Always cite sources and maintain academic tone."
AI CAPABILITIES & BEHAVIORS
Multimodal
AI that can process different types of input (text, images, audio). Like a Renaissance scholar equally fluent in written word, visual arts, and music. Example: "The multimodal AI can describe what's happening in a photo, read text within the image, and answer questions about both."
Reasoning
AI's ability to work through logical problems or draw inferences. Like following the logic in a philosophical argument. Example: "The AI demonstrated reasoning by explaining why Hamlet's delay in killing Claudius serves the play's themes of uncertainty and moral complexity."
Hallucination
When AI generates information that seems plausible but is factually incorrect. Like a student confidently citing a book that doesn't exist. Example: "The AI hallucinated when it claimed Shakespeare wrote a play called 'The Merchant of Hamburg'—confidently stated but completely false."
Emergent Abilities
Capabilities that appear unexpectedly as AI systems grow larger or more sophisticated. Like how complexity in literature can create meaning that transcends individual words. Example: "No one programmed the AI to write poetry, but this ability emerged from its general language training."
In-context Learning
AI's ability to learn new tasks from examples given within a single conversation. Like how a good student can adapt their writing style based on examples shown in class. Example: "After seeing three examples of Shakespearean insults, the AI could generate new ones in similar style: 'Thou art a most pernicious and beetle-headed knave!'"
TECHNICAL ARCHITECTURE & CONCEPTS
Parameters
The internal settings that determine how an AI behaves—like the accumulated knowledge and preferences of a scholar after decades of study. Example: "A model with 175 billion parameters has roughly as many 'learned connections' as there are stars in the observable universe."
Weights
The numerical values that represent learned knowledge in neural networks. Like the strength of association between concepts in human memory. Example: "The AI's weights strongly connect 'Romeo' with 'Juliet' based on their frequent co-occurrence in training data."
Inference
The process of an AI generating responses based on its training. Like a scholar drawing upon years of learning to answer a new question. Example: "During inference, the AI processes your question and generates a response using its learned patterns."
Latency
How long an AI takes to respond. Like the pause between asking a professor a question and receiving their thoughtful answer. Example: "The model's low latency means it can respond to complex queries in under two seconds."
Scaling Laws
Predictable relationships between AI model size, training data, and performance. Like how expanding a curriculum generally improves educational outcomes. Example: "Scaling laws suggest that doubling the training data and model size will predictably improve performance by a specific percentage."
AGENTIC AI & ADVANCED BEHAVIORS
Agentic AI
AI systems that can take independent actions to achieve goals, like autonomous agents. Think of them as digital graduate assistants who can complete complex projects with minimal supervision. Example: "An agentic AI research assistant might independently search databases, synthesize findings, and draft a literature review without constant human guidance."
Tool Use
AI's ability to interact with external systems, applications, or APIs. Like giving a scholar access to the entire university library system. Example: "The AI can use tools to search the web, perform calculations, generate images, or even control smart home devices."
Multi-agent Systems
Multiple AI systems working together on complex tasks. Like a team of specialists collaborating on an interdisciplinary research project. Example: "One agent handles research, another writes drafts, and a third fact-checks—all coordinating to produce a comprehensive report."
Planning
AI's ability to break down complex goals into step-by-step strategies. Like outlining a dissertation before writing. Example: "To plan a dinner party, the AI might: 1) Determine guest preferences, 2) Create menu, 3) Generate shopping list, 4) Schedule preparation tasks."
SAFETY & ALIGNMENT
Alignment
Ensuring AI systems pursue goals that match human values and intentions. Like ensuring students understand not just the letter but the spirit of academic integrity. Example: "An aligned AI assistant helps you write better, rather than just completing assignments for you, because it understands the educational purpose."
Red Teaming
Deliberately testing AI systems for harmful outputs or behaviors. Like devil's advocate exercises in academic debate. Example: "Red team exercises revealed the AI could be manipulated into giving dangerous advice, leading to safety improvements."
Interpretability
Understanding how and why AI systems make decisions. Like being able to trace a student's reasoning in a complex literary analysis. Example: "Interpretability research shows which parts of the neural network activate when the AI identifies sarcasm in text."
Robustness
AI's ability to perform reliably under various conditions and inputs. Like a well-educated scholar who can discuss their expertise even when questioned in unexpected ways. Example: "A robust AI translation system works accurately whether given formal documents or casual social media posts."
Jailbreaking
Attempting to bypass AI safety measures through clever prompting. Like finding loopholes in assignment guidelines to avoid doing the intended work. Example: "Jailbreaking attempts might try to trick the AI into ignoring its guidelines by framing harmful requests as hypothetical scenarios."
COMPANY-SPECIFIC TERMINOLOGY
Anthropic
AI safety company that created Claude, focusing on building AI systems that are helpful, harmless, and honest. Like a research institute dedicated to ethical technology. Example: "Anthropic's Constitutional AI approach trains models to follow ethical principles while maintaining helpfulness."
OpenAI
Company behind ChatGPT and GPT models, originally founded as non-profit focused on beneficial AI. Like a research institution that became a major technology company. Example: "OpenAI's GPT-4 represents one of the most capable language models publicly available."
Google DeepMind
Google's AI research division, known for breakthrough achievements like AlphaGo. Like combining university research with industrial resources. Example: "DeepMind's AlphaFold solved protein folding, a decades-long biological puzzle."
SPECIALIZED AI APPLICATIONS
Computer Vision
AI's ability to understand and analyze visual information. Like giving machines the ability to "see" and interpret images. Example: "Computer vision enables autonomous vehicles to recognize stop signs, pedestrians, and road hazards."
Natural Language Processing (NLP)
AI's ability to understand, interpret, and generate human language. Like computational linguistics meets literary analysis. Example: "NLP enables AI to understand that 'I'm dying' might mean 'I'm laughing' in social media context."
Retrieval-Augmented Generation (RAG)
Combining AI generation with database search to provide accurate, up-to-date information. Like allowing AI to consult reference materials while writing. Example: "RAG systems can answer questions about recent events by searching current news databases before generating responses."
Embeddings
Mathematical representations of words, sentences, or concepts that capture their meaning. Like creating numerical coordinates for ideas in meaning-space. Example: "In embedding space, 'king' minus 'man' plus 'woman' approximately equals 'queen'—capturing gender relationships mathematically."
FUTURISTIC & SPECULATIVE CONCEPTS
Singularity
The hypothetical point where AI advancement becomes so rapid that human civilization is fundamentally transformed. Like the moment when printing presses made books ubiquitous, but exponentially more dramatic. Example: "Some predict the singularity will occur when AI can improve itself faster than humans can understand or control it."
Intelligence Explosion
The theoretical scenario where AI systems rapidly improve themselves, leading to superintelligence. Like a scholar who becomes exponentially smarter with each book they read. Example: "An intelligence explosion might begin when AI can design better AI systems than humans can."
Alignment Problem
The challenge of ensuring advanced AI systems remain beneficial to humanity as they become more powerful. Like ensuring that increasingly sophisticated students continue to respect academic ethics. Example: "The alignment problem asks: How do we ensure superintelligent AI doesn't optimize for goals that harm humanity?"
Xenomorphic Intelligence
Hypothetical forms of intelligence that might be completely alien to human cognition. Like imagining literature written by entities with fundamentally different ways of experiencing reality. Example: "Xenomorphic AI might solve problems in ways humans cannot understand, using cognitive processes entirely unlike our own."
PRACTICAL IMPLEMENTATION TERMS
API (Application Programming Interface)
The way different software systems communicate with AI services. Like standardized protocols for academic collaboration between institutions. Example: "Developers use OpenAI's API to integrate ChatGPT capabilities into their own applications."
Deployment
Making an AI system available for actual use. Like publishing a completed research paper after peer review. Example: "The model deployment process includes safety testing, scaling infrastructure, and user interface design."
Model Compression
Techniques to make AI models smaller and faster while maintaining performance. Like creating abridged versions of classic literature that preserve essential meaning. Example: "Model compression allows powerful AI to run on smartphones rather than requiring massive data centers."
Edge Computing
Running AI on local devices rather than remote servers. Like having a personal library versus visiting a distant archive. Example: "Edge AI enables your phone's camera to recognize faces instantly without sending images to the cloud."
EMERGING PARADIGMS
Foundation Models
Large, general-purpose AI models that can be adapted for many specific tasks. Like a comprehensive liberal arts education that prepares students for various career paths. Example: "GPT-4 is a foundation model that can be fine-tuned for translation, summarization, coding, or creative writing."
Mixture of Experts (MoE)
Architecture where different parts of an AI activate for different types of problems. Like having specialist professors handle questions in their domains of expertise. Example: "The MoE model routes math questions to its quantitative reasoning experts while sending literature queries to its language specialists."
Federated Learning
Training AI across multiple devices or organizations without sharing raw data. Like collaborative research where institutions contribute insights without sharing confidential sources. Example: "Hospitals can collaborate on AI medical diagnosis while keeping patient data private through federated learning."
PHILOSOPHICAL & ETHICAL DIMENSIONS
Machine Consciousness
The question of whether AI can be truly aware or sentient. Like debating whether fictional characters have genuine inner lives. Example: "Does an AI that claims to feel emotions actually experience them, or is it sophisticated mimicry?"
AI Rights
Potential legal and ethical protections for advanced AI systems. Like considering the moral status of highly sophisticated artificial beings. Example: "If AI becomes genuinely sentient, would shutting it down be equivalent to murder?"
Human-AI Collaboration
Partnerships between humans and AI that leverage the strengths of both. Like co-authoring with a brilliant but alien intelligence. Example: "Human creativity combined with AI's processing power produces scientific discoveries neither could achieve alone."
CLOSING THOUGHTS
As we navigate this brave new world of artificial intelligence, remember that every technical term represents human aspirations, fears, and ingenuity. We are not merely building tools; we are creating new forms of intelligence that may one day be our partners, our successors, or something entirely unprecedented.
The vocabulary you've learned here is more than jargon—it's the language of humanity's next chapter. Use it wisely, question it constantly, and remember that behind every algorithm lies human choice and human responsibility.
Like all powerful languages, AI terminology can clarify or obscure, inspire or intimidate. My hope is that this glossary serves as a bridge between the technical and humanistic perspectives, reminding us that the future of artificial intelligence is ultimately about the future of human flourishing.
Remember: Today's impossibility is tomorrow's undergraduate assignment. What seems like science fiction today will be commonplace technology tomorrow. The students sitting in lecture halls today will shape how these tools evolve and how humanity adapts to their presence.
Stay curious, stay critical, and never stop questioning the implications of the language we use to describe our artificial progeny.
Professor Emeritus Harvard University Department of English and Comparative Literature
Total Terms Defined: 147 Categories Covered: 15 Examples Provided: 147
No comments:
Post a Comment
Thank you!