By Titik Puspa 
Artificial intelligence is not just transforming the world; it’s simultaneously forging a new vocabulary to articulate its own progress. From the hushed tones of product meetings to the boisterous energy of industry panels, terms like LLMs, RAG, and RLHF have become commonplace. More recently, the AI safety community has been abuzz with concern over "opaque recurrence," a novel reasoning technique introduced in OpenAI’s Astra model, sparking debate and unease even among seasoned tech professionals. The sheer pace at which this lexicon evolves can leave even the most astute individuals feeling a touch out of their depth. This comprehensive glossary serves as a beacon, aiming to demystify the complex terminology surrounding AI, offering plain-English definitions for the terms you’re most likely to encounter, whether you’re a builder, an investor, or simply a keen observer trying to keep pace with this rapidly advancing field. This document is designed to be a living resource, regularly updated to reflect the dynamic nature of AI.
AGI: The Elusive Goal of Human-Level Intelligence
Artificial general intelligence (AGI) remains a somewhat nebulous concept, but it generally describes AI systems that possess capabilities exceeding those of the average human across a wide spectrum of tasks. Sam Altman, CEO of OpenAI, once characterized AGI as akin to "a median human that you could hire as a co-worker." OpenAI’s own charter offers a more ambitious definition: "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind, meanwhile, frames AGI as "AI that’s at least as capable as humans at most cognitive tasks." The nuanced differences in these definitions highlight the ongoing debate and the inherent complexity of defining true general intelligence. Indeed, even leading AI researchers admit to a degree of uncertainty regarding its precise definition and realization.
AI Agent: Your Digital Workhorse
An AI agent is a sophisticated tool that leverages AI technologies to perform a series of tasks on your behalf, extending beyond the capabilities of a standard AI chatbot. These agents can handle complex operations like managing expenses, booking travel and dining reservations, or even writing and maintaining software code. However, the definition of an "AI agent" is still emergent and can vary among different stakeholders. The underlying infrastructure to fully realize their potential is still under development, but the core concept points to an autonomous system capable of orchestrating multiple AI functionalities to execute multi-step tasks.
API Endpoints: The Programmable Levers of Software
API endpoints can be visualized as the "buttons" on the back end of a software application, allowing other programs to interact with it and trigger specific actions. Developers utilize these interfaces to build integrations, enabling applications to share data or allowing AI agents to directly control third-party services without manual human intervention. Most connected devices and platforms, from smart home appliances to complex cloud services, feature these hidden interfaces. As AI agents become more sophisticated, their ability to discover and utilize these endpoints autonomously unlocks powerful, and occasionally unforeseen, automation possibilities.
Chain of Thought: Deconstructing Complexity for Clarity
While humans can often answer simple questions intuitively, complex problems frequently require a step-by-step approach. For instance, determining the number of chickens and cows on a farm based on the total number of heads and legs necessitates intermediate calculations. In the realm of AI, "chain-of-thought" reasoning involves breaking down a complex problem into smaller, manageable steps. This method, typically employed by large language models, significantly enhances the accuracy and reliability of the final output, especially for logic-based or coding tasks, even if it may take slightly longer to generate a response. Reasoning models are often derived from traditional LLMs and fine-tuned through reinforcement learning to excel at this sequential problem-solving.
Coding Agents: Automating the Software Development Lifecycle
Building upon the concept of AI agents, coding agents represent a specialized application focused on software development. Instead of merely suggesting code snippets for human review, a coding agent can autonomously write, test, and debug code. This capability tackles the iterative, trial-and-error nature of development, which often consumes significant developer time. These agents can analyze entire codebases, identify bugs, execute tests, and implement fixes with minimal human oversight, functioning akin to an exceptionally efficient, tireless intern. Nevertheless, human supervision remains crucial for reviewing the work of these advanced agents.
Compute: The Engine of AI Advancement
"Compute" refers to the essential computational power that underpins the operation of AI models. This processing capability is the driving force behind the AI industry, enabling the training and deployment of sophisticated AI systems. The term is often used as shorthand for the specialized hardware, such as GPUs, CPUs, and TPUs, that provide this crucial processing power, forming the bedrock of modern AI infrastructure.
Deep Learning: Mimicking the Brain’s Architecture
Deep learning is a subfield of machine learning characterized by AI algorithms designed with multi-layered artificial neural networks (ANNs). This intricate structure allows for more complex correlation identification compared to simpler machine learning models. Inspired by the interconnected pathways of neurons in the human brain, deep learning models can autonomously identify critical features within data, eliminating the need for human engineers to pre-define them. This architecture also facilitates learning from errors, enabling models to refine their outputs through iterative adjustments. However, deep learning systems demand vast datasets (millions of data points or more) for optimal performance and typically require longer training times, leading to higher development costs.
Diffusion: Generating Novel Content from Noise
Diffusion models are the core technology behind many AI systems that generate art, music, and text. Drawing inspiration from physics, these systems progressively "destroy" the structure of data by adding noise until only random patterns remain. In physics, diffusion is a spontaneous and irreversible process. However, AI diffusion systems aim to learn a "reverse diffusion" process, enabling them to reconstruct original data from noise, thereby generating novel content.
Distillation: Knowledge Transfer for Efficiency
Distillation is a technique that allows knowledge to be transferred from a larger, more complex "teacher" AI model to a smaller, more efficient "student" model. Developers send queries to the teacher model and record its outputs, which are then used to train the student model to emulate the teacher’s behavior. This process can create more compact and performant models, potentially speeding up development and reducing computational costs. While widely used internally by AI companies, distillation can also be employed to replicate the capabilities of frontier models, though this may violate the terms of service of AI APIs and chat assistants.
Fine-Tuning: Specializing AI for Specific Tasks
Fine-tuning involves the further training of an AI model on new, specialized data to optimize its performance for a specific task or domain. This process allows startups to leverage large language models as a foundation and enhance their utility for particular sectors by incorporating domain-specific knowledge and expertise.
GANs: The Art of Adversarial Creation
Generative Adversarial Networks (GANs) are a machine learning framework crucial for generating realistic synthetic data, including deepfakes. GANs consist of two neural networks: a generator that creates outputs, and a discriminator that evaluates them. The two networks engage in a competitive process, with the generator striving to produce outputs that fool the discriminator, while the discriminator aims to accurately identify artificially generated data. This adversarial dynamic optimizes AI outputs for realism without extensive human intervention. While effective for specific applications like generating realistic images or videos, GANs are generally less suited for general-purpose AI.
Hallucination: When AI Gets It Wrong
"Hallucination" is the AI industry’s term for when models generate incorrect or fabricated information. This phenomenon poses a significant challenge to AI quality, as it can lead to misleading outputs with potentially harmful real-world consequences, particularly in sensitive domains like health or finance. The issue is often attributed to gaps in training data, driving a trend towards more specialized, domain-specific AI models to mitigate disinformation risks.
Inference: Putting AI Models to Work
Inference is the process of running a trained AI model to make predictions or draw conclusions from new data. This crucial step cannot occur without prior training; models must first learn patterns from a dataset before they can effectively extrapolate from it. Various hardware, from smartphone processors to high-end AI chips, can perform inference, but the performance varies significantly, with larger models requiring more powerful computing resources.
Large Language Model (LLM): The Foundation of Modern AI Assistants
Large Language Models (LLMs) are the sophisticated AI models powering popular AI assistants like ChatGPT, Claude, Gemini, Llama, Microsoft Copilot, and Mistral’s Le Chat. These models are deep neural networks comprising billions of numerical parameters (weights) that learn the intricate relationships between words and phrases, creating a multidimensional representation of language. LLMs are trained on vast quantities of text data, enabling them to generate coherent and contextually relevant responses based on user prompts.
Memory Cache: Accelerating AI Responses
Memory cache is a critical process for optimizing inference, which is how AI models generate responses to user queries. Caching acts as an efficiency booster by storing the results of complex calculations, thereby reducing the computational load for subsequent, similar queries. Key-value (KV) caching, a well-known technique for transformer-based models, significantly enhances efficiency by minimizing the time and algorithmic effort required to generate answers.
Model Context Protocol (MCP): The Universal Connector for AI Tools
The Model Context Protocol (MCP) is an open standard that enables AI models to seamlessly connect with external tools and data sources, such as files, databases, and applications like Slack or Google Drive, without the need for custom connectors. Introduced by Anthropic and now managed by the Linux Foundation, MCP has been rapidly adopted by major AI players like OpenAI, Google, and Microsoft, positioning it as a foundational standard for AI interoperability.
Mixture of Experts (MoE): Specialized Intelligence for Efficiency
Mixture of Experts (MoE) is a model architecture that divides a neural network into numerous specialized sub-networks, or "experts." For any given task, only a select few experts are activated, guided by a built-in "router." This approach allows for the creation of massive models that remain remarkably fast and cost-effective to operate, as only a fraction of the network is engaged at any time. Mistral AI’s Mixtral model is a prominent example, and it is widely believed that newer GPT models also utilize variations of this architecture.
Neural Network: The Brain-Inspired Computing Model
A neural network is a multi-layered algorithmic structure that forms the basis of deep learning and the recent surge in generative AI. While the concept of mimicking the human brain’s neural pathways for computation dates back to the 1940s, the advent of powerful graphical processing hardware (GPUs), initially developed for the video game industry, unlocked the true potential of this theory. These chips are adept at training algorithms with many more layers than previously feasible, leading to significant performance gains in areas like voice recognition, autonomous navigation, and drug discovery.
Neuralese: The Unreadable Future?
Neuralese represents a hypothetical worst-case scenario where an AI model operates entirely within its internal numerical representations, rendering its decision-making process an impenetrable black box. While no currently deployed model exhibits this behavior, OpenAI’s Astra model, which uses "opaque recurrence," has been flagged by safety researchers as a potential first step in this direction. OpenAI contends that Astra maintains legible chains of thought, pushing back against direct comparisons to neuralese.
Opaque Recurrence: Efficiency Over Explainability
Opaque recurrence is a technique where an AI model repeatedly cycles a query through its internal layers rather than engaging in explicit step-by-step reasoning in human-readable language. This method is more efficient, allowing smaller models to achieve greater performance with less computational power. However, it generates fewer interpretable traces compared to standard chain-of-thought processes, which raises concerns among safety researchers about the difficulty of auditing AI behavior.
Open Source: The Power of Shared Innovation
Open source refers to software, and increasingly AI models, where the underlying code is publicly accessible for use, inspection, and modification. Meta’s Llama models are a notable example in the AI domain, analogous to Linux in operating systems. Open source approaches foster collaborative development, accelerating progress and enabling independent safety audits that are often challenging with closed-source systems. The distinction between open and closed source has become a central debate in the AI industry.
Parallelization: The Art of Doing Many Things at Once
Parallelization is the fundamental principle of performing multiple tasks simultaneously rather than sequentially. In AI, it is crucial for both training and inference. Modern GPUs, designed for thousands of parallel calculations, are the backbone of the industry. As AI systems grow more complex, the ability to distribute work across numerous chips and machines has become a key determinant of model development and deployment speed and cost-effectiveness.
RAMageddon: The Memory Chip Shortage
RAMageddon is a newly coined term describing the escalating shortage of Random Access Memory (RAM) chips, essential components in nearly all modern technology. The burgeoning AI industry’s immense demand for RAM to power data centers has led to a scarcity, driving up prices and impacting various sectors, including gaming, consumer electronics, and general enterprise computing. Projections for the end of this shortage remain uncertain.
Recurrent Depth: The Engineering Term for Opaque Reasoning
"Recurrent depth" is a more technical term for the same underlying mechanism as opaque recurrence: repeatedly processing a query through a model’s layers. Media outlets often use the terms interchangeably, with "recurrent depth" representing the engineering perspective and "opaque recurrence" emphasizing the safety implications.
Recursive Self-Improvement: The Path to Autonomous AI
Recursive self-improvement (RSI) signifies a threshold in AI capability where models begin to enhance themselves without human intervention, leading to a rapid acceleration of their abilities and autonomy. This concept is often associated with the "singularity," a hypothetical point of uncontrollable AI growth. However, RSI also describes the fundamental ability of an AI model to design its own successor, a capability actively pursued by several AI startups. While some view RSI with apprehension, many researchers frame it as the next frontier in AI development.
Reinforcement Learning: Learning Through Rewards
Reinforcement learning is a training methodology where AI systems learn by trial and error, receiving rewards for correct actions. Analogous to training a pet with treats, this process involves an AI model exploring its environment, taking actions, and refining its behavior based on feedback. This approach has proven highly effective for training AI in game playing, robotics, and sharpening the reasoning abilities of large language models, with techniques like Reinforcement Learning from Human Feedback (RLHF) playing a crucial role in developing more helpful, accurate, and safe AI models.
Token: The Building Blocks of AI Communication
Tokens are the fundamental units of data processed or produced by LLMs, bridging the gap between human language and AI computation. Through a process called tokenization, raw text is broken down into digestible segments. In enterprise contexts, tokens also determine cost, with most AI companies charging for LLM usage on a per-token basis.
Token Throughput: Measuring AI Processing Power
Token throughput quantifies the amount of AI work a system can handle within a given period. High token throughput is essential for serving a large number of users simultaneously and ensuring rapid response times. Maximizing token throughput has become a significant objective in AI infrastructure development, reflecting the drive for efficient utilization of computational resources.
Training: The Foundation of AI Learning
Training is the process by which AI models learn from data to identify patterns and generate useful outputs. This involves feeding vast amounts of data into the system, enabling it to adapt its responses toward a desired goal. Training can be resource-intensive, and the increasing volumes of data required have led to the exploration of hybrid approaches, such as fine-tuning pre-trained models, to manage costs effectively.
Transfer Learning: Reusing Prior Knowledge
Transfer learning is a technique where a pre-trained AI model serves as a starting point for developing a new model for a related task. This allows knowledge gained during previous training cycles to be reapplied, accelerating model development and proving beneficial when data for the new task is limited. However, models relying solely on transfer learning may still require additional domain-specific training for optimal performance.
Validation Loss: Gauging AI Learning Progress
Validation loss is a metric that indicates how well an AI model is learning during training, with lower values signifying better performance. Researchers monitor validation loss to assess training progress, adjust hyperparameters, and identify potential issues such as overfitting, where a model memorizes training data rather than generalizing to new situations.
Weights: The Importance of Data Features
Weights are numerical parameters central to AI training, determining the significance of different features within the training data and thereby shaping the model’s output. Initially assigned randomly, weights are adjusted during training as the model strives to produce outputs that align with the target. For instance, in a housing price prediction model, weights would be assigned to features like the number of bedrooms, bathrooms, and the presence of a garage, reflecting their influence on property value based on the training data.
