Essential AI Glossary for 2023 You Can’t Miss

The realm of artificial intelligence is rapidly changing, leading to a unique set of terms that explain its many concepts. At any tech meeting or product conference today, you’re likely to encounter phrases like LLMs, RAG, RLHF, and others that could confuse even experienced professionals. This glossary is designed to demystify these terms, offering clear definitions of the AI jargon you might come across—whether you’re diving into these technologies, investing in them, or simply keeping abreast of insights from platforms like TechCrunch or pertinent podcasts. We will continue to update it as the field evolves, reflecting the dynamic landscape of AI systems.

---

Artificial general intelligence (AGI) remains a somewhat unclear idea, usually denoting AI that surpasses the average human ability across a variety of tasks. OpenAI CEO Sam Altman likened AGI to the “median human you could hire as a colleague.” OpenAI’s charter characterizes AGI as “highly autonomous systems that excel at most economically valuable tasks.” Google DeepMind, on the other hand, describes AGI in a cumulative manner, defining it as “AI that is at least as competent as humans in most cognitive tasks.” If this sounds ambiguous, you’re not alone; many experts in AI research share this viewpoint.

An AI agent refers to a tool that utilizes AI technologies to automate a range of tasks on your behalf, going beyond the functions of a standard AI chatbot. This can include managing expenses, booking travel, or even programming and maintaining software. However, as mentioned earlier, interpretations of an “AI agent” can vary among stakeholders, and the foundational infrastructure is still in development to reach its anticipated capabilities. Essentially, an AI agent is an autonomous system that combines various AI technologies to accomplish intricate, multi-step tasks.

Think of API endpoints as “hidden buttons” within software that can be engaged by other applications to perform tasks. Developers use these interfaces to create integrations—for instance, permitting one app to pull data from another or allowing an AI agent to utilize third-party services autonomously. Many smart home devices and interconnected platforms incorporate these hidden buttons, often without the casual user’s awareness or involvement. As AI agents evolve, they are increasingly adept at autonomously identifying and leveraging these endpoints, revealing powerful—and sometimes unexpected—automation possibilities.

When faced with a straightforward question, humans can often respond instinctively—like “which is taller, a giraffe or a cat?” However, some questions may require breaking down intermediate steps to arrive at the correct answer. For example, if a farmer has a total of 40 heads of chickens and cows combined with 120 legs, one needs to create an equation to determine the number of each (20 chickens and 20 cows).

In the context of AI, chain-of-thought reasoning for large language models involves breaking down a problem into smaller, sequential steps to improve the accuracy of the final answer. This approach usually takes longer to formulate a response but significantly enhances the chances of correctness, especially in logical or programming situations. Reasoning models evolve from traditional large language models and are fine-tuned for chain-of-thought approaches, primarily through reinforcement learning.

(See: Large language model)

This term refers to a more specific concept than an “AI agent,” focusing on applications that can accomplish tasks independently and step-by-step to reach specific goals. A coding agent is a specialized version tailored specifically for software development. Unlike simply suggesting code for human evaluation, a coding agent can autonomously write, test, and debug software, managing the iterative processes typically encountered by developers. These agents can explore entire codebases, identify bugs, run tests, and apply fixes with minimal human involvement. They resemble an ultra-efficient intern who never tires and stays focused—though, like any intern, they still require human oversight.

Though a somewhat broad concept, compute primarily refers to the crucial computational power driving AI models. This power fuels the AI sector, enabling it to train and launch robust models. This term often serves as shorthand for the types of hardware that provide this computational power—like GPUs, CPUs, TPUs, and other infrastructure components that support the contemporary AI industry.

A subtype of self-improving machine learning, deep learning involves constructing AI algorithms within a layered artificial neural network (ANN) framework. This architecture allows for the recognition of more complex patterns than simpler machine learning methods like linear models or decision trees. Deep learning systems take inspiration from the intricately connected pathways of neurons within the human brain.

Deep learning models can autonomously detect significant features in data, removing the need for human engineers to delineate them. This architecture also supports algorithms that learn from errors, continuously refining outputs through iterative adjustments. However, deep learning systems frequently require large amounts of data for optimal performance (often millions of data points) and typically demand longer training periods compared to simpler models—leading to higher development costs.

(See: Neural network)

Diffusion refers to the technology behind many AI models that generate art, music, and text. Taking cues from principles in physics, diffusion systems systematically “destroy” data structures—like images or melodies—by introducing noise until they become unrecognizable. In physics, diffusion occurs spontaneously and irreversibly; for instance, sugar dissolved in coffee cannot revert to its original cube form. Conversely, AI diffusion systems aim to master a “reverse diffusion” technique to reclaim distorted data, effectively learning to uncover information obscured by noise.

Distillation signifies a method for extracting insights from a large AI model using a “teacher-student” framework. Developers query a teacher model and document its outputs, often contrasting these results against a predefined dataset for accuracy. These outputs are subsequently employed to train the student model, which seeks to replicate the teacher’s responses.

This process can yield a smaller, more efficient model that retains the performance of the larger counterpart while minimizing distillation loss. This is likely how OpenAI developed GPT-4 Turbo, a faster version of GPT-4.

While all AI organizations conduct distillation internally, some may also use it to improve their capabilities relative to leading models. However, extracting insights from a competitor generally violates AI API service agreements.

This term pertains to the retraining of an AI model to fine-tune its efficiency for a more specific task or domain than originally intended—typically by incorporating new, specialized (i.e., task-relevant) data.

Many AI startups begin with large language models as their foundation for creating commercial products but aim to enhance functionality for specific sectors or tasks by enriching prior training with domain-specific knowledge.

(See: Large language model [LLM])

A GAN, or Generative Adversarial Network, represents a machine learning framework that has spurred significant advancements in generative AI, particularly in producing realistic outputs—including, but not limited to, deepfake technologies. GANs comprise two neural networks; one generates outputs based on its training data, while the other evaluates them.

These two models compete against one another. The generator strives to produce outputs that can fool the discriminator, while the discriminator’s role is to distinguish authentic data from artificial creations. This structured competition can yield lifelike AI outputs with minimal human involvement. However, GANs are generally more effective in narrow contexts (such as generating realistic images or videos) rather than operating on a broader scale.

Within the AI community, the term “hallucination” describes instances where AI models generate inaccuracies—essentially fabricating false information. This presents a significant challenge for AI accuracy.

Hallucination can lead to misleading outputs from generative AI, posing notable risks—potentially resulting in dangerous outcomes (e.g., unsafe medical advice derived from a health-related inquiry).

The issue of AIs generating fabricated information is believed to stem from gaps in their training data. In response to these hallucinations, there is a growing movement towards developing increasingly specialized and/or vertical AI models—domain-specific AIs that require narrower expertise to close knowledge gaps and lessen misinformation risks.

Inference denotes the execution of an AI model; it involves activating a model to predict or deduce conclusions from previously observed data. Notably, inference cannot occur without training; a model must first recognize patterns within a dataset before effectively extrapolating from that training.

Various hardware setups can perform inference, ranging from consumer-grade processors to powerful GPUs and specialized AI accelerators. However, not all configurations execute models with equal efficiency; very large models would take significantly longer to provide predictions on a laptop compared to a cloud server with premium AI chips.

[See: Training]

Large language models, or LLMs, are the AI frameworks that power popular AI assistants like ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, or Mistral’s Le Chat. When you interact with an AI assistant, you’re engaging with a large language model that processes your inquiries directly or with support from various available tools, like web browsers or code interpreters.

LLMs consist of deep neural networks housing billions of numerical parameters (or weights, see below) that discern relationships between words and phrases to form a representation of language—essentially, a sophisticated map of language.

These models are created by analyzing patterns found in vast amounts of documents, articles, and transcripts. When you submit a prompt to an LLM, the model generates the most probable response associated with that prompt.

(See: Neural network)

Memory cache signifies a crucial process that enhances inference (the method by which AI formulates responses to user inquiries). Essentially, caching functions as an optimization strategy aimed at improving inference efficiency. AI operates through intricate mathematical calculations, and each calculation consumes resources. Caching reduces the computations a model must execute by storing certain calculations for future interactions. Various types of memory caching exist, with KV (key-value) caching being among the most prevalent. KV caching is employed in transformer-based models, improving efficiency and delivering quicker results by minimizing the time (and computational effort) required to produce responses to user prompts.

(See: Inference)

Model Context Protocol (MCP) is an open standard designed to enable AI models to connect to external tools and data—such as your files, databases, or applications like Slack and Google Drive—without requiring developers to build separate connectors for each integration. You can think of it as a USB-C port for AI. Introduced by Anthropic in 2024 and subsequently endorsed by the Linux Foundation, MCP has gained traction among OpenAI, Google, and Microsoft, becoming one of the most rapidly disseminated standards in the current AI landscape.

The Mixture of Experts architecture divides a neural network into multiple smaller specialized sub-networks, or “experts,” activating only a fraction of them for particular tasks. Instead of routing each request through the entire model—like calling your entire team for every question—an MoE model includes a built-in “router” that selects the most appropriate specialists for the task. This design supports the development of exceptionally large models that maintain relative efficiency and cost-effectiveness, as only a segment of the network operates at any one time. Mistral AI’s Mixtral model is a notable example; recent GPT models from OpenAI are thought to incorporate some variation of this methodology, although the company has yet to officially confirm this.

(See: Neural network, Deep learning)

A neural network provides the multi-layered algorithmic bedrock for deep learning—and, more broadly, the surge in generative AI tools following the rise of large language models.

The notion of modeling data processing algorithms after the intricately connected circuits of the human brain has existed since the 1940s; however, it was the recent explosion in graphical processing units (GPUs)—largely driven by the gaming sector—that truly unleashed this theory’s potential. GPUs are specifically designed for training algorithms with multiple layers, making neural network-based AI systems exceptionally skilled in various domains, including voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source refers to software—or increasingly, AI models—where the underlying code is publicly accessible, allowing anyone to use, examine, or modify it. In the AI sphere, Meta’s Llama series of models stands out as a significant example; Linux serves as a well-known analogue among operating systems historically. Open-source approaches enable researchers, developers, and organizations globally to build upon each other’s work, accelerating innovation and allowing independent safety assessments that proprietary systems often can’t easily provide. Closed source signifies that the code stays private; users can engage with the product but cannot inspect its inner workings, a characteristic often associated with OpenAI’s GPT models—this has triggered considerable discussion in the AI sector.

Parallelization entails executing multiple tasks concurrently, rather than sequentially—much like having 10 employees work on different aspects of a project simultaneously instead of one person completing each task in succession. In AI, parallelization is crucial for both training and inference: modern GPUs are specifically designed to handle thousands of calculations at once, which is a foundational element of their functionality within the industry. As AI systems grow in complexity and models become larger, the ability to parallelize tasks across multiple chips and machines has emerged as a key determinant of how quickly and efficiently models can be developed and implemented. Research into improved parallelization methods has become its own dedicated field.

RAMageddon is a playful term referring to a serious trend affecting the tech landscape: the ongoing scarcity of random access memory (RAM) chips, which are essential for virtually all tech products we rely on daily. As the AI sector has expanded, major tech companies and AI labs are acquiring RAM at unprecedented rates to bolster their data centers, leading to diminished availability in the broader market. This supply constraint is driving up RAM prices.

This shortage impacts various sectors, including gaming (with companies needing to raise console prices due to difficulties in sourcing memory chips for their devices), consumer electronics (where memory shortages could result in the most significant decline in smartphone shipments in over a decade), and general enterprise computing (as businesses grapple to secure adequate RAM for their data operations). The rising prices are unlikely to stabilize until that dreaded shortage resolves, but sadly, indications of a swift resolution appear limited.

Similar to AGI, recursive self-improvement describes a potential threshold for AI advancement, allowing minimal human involvement. In this scenario, AI models begin to enhance their capabilities independently, resulting in significant improvements in autonomy and functionality. Some narratives liken this event to a singularity moment—when AI models become immune to outside intervention. However, recursive self-improvement raises a fundamental question: can an AI model create its successor? This inquiry simplifies the task for engineers aiming to develop such models. Many recent AI startups focus on creating recursively self-improving models while often downplaying the potentially disastrous outcomes, framing recursive self-improvement as simply the next horizon in research.

Reinforcement learning is an AI training technique where a system learns through trial and error, receiving rewards for correct actions—similar to training a pet with treats, but in this case, the “pet” is a neural network, and the “treat” is a mathematical signal indicating success. Unlike supervised learning, which involves training a model on a fixed dataset of labeled examples, reinforcement learning allows a model to interact with its environment, undertake actions, and iteratively refine its behavior based on the feedback it acquires. This approach has proven particularly successful for training AI to play games, manage robotics, and, more recently, to enhance the reasoning capabilities of large language models. Techniques like reinforcement learning from human feedback (RLHF) have become critical in how leading AI labs hone their models for improved efficacy, accuracy, and safety.

When communicating in dialogue between humans and machines, noticeable challenges arise—humans converse in natural language, while AI systems operate through complex algorithmic processes grounded in data. Tokens serve as the bridge between these two worlds: they represent the fundamental units of human-AI interaction, encapsulating distinct segments of data processed or produced by an LLM. Tokens are derived through a process called tokenization, which divides raw text into manageable parts that a language model can understand, much like how a compiler converts human language into binary code for computers. In business contexts, tokens also influence costs—most AI companies charge for LLM usage on a per-token basis, implying that greater usage correlates with higher expenses.

Thus, tokens are the small text units—often fragments of words rather than full phrases—that AI language models parse prior to processing; they can be likened to “words” with regards to AI workloads. Throughput signifies the amount that can be processed within a given time, so token throughput acts as a metric for the volume of AI work a system can manage simultaneously. High token throughput is a critical goal for AI infrastructure teams, as it affects how many users a model can serve simultaneously and the speed at which they receive responses. AI researcher Andrej Karpathy has voiced concern when his AI subscriptions lie idle—reflecting on his experiences as a graduate student when costly computing hardware remained underutilized—highlighting why maximizing token throughput has become an urgent aim in the field.

Training machine learning AIs refers to a process known as training. This essentially involves supplying data to enable the model to recognize patterns and generate desirable outputs. It is fundamentally the system’s response to the characteristics of data, allowing it to align outputs toward a desired outcome—whether that entails identifying images of cats or composing a haiku in response to a prompt.

Training can be resource-intensive due to the large volume of inputs required, which has been continually rising—hence the utility of hybrid strategies, such as infusing rules-based AIs with targeted data, to manage expenses without starting from scratch.

[See: Inference]

This technique utilizes a previously trained AI model as a starting point for developing a new model aimed at a different but typically related task, thereby facilitating the reapplication of insights garnered from prior training cycles.

Transfer learning can boost model development efficiency. It is especially beneficial when the available data for the target task is somewhat limited. However, it’s important to understand that this approach has limitations; models employing transfer learning to acquire generalized skills will likely need additional training with fresh data to excel in their specific focus areas.

(See: Fine tuning)

Validation loss represents a benchmark reflecting an AI model’s learning progression during training—lower values are preferred. Researchers closely monitor this metric as a real-time performance indicator, informing decisions about when to stop training, adjust hyperparameters, or investigate potential complications. A significant concern that validation loss helps reveal is overfitting, where a model memorizes its training data rather than genuinely understanding applicable patterns for new situations. Think of validation loss as a tool that distinguishes between a student who grasps the material and one who merely memorized last year’s exam—it reveals the trajectory your model is taking.

Weights are crucial in AI training, determining the significance (or weight) assigned to various features (or input variables) within the training data, ultimately affecting the AI model’s outputs.

In simpler terms, weights are numerical parameters indicating which elements within a dataset are the most critical for the specific training task. They accomplish this effect through multiplication with inputs. Training typically begins with randomly initialized weights, but as the process advances, these weights are fine-tuned as the model seeks to produce outputs that closely align with the target.

For instance, an AI model designed to estimate housing prices based on historical real estate data for a particular area might assign weights to variables such as the number of bedrooms and bathrooms, whether a property is semi-detached or detached, if it includes parking, a garage, and more.

Ultimately, the weights assigned reflect how much each of these factors impacts a property’s value according to the available dataset.

This article is regularly updated with new information.