Essential AI Terminology to Understand in 2023

The realm of artificial intelligence is progressing rapidly, introducing a plethora of terminology to describe its concepts. At modern tech events and product launches, terms like LLMs, RAG, RLHF, and others can confuse even seasoned professionals. This glossary aims to clarify these terms by providing straightforward definitions of AI-related language you may come across—whether you’re exploring these technologies, investing in them, or keeping up-to-date through platforms like TechCrunch or relevant podcasts. We’ll make sure it evolves along with the dynamic AI landscape.

---

Artificial general intelligence (AGI) is a sophisticated idea generally pertaining to AIs that exceed human capabilities across a wide range of tasks. OpenAI’s CEO, Sam Altman, likens AGI to “the average employee you might hire.” OpenAI defines AGI as “highly autonomous systems that outperform in most economically valuable tasks.” Conversely, Google DeepMind describes it as “AI capable of performing most cognitive tasks at a human level.” If this explanation feels ambiguous, you’re not alone; many in the AI research field share similar views.

An AI agent is a system that utilizes AI technologies to automate a variety of functions, far exceeding standard chatbot tasks. Duties can range from managing expenses and planning travel to programming and maintaining software. As previously noted, interpretations of “AI agent” differ among stakeholders, and its foundational framework continuously adapts to meet rising expectations. In essence, an AI agent serves as an autonomous system that integrates multiple AI technologies to tackle intricate, multi-step tasks.

Think of API endpoints as “invisible buttons” in software that can be triggered by other applications to perform actions. Developers utilize these interfaces to enable integrations—like allowing one application to pull data from another or enabling an AI agent to operate independently with third-party services. Many smart home devices and interconnected systems include these subtle controls, often going unnoticed by most users. As AI agents advance, their ability to discover and utilize these endpoints grows, unveiling remarkable—and sometimes unexpected—automation capabilities.

When faced with simple questions, humans typically respond instinctively—like asking, “Which is taller, a giraffe or a cat?” However, some queries require breaking down intermediate steps to arrive at correct answers. For instance, if a farmer has 40 animals (chickens and cows) with a total of 120 legs, calculations must be made to determine the number of each type of animal (20 chickens and 20 cows).

In the context of AI, the chain-of-thought reasoning used by large language models involves deconstructing problems into smaller, sequential steps to increase the accuracy of the final solution. While this method may be slower, it greatly enhances correctness, especially in logical or programming contexts. Reasoning models evolve from traditional large language models and are fine-tuned for chain-of-thought tasks primarily through reinforcement learning.

(See: Large language model)

This term is more specific than “AI agent,” focusing on applications that can independently execute tasks in sequence to reach defined goals. A coding agent is a specific type concentrating exclusively on software development. Rather than just generating code for human review, a coding agent can autonomously write, test, and debug software, addressing the iterative challenges developers often encounter. These agents can navigate entire codebases, spot bugs, run tests, and apply fixes with minimal human oversight, functioning like a dedicated intern—diligent and tireless, yet still requiring human supervision.

While somewhat broad, compute primarily refers to the computational power fueling AI models. This capability is crucial to the AI industry, enabling the training and deployment of advanced models. It generally relates to the types of hardware that provide this computational strength—such as GPUs, CPUs, TPUs, and other essential infrastructure vital to today’s AI ecosystem.

Deep learning, a subset of self-improving machine learning, centers on developing AI algorithms with a layered artificial neural network (ANN) structure. This approach allows for the detection of more intricate patterns compared to traditional machine learning methods like linear models or decision trees. Deep learning systems draw inspiration from the complex pathways of neurons in the human brain.

Deep learning models autonomously recognize significant features in data, eliminating the need for human engineers to specify them. This architecture also enables algorithms to learn from mistakes, progressively improving outputs through iterative adjustments. However, deep learning systems typically require vast datasets for optimal performance (often millions of data points) and usually involve longer training times than simpler models—resulting in higher development costs.

(See: Neural network)

Diffusion refers to techniques used by various AI models to create art, music, and text. Grounded in physics, diffusion systems systematically “degrade” data structures—like images or melodies—by introducing noise until they become indistinct. In physics, diffusion happens naturally and irreversibly; for example, sugar dissolved in coffee cannot revert to its original form. However, AI diffusion systems aim to master a “reverse diffusion” process to recover distorted data, learning to extract information obscured by noise.

Distillation describes a method for extracting insights from a large AI model using a “teacher-student” paradigm. Developers interact with a teacher model and document its outputs, often benchmarking these results against a predefined dataset for accuracy. These outputs are then utilized to train a student model, intended to replicate the teacher’s responses.

This method can produce a more compact, efficient model that retains the capabilities of its larger counterpart while minimizing distillation loss. It is widely believed that this technique contributed to the development of GPT-4 Turbo, a faster variant of GPT-4 from OpenAI.

While all AI organizations conduct distillation internally, some may use it to enhance their features relative to existing models. However, extracting insights from a competitor could violate AI API service agreements.

This term refers to the retraining of an AI model to enhance its performance for a more specific task or domain than it was originally designed for—usually through the incorporation of new data relevant to the task.

Many AI startups commence with large language models as the foundation for their products, aiming to refine functionality for particular sectors or tasks by integrating domain-specific knowledge into existing training.

(See: Large language model [LLM])

A GAN, or Generative Adversarial Network, is a machine learning framework that has driven generative AI, particularly in producing realistic outputs—including deepfake technologies. GANs consist of two neural networks; one generates outputs based on its training, while the other assesses them.

These two networks engage in a competitive dynamic; the generator aims to create outputs that can deceive the discriminator, while the discriminator works to distinguish real data from fabricated content. This structured competition can result in realistic AI-generated outputs with minimal human intervention. However, GANs typically excel only under specific conditions (such as creating authentic images or videos), rather than universally.

In the AI community, “hallucination” refers to instances where AI models generate inaccuracies—essentially fabricating incorrect information. This presents a significant obstacle to AI accuracy.

Hallucinations can lead to misleading outputs from generative AI, carrying serious risks—potentially resulting in harmful outcomes (for example, unsafe medical advice stemming from a health inquiry).

The phenomenon of AIs producing fictitious data is thought to arise from gaps in their training information. Consequently, there is a growing demand for more specialized and/or vertical AI models—domain-specific AIs that need narrower expertise to bridge knowledge gaps and reduce the risks of misinformation.

Inference signifies the functioning of an AI model, engaging it to predict or deduce conclusions based on previously observed data. Crucially, inference cannot occur without prior training; a model must first learn patterns within a dataset before it can effectively extrapolate from that learning.

Various hardware configurations can support inference, ranging from consumer-grade processors to powerful GPUs and specialized AI accelerators. However, not all setups process models with the same efficiency; for instance, exceedingly large models would take significantly longer to produce predictions on a laptop compared to a cloud server outfitted with advanced AI chips.

[See: Training]

Large language models, or LLMs, underpin popular AI assistants, including ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, and Mistral’s Le Chat. When you engage with an AI assistant, you’re utilizing a large language model that addresses your queries or employs various available tools, such as web browsers or code interpreters.

LLMs consist of deep neural networks with billions of numerical parameters (or weights) that discern relationships among words and phrases to create a representation of language—essentially, a sophisticated mapping of language.

These models undergo training by examining patterns identified in extensive datasets of documents, articles, and transcripts. When you give a prompt to an LLM, the model formulates the most probable response related to that input.

(See: Neural network)

Memory cache is an essential process that enhances inference (how AI formulates responses to user prompts). Essentially, caching acts as an optimization strategy aimed at improving inference performance. AI operations involve complex mathematical computations, each consuming resources. Caching diminishes the number of computations a model must perform by retaining specific calculations for future use. Various types of memory caching exist, with KV (key-value) caching being one of the most frequently used. KV caching is applied in transformer-based models, enhancing efficiency and yielding quicker results by reducing the time (and computational effort) required to respond to user queries.

(See: Inference)

Model Context Protocol (MCP) is an open standard introduced to streamline interactions between AI models and external tools and data—such as your files, databases, or applications like Slack and Google Drive—without necessitating developers to create custom connectors for every integration. Think of it as a USB-C port for AI. Launched by Anthropic in 2024 and later endorsed by the Linux Foundation, MCP has quickly gained traction among OpenAI, Google, and Microsoft, becoming one of the most widely adopted standards in today’s AI ecosystem.

The Mixture of Experts architecture divides a neural network into multiple smaller, specialized sub-networks, or “experts,” activating only a chosen subset of them for specific tasks. Rather than routing every request through the entire model—akin to consulting an entire team for each task—the MoE model utilizes an internal “router” that selects the most relevant specialists for the job. This design allows for the creation of exceptionally large models while maintaining efficiency and cost-effectiveness, as only a limited portion of the network operates at any given time. Mistral AI’s Mixtral model is a notable example; recent GPT models from OpenAI are rumored to employ a variation of this methodology, although the company has yet to publicly confirm this.

(See: Neural network, Deep learning)

A neural network serves as the multilayered algorithmic foundation for deep learning—and, more broadly, for the evolution of generative AI tools following the introductions of large language models.

The concept of modeling data-processing algorithms after the intricately interconnected circuits of the human brain has existed since the 1940s; however, it was the recent influx of graphical processing units (GPUs)—largely driven by the gaming sector—that truly unlocked this theory’s potential. GPUs are specifically tailored for training algorithms with multiple layers, enabling neural network-based AI systems to excel in various fields, including voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source pertains to software—or increasingly, AI models—where the source code is publicly available, permitting anyone to use, scrutinize, or alter it. In the AI domain, Meta’s Llama series of models stands out as a notable example, while Linux is acknowledged as a benchmark among operating systems. Open-source practices empower researchers, developers, and organizations worldwide to benefit from one another’s contributions, fostering innovation while allowing independent safety assessments often lacking in closed systems. Closed source indicates that the code remains proprietary; users can interact with the product but cannot explore its internal mechanisms, a trait often associated with OpenAI’s GPT models—which has ignited much debate within the AI community.

Parallelization involves executing multiple tasks at the same time rather than one after another—similar to engaging ten individuals to address different aspects of a project simultaneously, instead of having one person complete each task in succession. In AI, parallelization is essential for both training and inference; modern GPUs are specifically engineered to manage thousands of calculations concurrently, being the backbone of their role in the field. As AI systems grow more complex and models expand, the capacity to parallelize tasks across multiple chips and machines has become critical in influencing the speed and efficiency of model development and deployment. Research into improved parallelization techniques has emerged as a specialized area of study.

RAMageddon is a playful term reflecting a significant trend affecting the tech industry: the ongoing shortage of random access memory (RAM) chips, which are crucial for nearly all technology products. With the AI sector booming, major tech companies and AI laboratories have been acquiring RAM on an unprecedented scale to enhance their data centers, leading to reduced availability in the broader market. This supply crunch is driving up RAM prices.

This shortage impacts various sectors, such as gaming (where companies must hike console prices due to challenges in sourcing memory chips for their devices), consumer electronics (where memory constraints could lead to the lowest smartphone shipments in over a decade), and overall corporate computing (as businesses strive to secure sufficient RAM for their data operations). Price increases are unlikely to stabilize until this anticipated shortfall is resolved, although regrettably, signs of a swift resolution are scant.

Similar to AGI, recursive self-improvement suggests a potential breakthrough in AI development, enabling minimal human oversight. In this scenario, AI models begin to autonomously enhance their capabilities, leading to significant advancements in autonomy and performance. Some narratives liken this milestone to a singularity moment—when AI models become impervious to external influence. However, recursive self-improvement raises a critical question: can an AI model create its successor? This dilemma simplifies the challenges engineers face in developing such systems. Numerous emerging AI startups are aiming to create models capable of recursive self-improvement while often downplaying the potential catastrophic consequences, presenting recursive self-improvement simply as the next frontier of research.

Reinforcement learning is a training strategy for AI systems where a model learns via trial and error, receiving rewards for correct actions—similar to training a pet with treats, although here, the “pet” is a neural network, and the “treat” is a mathematical signal indicating success. In contrast to supervised learning, which involves training a model on a specific dataset of labeled examples, reinforcement learning allows a model to engage with its environment, take actions, and iteratively refine its behavior based on feedback received. This technique has proven particularly effective for training AI in games, robotics control, and, more recently, improving reasoning capabilities for large language models. Approaches like reinforcement learning from human feedback (RLHF) have become essential in how leading AI labs fine-tune their models for superior performance, accuracy, and safety.

In the interaction between humans and machines, notable challenges arise—humans communicate in natural language while AI systems operate through intricate algorithmic processes grounded in data. Tokens serve as the bridge connecting these two worlds: they represent the basic units of human-AI interaction, embodying discrete segments of data processed or generated by an LLM. Tokens are formed through tokenization, which divides raw text into manageable parts that a language model can interpret, akin to how a compiler transforms human language into binary for computers. In commercial contexts, tokens also have cost implications—many AI firms charge for LLM usage on a per-token basis, meaning that greater usage leads to higher expenses.

Thus, tokens are small text units—typically fragments of words instead of complete phrases—that AI language models analyze prior to processing; they can be compared to “words” in the context of AI workloads. Throughput indicates the volume that can be processed within a certain timeframe, making token throughput a measure of how much AI work a system can handle simultaneously. High token throughput is a crucial goal for AI infrastructure teams, as it affects how many users a model can support at once and the speed at which they receive responses. AI researcher Andrej Karpathy has expressed concerns regarding his AI subscriptions being underutilized—reflecting on his experiences as a graduate student when expensive computing hardware was inefficiently used—highlighting why optimizing token throughput has become an urgent focus in the field.

Training machine learning AIs refers to a process known as training. This involves supplying data to the model to assist it in recognizing patterns and producing desired outputs. It signifies the system’s response to the attributes of the data, aligning its outputs toward a specific objective—whether that means identifying images of cats or composing a haiku in response to a prompt.

Training can require substantial resources due to the vast quantities of input needed, which have been consistently increasing—especially with the significance of hybrid strategies, such as combining rules-based AIs with targeted data, to manage costs without starting the training process from scratch.

[See: Inference]

This technique employs a previously trained AI model as a foundation for developing a new model targeting a different but often related task, enabling the reapplication of insights gained from earlier training cycles.

Transfer learning can enhance the efficiency of model development. It is particularly advantageous when the available data for the target task is limited. However, it’s essential to note that this method also has limitations; models applying transfer learning to acquire general capabilities will likely need additional training with new data to excel in their specific focus areas.

(See: Fine-tuning)

Validation loss serves as a metric indicating an AI model’s learning progress during training—ideally, lower values are more desirable. Researchers monitor this measure closely as a real-time performance indicator, guiding decisions on when to conclude training, adjust hyperparameters, or investigate potential issues. A primary concern that validation loss helps identify is overfitting, where a model memorizes its training data rather than genuinely understanding applicable patterns for new situations. Consider validation loss as a tool that differentiates between a student who truly comprehends the material and one who merely memorizes past exam content—it encapsulates your model’s developmental journey.

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

In essence, weights are numerical parameters signifying which elements within a dataset hold the most importance for the specified training task. They achieve this through multiplication with inputs. Training typically starts with randomly assigned weights, but as the process advances, these weights are adjusted to produce outputs that closely align with desired results.

For example, an AI model designed to forecast housing prices based on past real estate data for a specific area might assign weights to factors like the number of bedrooms and bathrooms, whether a property is semi-detached or detached, and if it includes parking or a garage.

Ultimately, the assigned weights reflect the significance of each of these factors in determining a property’s value as per the available dataset.

This article is regularly updated with new information.