Essential AI Terminology to Understand in 2023

The field of artificial intelligence is rapidly evolving, giving rise to a plethora of new terms that aim to clarify its concepts. At modern tech conferences and product unveilings, acronyms such as LLMs, RAG, RLHF, and others may confuse even seasoned experts. This glossary is designed to simplify these terms by providing straightforward definitions for the AI vocabulary you may encounter—whether you’re exploring these technologies, investing in them, or keeping up with news from sources like TechCrunch or pertinent podcasts. We will ensure it adapts to the swiftly changing AI landscape.

---

Artificial general intelligence (AGI) represents a sophisticated concept typically describing AIs that can outperform human capabilities across a variety of tasks. OpenAI’s CEO, Sam Altman, likens AGI to “the average employee you might hire.” OpenAI defines AGI as “highly autonomous systems that excel in most economically valuable tasks,” while Google DeepMind describes it as “AI that can perform most cognitive tasks at a human level.” If this definition seems unclear, you’re not alone; many within the AI research community share similar views.

An AI agent refers to a system utilizing AI technologies to perform a wide array of tasks, vastly surpassing the capabilities of a basic chatbot. Duties can include budget management, travel arrangement, and coding, among other tasks. As mentioned earlier, interpretations of “AI agent” differ among stakeholders, and its underlying framework continually evolves to meet growing expectations. In essence, an AI agent functions as an autonomous system integrating various AI technologies to tackle complex, multi-step tasks.

Think of API endpoints as “invisible buttons” in software that can be activated by other applications to perform different functions. Developers use these interfaces for seamless integrations—allowing one application to extract data from another or enabling an AI agent to operate independently with third-party services. Numerous smart home devices and interconnected systems incorporate these subtle controls, often overlooked by the average user. As AI agents advance, their ability to discover and make use of these endpoints will expand, revealing remarkable—and sometimes surprising—automation possibilities.

When faced with straightforward questions, humans usually respond instinctively—like asking, “Which is taller, a giraffe or a cat?” However, certain queries require breaking down intermediate steps to arrive at accurate answers. For instance, if a farmer has 40 animals (chickens and cows) that total 120 legs, one must perform calculations to identify the type of each animal (20 chickens and 20 cows).

In AI, chain-of-thought reasoning employed by large language models involves dissecting issues into smaller, sequential steps to improve the accuracy of the final answer. While this strategy may take more time, it significantly enhances correctness, especially in logical or programming contexts. Reasoning models evolve from standard large language models and are fine-tuned for chain-of-thought tasks mainly through reinforcement learning.

(See: Large language model)

This term is more specific than “AI agent,” focusing on applications that can autonomously execute tasks in a sequence to achieve defined objectives. A coding agent is a specialized type of AI primarily aimed at software development. Rather than simply generating code for human review, a coding agent can independently write, test, and debug software, addressing the iterative challenges faced by developers. These agents can traverse entire codebases, spot bugs, execute tests, and implement fixes with minimal human intervention, operating much like diligent interns—industrious and persistent, yet still requiring human guidance.

While somewhat broad, compute generally pertains to the computational power that supports AI models. This capacity is crucial in the AI arena, enabling the training and execution of complex models. It typically relates to the types of hardware that provide this computational strength—such as GPUs, CPUs, TPUs, and other infrastructures vital to today’s AI ecosystem.

Deep learning, a branch of self-improving machine learning, concentrates on crafting AI algorithms using a layered artificial neural network (ANN) architecture. This methodology allows for the identification of more complex patterns compared to traditional training methods like linear models or decision trees. Deep learning systems are inspired by the intricate pathways of neurons in the human brain.

Deep learning models can autonomously discern significant features within data, reducing the need for human engineers to identify them. This framework also empowers algorithms to learn from mistakes, continually refining outputs through iterative adjustments. Nonetheless, deep learning systems often require extensive datasets to function optimally (often millions of data points) and generally entail longer training times than simpler models—resulting in higher development costs.

(See: Neural network)

Diffusion refers to techniques employed by various AI models to produce art, music, and text. Stemming from physics, diffusion systems systematically “degrade” data structures—like images or sounds—by adding noise until they lose clarity. In physics, diffusion occurs naturally and irreversibly; think of how sugar dissolves in coffee, making it impossible to return to its unaltered state. However, AI diffusion systems seek to master a “reverse diffusion” technique to reverse the degradation of data, learning to retrieve information obscured by noise.

Distillation defines a process for extracting insights from a large AI model utilizing a “teacher-student” framework. Developers interact with a teacher model and meticulously document its outputs, often benchmarking these against a predefined dataset for accuracy. These results are then used to train a student model that emulates the teacher’s responses.

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

While all AI organizations perform distillation internally, some might utilize it to enhance their features compared to existing models. However, extracting insights from a competitor could violate AI API service agreements.

This term denotes the retraining of an AI model to enhance its performance for a more specific task or domain than it was originally developed for—typically achieved by incorporating additional data pertinent to the task.

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

(See: Large language model [LLM])

A GAN, or Generative Adversarial Network, is a machine learning framework that has been crucial in advancing generative AI, particularly for creating realistic outputs—including deepfakes. GANs comprise two neural networks: one produces outputs based on its training, while the other evaluates them.

These two networks maintain a competitive relationship; the generator aims to create outputs that can mislead the discriminator, while the discriminator seeks to identify real data from generated content. This structured rivalry can produce lifelike AI-generated outputs with minimal human intervention. However, GANs generally excel only under specific conditions (for instance, when generating authentic images or videos) rather than universally.

Within the AI community, “hallucination” refers to instances where AI models generate inaccuracies—essentially producing false information. This presents a significant challenge to AI accuracy.

Hallucinations may result in misleading outputs from generative AI, posing serious risks—potentially leading to dangerous consequences (such as unsafe medical advice following a health inquiry).

The phenomenon of AIs generating fictitious data is believed to stem from gaps in their training data. Consequently, there is a growing demand for more specialized and/or vertical AI models—domain-specific AIs requiring narrower expertise to bridge knowledge gaps and mitigate misinformation risks.

Inference refers to the operation of an AI model, enabling it to make predictions or draw conclusions based on previously observed data. Importantly, inference cannot occur without prior training; a model must first recognize patterns within a dataset before it can effectively extrapolate from that learning.

Various hardware configurations can support inference, ranging from consumer-grade processors to high-end GPUs and specialized AI accelerators. However, not all setups process models with the same efficiency; for example, exceptionally large models would take significantly longer to produce predictions on a laptop compared to a cloud server equipped with state-of-the-art AI chips.

[See: Training]

Large language models, or LLMs, are the backbone for widely used AI assistants, including ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, and Mistral’s Le Chat. When you interact with an AI assistant, you’re engaging with a large language model that responds to your queries or utilizes various available tools, such as web browsers or code interpreters.

LLMs consist of deep neural networks featuring billions of numerical parameters (or weights) that interpret relationships among words and phrases to construct representations of language—essentially, sophisticated mappings of language.

These models are trained by analyzing patterns discovered within extensive datasets of documents, articles, and transcripts. When you prompt an LLM, the model produces the most likely response relevant to that input.

(See: Neural network)

Memory cache is a vital process that enhances inference (how AI formulates responses to user prompts). Essentially, caching serves as an optimization strategy aimed at improving inference performance. AI operations require intricate mathematical calculations, each taxing resources. Caching reduces the number of computations a model must carry out by retaining certain calculations for future use. Various forms of memory caching exist, with KV (key-value) caching being one of the most prevalent. KV caching is utilized in transformer-based models, increasing efficiency and yielding faster results by minimizing the time (and computational effort) needed to respond to user inquiries.

(See: Inference)

Model Context Protocol (MCP) represents an open standard designed to streamline interactions between AI models and external tools and data—such as your files, databases, or applications like Slack and Google Drive—without requiring developers to create custom connectors for each integration. Picture it as a USB-C port for AI. Introduced by Anthropic in 2024 and later adopted by the Linux Foundation, MCP has quickly gained traction among OpenAI, Google, and Microsoft, evolving into one of the most recognized standards in the present AI ecosystem.

The Mixture of Experts architecture divides a neural network into multiple smaller, specialized sub-networks, or “experts,” activating only a selected subset for particular tasks. Rather than processing every request through the entire model—similar to consulting a full team for each task—the MoE model employs an internal “router” to select the most relevant specialists for a specific job. This design allows for the creation of exceedingly large models while retaining efficiency and cost-effectiveness since only a limited portion of the network operates at any given time. Mistral AI’s Mixtral model showcases this; recent GPT models from OpenAI are rumored to implement a variation of this method, although the company has not publicly confirmed this.

(See: Neural network, Deep learning)

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

The concept of modeling data-processing algorithms after the intricately interconnected circuits of the human brain has been around since the 1940s; however, the recent rise of graphical processing units (GPUs)—largely driven by the gaming industry—has truly unlocked the potential of this theory. GPUs are specifically designed for training algorithms with multiple layers, enabling AI systems based on neural networks to excel in diverse fields, including voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source refers to software—or increasingly, AI models—where the source code is publicly accessible, allowing anyone to use, examine, or modify it. In the AI realm, Meta’s Llama series of models stands out as a significant example, while Linux is recognized as a benchmark among operating systems. Open-source methodologies empower researchers, developers, and organizations globally to reciprocate each other’s contributions, fostering innovation and enabling independent safety evaluations often absent in closed systems. Closed source indicates that the code remains proprietary; users can utilize the product but cannot inspect its internal workings, a characteristic frequently linked to OpenAI’s GPT models—which has sparked considerable debate within the AI community.

Parallelization involves executing multiple tasks simultaneously rather than one after the other—similar to engaging ten individuals to tackle different aspects of a project concurrently, rather than one person completing each task in sequence. In AI, parallelization is essential for both training and inference; modern GPUs are specifically engineered to handle thousands of calculations at once, which is foundational to their role in the field. As AI systems become more complex and models expand, the ability to parallelize tasks across multiple chips and machines has become crucial in defining the speed and efficiency of model development and deployment. Research into improved parallelization techniques has emerged as a specialized area of inquiry.

RAMageddon is a lighthearted term encapsulating a prominent trend impacting the tech industry: the ongoing shortage of random access memory (RAM) chips, which are integral to nearly all technology products. As the AI landscape surges, major tech firms and AI labs have been acquiring RAM at an unparalleled rate to strengthen their data centers, resulting in diminished availability in the broader market. This supply crunch is driving up RAM prices.

This shortage impacts various sectors, such as gaming (where companies need to raise console prices due to difficulties 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 adequate RAM for their data operations). Price increases are unlikely to stabilize until this anticipated shortage resolves; however, regrettably, no quick resolution is in sight.

Similar to AGI, recursive self-improvement hints at a potential breakthrough in AI development, allowing minimal human oversight. In this context, AI models begin to autonomously enhance their capabilities, leading to substantial advancements in autonomy and performance. Some narratives liken this milestone to a singularity moment—when AI models become unaffected by external influence. However, recursive self-improvement presents a critical question: can an AI model create its successor? This dilemma underscores the challenges engineers face when developing such systems. Numerous emerging AI startups are striving to create models capable of recursive self-improvement while often downplaying potential catastrophic consequences, framing recursive self-improvement merely as the next frontier of research.

Reinforcement learning is a training strategy for AI systems where a model learns through trial and error, receiving rewards for favorable actions—similar to training a pet with treats, where the “pet” is a neural network, and the “treat” is a mathematical signal indicating success. Conversely, supervised learning involves training a model on a specific dataset of labeled examples, while reinforcement learning enables a model to interact with its environment, perform actions, and iteratively refine its behavior based on the feedback it receives. This approach has proven particularly effective in training AI for games, robotics, and, more recently, enhancing reasoning capabilities in large language models. Techniques such as reinforcement learning from human feedback (RLHF) have become integral to how leading AI labs refine their models for improved performance, accuracy, and safety.

In the interaction between humans and machines, considerable challenges arise—humans converse in natural language, while AI systems function through complex algorithmic processes grounded in data. Tokens serve as the bridge uniting these two realms: they represent the basic units of human-AI interaction, embodying discrete segments of data processed or generated by an LLM. Tokens are created through tokenization, which disassembles raw text into manageable parts that a language model can comprehend, much like a compiler translates human language into binary for computers. In commercial contexts, tokens also carry cost implications—many AI companies charge for LLM usage on a per-token basis, meaning greater consumption leads to higher expenses.

Thus, tokens are small textual units—often fragments of words instead of entire phrases—that AI language models analyze prior to processing; they can be likened to “words” in AI tasks. Throughput indicates the quantity that can be processed within a specific timeframe, with token throughput serving as a measure of the extent of AI work a system can manage simultaneously. High token throughput is a vital goal for AI infrastructure teams, as it affects how many users a model can support concurrently and the speed at which they receive responses. AI researcher Andrej Karpathy has raised concerns about his AI subscriptions being underutilized—reflecting on his experiences as a graduate student when costly computing hardware was ineffectively employed—emphasizing why optimizing token throughput has become an urgent focus in the industry.

Training machine learning AIs refers to a process known as training. This involves providing data to the model to enable it to identify patterns and generate desired outputs. It signifies the system’s response to the attributes of the data, guiding its outputs toward a specific objective—whether that includes classifying images of cats or crafting a haiku in reaction to a prompt.

Training can demand substantial resources due to the vast amounts of input required, which have been continually increasing—especially with the notion of hybrid strategies, such as merging rule-based AIs with curated data, to manage expenses without initiating the training process anew.

[See: Inference]

This approach utilizes a previously trained AI model as a foundation for generating a new model aimed at a different but often related task, facilitating the reapplication of insights gained from earlier training cycles.

Transfer learning can enhance model development efficiency. It is particularly advantageous when the data available for the target task is limited. Nevertheless, it’s important to acknowledge that this method also has drawbacks; models using transfer learning to gain general capabilities likely require further training with new data to excel in their specific focus areas.

(See: Fine-tuning)

Validation loss serves as a metric that reflects the learning trajectory of an AI model during training—ideally, lower values are preferable. Researchers closely monitor this metric as a real-time performance gauge, informing decisions on when to stop training, alter hyperparameters, or investigate potential issues. A primary concern that validation loss helps identify is overfitting, where a model memorizes its training data instead of genuinely understanding applicable patterns for new contexts. Consider validation loss as a tool that differentiates between a student who comprehends the material and one who merely memorizes prior exam content—it encapsulates your model’s developmental path.

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

In essence, weights are numerical parameters indicating which elements within a dataset hold the highest relevance for the specified training task. They achieve this through multiplication with inputs. Training usually starts with randomly assigned weights, but as the process progresses, these weights are refined to yield outputs closely aligned with the desired outcomes.

For example, an AI model designed to estimate housing prices based on historical real estate data for a particular area might assign weights to factors such as 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 importance of each of these factors in determining a property’s value according to the available dataset.

This article is regularly updated with new information.