Essential AI Terminology to Understand in 2023

The realm of artificial intelligence is rapidly evolving, introducing a host of new terminology to clarify its principles. At modern tech conferences and product launches, expressions like LLMs, RAG, RLHF, and others can bewilder even seasoned professionals. This glossary aims to clarify these terms by providing straightforward definitions for the AI-related jargon you may encounter—be it as you explore these technologies, invest in them, or keep up-to-date with sources like TechCrunch or pertinent podcasts. We’ll ensure it adapts alongside the swiftly transforming AI landscape.

---

Artificial general intelligence (AGI) is a nuanced concept generally describing AIs that exceed human capabilities across a variety of tasks. OpenAI’s CEO, Sam Altman, likens AGI to “the average employee you might hire.” According to OpenAI, AGI is “highly autonomous systems that excel in most economically valuable tasks.” In contrast, Google DeepMind describes it as “AI capable of performing most cognitive tasks at a human level.” If this definition strikes you as ambiguous, you’re not alone; many in the AI research field share similar feelings.

An AI agent refers to a system that utilizes AI technologies to perform a wide range of tasks, far surpassing simple chatbot functions. Duties may include budget management, travel organization, and coding, among others. As previously mentioned, interpretations of “AI agent” differ among stakeholders, and its underlying structure continually adapts to meet higher expectations. Essentially, an AI agent functions as an autonomous system that integrates multiple AI technologies to tackle complex, multi-step tasks.

Consider API endpoints as “invisible buttons” in software that can be activated by other applications to perform various functions. Developers use these interfaces for seamless integrations—such as enabling one application to retrieve data from another or allowing an AI agent to operate independently with third-party services. Many smart home devices and interconnected systems incorporate these subtle controls, often unnoticed by average users. As AI agents mature, their ability to discover and utilize these endpoints will expand, revealing remarkable—and occasionally unexpected—automation capabilities.

When faced with straightforward questions, humans usually respond intuitively—such as asking, “Which is taller, a giraffe or a cat?” However, some questions require breaking down intermediate steps to obtain accurate answers. For instance, if a farmer has 40 animals (chickens and cows) with a total of 120 legs, calculations must be performed to determine the number of each type of animal (20 chickens and 20 cows).

In AI, the chain-of-thought reasoning employed by large language models involves breaking problems down into smaller, sequential steps to enhance the accuracy of the final answer. Although this method may be lengthier, it significantly improves 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 autonomously perform tasks in a sequence to achieve defined goals. A coding agent is a particular kind of AI that concentrates solely on software development. Instead of merely generating code for human evaluation, a coding agent can independently write, test, and debug software, tackling the iterative challenges faced by developers. These agents can navigate entire codebases, identify bugs, run tests, and implement fixes with minimal human intervention, acting like diligent interns—industrious and relentless, yet still requiring human guidance.

Although somewhat broad, compute generally pertains to the computational power that underpins AI models. This capability is essential to the AI field, enabling the training and deployment of advanced models. It usually relates to the types of hardware providing this computational capacity—such as GPUs, CPUs, TPUs, and other fundamental infrastructures crucial to today’s AI ecosystem.

Deep learning, a subdivision of self-improving machine learning, centers around creating AI algorithms using a layered artificial neural network (ANN) structure. This method allows for the detection of more sophisticated patterns compared to traditional training methods like linear models or decision trees. Deep learning systems take inspiration from the complex pathways of human brain neurons.

Deep learning models autonomously recognize significant features within data, reducing the need for human engineers to identify them. This architecture also allows algorithms to learn from mistakes, continuously refining outputs through iterative adjustments. However, deep learning systems often require extensive datasets to function optimally (often millions of data points) and generally involve longer training periods compared to simpler models—resulting in increased development costs.

(See: Neural network)

Diffusion refers to techniques various AI models utilize to generate art, music, and text. Grounded in physics, diffusion systems systematically “degrade” data structures—like images or tunes—by incorporating noise until they lose clarity. In physics, diffusion occurs naturally and irreversibly; for instance, once sugar dissolves in coffee, it cannot revert to its initial form. Nonetheless, AI diffusion systems aim to perfect a “reverse diffusion” method to recover distorted data, learning to retrieve information hidden by noise.

Distillation describes a technique for deriving insights from a large AI model using a “teacher-student” paradigm. Developers interact with a teacher model and document its outputs, often benchmarking these against a predefined dataset for accuracy. These outcomes are then used to train a student model aimed at mimicking the teacher’s responses.

This approach can result in a more compact, efficient model that retains the capabilities of its larger counterpart while minimizing distillation loss. It’s widely believed that this method contributed to the development of GPT-4 Turbo, a faster variation of GPT-4 from OpenAI.

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

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

Numerous AI startups start 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 been instrumental in advancing generative AI, particularly for producing realistic outputs—including deepfakes. GANs consist of two neural networks; one generates outputs based on its training, while the other evaluates them.

These two networks engage in a competitive relationship; the generator strives to create outputs that can fool the discriminator, while the discriminator attempts to distinguish real data from fabricated content. This structured competition can lead to lifelike AI-generated outputs with minimal human intervention. However, GANs tend to excel only under certain conditions (for instance, when generating authentic images or videos) rather than universally.

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

Hallucinations can result in misleading outputs from generative AI, carrying serious risks—potentially leading to harmful outcomes (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 information. Hence, there is an increasing demand for more specialized and/or vertical AI models—domain-specific AIs that require narrower expertise to bridge knowledge gaps and reduce the risks of misinformation.

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

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 equal efficiency; for instance, exceptionally large models would take significantly longer to generate predictions on a laptop compared to a cloud server equipped with cutting-edge AI chips.

[See: Training]

Large language models, or LLMs, serve as the backbone for widely utilized 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 applying a large language model that responds to your inquiries or utilizes various available tools, such as web browsers or code interpreters.

LLMs comprise 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 are trained by analyzing patterns identified in extensive datasets of documents, articles, and transcripts. When you prompt an LLM, the model formulates the most probable response related to that input.

(See: Neural network)

Memory cache is a vital process that improves inference (how AI formulates responses to user prompts). Essentially, caching serves as an optimization strategy aimed at enhancing inference performance. AI operations involve complex mathematical calculations, each consuming resources. Caching reduces the number of computations a model must perform by retaining specific calculations for future use. Various forms of memory caching exist, with KV (key-value) caching being one of the most widely used. KV caching is employed in transformer-based models, enhancing efficiency and delivering faster results by minimizing the time (and computational effort) necessary to respond to user inquiries.

(See: Inference)

Model Context Protocol (MCP) is an open standard created 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 construct custom connectors for each integration. Think of it as a USB-C port for AI. Introduced by Anthropic in 2024 and later endorsed by the Linux Foundation, MCP has rapidly gained traction among OpenAI, Google, and Microsoft, evolving into one of the most widely accepted standards in the current AI ecosystem.

The Mixture of Experts architecture divides a neural network into numerous smaller, specialized sub-networks, or “experts,” activating only a selected subset for specific tasks. Rather than processing every request through the entire model—similar to consulting a whole team for each task—the MoE model uses an internal “router” to select the most relevant specialists for the particular job. This design allows for the creation of exceedingly 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 prime example; recent GPT models from OpenAI are rumored to employ a variation of this methodology, though the company has not publicly confirmed 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 launches of large language models.

Modeling data-processing algorithms after the intricately interconnected circuits of the human brain has been a concept since the 1940s; however, the recent surge of graphical processing units (GPUs)—largely spurred by the gaming industry—has genuinely unlocked this theory’s capacity. GPUs are specifically engineered for training algorithms with multiple layers, allowing neural network-based AI systems to excel in various sectors, 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 available, allowing anyone to utilize, examine, or modify it. In the AI sphere, 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 while enabling independent safety evaluations often lacking in closed systems. Closed source indicates that the code remains proprietary; users can engage with the product but cannot delve into its internal mechanisms, a characteristic frequently associated with OpenAI’s GPT models—which has generated considerable debate within the AI community.

Parallelization entails executing multiple tasks simultaneously, rather than sequentially—comparable to engaging ten individuals to address different aspects of a project concurrently, rather than having one individual complete each task in order. In AI, parallelization is crucial for both training and inference; contemporary GPUs are specifically designed to handle thousands of calculations at once, forming the backbone of their role in the field. As AI systems become more sophisticated and models extend, the ability to parallelize tasks across multiple chips and machines has become essential in determining 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 whimsical term reflecting a notable trend affecting the tech industry: the ongoing shortage of random access memory (RAM) chips, which are essential for almost all technology products. With the AI sector booming, major tech companies and AI labs have been acquiring RAM at an unprecedented pace to strengthen their data centers, leading to diminished 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 difficulties sourcing memory chips for their devices), consumer electronics (where memory limitations could result in the lowest smartphone shipments in over a decade), and overall corporate computing (as businesses aim to secure sufficient RAM for their data operations). Price increases are unlikely to stabilize until this anticipated shortage is resolved; however, regrettably, signs of a swift resolution are scant.

Similar to AGI, recursive self-improvement suggests a potential breakthrough in AI development, allowing for 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 resistant to external influence. However, recursive self-improvement raises a crucial question: can an AI model create its successor? This dilemma underscores the challenges engineers face when developing such systems. Numerous emerging AI startups are working to create models capable of recursive self-improvement while often downplaying the potential catastrophic consequences, presenting recursive self-improvement merely as the next frontier of research.

Reinforcement learning is a training strategy for AI systems wherein a model learns through trial and error, receiving rewards for correct 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 allows a model to interact with its environment, perform actions, and iteratively refine its behavior based on received feedback. This approach has proven particularly effective in training AI for games, robotics, and, more recently, enhancing reasoning capabilities for 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, notable challenges arise—humans communicate in natural language, while AI systems function through intricate algorithmic processes grounded in data. Tokens serve as the bridge connecting 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 formed through tokenization, which breaks down raw text into manageable parts that a language model can comprehend, much like how a compiler translates human language into binary for computers. In commercial contexts, tokens also carry cost implications—many AI firms charge for LLM usage on a per-token basis, meaning that increased utilization leads to higher expenses.

Therefore, tokens are small textual units—often fragments of words rather than full phrases—that AI language models analyze before processing; they can be compared to “words” in the context of AI tasks. Throughput signifies the quantity that can be processed within a defined timeframe, with token throughput serving as a measurement of the extent of AI work a system can handle concurrently. High token throughput is a critical goal for AI infrastructure teams, as it affects how many users a model can accommodate at the same time and the speed at which they receive responses. AI researcher Andrej Karpathy has expressed concerns about his AI subscriptions being underutilized—reflecting on his experiences as a graduate student when costly computing hardware was not efficiently employed—highlighting why optimizing token throughput has become an urgent focus in the industry.

Training machine learning AIs pertains to a process known as training. This involves supplying data to the model to enable it to recognize patterns and generate desired outputs. It signifies the system’s response to the properties of the data, aligning its outputs toward a specific objective—whether that involves classifying images of cats or composing a haiku in response to a prompt.

Training can require significant resources due to the vast amounts of input needed, which have been steadily increasing—especially with the emphasis on hybrid strategies, such as combining rule-based AIs with curated data, to manage expenses without starting the training process anew.

[See: Inference]

This technique utilizes a previously trained AI model as a foundation for generating a new model aimed at a different but often related task, enabling the reapplication of insights obtained from prior training cycles.

Transfer learning can enhance model development efficiency. It is especially useful when the available data for the target task is limited. However, it’s essential to recognize that this method also has drawbacks; models using transfer learning to acquire general capabilities likely require additional training with new data to excel in their specific areas of focus.

(See: Fine-tuning)

Validation loss serves as a metric indicating an AI model’s learning progression during training—ideally, lower values are desired. Researchers closely monitor this metric as a real-time performance indicator, guiding decisions on when to halt training, adjust hyperparameters, or explore potential issues. A primary concern that validation loss helps flag is overfitting, where a model memorizes its training data rather than truly understanding applicable patterns for new contexts. Think of validation loss as a tool that distinguishes between a student who genuinely grasps the material and one who merely memorizes past exam content—it encapsulates your model’s developmental journey.

Weights are foundational to AI training, determining the importance (or weight) assigned to various features (or input variables) within the training data, ultimately influencing the outputs generated by the AI model.

In essence, weights are numerical parameters indicating which elements within a dataset hold the highest significance for the specific training task. They achieve this through multiplication with inputs. Training typically begins with randomly assigned weights, but as the process unfolds, these weights are refined to generate outputs closely aligned with the desired results.

For example, an AI model designed to predict 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 elements in determining a property’s value according to the available dataset.

This article is regularly updated with new information.