Essential AI Terms to Understand in 2023

The realm of artificial intelligence is rapidly evolving, introducing a plethora of terminology to frame its concepts. At modern tech conferences and product launches, abbreviations such as LLMs, RAG, RLHF, and others can perplex even experienced experts. This glossary aims to clarify these terms by providing straightforward definitions of AI jargon that you might encounter—whether you’re exploring these technologies, investing in them, or keeping pace with media such as TechCrunch or related podcasts. We will ensure it remains pertinent to reflect the continually changing environment of AI systems as the industry advances.

---

Artificial general intelligence (AGI) is a complex notion that generally denotes AI that surpasses human abilities across a wide range of tasks. OpenAI’s CEO, Sam Altman, compares AGI to the “average person you might employ as a coworker.” OpenAI defines AGI as “highly autonomous systems that perform exceptionally well in most economically relevant tasks.” Meanwhile, Google DeepMind describes it as “AI proficient in executing most cognitive tasks at a human level.” If this description seems unclear, you’re not alone; many in the AI research community share similar views.

An AI agent is a tool that utilizes AI technologies to automate a variety of tasks, extending beyond traditional chatbot roles. These tasks can include expense tracking, travel planning, or even coding and software maintenance. As previously noted, the term “AI agent” can differ across stakeholders, and its foundational structure is consistently evolving to meet rising expectations. Fundamentally, an AI agent functions as an autonomous system that integrates multiple AI technologies to manage complex, multi-step processes.

Think of API endpoints as “hidden buttons” within software that other applications can trigger to carry out functions. Developers utilize these interfaces to create 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 networked systems include these inconspicuous buttons, which are often overlooked by the average user. As AI agents progress, their ability to identify and utilize these endpoints improves, unveiling impressive—and sometimes surprising—automation features.

When faced with simple questions, humans typically respond instinctively—such as asking, “Which is taller, a giraffe or a cat?” However, some questions require breaking down intermediate steps to arrive at accurate answers. For instance, if a farmer has 40 animals (chickens and cows) summing up to 120 legs, one must create an equation to determine how many of each animal exist (20 chickens and 20 cows).

In AI, chain-of-thought reasoning employed by large language models involves breaking a problem down into smaller, sequential steps to improve the accuracy of the final answer. Although this approach may take longer, it significantly enhances correctness, particularly in logical or programming situations. Reasoning models arise from classical large language models and are honed for chain-of-thought processes, primarily through reinforcement learning.

(See: Large language model)

This term is more specific than “AI agent,” highlighting applications that can autonomously execute tasks in a series to achieve intended results. A coding agent is a specialized type focusing solely on software development. Rather than just generating code for human review, a coding agent can independently write, test, and debug software, addressing the iterative challenges developers often face. These agents can navigate entire codebases, spot bugs, run tests, and apply fixes with minimal human intervention, effectively acting like a diligent intern—hardworking and tireless, yet still requiring human guidance.

Though somewhat broad, compute primarily pertains to the computational power that drives AI models. This capability underpins the AI sector, enabling the training and deployment of robust models. It often relates to the types of hardware providing this computational strength—like GPUs, CPUs, TPUs, and other crucial infrastructure essential to today’s AI ecosystem.

Deep learning, a subset of self-improving machine learning, centers on developing AI algorithms within a layered artificial neural network (ANN) structure. This design allows for the identification of more complex patterns compared to traditional machine learning 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 autonomously discover significant features within data, negating the need for human engineers to define them. This structure also enables algorithms to learn from mistakes, gradually improving outputs through iterative adjustments. However, deep learning systems typically require vast amounts of data for optimal performance (often millions of data points) and often endure longer training times compared to simpler models—resulting in higher development costs.

(See: Neural network)

Diffusion refers to techniques utilized by various AI models to create art, music, and text. Originating from physics, diffusion systems systematically “destroy” data structures—like images or melodies—by adding noise until they become indistinguishable. In physics, diffusion occurs naturally and irreversibly; for instance, sugar dissolved in coffee cannot revert to its original form. Conversely, AI diffusion systems strive to master a “reverse diffusion” process to recover distorted data, effectively learning to extract information obscured by noise.

Distillation describes a method of extracting insights from a large AI model through a “teacher-student” framework. Developers interact with a teacher model and document its outputs, often measuring these results against a pre-defined dataset for accuracy. These outputs are then used to train a student model, aiming to reproduce the teacher’s responses.

This technique can result in a more compact, efficient model that retains the efficacy of its larger counterpart while reducing distillation loss. It’s widely thought that this method contributed to the development of GPT-4 Turbo, a faster version of GPT-4 from OpenAI.

While all AI organizations perform distillation internally, some may exploit it to enhance their features in comparison to dominant 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 initially designed for—typically through the inclusion of new data relevant to the task.

Many AI startups begin with large language models as the foundation for their products, aiming to improve functionality for specific sectors or tasks by adding domain-specific knowledge onto prior training.

(See: Large language model [LLM])

A GAN, or Generative Adversarial Network, is a machine learning framework that has propelled generative AI, especially in producing realistic outputs—including deepfake technologies. 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 aims to produce outputs that can deceive the discriminator, while the discriminator attempts to differentiate real data from fake content. This structured rivalry can yield realistic AI-generated outputs with minimal human intervention. Nonetheless, GANs typically perform well only under specific conditions (e.g., generating authentic images or videos), rather than universally.

Within the AI community, “hallucination” refers to instances when AI models generate inaccuracies—essentially fabricating incorrect information. This poses a formidable challenge to AI accuracy.

Hallucinations may produce misleading outputs from generative AI, carrying significant risks—potentially resulting in harmful consequences (e.g., unsafe medical advice arising from a health inquiry).

The phenomenon of AIs generating fictional information is thought to stem from gaps in their training data. Consequently, there is an increasing demand for more specialized and/or vertical AI models—domain-specific AIs that require narrower expertise to close knowledge gaps and reduce misinformation risks.

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

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

[See: Training]

Large language models, or LLMs, act as the foundational AI frameworks behind popular 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 utilizing a large language model that responds to your inquiries or employs various available tools, like web browsers or code interpreters.

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

These models are trained by analyzing patterns found in vast amounts of documents, articles, and transcripts. When you provide a prompt to an LLM, the model generates the most probable response related to that input.

(See: Neural network)

Memory cache represents a vital process that enhances inference (how AI crafts responses to user prompts). Essentially, caching acts as an optimization method aimed at improving inference performance. AI functions through intricate mathematical calculations, each consuming resources. Caching reduces the number of computations a model needs to perform by retaining certain calculations for future use. There are various types of memory caching, with KV (key-value) caching being among the most frequently utilized. KV caching is employed in transformer-based models, enhancing efficiency and delivering quicker results by minimizing the time (and computational effort) required to answer user prompts.

(See: Inference)

Model Context Protocol (MCP) is an open standard designed to streamline AI models’ interactions with external tools and data—such as your files, databases, or applications like Slack and Google Drive—without requiring developers to create individual connectors for each integration. Envision 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, becoming one of the most widely adopted standards in the current AI ecosystem.

The Mixture of Experts architecture divides a neural network into several smaller specialized sub-networks, or “experts,” activating only a chosen subset of them for specific tasks. Instead of routing every request through the entire model—similar to consulting an entire team for every question—the MoE model utilizes an intrinsic “router” that chooses the most relevant specialists for the job. This configuration enables the construction of exceptionally large models while retaining efficiency and cost-effectiveness, as only a fraction of the network remains active at any one time. Mistral AI’s Mixtral model is a notable example; recent GPT models from OpenAI are rumored to use a variation of this method, although the company has not yet confirmed it publicly.

(See: Neural network, Deep learning)

A neural network serves as the multilayered algorithmic foundation for deep learning—and, more broadly, for the rise of generative AI tools following the advent 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, it was the recent increase in graphical processing units (GPUs)—largely propelled by the gaming industry—that truly unlocked the potential of this theory. GPUs are specifically crafted for training algorithms with multiple layers, enabling neural network-based AI systems to excel across various domains, such as voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source relates to software—or increasingly, AI models—where the core code is accessible to the general populace, allowing anyone to employ, inspect, or alter it. In the AI landscape, Meta’s Llama series of models stands out as a leading example; Linux is recognized as a benchmark among operating systems. Open-source principles enable researchers, developers, and organizations worldwide to build upon each other’s work, fostering innovation while allowing independent safety assessments that closed systems often lack. Closed source indicates that the code remains proprietary; users can interact with the product but cannot examine its internal workings, a characteristic often associated with OpenAI’s GPT models—which has sparked considerable discussion within the AI community.

Parallelization involves executing multiple tasks at the same time rather than sequentially—much like employing ten individuals to tackle various aspects of a project simultaneously, rather than having one person complete each task in sequence. In AI, parallelization is vital for both training and inference; modern GPUs are specifically engineered to handle thousands of calculations simultaneously, which forms the foundation of their role in the sector. As AI systems grow more complex and models expand, the capability to parallelize tasks across multiple chips and machines has become crucial in determining how quickly and effectively models can be developed and deployed. Research into improved parallelization techniques has evolved into a specialized field.

RAMageddon is a lighthearted term illustrating a significant trend affecting the tech landscape: the ongoing scarcity of random access memory (RAM) chips, which are essential for nearly all technology products. As the AI domain surges, major tech firms and AI labs have been securing RAM at unprecedented levels to bolster their data centers, resulting in reduced availability in the broader market. This supply crunch is driving up RAM prices.

This shortage influences various sectors, including gaming (where companies are forced to increase console prices due to difficulties in obtaining memory chips for their devices), consumer electronics (where memory limitations could lead to the lowest smartphone shipments in over a decade), and general corporate computing (as businesses strive to secure adequate RAM for their data operations). Price increases are unlikely to stabilize until this anticipated shortage resolves, but unfortunately, 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 context, AI models begin to autonomously enhance their capabilities, leading to considerable advancements in autonomy and functionality. Some narratives equate this scenario to a singularity moment—when AI models become immune to external influence. However, recursive self-improvement raises a fundamental question: can an AI model create its successor? This dilemma simplifies the challenges engineers face in developing such systems. Numerous emerging AI startups aim to create models capable of recursive self-improvement while often downplaying the potential catastrophic implications, presenting recursive self-improvement merely as the next frontier in research.

Reinforcement learning is an AI training method whereby a system learns through trial and error, receiving rewards for correct actions—similar to instructing a pet with treats, although here, 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, take actions, and iteratively improve its behavior based on feedback received. This method has proven particularly effective for training AI in games, controlling robotics, and, more recently, enhancing reasoning abilities for large language models. Techniques such as reinforcement learning from human feedback (RLHF) have become critical in how leading AI labs fine-tune their models for improved performance, precision, and safety.

In the exchange between humans and machines, notable challenges arise—humans communicate in natural language, while AI systems operate through intricate algorithmic processes based on data. Tokens serve as the bridge between these two realms: they represent the fundamental units of human-AI interaction, encapsulating distinct segments of data processed or generated by an LLM. Tokens are produced through a process called tokenization, which breaks down raw text into manageable pieces that a language model can interpret, similar to how a compiler translates human language into binary for computers. In commercial contexts, tokens also influence costs—many AI companies charge for LLM usage on a per-token basis, indicating that increased utilization correlates with higher expenses.

Thus, tokens represent small text units—typically fragments of words rather than complete phrases—that AI language models analyze prior to processing; they can be likened to “words” in terms of AI workloads. Throughput refers to the quantity that can be processed within a given timeframe, rendering token throughput a measure of how much AI work a system can manage simultaneously. High token throughput is a crucial aim for AI infrastructure teams since it affects how many users a model can assist at once 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 inefficiently used—highlighting why optimizing token throughput has emerged as an immediate focus in the field.

Training machine learning AIs encompasses a process known as training. This primarily involves supplying data to enable the model to identify patterns and produce desired outputs. It signifies the system’s response to the characteristics of the data, aligning its outputs toward a desired outcome—whether that pertains to identifying images of cats or composing a haiku in response to a prompt.

Training can demand significant resources due to the extensive amounts of input needed, which have been consistently growing—especially with the significance of hybrid strategies, like merging rules-based AIs with targeted data, to manage costs without restarting the process from scratch.

[See: Inference]

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

Transfer learning can improve model development efficiency. It proves especially beneficial when the data available for the target task is limited. However, it’s important to acknowledge that this approach also has limitations; models utilizing transfer learning to acquire generalized skills will likely require further training with new data to excel in their specific focus areas.

(See: Fine tuning)

Validation loss serves as a metric representing an AI model’s learning progress throughout training—ideally, lower values are more favorable. Researchers monitor this indicator closely as a real-time performance measure, assisting decisions about when to halt training, adjust hyperparameters, or investigate potential issues. A significant concern that validation loss helps to identify is overfitting, where a model memorizes its training data rather than genuinely understanding applicable patterns for new scenarios. Think of validation loss as a tool that distinguishes between a student who comprehends the material and one who merely memorizes previous exam content—it encapsulates your model’s developmental journey.

Weights are essential 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 simpler terms, weights are numerical parameters that indicate which elements within a dataset are most critical for the specific training task. They achieve this through multiplication with inputs. Training typically begins with randomly assigned weights, but as the process progresses, these weights are refined to produce outputs that closely align with desired outcomes.

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

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

This article is regularly updated with new information.