Essential AI Terms to Know in 2023

The realm of artificial intelligence is evolving swiftly, bringing forth a unique vocabulary to elucidate its various concepts. At contemporary tech conferences or product demonstrations, jargon such as LLMs, RAG, RLHF, and others can perplex even seasoned professionals. This glossary seeks to demystify these terms by offering clear definitions of AI terminology you may come across—whether you’re delving into these technologies, investing, or staying informed with outlets like TechCrunch or similar podcasts. We will continuously update it to mirror the ever-changing landscape of AI systems as the industry progresses.

---

Artificial general intelligence (AGI) remains an elusive notion, typically indicating AI that surpasses human average capabilities across diverse tasks. OpenAI CEO Sam Altman likened AGI to the “average human one could recruit as a coworker.” OpenAI defines AGI as “highly autonomous systems that excel in most economically significant activities.” Contrarily, Google DeepMind characterizes it as “AI that can competently perform most cognitive tasks at a level comparable to humans.” If this seems ambiguous, you’re not alone; many within the AI research sphere share this perspective.

An AI agent refers to a tool that utilizes AI technologies to automate a variety of tasks for you, expanding beyond traditional AI chatbot capabilities. These tasks may include expense monitoring, travel planning, or even writing and maintaining software. As previously mentioned, the meaning of an “AI agent” can vary among different stakeholders, and its foundational framework is continually evolving toward its expected capabilities. Essentially, an AI agent is an autonomous system that incorporates various AI technologies to address complex, multi-step tasks.

Consider API endpoints as “hidden buttons” within software that other applications can trigger to execute tasks. Developers utilize these interfaces to create integrations—such as enabling one application to retrieve data from another or allowing an AI agent to independently interface with third-party services. Many smart home gadgets and interconnected systems contain these hidden buttons, often unnoticed by the average user. As AI agents advance, their ability to autonomously discover and utilize these endpoints is improving, revealing powerful—and sometimes unexpected—automation features.

When faced with straightforward questions, humans often respond instinctively—like, “which is taller, a giraffe or a cat?” However, some questions necessitate 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 must formulate an equation to determine how many of each there are (20 chickens and 20 cows).

In AI, chain-of-thought reasoning for large language models involves segmenting a problem into smaller, sequential steps to enhance the accuracy of the final answer. This method usually takes longer to reach a solution but significantly increases the probability of correctness, especially in logical or programming contexts. Reasoning models develop from standard large language models and are fine-tuned for chain-of-thought techniques, largely through reinforcement learning.

(See: Large language model)

This term is more focused than “AI agent,” honing in on applications adept at independently executing tasks in sequence to achieve specified results. A coding agent is a specific type aimed at software development. Unlike merely generating code for human review, a coding agent can autonomously write, test, and debug software, addressing the iterative challenges developers often face. These agents can navigate entire codebases, identify bugs, perform tests, and apply fixes with minimal human oversight, functioning like a remarkably efficient intern—diligent and tireless, yet still necessitating human direction.

Though somewhat broad, compute primarily pertains to the computational power that underpins AI models. This capability propels the AI sector, enabling the training and deployment of powerful models. It often serves as shorthand for the types of hardware that deliver this computational power—like GPUs, CPUs, TPUs, and other foundational infrastructure crucial to today’s AI ecosystem.

Deep learning, a subdivision of self-improving machine learning, focuses on creating AI algorithms within a layered artificial neural network (ANN) framework. This architecture enables the recognition of more sophisticated patterns than traditional machine learning techniques such as linear models or decision trees. Deep learning systems are inspired by the complex pathways of neurons in the human brain.

Deep learning models can independently discover important features within data, eliminating the need for human engineers to specify them. This framework also allows algorithms to learn from mistakes, continuously improving outputs through iterative updates. However, deep learning systems generally require vast amounts of data for optimal performance (often millions of data points) and typically experience longer training periods compared to simpler models—resulting in higher developmental costs.

(See: Neural network)

Diffusion refers to the methods employed by various AI models to generate art, music, and text. Derived from physics, diffusion systems systematically “destroy” data structures—like images or melodies—by introducing noise until they are indistinguishable. In physics, diffusion unfolds spontaneously and irreversibly; for instance, sugar dissolved in coffee cannot revert to its original form. In contrast, AI diffusion systems strive to master a “reverse diffusion” technique to recover distorted data, effectively learning to uncover information obscured by noise.

Distillation represents a technique for deriving insights from a large AI model using a “teacher-student” framework. Developers engage with a teacher model and document its outputs, frequently benchmarking these results against a predefined dataset for accuracy. These outputs train the student model, which aims to replicate the teacher’s responses.

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

While all AI organizations conduct distillation internally, some may utilize it to enhance their features in comparison to leading models. However, extracting insights from a competitor typically infringes upon AI API service agreements.

This term denotes the retraining of an AI model to enhance its effectiveness for a more specialized task or domain than initially intended—usually by integrating new, task-relevant data.

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

(See: Large language model [LLM])

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

These two models challenge each other; the generator aims to produce outputs that can deceive the discriminator, while the discriminator works to distinguish genuine data from fabricated constructs. This structured competition can yield realistic AI outputs with minimal human input. However, GANs tend to perform optimally under specific conditions (like generating authentic images or videos) rather than being universally applicable.

Within the AI community, “hallucination” refers to instances where AI models produce inaccuracies—essentially fabricating incorrect information. This poses a significant barrier to AI accuracy.

Hallucinations can result in misleading outputs from generative AI, leading to considerable risks—potentially causing harmful consequences (e.g., unsafe medical advice from a health-related question).

The issue of AIs generating fictitious information is believed to stem from inadequacies in their training data. Consequently, there is a rising push toward developing more specialized and/or vertical AI models—domain-specific AIs that require narrower expertise to bridge knowledge gaps and diminish misinformation risks.

Inference signifies the execution of an AI model; it involves activating a model to predict or derive conclusions from previously observed data. Importantly, inference cannot transpire without training; a model must first identify patterns within a dataset before effectively extrapolating from that training.

Various hardware setups can execute inference, ranging from consumer-grade processors to high-performance GPUs and specialized AI accelerators. However, not all configurations perform models with equal efficacy; for instance, exceptionally large models would take significantly longer to produce predictions on a laptop compared to a cloud server equipped with advanced AI chips.

[See: Training]

Large language models, or LLMs, are the AI frameworks powering well-known AI assistants like ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, and Mistral’s Le Chat. When you interact with an AI assistant, you engage with a large language model that directly addresses your queries or utilizes various available tools, such as web browsers or code interpreters.

LLMs comprise deep neural networks with billions of numerical parameters (or weights, see below) that identify relationships among words and phrases to create a representation of language—essentially, a sophisticated map of language.

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

(See: Neural network)

Memory cache represents a crucial process that enhances inference (how AI formulates responses to user prompts). Essentially, caching serves as an optimization technique aimed at boosting inference performance. AI operates through intricate mathematical computations, where each calculation consumes resources. Caching mitigates the number of computations a model must perform by retaining certain calculations for future interactions. Numerous forms of memory caching exist, with KV (key-value) caching being among the most prevalent. KV caching is utilized in transformer-based models, enhancing efficiency and yielding quicker outcomes by minimizing the time (and computational effort) needed to generate responses to user prompts.

(See: Inference)

Model Context Protocol (MCP) is an open standard designed to facilitate AI models’ interaction 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. You can think of it as a USB-C port for AI. Initiated by Anthropic in 2024 and later supported by the Linux Foundation, MCP has gained traction among OpenAI, Google, and Microsoft, swiftly becoming one of the most widely embraced standards in the current AI ecosystem.

The Mixture of Experts architecture divides a neural network into various smaller specialized sub-networks, or “experts,” activating only a subset of them for specific tasks. Instead of routing each request through the entire model—akin to consulting your whole team for every question—the MoE model employs a built-in “router” that selects the most relevant specialists for the task. This design allows for the development of exceptionally large models while maintaining relative efficiency and cost-effectiveness, as only a fraction of the network is active at any moment. Mistral AI’s Mixtral model is a notable example; recent GPT models from OpenAI are rumored to employ some variation of this approach, although the company has yet to confirm it officially.

(See: Neural network, Deep learning)

A neural network serves as the multi-layered algorithmic foundation for deep learning—and more broadly, for the emergence 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 existed since the 1940s; however, it was the recent increase in graphical processing units (GPUs)—primarily driven by the gaming industry—that truly unlocked this theory’s potential. GPUs are specifically engineered for training algorithms with multiple layers, enabling neural network-based AI systems to excel across various fields, including voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source signifies software—or increasingly, AI models—where the underlying code is publicly accessible, allowing anyone to use, inspect, or modify it. In the AI realm, Meta’s Llama series of models stands out as a significant instance; Linux is a well-known benchmark among operating systems. Open-source principles empower researchers, developers, and organizations worldwide to build upon each other’s contributions, accelerating innovation and facilitating independent safety evaluations that closed systems often cannot provide. Closed source indicates that the code remains private; users can interact with the product but cannot scrutinize its internal workings, a feature often associated with OpenAI’s GPT models—which has sparked considerable debate within the AI community.

Parallelization involves executing multiple tasks concurrently rather than sequentially—similar to employing 10 individuals to tackle different aspects of a project simultaneously instead of having one person complete each task consecutively. In AI, parallelization is crucial for both training and inference; modern GPUs are specially designed to handle thousands of calculations simultaneously, forming the cornerstone of their role in the industry. As AI systems grow more complex and models expand, the ability to parallelize tasks across multiple chips and machines has become a vital factor in determining how swiftly and efficiently models can be developed and deployed. Research into improved parallelization techniques has developed into a focused field.

RAMageddon is a humorous term reflecting a significant trend impacting the tech landscape: the ongoing shortage of random access memory (RAM) chips, which are essential for virtually all technology products. As the AI sector has surged, major tech corporations and AI laboratories have been acquiring RAM at unprecedented levels to bolster their data centers, leading to decreased availability in the broader market. This supply crunch is driving up RAM prices.

This shortage impacts various sectors, such as gaming (where companies are compelled to raise console prices due to challenges in sourcing 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 sufficient RAM for their data operations). The escalating prices are unlikely to stabilize until this anticipated shortage resolves, but unfortunately, signs of a swift resolution appear minimal.

Comparable to AGI, recursive self-improvement signifies a potential milestone in AI evolution, enabling minimal human intervention. In this scenario, AI models begin to enhance their capabilities autonomously, resulting in significant advancements in autonomy and function. Some narratives equate this phenomenon to a singularity moment—when AI models become impervious to external influence. However, recursive self-improvement raises a fundamental question: can an AI model generate its successor? This inquiry simplifies the challenge for engineers attempting to design such models. Numerous emerging AI startups focus on developing recursively self-improving models while often downplaying the potentially catastrophic outcomes, presenting recursive self-improvement merely as the next frontier in research.

Reinforcement learning is an AI training methodology where a system learns through trial and error, receiving rewards for correct actions—akin to training a pet with treats, though in this context, 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 engage with its environment, take actions, and iteratively refine its behavior based on received feedback. This technique has shown particular efficacy for training AI to compete in games, control robotics, and, more recently, enhance the reasoning capabilities of large language models. Approaches like reinforcement learning from human feedback (RLHF) have become integral in how leading AI labs fine-tune their models for improved effectiveness, precision, and safety.

In interactions between humans and machines, significant challenges arise—humans communicate in natural language, while AI systems operate through complex algorithmic processes grounded in data. Tokens bridge 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 constructed through a process known as tokenization, which divides raw text into manageable parts that a language model can comprehend, similar to how a compiler transforms human language into binary code for computers. In business environments, tokens also impact costs—numerous AI companies charge for LLM usage on a per-token basis, indicating that increased utilization correlates with higher expenses.

Thus, tokens are the small text units—often fragments of words rather than complete phrases—that AI language models analyze prior to processing; they can be likened to “words” concerning AI workloads. Throughput denotes the quantity that can be processed within a specified timeframe, making token throughput a measure of the volume of AI work a system can manage concurrently. High token throughput is a critical objective for AI infrastructure teams, as it influences how many users a model can serve simultaneously and the speed at which they receive responses. AI researcher Andrej Karpathy has voiced concerns about his AI subscriptions remaining underutilized—reflecting on his experiences as a graduate student when costly computing hardware was employed inefficiently—highlighting why optimizing token throughput has become an urgent objective in the field.

Training machine learning AIs involves a process known as training. This primarily consists of providing data to enable the model to recognize patterns and generate desired outputs. It embodies the system’s response to the characteristics of the data, allowing it to align outputs toward a desired goal—whether that entails identifying images of cats or creating a haiku in response to a prompt.

Training can be resource-intensive due to the immense quantities of inputs required, which have been consistently increasing—especially the significance of hybrid strategies, like merging rules-based AIs with targeted data, to manage costs without restarting the process.

[See: Inference]

This method utilizes a previously trained AI model as a baseline for developing a new model aimed at a different but typically related task, allowing for the reapplication of insights gained from earlier training cycles.

Transfer learning can enhance model development efficiency. It is particularly beneficial when the available data for the target task is limited. However, it’s essential to recognize that this approach has limitations; models leveraging transfer learning to gain generalized skills will likely require additional training with new data to excel in their specific focus areas.

(See: Fine tuning)

Validation loss acts as a metric indicating an AI model’s learning progress during training—lower values are preferred. Researchers diligently monitor this indicator as a real-time performance gauge, aiding decisions on when to halt training, modify hyperparameters, or investigate potential issues. A key concern that validation loss helps 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 distinguishing between a student who comprehends the material and one who simply memorizes last year’s exam—it illustrates the developmental path of your model.

Weights are foundational 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 indicating which elements within a dataset are most crucial 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 refined as the model aims to produce outputs that closely align with the desired outcomes.

For instance, an AI model designed to predict housing prices based on historical real estate data for a specific area might assign weights to variables 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 weights assigned reflect how significantly each of these factors impacts a property’s value according to the available dataset.

This article is regularly updated with new information.