Artificial intelligence is transforming the world and concurrently creating an entirely new set of terminology to articulate its process. Attend any product meeting, pitch, or panel discussion these days, and you’ll encounter terms like LLMs, RAG, RLHF, and many others that can leave even the most knowledgeable individuals in the tech field feeling slightly inadequate. This glossary is our effort to address that: straightforward definitions of the AI terminology you’re most likely to encounter, whether you’re working with this technology, investing in it, or merely trying to stay informed by reading TechCrunch or engaging with relevant podcasts. We keep it updated regularly as the domain progresses, so think of it as a dynamic document, much like the AI technologies it explains.

---

Artificial general intelligence, abbreviated AGI, is a vague term. It generally denotes AI that surpasses the average human in numerous, if not all, activities. OpenAI CEO Sam Altman once referred to AGI as the “equivalent of a median human that you could hire as a co-worker.” In contrast, OpenAI’s charter characterizes AGI as “highly autonomous systems that outshine humans in most economically valuable tasks.” Google DeepMind has a slightly different interpretation; it sees AGI as “AI that can perform at least as well as humans in most cognitive tasks.” Confused? Don’t be concerned — experts leading AI research are just as puzzled.

An AI agent denotes a tool that employs AI technologies to carry out a series of tasks on your behalf — surpassing the capabilities of a basic AI chatbot — such as processing expenses, reserving tickets or tables at restaurants, or even producing and managing code. Nevertheless, as we’ve previously articulated, there are many components in this emerging space, so “AI agent” may signify different things to various individuals. The infrastructure is still being constructed to fulfill its expected functions. However, the fundamental idea suggests an autonomous system that may utilize several AI systems to execute multistep tasks.

Consider API endpoints as “buttons” located behind a software application that other programs can engage to trigger actions. Developers utilize these interfaces to establish integrations — for example, allowing one application to retrieve data from another, or enabling an AI agent to directly control third-party services without human manual intervention at each interface. Most smart home gadgets and connected platforms possess these concealed buttons, even if everyday users never notice or engage with them. As AI agents become more proficient, they increasingly can identify and utilize these endpoints independently, revealing robust — and sometimes unforeseen — opportunities for automation.

In response to a simple inquiry, a human brain can provide an answer without much deliberation — queries like “which animal is taller, a giraffe or a cat?” However, in many instances, you often require pen and paper to arrive at the correct response because there are intermediate steps involved. For example, if a farmer has chickens and cows, and collectively they possess 40 heads and 120 legs, you might need to jot down a simple equation to deduce the answer (20 chickens and 20 cows).

In reference to AI, chain-of-thought reasoning for large language models indicates the process of decomposing a problem into smaller, intermediate steps to enhance the quality of the final output. It typically takes more time to arrive at an answer, but the result is more likely to be accurate, notably in contexts involving logic or coding. Reasoning models are derived from conventional large language models and fine-tuned for chain-of-thought reasoning through reinforcement learning.

(See: Large language model)

This is a more precise notion than an “AI agent,” signifying a program capable of independently taking sequential actions to achieve a goal. A coding agent represents a specific instance applied to software development. Rather than merely suggesting code for a human to review and input, a coding agent can autonomously write, test, and debug code, managing the type of iterative, trial-and-error tasks that typically occupy a developer’s time. These agents can traverse entire codebases, identifying bugs, executing tests, and applying fixes with minimal human oversight. Envision it as hiring an exceptionally fast intern who never tires and remains focused — though, as with any intern, a human must still evaluate the work.

Though a somewhat versatile term, compute generally pertains to the essential computational power that enables AI models to function. This processing capability fuels the AI industry, providing the capacity to train and deploy its powerful models. The term often serves as shorthand for the types of hardware supplying the computational power — such as GPUs, CPUs, TPUs, and other forms of infrastructure that form the foundation of the contemporary AI sector.

A subset of self-enhancing machine learning in which AI algorithms are designed with a multi-layered, artificial neural network (ANN) structure. This configuration enables them to establish more intricate correlations compared to simpler machine learning models, like linear models or decision trees. The framework of deep learning algorithms is inspired by the interconnected pathways of neurons within the human brain.

Deep learning AI models can autonomously identify significant features in data, rather than depending on human engineers to define these attributes. The structure also supports algorithms capable of learning from mistakes and, through repetition and modification, improving their outputs. However, deep learning systems necessitate a substantial amount of data points to produce satisfactory results (millions or more). They also generally require longer training times compared to simpler machine learning algorithms — resulting in higher development costs.

(See: Neural network)

Diffusion is the underlying technology for many art-, music-, and text-generating AI models. Inspired by physics, diffusion systems gradually “deconstruct” data structures — such as photographs, songs, and so forth — by adding noise until nothing remains. In physics, diffusion is a spontaneous and irreversible process — sugar dissolved in coffee cannot revert to its cubical form. Conversely, diffusion systems in AI aim to learn a form of “reverse diffusion” process to reconstruct the deconstructed data, acquiring the capability to retrieve data from noise.

Distillation is a method used to extract knowledge from a large AI model utilizing a ‘teacher-student’ framework. Developers issue requests to a teacher model and record the responses. These responses are sometimes compared with a dataset to assess their accuracy. The outputs are then employed to train the student model, which is designed to mimic the behavior of the teacher.

Distillation can facilitate the creation of a smaller, more efficient model based on a larger model with minimal distillation loss. This is likely the process through which OpenAI developed GPT-4 Turbo, a quicker variant of GPT-4.

While all AI companies utilize distillation internally, it may have also been adopted by some AI firms to catch up with leading models. Distillation from a competitor usually breaches the service agreements of AI APIs and chat assistants.

This refers to the additional training of an AI model to optimize its performance for a more specific task or domain than what was previously emphasized during its training — typically by providing new, specialized (i.e., task-specific) data. 

Numerous AI startups are using large language models as a foundation to develop commercial products but are striving to enhance utility for a particular sector or task by supplementing initial training phases with fine-tuning based on their specific domain knowledge and expertise.

(See: Large language model [LLM])

A GAN, or Generative Adversarial Network, is a machine learning framework that supports significant advancements in generative AI related to generating realistic data — including but not limited to deepfake technologies. GANs employ a pair of neural networks, one which leverages its training data to produce an output that is then evaluated by the other model.

The two models are essentially programmed to compete against each other. The generator strives to produce outputs that the discriminator cannot identify as artificially created data, while the discriminator’s objective is to detect such data. This structured competition can enhance AI outputs to appear more realistic without the necessity for further human involvement. Although GANs are most effective for specific applications (such as creating realistic images or videos), they are less suited for general-purpose AI.

Hallucination is the term preferred by the AI industry to describe instances where AI models fabricate information — essentially generating incorrect data. This presents a significant challenge for AI quality. 

Hallucinations can produce outputs from GenAI that may mislead and could potentially lead to real-world dangers — such as harmful medical advice resulting from a health query.

The issue of AIs concocting information is believed to stem from gaps in their training data. Hallucinations are driving a movement toward more specialized and/or vertical AI models — that is, domain-specific AIs that necessitate narrower expertise — as a method to minimize the chances of knowledge gaps and mitigate disinformation risks.

Inference is the execution of an AI model. It involves unleashing a model to make predictions or derive conclusions from previously seen data. To clarify, inference cannot occur without prior training; a model must learn patterns from a data set before it can effectively infer from this training data.

Various types of hardware are capable of performing inference, ranging from smartphone processors to powerful GPUs and custom-designed AI accelerators. However, not all can operate models equally efficiently. Very large models would take considerable time to predict outcomes on a laptop compared to a cloud server equipped with top-tier AI chips.

[See: Training]

Large language models, or LLMs, are the AI architectures utilized by well-known AI assistants, like ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, or Mistral’s Le Chat. When you interact with an AI assistant, you are engaging with a large language model that processes your request directly or with the aid of various tools available, such as web browsing or code interpreters.

LLMs are deep neural networks consisting of billions of numerical parameters (or weights, as explained below) that learn the relationships between words and phrases and formulate a representation of language, akin to a multidimensional map of words.

These models are developed by encoding the patterns they detect in billions of books, articles, and transcripts. Upon prompting an LLM, the model produces the most probable pattern that corresponds to the prompt.

(See: Neural network)

Memory cache signifies a crucial process that enhances inference (the process by which AI generates responses to user inquiries). Essentially, caching represents an optimization strategy aimed at increasing inference efficiency. AI fundamentally relies on high-level mathematical calculations, and every time those calculations are performed, they consume more energy. Caching is intended to decrease the number of calculations a model may need by retaining specific computations for future queries and operations. There are various forms of memory caching, with KV (or key-value) caching being one of the more recognized. KV caching functions in transformer-based models, enhancing efficiency and delivering quicker results by minimizing the time (and algorithmic effort) necessary to generate answers to user questions.   

(See: Inference)  

Model Context Protocol, or MCP, is an open standard enabling AI models to connect to external tools and data — your files, databases, or applications like Slack and Google Drive — without the need for developers to create custom connectors for every individual pairing. Think of it as a USB-C port for AI. Anthropic introduced MCP in 2024 and later transferred it to the Linux Foundation, and it has since been adopted by OpenAI, Google, and Microsoft, making it one of the fastest-spreading standards in recent AI history.

Mixture of Experts is a model architecture that divides a neural network into numerous smaller specialized sub-networks, or “experts,” and activates only a select few for any specific task. Instead of routing every request through the entire model — much like gathering your entire office for every inquiry — an MoE model possesses an internal “router” that selects the appropriate specialists for the task at hand. This capability allows for the construction of large models that remain relatively swift and economical to operate, since only a fraction of the network is active at any given time. Mistral AI’s Mixtral model serves as a well-known example; OpenAI’s newer GPT models are also broadly believed to utilize some variation of this methodology, though the company has never officially confirmed it.

(See: Neural network, Deep learning)

A neural network pertains to the multi-layered algorithmic framework that supports deep learning — and, more broadly, the entire surge in generative AI tools following the emergence of large language models. 

Although the concept of deriving inspiration from the intricately interconnected pathways of the human brain as a design model for data processing algorithms extends back to the 1940s, the recent advent of graphical processing units (GPUs) — driven by the gaming industry — truly unlocked the potential of this theory. These chips proved exceptionally efficient for training algorithms with many more layers than what was feasible in earlier periods — enabling neural network-based AI systems to achieve significantly better performance across multiple fields, including voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source pertains to software — or, increasingly, AI models — whose underlying code is made publicly accessible for anyone to use, examine, or alter. In the realm of AI, Meta’s Llama family of models is a notable example; Linux serves as the famous historical counterpart in operating systems. Open source methodologies permit researchers, developers, and companies globally to build upon each other’s work, expediting progress and enabling independent safety audits that proprietary systems cannot easily facilitate. Closed source refers to situations where the code is proprietary — you can utilize the product but not understand how it operates, as is the case with OpenAI’s GPT models — a distinction that has evolved into one of the hallmark debates in the AI sector.

Parallelization signifies performing multiple tasks simultaneously rather than sequentially — akin to having 10 employees working on different aspects of a project concurrently instead of a single employee handling everything one after the other. In AI, parallelization is essential for both training and inference: modern GPUs are specifically engineered to execute thousands of calculations in parallel, which significantly contributes to their establishment as the hardware backbone of the industry. As AI systems increase in complexity and models expand in size, the capability to parallelize processes across numerous chips and machines has become one of the most vital factors in determining the speed and cost-effectiveness of model development and deployment. Research into enhanced parallelization strategies is now a specialized area of study in its own right.

RAMageddon is the catchy new term describing an increasingly troublesome trend affecting the tech sector: a persistent shortage of random access memory, or RAM chips, which are integral to nearly all the tech products we rely on daily. With the growth of the AI industry, major tech companies and AI laboratories — all competing to build the most powerful and efficient AI — are purchasing vast amounts of RAM to support their data centers, leaving scant supply for everyone else. This supply bottleneck results in rising prices for what remains.

This shortage extends to industries like gaming (where prominent companies had to hike console prices because of difficulties in sourcing memory chips for their devices), consumer electronics (where the memory shortage could lead to the largest decline in smartphone shipments in over a decade), and general enterprise computing (due to insufficient RAM for their own data centers). The price surge is anticipated to stabilize only once the feared shortage resolves, but, unfortunately, there is currently little indication that this will happen anytime soon.  

Similar to AGI, recursive self-improvement signifies a boundary for how intelligent AI might become and to what extent it may minimize reliance on humans. In the scenario of RSI, AI models begin to enhance themselves autonomously without human input, resulting in a rapid increase in capabilities and independence. In certain narratives, this represents a cataclysmic event akin to the singularity, a moment when AI models become impervious to external intervention. However, RSI also describes a fundamental ability — can an AI model create its own successor? — which simplifies the task for engineers aiming to construct it. Several up-and-coming AI startups are focused on developing recursively self-improving models, yet most downplay any apocalyptic consequences, presenting RSI merely as the next research frontier.

Reinforcement learning constitutes a method for training AI where a system learns by experimenting and receiving rewards for correct responses — somewhat like training a beloved pet with treats, except in this case, the “pet” is a neural network and the “treat” is a mathematical signal indicating success. Divergent from supervised learning, where a model is taught on a predetermined dataset of labeled examples, reinforcement learning enables a model to explore its environment, take actions, and adjust its behavior based on received feedback. This approach has proven particularly effective for training AI to play games, control robots, and, more recently, to enhance the reasoning capabilities of large language models. Techniques like reinforcement learning from human feedback, or RLHF, are now integral to how leading AI laboratories refine their models for improved helpfulness, accuracy, and safety.

When considering human-machine interaction, several evident challenges arise — individuals communicate using human language, whereas AI systems perform tasks through complex algorithmic procedures influenced by data. Tokens serve to bridge that divide: they represent the fundamental units of interaction between humans and AI, corresponding to discrete segments of data that have been processed or generated by an LLM. They are produced through a process known as tokenization, which segments raw text into manageable units that a language model can interpret, akin to how a compiler translates human language into binary code that a computer can comprehend. In enterprise environments, tokens also dictate costs — most AI businesses charge for LLM usage on a per-token basis, implying that the more a business utilizes the system, the more it incurs in expenses.

Thus, tokens refer to the small units of text — often fragments of words rather than complete ones — that AI language models disassemble language into before processing; they are broadly comparable to “words” for the sake of comprehending AI workloads. Throughput denotes how much can be processed within a certain timeframe, hence token throughput essentially functions as a gauge of the amount of AI work a system can manage simultaneously. High token throughput is a crucial objective for AI infrastructure teams, as it determines how many users a model can serve concurrently and how quickly each can expect a response. AI researcher Andrej Karpathy has expressed feeling anxious when his AI subscriptions remain idle — echoing the concern he experienced during his graduate studies when costly computer hardware wasn’t maximally utilized — a sentiment that encapsulates why maximizing token throughput has become a focal point in the field.

Developing machine learning AIs involves a process termed training. Simply put, this refers to data being provided to enable the model to learn from patterns and produce useful outputs. Essentially, it involves the system responding to characteristics in the data, which allows it to adapt outputs towards a desired objective — whether that involves identifying images of cats or generating a haiku upon request.

Training can incur significant costs because it demands numerous inputs, and the required volumes have been steadily increasing — which is why hybrid methods, such as fine-tuning a rules-based AI with targeted data, can help manage expenses without starting entirely from zero.

[See: Inference]

A technique in which a previously trained AI model is utilized as the base for developing a new model aimed at a different but typically related task — permitting the knowledge acquired in prior training cycles to be reapplied. 

Transfer learning can promote efficiency by expediting model development. It can also be advantageous when there is limited data available for the task for which the model is being crafted. However, it’s imperative to recognize that this approach has limitations. Models utilizing transfer learning to gain generalized competencies will likely need additional data training to excel in their specific focus area.

(See: Fine tuning)

Validation loss signifies a metric that indicates how effectively an AI model is learning during training — with a lower value being preferable. Researchers monitor it attentively as a form of real-time assessment, using it to determine when to cease training, when to modify hyperparameters, or whether to investigate potential issues. One of the main concerns it helps highlight is overfitting, a circumstance where a model memorizes its training data instead of genuinely learning patterns it can generalize to new situations. Think of it as distinguishing a student who comprehensively grasps the material from one who merely memorized last year’s exam — validation loss aids in revealing which type your model is developing into.

Weights are fundamental to the training of AI, as they ascertain the significance (or weight) assigned to differing features (or input variables) within the training data — thereby molding the AI model’s output. 

To rephrase, weights are numerical parameters that delineate what is most salient in a dataset for the designated training task. They fulfill their role by applying multiplication to inputs. Model training typically initiates with weights that are assigned randomly, but as the training progresses, the weights adapt as the model aims to arrive at an output that aligns more closely with the target.

For instance, an AI model designed to predict housing prices based on historical real estate data for a specified location could incorporate weights for attributes such as the number of bedrooms and bathrooms, whether a property is detached or semi-detached, and whether it includes parking, a garage, among others. 

Ultimately, the weights the model attaches to each of these inputs represent how significantly they influence the value of a property based on the given dataset.

This article is updated regularly with new information.