Essential AI Glossary for 2023

The realm of artificial intelligence is rapidly progressing, bringing forth a unique vocabulary that clarifies its various concepts. At contemporary tech events or product conferences, terms like LLMs, RAG, RLHF, and others may confuse even the most experienced experts. This glossary seeks to demystify these terms, offering clear definitions of the AI terminology you might come across—whether you’re delving into these technologies, investing in them, or keeping up with updates via platforms like TechCrunch or pertinent podcasts. We will ensure it stays current as the field evolves, reflecting the dynamic nature of AI systems.

---

Artificial general intelligence (AGI) is still a somewhat ambiguous idea, generally signifying AI that surpasses average human abilities across various tasks. OpenAI CEO Sam Altman likened AGI to the “typical human you could hire as a co-worker.” According to OpenAI’s charter, AGI is described as “highly autonomous systems that perform exceptionally in most economically valuable tasks.” In contrast, Google DeepMind defines AGI as “AI that performs at least as competently as humans in most cognitive tasks.” If this sounds vague, you’re not alone; many in the AI research community feel similarly.

An AI agent is a tool that harnesses AI technologies to automate a variety of tasks on your behalf, going beyond the functionalities of a standard AI chatbot. This can include handling expenses, making travel arrangements, or even coding and maintaining software. As noted earlier, the definition of an “AI agent” can vary among stakeholders, and the foundational infrastructure is still evolving toward its anticipated capabilities. In essence, an AI agent is an autonomous system that integrates different AI technologies to complete complex, multi-step tasks.

Think of API endpoints as “hidden buttons” within software that other applications can press to perform tasks. Developers utilize these interfaces to build integrations—for example, facilitating one app to pull data from another or allowing an AI agent to independently access third-party services. Numerous smart home devices and interconnected platforms include these hidden buttons, frequently without the casual user’s awareness. As AI agents progress, their ability to autonomously find and utilize these endpoints is increasing, revealing powerful—and sometimes unforeseen—automation capabilities.

When faced with a simple question, humans often instinctively respond—like “which is taller, a giraffe or a cat?” However, some queries may require breaking down intermediate steps to reach the correct answer. For instance, 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 breaking a problem down into smaller, sequenced steps to improve the precision of the final answer. This approach usually takes longer to generate a response but significantly increases the likelihood of accuracy, especially in logical or programming contexts. Reasoning models evolve from conventional large language models and are refined for chain-of-thought methods, largely through reinforcement learning.

(See: Large language model)

This term is more specific than “AI agent,” focusing on applications capable of performing tasks independently and sequentially to achieve specific results. A coding agent is a specialized type designed for software development. Unlike merely suggesting code for human review, a coding agent can autonomously craft, test, and debug software, handling the iterative challenges that developers typically face. These agents can navigate entire codebases, identify bugs, perform tests, and implement fixes with minimal human intervention. They operate like an ultra-efficient intern—focused and tireless, yet they still need human supervision.

While somewhat broad, compute mainly refers to the computational power behind AI models. This capacity drives the AI industry, enabling the training and deployment of powerful models. It often serves as shorthand for the types of hardware that provide this computational capability—like GPUs, CPUs, TPUs, and other infrastructure components essential to the contemporary AI landscape.

Deep learning, a subset of self-improving machine learning, focuses on developing AI algorithms within a layered artificial neural network (ANN) framework. This structure allows for the identification of more intricate patterns than basic machine learning methods such as linear models or decision trees. Deep learning systems draw inspiration from the complex pathways of neurons in the human brain.

Deep learning models can independently identify critical features within data, removing the need for human engineers to outline them. This structure also enables algorithms to learn from errors, continuously refining outputs through iterative adjustments. Yet, deep learning systems generally require substantial amounts of data for optimal performance (often millions of data points) and usually demand longer training times compared to simpler models—leading to higher development costs.

(See: Neural network)

Diffusion describes the technology utilized by numerous AI models to create art, music, and text. Drawing inspiration from physics, diffusion systems systematically “destroy” data structures—like images or melodies—by adding noise until they become unrecognizable. In physics, diffusion occurs spontaneously and irreversibly; sugar dissolved in coffee cannot turn back into its original cube form. Conversely, AI diffusion systems aim to master a “reverse diffusion” technique to recover distorted data, effectively learning how to reveal information obscured by noise.

Distillation represents a method of extracting insights from a large AI model using a “teacher-student” approach. Developers query a teacher model and document its outputs, often contrasting these results against a predetermined dataset for precision. These outputs are then employed to train the student model, which aims to mimic the teacher’s responses.

This process can produce a smaller, more efficient model that retains the performance of its larger counterpart while minimizing distillation loss. It is believed that this is how OpenAI developed GPT-4 Turbo, a quicker version of GPT-4.

While all AI organizations internally perform distillation, some may utilize it to enhance their functionalities relative to leading models. However, extracting insights from a competitor typically violates AI API service agreements.

This term refers to the retraining of an AI model to improve its efficacy for a more specific task or domain than initially intended—usually by incorporating new, specialized (task-relevant) data.

Many AI startups begin with large language models as their foundation for developing commercial products but aspire to enhance functionality for specific sectors or tasks by enriching prior training with domain-specific knowledge.

(See: Large language model [LLM])

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

These two models compete against one another. The generator attempts to produce outputs that can fool the discriminator, while the discriminator’s job is to distinguish genuine data from artificial creations. This structured competition can yield realistic AI outputs with minimal human intervention. However, GANs tend to be more effective in specific contexts (such as generating realistic images or videos) rather than being applicable on a broader scale.

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

Hallucination could lead to misleading outputs from generative AI, creating notable risks—possibly resulting in harmful consequences (e.g., unsafe medical advice stemming from a health-related question).

The issue of AIs generating fabricated information is thought to originate from gaps in their training data. In response to these hallucinations, there is a growing push towards developing increasingly specialized and/or vertical AI models—domain-specific AIs that require narrower expertise to fill knowledge gaps and mitigate 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 happen without training; a model must initially identify patterns within a dataset before effectively extrapolating from that training.

Different hardware configurations can perform inference, ranging from consumer-grade processors to powerful GPUs and specialized AI accelerators. However, not all setups execute models with the same level of efficiency; very large models would take significantly longer to provide 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 are engaging with a large language model that handles your inquiries directly or with support from various available tools, such as web browsers or code interpreters.

LLMs consist of deep neural networks containing billions of numerical parameters (or weights, see below) that discern relationships between words and phrases to create a representation of language—essentially, a sophisticated map of language.

These models are constructed by analyzing patterns found in vast amounts of documents, articles, and transcripts. When you submit a prompt to an LLM, the model generates the most likely response linked to that prompt.

(See: Neural network)

Memory cache denotes an essential process that enhances inference (how AI forms responses to user inquiries). Essentially, caching serves as an optimization technique aimed at improving inference efficiency. AI functions through intricate mathematical computations, and each calculation consumes resources. Caching reduces the number of computations a model must perform by retaining certain calculations for future interactions. Various types of memory caching exist, with KV (key-value) caching being among the most prevalent. KV caching is utilized in transformer-based models, boosting efficiency and yielding quicker results by minimizing the time (and computational effort) necessary to produce responses to user prompts.

(See: Inference)

Model Context Protocol (MCP) is an open standard created to allow AI models to connect with external tools and data—such as your files, databases, or applications like Slack and Google Drive—without requiring developers to create separate connectors for each integration. You can liken it to a USB-C port for AI. Introduced by Anthropic in 2024 and later endorsed by the Linux Foundation, MCP has gained traction among OpenAI, Google, and Microsoft, becoming one of the most rapidly adopted standards in today’s AI landscape.

The Mixture of Experts architecture segments a neural network into various smaller specialized sub-networks, or “experts,” activating only a portion of them for specific tasks. Rather than routing each request through the entire model—similar to calling your entire team for every question—the MoE model incorporates a built-in “router” that selects the most relevant specialists for the job. This design facilitates the creation of exceptionally large models that sustain relative efficiency and cost-effectiveness, as only a fraction of the network is operational at any given time. Mistral AI’s Mixtral model is a notable example; recent GPT models from OpenAI are believed to utilize some variation of this methodology, although the company has not confirmed it officially.

(See: Neural network, Deep learning)

A neural network offers the multi-layered algorithmic foundation for deep learning—and, more broadly, the proliferation of generative AI tools following the emergence of large language models.

The concept of modeling data processing algorithms after the intricately connected circuits of the human brain has existed since the 1940s; however, it was the recent explosion in graphical processing units (GPUs)—primarily driven by the gaming sector—that truly unleashed this theory’s potential. GPUs are designed specifically for training algorithms with multiple layers, enabling neural network-based AI systems to excel in various domains, including voice recognition, autonomous navigation, and drug discovery.

(See: Large language model [LLM])

Open source refers to software—or increasingly, AI models—where the underlying code is publicly available, allowing anyone to use, examine, or modify it. In the AI sphere, Meta’s Llama series of models stands out as a crucial example; Linux serves as a well-known benchmark among operating systems historically. Open-source methodologies empower researchers, developers, and organizations worldwide to build upon each other’s work, accelerating innovation and facilitating independent safety evaluations that closed systems often cannot easily provide. Closed source signifies that the code remains private; users can engage with the product but cannot inspect its internal mechanics, a characteristic often associated with OpenAI’s GPT models—which has sparked considerable debate within the AI community.

Parallelization entails executing multiple tasks simultaneously, as opposed to sequentially—similar to employing 10 employees to work on different facets of a project concurrently instead of having one person finish each task one after the other. In AI, parallelization is critical for both training and inference; modern GPUs are specifically structured to manage thousands of calculations all at once, which is foundational to their role within the industry. As AI systems grow in complexity and models expand, the ability to parallelize tasks across multiple chips and machines has emerged as a key factor in determining how swiftly and efficiently models can be developed and deployed. Research into improved parallelization methods has evolved into its own dedicated 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 nearly all technology products we rely on daily. As the AI sector has surged, major tech companies and AI labs have been acquiring RAM at unprecedented rates to bolster their data centers, leading to decreased availability in the broader market. This supply limitation is driving up RAM prices.

This shortage influences various sectors, such as gaming (where companies are forced to raise console prices due to challenges in sourcing memory chips for their devices), consumer electronics (where memory scarcities could potentially lead to the lowest smartphone shipments in over a decade), and general corporate computing (as businesses wrestle to secure sufficient RAM for their data operations). The rising prices are unlikely to stabilize until that feared shortage resolves, but sadly, indications of a swift resolution appear minimal.

Similar to AGI, recursive self-improvement denotes a potential milestone for AI advancement, allowing minimal human involvement. In this scenario, AI models begin enhancing their capabilities independently, yielding substantial improvements in autonomy and functionality. Some narratives compare this occurrence to a singularity moment—when AI models become impervious to external influence. However, recursive self-improvement raises a fundamental question: can an AI model develop its successor? This query streamlines the task for engineers attempting to create such models. Many emerging AI startups center their efforts on building recursively self-improving models while often downplaying the potentially catastrophic outcomes, presenting recursive self-improvement simply as the next frontier in research.

Reinforcement learning is an AI training method where a system learns through trial and error, receiving rewards for correct actions—akin to training a pet with treats, though in this instance, 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 refine its behavior based on the feedback it gathers. This method has proven especially effective for training AI to play games, manage robotics, and, more recently, enhance the reasoning capabilities of large language models. Techniques such as reinforcement learning from human feedback (RLHF) have become vital in how leading AI labs fine-tune their models for enhanced effectiveness, precision, and safety.

When navigating conversations between humans and machines, significant challenges emerge—humans communicate in natural language, while AI systems function through intricate algorithmic processes grounded in data. Tokens act as the bridge between these two domains: 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 divides raw text into manageable parts that a language model can understand, much like a compiler translates human language into binary code for computers. In business scenarios, tokens also influence costs—most AI companies charge for LLM usage on a per-token basis, implying that increased usage correlates with heightened expenses.

Thus, tokens are the small text units—often parts of words rather than complete phrases—that AI language models analyze before processing; they can be likened to “words” concerning AI workloads. Throughput denotes the amount that can be processed in a given timeframe, so token throughput serves as a metric for the volume of AI work a system can handle simultaneously. High token throughput is a crucial objective for AI infrastructure teams, as it affects how many users a model can serve simultaneously and the speed at which they receive responses. AI researcher Andrej Karpathy has articulated concerns when his AI subscriptions remain idle—reflecting on his experiences as a graduate student when costly computing hardware was underutilized—highlighting why maximizing token throughput has become an urgent goal in the field.

Training machine learning AIs involves a process known as training. This primarily entails supplying data to allow the model to recognize patterns and generate desired outputs. It represents the system’s response to the data’s characteristics, enabling it to align outputs towards a sought-after outcome—whether that involves identifying images of cats or composing a haiku in response to a prompt.

Training can be resource-heavy due to the vast amounts of inputs needed, which have been consistently increasing—hence the value of hybrid strategies, such as combining rules-based AIs with targeted data, to manage costs without starting from scratch.

[See: Inference]

This technique applies a previously trained AI model as a foundation for developing a new model targeted at a different but typically related task, facilitating the reapplication of insights gathered from earlier training cycles.

Transfer learning can enhance model development efficiency. It is especially advantageous when the data available for the target task is somewhat limited. However, it is important to note that this approach has limitations; models employing transfer learning to acquire generalized skills will likely require additional training with fresh data to excel in their specific areas of focus.

(See: Fine tuning)

Validation loss serves as a metric reflecting an AI model’s learning progress during training—lower values are more desirable. Researchers closely track this indicator as a real-time performance gauge, informing decisions about when to stop training, modify hyperparameters, or investigate potential issues. A significant concern that validation loss helps reveal is overfitting, whereby a model memorizes its training data rather than genuinely grasping applicable patterns for new scenarios. Consider validation loss as a tool distinguishing between a student who understands the material and one who merely memorized last year’s exam—it illustrates the development trajectory of your model.

Weights are integral to AI training, determining the importance (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 signifying which elements within a dataset are most critical for the specific training task. They achieve this effect through multiplication with inputs. Training generally commences with randomly initialized weights, but as the process unfolds, these weights are fine-tuned as the model aims to produce outputs that closely match the desired outcomes.

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

Ultimately, the weights assigned reflect how much each of these factors affects a property’s value according to the available dataset.

This article is regularly updated with new information.