Essential AI Concepts to Understand in 2023

The field of artificial intelligence is advancing rapidly, bringing forth a multitude of new terms designed to clarify its intricacies. Recent technology events and product launches have highlighted concepts such as LLMs, RAG, and RLHF, leaving even seasoned professionals perplexed. This glossary aims to demystify these terms by providing straightforward definitions of the AI terminology you may encounter—whether you’re researching these technologies, investing in them, or keeping up with news from sources like TechCrunch or various podcasts. We will continuously update this glossary to keep pace with the evolving AI landscape.

---

Artificial general intelligence (AGI) refers to AI systems that outperform humans across a broad spectrum of tasks. Sam Altman, CEO of OpenAI, equates AGI to “the average employee you might hire.” OpenAI describes AGI as “highly autonomous systems that excel in most economically relevant tasks,” while Google DeepMind defines it as “AI capable of performing a majority of cognitive tasks at human levels.” If this explanation appears vague, you are not alone; many in the AI research community share a similar view.

An AI agent is a system that employs AI technologies to accomplish various tasks, extending beyond the basic functionalities of a chatbot. These tasks may include budget management, itinerary organization, and code generation. Definitions of “AI agent” can differ among various stakeholders, and its underlying architecture is continually evolving to address increasingly complex challenges. In essence, an AI agent operates as an autonomous entity that integrates different AI technologies to tackle intricate, multi-layered tasks.

API endpoints serve as “invisible buttons” within software that enable other applications to perform different functions. Developers utilize these interfaces for seamless integration, allowing one application to access data from another or permitting an AI agent to interact independently with external services. Numerous smart home devices and interconnected systems use these subtle controls, frequently without the average user’s awareness. As AI agents mature, their capability to recognize and engage with these endpoints will enhance, unlocking extraordinary—and occasionally unforeseen—automation possibilities.

Humans typically react instinctively to straightforward questions—such as “Which is taller, a giraffe or a cat?” However, certain inquiries require breaking down intermediary steps to find the correct answer. For instance, if a farmer has 40 animals (chickens and cows) that total 120 legs, calculations are necessary to determine the number of each type of animal (20 chickens and 20 cows).

In AI, chain-of-thought reasoning employed by large language models involves dissecting problems into smaller, sequential steps to enhance the accuracy of the final outcome. While this approach may take more time, it significantly boosts precision, especially for logical or programming tasks. Reasoning models represent a shift from traditional large language models and are optimized for chain-of-thought tasks mainly through reinforcement learning.

(See: Large language model)

This term is narrower than “AI agent,” focusing on applications that can autonomously complete tasks to achieve specific objectives. A coding agent is an example of an AI designed for software development. Instead of merely generating code for human review, a coding agent can independently execute, test, and debug software, addressing the iterative challenges developers face. These agents can traverse entire codebases, detect bugs, run tests, and implement fixes with minimal human involvement, akin to diligent interns—committed and tenacious, yet still requiring some oversight.

Though somewhat broad, compute generally refers to the computational power propelling AI models. This capability is essential in the AI domain, enabling the training and functioning of intricate models. It often relates to specific hardware types that provide this computational muscle—such as GPUs, CPUs, TPUs, and other infrastructures pivotal to the current AI landscape.

Deep learning, a subset of self-improving machine learning, revolves around developing AI algorithms that utilize a layered artificial neural network (ANN) framework. This approach facilitates the recognition of more complex patterns compared to traditional training methodologies, such as linear models or decision trees. Deep learning systems are inspired by the intricate neuronal structures found in the human brain.

Deep learning models can autonomously identify critical features within data, lessening the need for human engineers to specify them. This architecture also allows algorithms to learn from mistakes, systematically refining outputs through iterative improvements. However, deep learning systems typically require extensive datasets to function effectively (often made up of millions of data points) and generally endure prolonged training durations compared to simpler models, which can escalate development costs.

(See: Neural network)

Diffusion refers to techniques employed by various AI models to generate art, music, and text. Borrowed from physics, diffusion systems methodically “degrade” data structures—such as images or sounds—by adding noise until they lose clear definition. In physics, diffusion is a natural, irreversible process; imagine how sugar dissolves in coffee, making it impossible to revert to its initial state. Conversely, AI diffusion systems aim to master a “reverse diffusion” technique to restore clarity from noise-affected data.

Distillation is the method of extracting insights from a large AI model through a “teacher-student” framework. Developers interact with a teacher model and diligently document its outputs, often evaluating them against a specified dataset for accuracy. These insights are then used to train a student model that mimics the teacher’s responses.

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

While all AI organizations utilize distillation internally, some may implement it to refine their features compared to existing models. However, gaining insights from a competitor may violate AI API service agreements.

This term indicates the retraining of an AI model to enhance its performance on a particular task or domain that was not initially its focus—usually accomplished by integrating relevant data into the training process.

Many AI startups begin with large language models as the core of their products, aiming to enhance functionalities for specific sectors or tasks by incorporating domain-specific insights into their previous training.

(See: Large language model [LLM])

GAN, or Generative Adversarial Network, is a machine learning framework vital for developing generative AI, particularly for creating realistic outputs, including deepfakes. GANs consist of two neural networks: one generates outputs based on its training, while the other evaluates them.

These networks maintain a competitive dynamic; the generator strives to create outputs that can deceive the discriminator, while the discriminator tries to differentiate between authentic data and generated content. This structured rivalry can yield remarkably lifelike AI-generated results with minimal human intervention. However, GANs typically excel under certain conditions (like generating realistic images or videos) rather than universally.

Within the AI community, “hallucination” refers to instances when AI models produce inaccuracies—essentially fabricating false information. This raises significant reliability concerns regarding AI.

Hallucinations can lead to misleading outputs from generative AI, posing serious risks—such as providing unsafe medical advice in response to health-related questions.

The tendency for AIs to generate fictitious data is thought to stem from gaps in their training datasets. As a result, there is an increasing demand for more specialized and/or vertical AI models—domain-specific AIs that require refined expertise to address knowledge gaps and mitigate misinformation challenges.

Inference signifies how an AI model operates, enabling it to make predictions or draw conclusions based on previously recognized data. Importantly, inference cannot occur without prior training; a model must first learn to identify patterns within a dataset before effectively applying that knowledge.

Various hardware configurations can facilitate inference, ranging from consumer-grade processors to high-performance GPUs and specialized AI accelerators. However, not all configurations can process models with the same efficiency; particularly 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, underpin popular AI assistants such as 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 explicitly designed to respond to your queries or utilize available tools like web browsers or code interpreters.

LLMs consist of deep neural networks with billions of numerical parameters (or weights) that analyze relationships between words and phrases, generating language representations—essentially detailed mappings of language.

These models are trained by analyzing patterns in extensive datasets of documents, articles, and transcripts. When prompted, an LLM provides what it interprets as the most relevant response to that input.

(See: Neural network)

Memory cache is a crucial process that optimizes inference (how AI formulates responses to user prompts). Generally, caching acts as an optimization tactic designed to enhance inference performance. AI tasks involve complex mathematical operations, each requiring resources. Caching decreases the number of computations a model must execute by storing specific calculations for future use. There are various forms of memory caching, with KV (key-value) caching being among the most prevalent. KV caching is employed in transformer-based models to increase efficiency and deliver quicker results by minimizing the time (and computational effort) needed to respond to user inquiries.

(See: Inference)

The Model Context Protocol (MCP) is an open standard created to simplify interactions between AI models and external tools and data—such as your files and applications like Slack and Google Drive—without requiring developers to construct custom connectors for every integration. Think of it as a USB-C port for AI. Introduced by Anthropic in 2024 and later adopted by the Linux Foundation, MCP has quickly gained popularity among OpenAI, Google, and Microsoft, evolving into a recognized standard in today’s AI ecosystem.

The Mixture of Experts (MoE) architecture segments a neural network into smaller, specialized sub-networks, or “experts,” activating only a selected subset for specific tasks. Instead of routing every request through the full model—similar to consulting an entire team for each task—the MoE model uses an internal “router” to identify the most appropriate experts for a task. This strategy facilitates the creation of exceptionally large models while maintaining efficiency and cost-effectiveness, as only a fraction of the network is operational at any time. Mistral AI’s Mixtral model exemplifies this; recent GPT models from OpenAI are rumored to utilize a variation of this technique, though the company has not formally confirmed it.

(See: Neural network, Deep learning)

A neural network serves as the multilayered algorithmic framework for deep learning—and more broadly, for the advent of generative AI tools following the introduction of large language models.

The concept of modeling data processing algorithms on the complex interconnections of the human brain has existed since the 1940s; however, the recent surge in graphical processing units (GPUs)—primarily driven by the gaming industry—has truly unlocked this potential. GPUs are specifically tailored to train algorithms with multiple layers, enabling AI systems reliant on neural networks to excel in areas like voice recognition, autonomous driving, and drug discovery.

(See: Large language model [LLM])

Open source describes software—or increasingly, AI models—whose source code is available to the public, allowing anyone to use, inspect, or modify it. In the AI domain, Meta’s Llama series of models stands out as a prominent example, while Linux serves as a benchmark operating system. Open-source practices encourage collaboration among global researchers, developers, and organizations, fostering innovation and facilitating independent safety assessments often absent in closed systems. Closed source refers to proprietary software; users can utilize the product but cannot examine its internal mechanisms, a characteristic often associated with OpenAI’s GPT models—sparking considerable debate within the AI community.

Parallelization involves executing multiple tasks concurrently rather than sequentially—like assigning ten individuals to manage different project elements at the same time instead of relying on one person to finish each task in succession. In AI, parallelization is critical for both training and inference; modern GPUs are designed to handle thousands of computations simultaneously, highlighting their pivotal role in the industry. As AI systems increase in complexity and models expand, the capability to parallelize tasks across numerous chips and machines has become vital for the speed and efficiency of model development and deployment. Research into advanced parallelization techniques has emerged as a significant area of study.

RAMageddon is a playful term that captures a major trend affecting the tech industry: the ongoing shortage of random access memory (RAM) chips, which are essential for nearly every tech product. As the AI sector flourishes, leading tech companies and AI laboratories have been acquiring RAM at unprecedented levels to enhance their data centers, leading to diminished availability in the broader market. This supply limitation is driving up RAM prices.

This shortage impacts numerous sectors, including gaming (where companies are grappling with increased console prices due to difficulties in sourcing memory chips for their devices), consumer electronics (where memory constraints could result in the lowest smartphone shipments in over a decade), and overall corporate computing (as businesses strive to secure adequate RAM for their data tasks). Price hikes are projected to continue until this anticipated shortage is resolved; unfortunately, a swift resolution does not appear forthcoming.

Similar to AGI, recursive self-improvement signals a potential breakthrough in AI development, allowing for minimal human oversight. In this context, AI models start to independently enhance their functionalities, leading to notable improvements in autonomy and performance. Some narratives liken this scenario to a singularity moment—when AI systems achieve self-sufficiency. However, recursive self-improvement raises a critical question: can an AI model create its successor? This dilemma highlights the challenges engineers face in the creation of such systems. Many emerging AI startups are working towards building models capable of achieving recursive self-improvement, frequently aiming to minimize potential catastrophic outcomes, viewing it as the next frontier in research.

Reinforcement learning is a training methodology for AI where a model learns through trial and error, receiving rewards for successful actions—similar to how one might train a pet with treats, where the “pet” signifies a neural network and the “treat” represents a mathematical signal indicating success. In contrast, supervised learning involves training a model on a particular dataset of labeled instances, while reinforcement learning allows a model to interact with its environment, perform actions, and iteratively adjust its behavior based on received feedback. This method has proven particularly effective for training AI in games, robotics, and, most recently, in enhancing reasoning capabilities in large language models. Techniques like reinforcement learning from human feedback (RLHF) are now central to how leading AI laboratories refine their models for improved performance, accuracy, and safety.

In human-machine interactions, notable challenges emerge—humans communicate using natural language, while AI systems function through complex algorithmic processes rooted in data. Tokens act as the bridge connecting these two domains: they represent the fundamental units of human-AI interaction, encompassing distinct data segments processed or generated by an LLM. Tokens are created through tokenization, breaking down raw text into manageable components that a language model can analyze, akin to how a compiler translates human language into binary for computers. In commercial scenarios, tokens also carry financial implications—many AI companies charge for LLM usage based on token consumption, meaning increased usage leads to higher costs.

Therefore, tokens are small units of text—often parts of words rather than complete phrases—that AI language models analyze prior to processing; they can be likened to “words” in AI applications. Throughput denotes the amount that can be processed within a certain timeframe, with token throughput serving as a metric for how much AI workload a system can manage concurrently. Achieving high token throughput is a critical goal for AI infrastructure teams, influencing the number of users a model can support simultaneously and the speed at which they receive responses. AI researcher Andrej Karpathy has raised concerns regarding underutilized AI subscriptions—drawing from his experiences as a graduate student when costly computing hardware was misapplied—emphasizing why optimizing token throughput has become an urgent focus in the industry.

Training machine learning AIs is referred to as training. This process involves inputting data into the model, enabling it to identify patterns and produce desired outputs. It signifies the system’s response to the data characteristics, steering its outputs toward a specific aim—whether that entails categorizing images of cats or composing a haiku in response to a prompt.

Training can demand significant resources due to the extensive volumes of input required, which have continually escalated—especially with the emergence of hybrid techniques, such as integrating rule-based AIs with curated data, to manage costs without restarting the training cycle.

[See: Inference]

This method utilizes an already trained AI model as a foundation to create a new model tailored for a different but often related task, allowing for the reapplication of insights gained from previous training cycles.

Transfer learning can enhance model development efficiency. It proves especially useful when data availability for the target task is limited. However, it’s important to acknowledge that this approach also has constraints; models that employ transfer learning to acquire general skills will likely require additional training with new data to excel in their particular fields of expertise.

(See: Fine-tuning)

Validation loss represents a metric that reflects an AI model’s learning progress during training—ideally, lower values indicate better performance. Researchers closely monitor this metric as a real-time performance indicator, guiding decisions on when to conclude training, adjust hyperparameters, or investigate potential issues. A primary concern that validation loss aids in clarifying is overfitting, where a model memorizes its training data instead of genuinely comprehending applicable patterns for new situations. Consider validation loss as a tool that distinguishes between a student who grasps the content and one who merely memorizes past exam questions—it illustrates your model’s growth trajectory.

Weights are pivotal components of AI training, determining the significance (or weight) assigned to various attributes (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 are most relevant to the specified training task. They operate through multiplication with inputs. Training typically begins with randomly assigned weights, but as the process evolves, these weights are adjusted to yield outputs that closely align with the desired results.

For example, an AI model tasked with predicting housing prices based on historical real estate data for a specific area might assign weights to factors such as the number of bedrooms and bathrooms, whether a property is semi-detached or detached, and the presence of parking or a garage.

Ultimately, the assigned weights reflect the importance of each of these factors in determining a property’s value based on the available dataset.

This article is periodically updated with new insights.