Essential AI Concepts to Understand in 2023

The field of artificial intelligence is evolving rapidly, leading to an array of new terms that aim to clarify its intricacies. Recent tech conferences and product launches have emphasized terms like LLMs, RAG, and RLHF, leaving even seasoned professionals confused. This glossary aims to demystify these terms by providing straightforward definitions of the AI jargon you may encounter—whether you’re researching these technologies, investing in them, or staying informed with updates from outlets like TechCrunch or various podcasts. We will continuously update this glossary to keep in line with the changing AI landscape.

---

Artificial general intelligence (AGI) generally refers to AIs that outperform humans across a broad spectrum of tasks. Sam Altman, CEO of OpenAI, compares AGI to “the average employee you might hire.” OpenAI describes AGI as “highly autonomous systems that excel in most economically important tasks,” while Google DeepMind defines it as “AI capable of performing a majority of cognitive tasks at human levels.” If this definition feels somewhat ambiguous, you’re not alone; many in the AI research community feel similarly.

An AI agent is a system that utilizes AI technologies to carry out various tasks, extending beyond mere chatbot functions. These tasks may include budget management, itinerary organization, and code generation. The definition of “AI agent” can differ among various stakeholders, and its underlying architecture is continually developing to meet increasingly demanding challenges. Essentially, an AI agent operates as an independent entity that incorporates diverse AI technologies to tackle complex, multi-dimensional tasks.

API endpoints act like “invisible buttons” within software that other applications use to execute different operations. Developers leverage these interfaces for seamless integration, allowing one application to pull data from another or enabling an AI agent to function autonomously with external services. A multitude of smart home devices and interconnected systems utilize these subtle controls, often without the average user’s awareness. As AI agents progress, their ability to identify and interact with these endpoints will broaden, unlocking remarkable—and sometimes unexpected—automation capabilities.

Humans usually respond instinctively to simple queries—like “Which is taller, a giraffe or a cat?” Yet, some queries require breaking down intermediary steps to arrive at the correct answer. For instance, if a farmer has 40 animals (chickens and cows) resulting in a total of 120 legs, calculations are needed to determine the number of each animal type (20 chickens and 20 cows).

In AI, chain-of-thought reasoning utilized by large language models involves decomposing problems into smaller, sequential steps to improve the accuracy of the final answer. Although this approach may take longer, it substantially enhances precision, especially for logical or programming tasks. Reasoning models signify a shift from traditional large language models and are fine-tuned for chain-of-thought tasks largely via reinforcement learning.

(See: Large language model)

This term is narrower than “AI agent,” focusing on applications that can independently perform tasks to achieve specific objectives. A coding agent serves as a prime example of an AI designed for software development. Rather than merely generating code for human review, a coding agent can execute, test, and debug software autonomously, tackling the iterative challenges developers encounter. These agents can navigate entire codebases, identify bugs, run tests, and apply fixes with minimal human input, akin to diligent interns—hardworking and persistent, yet still requiring some oversight.

While somewhat broad, compute typically refers to the computational power driving AI models. This capability is essential in the AI field, enabling the training and operation of intricate models. It often pertains to the types of hardware that provide this computational strength—such as GPUs, CPUs, TPUs, and other infrastructures critical to today’s AI landscape.

Deep learning, a subset of self-improving machine learning, focuses on developing AI algorithms that employ a layered artificial neural network (ANN) structure. This method allows for the identification of more complex patterns compared to traditional training techniques, like linear models or decision trees. Deep learning systems draw inspiration from the complex neuronal networks present in the human brain.

Deep learning models can autonomously identify key features within data, minimizing the necessity for human engineers to specify them. This architecture also enables algorithms to learn from errors, systematically refining outputs through iterative modifications. However, deep learning systems typically require large datasets to function effectively (often including millions of data points) and generally undergo longer training durations compared to simpler models, which can escalate development costs.

(See: Neural network)

Diffusion refers to the techniques employed by various AI models to create art, music, and text. Borrowed from physics, diffusion systems systematically “degrade” data structures—like images or sounds—by introducing noise until they lose definition. In physics, diffusion is a natural, irreversible process; think of how sugar dissolves in coffee, rendering it impossible to revert to its original state. However, AI diffusion systems aim to master a “reverse diffusion” technique to restore clarity from noise-obscured data.

Distillation is the process of extracting insights from a large AI model using a “teacher-student” approach. Developers collaborate with a teacher model and meticulously document its outputs, often assessing them against a specific dataset for accuracy. These insights are then used to train a student model that emulates the teacher’s responses.

This process can create a more compact, efficient model that retains the capabilities of its larger counterpart while minimizing distillation loss. It is widely believed that this method contributed to the development of GPT-4 Turbo, a faster version of GPT-4 crafted by OpenAI.

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

This term signifies the retraining of an AI model to enhance its performance on a specific task or domain that was not its original focus—typically achieved by incorporating relevant data into the training process.

Many AI startups commence with large language models as the backbone of their products, seeking to enhance functionalities for specific sectors or tasks by integrating domain-specific insights into their existing training.

(See: Large language model [LLM])

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

These networks maintain competitive dynamics; the generator aims to produce outputs that can deceive the discriminator, while the discriminator strives to distinguish between authentic data and generated content. This structured rivalry can yield convincingly authentic AI-generated results with minimal human intervention. However, GANs tend to perform better under specific conditions (like generating realistic images or videos) rather than universally.

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

Hallucinations can result in misleading outputs from generative AI, presenting serious risks—like offering unsafe medical advice in response to health-related inquiries.

The tendency for AIs to generate fictitious data is believed to stem from gaps in their training datasets. Consequently, there is a growing demand for more specialized and/or vertical AI models—domain-specific AIs necessitating refined expertise to address knowledge gaps and mitigate misinformation challenges.

Inference signifies the operation of an AI model, allowing it to make predictions or draw conclusions based on previously recognized data. Notably, inference cannot occur without prior training; a model must first learn to recognize patterns within a dataset before it can effectively apply 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; for instance, particularly large models would take significantly longer to yield predictions on a laptop than on a cloud server equipped with advanced AI chips.

[See: Training]

Large language models, or LLMs, are foundational to 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 engage with a large language model specifically designed to respond to your queries or utilize available tools like web browsers or code interpreters.

LLMs consist of deep neural networks containing 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 examining patterns in vast 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). Essentially, caching serves as an optimization technique aimed at improving inference performance. AI tasks involve complex mathematical operations, each demanding resources. Caching reduces the number of computations a model needs to perform by storing specific calculations for future use. Various forms of memory caching exist, with KV (key-value) caching being one of the most prevalent. KV caching is utilized in transformer-based models to enhance efficiency and deliver faster results by minimizing the time (and computational effort) required to respond to user inquiries.

(See: Inference)

The Model Context Protocol (MCP) is an open standard designed to streamline interactions between AI models and external tools and data—such as your files and applications like Slack and Google Drive—without necessitating developers to create custom connectors for each 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 traction among OpenAI, Google, and Microsoft, evolving into an established standard in today’s AI ecosystem.

The Mixture of Experts (MoE) architecture divides a neural network into smaller, specialized sub-networks, or “experts,” activating only a selected subset for particular tasks. Instead of routing every request through the entire model—similar to consulting a full team for each task—the MoE model utilizes an internal “router” to identify the most suitable experts for a task. This approach facilitates the creation of exceptionally large models while maintaining efficiency and cost-effectiveness, as only a small portion of the network is operational at any given time. Mistral AI’s Mixtral model exemplifies this; recent GPT models from OpenAI are rumored to employ a variant of this technique, although the company has not publicly validated it.

(See: Neural network, Deep learning)

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

The concept of modeling data processing algorithms on the intricately connected circuits of the human brain has existed since the 1940s; however, the recent rise in graphical processing units (GPUs)—primarily driven by the gaming industry—has genuinely unleashed this potential. GPUs are specifically architected to train algorithms with multiple layers, empowering AI systems reliant on neural networks to excel in applications like voice recognition, autonomous driving, and drug discovery.

(See: Large language model [LLM])

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

Parallelization involves executing multiple tasks simultaneously rather than sequentially—like assigning ten individuals to manage various project components concurrently instead of relying on one person to complete each task in line. In AI, parallelization is essential for both training and inference; modern GPUs are engineered to handle thousands of computations simultaneously, underpinning their crucial role in the industry. As AI systems escalate in complexity and models expand, the ability to parallelize tasks across numerous chips and machines has become vital for the speed and efficiency of model development and implementation. Research into advanced parallelization techniques has emerged as a significant area of study.

RAMageddon is a humorous term that captures a notable trend impacting the tech industry: the ongoing shortage of random access memory (RAM) chips, which are crucial for nearly every tech product. As the AI sector flourishes, leading tech firms and AI labs have been acquiring RAM at unprecedented levels to enhance their data centers, resulting in reduced availability in the broader market. This supply constraint is driving up RAM prices.

This shortage affects various sectors, including gaming (where companies face increased console prices due to challenges in obtaining memory chips for their devices), consumer electronics (where memory limitations could lead to the lowest smartphone shipments in over a decade), and overall corporate computing (as businesses strive to ensure adequate RAM for their data tasks). Price hikes are likely to continue until this anticipated shortage is resolved; unfortunately, a swift solution does not seem imminent.

Similar to AGI, recursive self-improvement indicates a potential breakthrough in AI progress, enabling minimal human oversight. In this context, AI models begin to independently enhance their functionalities, leading to significant improvements in autonomy and performance. Some narratives liken this scenario to a singularity moment—when AI systems become self-sufficient. However, recursive self-improvement raises a critical question: can an AI model create its successor? This dilemma highlights the challenges that engineers face in developing such systems. Numerous emerging AI startups aim to construct models capable of achieving recursive self-improvement, often striving to minimize potential catastrophic outcomes, viewing it instead as the next frontier in research.

Reinforcement learning is a training methodology for AI systems wherein 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” symbolizes a neural network and the “treat” signifies a mathematical signal indicating success. In contrast, supervised learning involves training a model on a specific dataset of labeled instances, whereas reinforcement learning allows a model to engage with its environment, carry out actions, and iteratively modify its behavior based on the feedback received. This method has proven particularly effective for training AI in games, robotics, and, most recently, enhancing reasoning capabilities in large language models. Strategies like reinforcement learning from human feedback (RLHF) are now central to how leading AI labs fine-tune their models for optimized performance, accuracy, and safety.

In human-machine interactions, notable challenges arise—humans communicate using natural language, while AI systems function through complex algorithmic processes rooted in data. Tokens act as the bridge between these two realms: they encompass the fundamental units of human-AI interaction, representing distinct segments of data processed or generated by an LLM. Tokens are produced 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 contexts, tokens also carry financial implications—many AI firms charge for LLM usage based on token consumption, meaning increased usage results in higher costs.

Thus, tokens are small pieces of text—often portions 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 of how much AI workload a system can handle simultaneously. Achieving high token throughput is a crucial goal for AI infrastructure teams, influencing the number of users a model can support concurrently and the speed at which they receive responses. AI researcher Andrej Karpathy has expressed concerns about underutilized AI subscriptions—drawing from his experiences as a graduate student when expensive computing hardware was applied inefficiently—emphasizing why optimizing token throughput has become a pressing priority in the industry.

Training machine learning AIs is known as training. This involves inputting data into the model, enabling it to recognize patterns and produce desired outputs. It represents the system’s response to the data features, directing its outputs towards a specific goal—whether that entails classifying images of cats or crafting a haiku in response to a prompt.

Training can demand significant resources due to the vast amounts of input required, which have continually escalated—especially with the advent of hybrid approaches, such as combining rule-based AIs with curated data, to manage costs without restarting the training cycle.

[See: Inference]

This method employs an already trained AI model as a foundation to create a new model tailored for a different yet often related task, permitting the reapplication of insights gained from prior training cycles.

Transfer learning can enhance model development efficiency. It proves especially beneficial when data availability for the target task is limited. However, it’s essential to note that this approach also has downsides; models that utilize transfer learning to acquire general skills will likely require further training with new data to excel in their specific areas of expertise.

(See: Fine-tuning)

Validation loss denotes a metric that reflects an AI model’s learning journey during training—ideally, lower values indicate better performance. Researchers carefully monitor this metric as a real-time performance gauge, impacting decisions regarding when to stop training, modify hyperparameters, or investigate potential issues. A primary concern that validation loss helps clarify is overfitting, where a model memorizes its training data instead of genuinely understanding applicable patterns for new situations. Think of validation loss as a tool that distinguishes between a student who grasps the material and one who merely memorizes past exam questions—it illustrates your model’s development pathway.

Weights are crucial components of AI training, determining the significance (or weight) assigned to various attributes (or input variables) within the training data, ultimately affecting the outputs generated by the AI model.

In essence, weights are numerical parameters indicating which elements within a dataset are most relevant to the assigned training task. They operate by multiplication with inputs. Training typically starts with randomly assigned weights, but as the process progresses, these weights are adjusted to produce 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 like 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 regularly updated with new insights.