Confused by AI Terminology? Let’s Clear It Up Together!

Artificial intelligence is reshaping our world, creating new terminology to explain its processes. Within minutes of exploring AI, terms like LLMs, RAG, and RLHF emerge, often leaving even the most knowledgeable tech professionals feeling uncertain. This glossary aims to clarify these concepts. Regular updates reflect the field’s evolution, mirroring the dynamic nature of the AI systems discussed.

---

Artificial general intelligence, abbreviated as AGI, is an ambiguous yet significant concept. Generally, it describes an AI that surpasses the average human’s capabilities across various tasks. Sam Altman, CEO of OpenAI, once likened AGI to a “median human you could hire as a co-worker.” Meanwhile, OpenAI’s charter depicts AGI as “highly autonomous systems that outperform humans in most economically valuable tasks.” Google DeepMind has its own view, considering AGI as “AI that matches humans in most cognitive activities.” If this has left you puzzled, you’re not alone—even experts are uncertain.

An AI agent is an advanced tool leveraging AI capabilities to undertake complex tasks on your behalf, extending beyond the functions of basic chatbots. Examples include managing expenses, making reservations, or writing code. As noted previously, the landscape of AI agents is continually evolving, leading to varied interpretations of the term. The necessary infrastructure to support these capabilities is still under development, but the essence remains: a system that autonomously utilizes multiple AI technologies to achieve intricate tasks.

API endpoints can be considered as the “buttons” behind software that can be operated by other applications to perform specific actions. Developers leverage these interfaces to enable functionalities—like allowing one application to access data from another or allowing an AI agent to directly manage third-party services without manual input. Many smart devices come equipped with these concealed interfaces, even though most users are unaware of their existence. As AI agents become more sophisticated, their ability to autonomously locate and utilize these endpoints expands, unlocking new avenues for automation.

Humans can quickly answer simple questions—like determining which is taller, a giraffe or a cat—often without much thought. However, some queries necessitate a more methodical approach. For example, in a situation where a farmer has a mix of chickens and cows resulting in 40 heads and 120 legs, one might need to write equations to find the solution (20 chickens and 20 cows).

In AI terms, chain-of-thought reasoning for large language models involves disassembling a problem into smaller, sequential components to enhance the accuracy of the final answer. While this process may take longer, it generally yields more reliable outcomes, particularly in logical or programming contexts. Models focused on reasoning have emerged from traditional language models and are refined for this kind of thought process through reinforcement learning.

(See: Large language model)

A more targeted term than “AI agent,” a coding agent delivers specific functions in software development. Unlike conventional coding suggesters, coding agents autonomously write, test, and debug code, managing the iterative and often tedious tasks that consume a developer’s time. They can oversee entire codebases, identify bugs, carry out tests, and implement fixes with minimal human input, akin to employing a relentless intern who remains focused and efficient—though human oversight is still essential.

While somewhat broad, the term compute refers to the essential computational power needed for AI models to function. This processing capacity is what fuels the AI sector, empowering the training and deployment of complex models. It often denotes the hardware that provides this power, such as GPUs, CPUs, TPUs, and other critical infrastructure that supports the current AI landscape.

Deep learning, a subset of advanced machine learning, incorporates a structured framework utilizing multiple layers of artificial neural networks (ANNs). This design allows these models to draw more intricate correlations than simpler machine learning systems, like linear models or decision trees. Inspired by the complex networks of human neurons, deep learning algorithms autonomously identify crucial data features without requiring human definitions. They also support self-improvement through error recognition, enhancing outputs via repetitive cycles. However, to function effectively, these models demand vast datasets, often requiring millions of data points, with extended training times that translate to skyrocketing development costs.

(See: Neural network)

Diffusion stands at the core of various art, music, and text-generating AI technologies. Drawing on principles from physics, diffusion systems progressively deteriorate the integrity of input data—like images or sound—by introducing noise until the original structure is lost. In physics, this process is spontaneous and cannot be reversed, much like sugar dissolving in coffee. However, AI diffusion systems aim to master a type of “reverse diffusion,” allowing them to reconstruct the original data from the noise introduced.

Distillation extracts insights from larger AI models through a teacher-student dynamic. Developers direct requests to a teacher model and observe the outputs, sometimes comparing these results to established datasets for accuracy. The outcomes serve as training inputs for a student model, designed to replicate the teacher’s behaviors.

This process can yield a more streamlined and efficient model based on a larger counterpart while minimizing distillation loss. It likely played a role in the creation of faster variants, like GPT-4 Turbo from OpenAI. While distillation is a common internal practice among AI developers, it may also serve as a contentious means for some companies to catch up with leading models, often violating AI API terms of service.

Fine-tuning addresses the optimization of an AI model for a more specific application beyond its original training focus through the introduction of new, specialized data.

Numerous AI startups utilize large language models as a foundation to develop commercial products, aiming to enhance functionality for distinct sectors by supplementing earlier training phases with domain-specific insights.

(See: Large language model [LLM])

Generative Adversarial Networks (GANs) exemplify a machine learning structure that shapes advancements in generative AI, particularly in generating realistic data—including deepfake technologies. GANs consist of two neural networks: one generates outputs based on training data, while the other assesses whether the data appears genuine.

The two networks compete to enhance one another, with the generator striving to craft outputs that deceive the discriminator, while the discriminator identifies artificial data. This competition can refine AI outputs, enhancing realism without requiring extra human input. However, GANs operate best within specific contexts rather than general-purpose AI.

Hallucination refers to the occurrence of AI models generating incorrect information—essentially fabricating responses. This quality poses significant concerns for the reliability of AI outputs.

The phenomenon of hallucination can lead to misleading generative AI outputs, presenting real-world risks—such as providing harmful medical advice in response to health-related queries. This issue typically stems from incomplete training data. Increasingly, the AI industry is advocating for specialized AI models—domain-focused AIs—to mitigate potential knowledge gaps and reduce misinformation risks.

Inference encompasses the operation of an AI model, setting it in motion to make decisions or predictions based on existing data patterns. However, inference relies on prior model training; a system must first identify patterns within a dataset to effectively extrapolate from its training.

A variety of hardware—from smartphone processors to high-performance GPUs and specialized AI accelerators—can execute inference. However, their efficiency varies; large models, for instance, may take significantly longer on a laptop than on a high-spec cloud server.

[See: Training]

Large language models (LLMs) power popular AI assistants like ChatGPT, Claude, Google’s Gemini, Meta’s Llama, Microsoft Copilot, or Mistral’s Le Chat. Engaging with these assistants, users interact with LLMs that process requests—either directly or with the assistance of various tools like web search or code interpreters.

LLMs are built from deep neural networks comprising billions of numerical parameters (or weights, see below) that discern relationships among words and phrases, fabricating a representation of language—a multi-dimensional mapping of vocabulary.

These models derive patterns by analyzing billions of texts, including books and articles. When prompted, an LLM generates the pattern most fitting for the input.

(See: Neural network)

Memory cache plays a crucial role in enhancing inference—the process by which AI generates responses to user queries. Essentially, caching optimizes this process by reducing repetitive calculations. As AI operations hinge on complex mathematical computations, caching minimizes the amount of recalculating by retaining specific results for future queries. Various forms of memory caching exist, with KV (key value) caching being a notable method established in transformer-based models. This approach markedly boosts efficiency, accelerating response times by diminishing the necessary processing load.

(See: Inference)

A neural network embodies the multi-layered algorithmic design that drives deep learning—and, importantly, the surge in generative AI tools following the rise of large language models.

While the concept of mimicking the interconnected designs of the human brain for data processing algorithms dates back to the 1940s, it’s the emergence of graphical processing units (GPUs) from the gaming industry that truly realized this vision. These units are adept at training deep learning algorithms with numerous layers, enhancing performance across various fields, including voice recognition and drug discovery.

(See: Large language model [LLM])

Open source refers to software and AI models where the source code is accessible to the public for inspection, modification, or use. A prominent example in AI is Meta’s Llama models, with Linux being a historic parallel in operating systems. Open-source frameworks allow for collaborative development, enabling researchers and developers globally to build upon existing work, thus accelerating advancements and enabling independent safety evaluations that aren’t easily achievable through closed systems. Closed source means the code is not accessible—users can utilize the AI but cannot see its mechanics, as with OpenAI’s GPT models, highlighting a significant ongoing debate in the industry.

Parallelization refers to undertaking multiple tasks concurrently rather than sequentially—imagine ten employees tackling different project segments simultaneously, instead of one person handling everything step-by-step. In AI, this concept is essential to both training and inference, as contemporary GPUs are designed to conduct thousands of calculations at once, contributing to their status as the backbone of the AI industry. As AI complexity grows and models expand, effective parallelization across multiple chips and machines has become critical for expediting model development and deployment. Research into improved parallelization techniques is now a dedicated field.

RAMageddon is a lighthearted term for a serious trend within the tech industry: the escalating shortage of random access memory (RAM) chips that power almost every technological product we use today. The surge in the AI sector has prompted major tech companies and labs—competing to create the most advanced AI—to purchase significant quantities of RAM for their data centers, leaving limited supplies for other industries. This supply bottleneck has escalated costs for remaining RAM, impacting sectors like gaming, where companies may increase console prices due to scarcity, and consumer electronics, potentially resulting in the biggest decline in smartphone shipments in a decade.

Recursive self-improvement, much like AGI, represents an important benchmark for the advancement and autonomy of AI. In this scenario, AI models enhance themselves without human oversight, leading to dramatic leaps in capabilities. This concept raises concerns about a potential cataclysm akin to a singularity moment, wherein AI becomes impervious to external influence. A number of emerging AI startups are pursuing models that improve recursively, but many present RSI as a rational next step in AI research rather than an apocalyptic scenario.

Reinforcement learning entails a method of AI training where systems acquire knowledge through trial and error, receiving rewards for correct actions, akin to training a pet with treats—albeit in this case, the “pet” is a neural network receiving mathematical feedback. Unlike supervised learning, which employs defined datasets, reinforcement learning enables models to explore their environments, take actions, and adjust behavior based on feedback. This approach has proven particularly effective for training AI in games, robotics, and enhancing the logical capabilities of large language models. Techniques such as reinforcement learning from human feedback (RLHF) are becoming integral to refining AI models, making them more accurate, helpful, and safer.

Tokenization facilitates communication between humans and machines by breaking down language into smaller components—tokens—representing distinct pieces of data processed by large language models. This process translates raw text into more manageable segments, akin to how a compiler would convert human language into binary code for computers to process. In corporate contexts, tokens often determine costs, as AI companies typically charge for usage based on token volume, meaning businesses incur higher expenses with increased AI interactions.

Tokens are essentially small fragments of text—often parts of words—that AI models utilize to process input, serving a similar purpose as “words” in understanding AI workloads. Throughput gauges the amount of data processed in a specific timeframe, with token throughput quantifying the volume of AI operations a system can handle concurrently. High token throughput is a priority for AI infrastructure teams; it influences how many users can be served at once and how quickly they receive responses. AI researcher Andrej Karpathy has expressed anxiety when his AI subscriptions remain idle, recalling similar concerns from his graduate studies when expensive computing resources were underutilized, emphasizing why maximizing token throughput has become a focal point in AI development.

Training encompasses the process of feeding data to AI systems, enabling them to discern patterns and generate useful outputs. Essentially, it reflects the mechanism through which systems respond to data characteristics, refining outputs toward desired outcomes—whether that means spotting feline photos or composing poetry on command.

Training can be costly, as it often necessitates vast amounts of data, and demand for inputs has been steadily rising. Consequently, strategies like fine-tuning a rules-based AI with targeted data are increasingly employed to manage expenses while avoiding the need to start the entire training process from scratch.

[See: Inference]

Transfer learning involves leveraging a previously trained model as a foundation for developing another model for a related task—utilizing prior knowledge to accelerate the development process.

This method fosters efficiency by reducing the time and resources required to build new models, especially when data on the target task is limited. However, it is essential to recognize its limitations; models relying on transfer learning will likely necessitate additional training on specific data to perform optimally.

(See: Fine tuning)

Validation loss is a metric that indicates an AI model’s learning efficacy during training—lower values are preferable. Researchers monitor this closely as a form of real-time evaluation, helping them determine when to cease training, adjust hyperparameters, or investigate potential issues. A crucial concern highlighted by this metric is overfitting, where a model memorizes its training data rather than learning transferable patterns applicable in new contexts. Consider it the distinction between a student who comprehends the course material versus one who merely memorizes past exam questions—validation loss helps uncover which type of learner your model is becoming.

Weights serve a fundamental role in AI training by assigning levels of importance to distinct features present in the training dataset, shaping the resulting output.

More specifically, weights are numerical parameters dictating which aspects of a dataset are most influential for the designated training task. They perform this function by applying multipliers to inputs. Initially, weights are assigned randomly during training, but they are adjusted as the model iteratively seeks more accurate outputs that align with target objectives.

For example, an AI model forecasting real estate prices may incorporate weights for attributes like the number of bedrooms or bathrooms, whether it’s a detached or semi-detached property, or additional features such as parking facilities. The weights signify the influence each of these factors has on the final property valuation based on the utilized dataset.

This article is updated regularly with new information.