Artificial intelligence is a complex field where specialists often use technical language. Since our reporting on the AI industry frequently includes these terms, we’ve compiled this glossary to define key words and phrases found in our articles.
This glossary will be updated regularly to include new terms as AI researchers develop innovative methods and identify new safety challenges.
AGI
Artificial General Intelligence (AGI) is a concept that lacks a precise definition, but it broadly signifies AI systems capable of performing most tasks better than an average human. OpenAI CEO Sam Altman characterized AGI as being comparable to a “median human co-worker.” OpenAI’s charter describes AGI as “highly autonomous systems that outperform humans at most economically valuable work,” while Google DeepMind considers it “AI that’s at least as capable as humans at most cognitive tasks.” If these definitions seem unclear, rest assured that even leading AI experts share this uncertainty.
AI agent
An AI agent is an advanced AI-powered tool designed to autonomously execute a sequence of tasks for a user, surpassing the capabilities of a standard chatbot. Examples include managing expenses, making reservations, or even developing and maintaining software code. As this field is rapidly evolving, the term “AI agent” can have varied interpretations, and the necessary infrastructure to realize its full potential is still under development. Fundamentally, it represents an independent system that can leverage multiple AI components to complete complex, multi-step operations.
Chain of thought
When faced with straightforward questions, such as “Is a giraffe or a cat taller?”, humans often respond instinctively. However, for more complex problems, like determining the number of chickens and cows a farmer has given a total of 40 heads and 120 legs (answer: 20 chickens and 20 cows), intermediary steps or calculations are typically required.
Within AI, particularly for large language models, “chain-of-thought” reasoning involves dissecting a problem into smaller, sequential steps to enhance the accuracy of the final output. Although this process may extend the time required to generate a response, it significantly increases the likelihood of a correct answer, especially in scenarios involving logic or coding. These reasoning models are built upon conventional large language models and refined for chain-of-thought capabilities through reinforcement learning.
(See: Large language model)
Compute
“Compute,” though a versatile term, primarily denotes the essential processing capability that enables AI models to function. This computational strength is crucial for the AI sector, facilitating the training and deployment of sophisticated models. It frequently serves as a concise way to refer to the hardware that delivers this power, including GPUs, CPUs, TPUs, and other foundational infrastructure vital to the contemporary AI industry.
Deep learning
Deep learning is a specialized area of self-improving machine learning that utilizes AI algorithms structured with multiple layers, forming an artificial neural network (ANN). This architecture enables them to uncover more intricate relationships in data than simpler machine learning methods like linear models or decision trees. The design of deep learning algorithms mirrors the complex, interconnected neural pathways found in the human brain.
Deep learning AI models can independently detect crucial features within data, eliminating the need for human engineers to explicitly define them. This design also allows algorithms to learn from mistakes, iteratively refining their outputs. Nevertheless, deep learning systems demand extensive datasets (millions or more) to achieve optimal performance. Their training processes are also generally more time-consuming than those of simpler machine learning algorithms, which often results in higher development expenses.
(See: Neural network)
Diffusion
Diffusion technology underpins numerous AI models that generate art, music, and text. Drawing inspiration from physics, diffusion systems gradually “corrupt” data’s structure—such as images or audio—by introducing noise until it’s completely obscured. While physical diffusion is a spontaneous, irreversible process (like dissolving sugar in coffee), AI diffusion systems learn a “reverse diffusion” technique to reconstruct the original data, thereby recovering information from noise.
Distillation
Distillation is a method for transferring knowledge from a large “teacher” AI model to a smaller “student” model. Developers query the teacher model, record its responses, and sometimes evaluate their accuracy against a separate dataset. These recorded outputs then serve as training data for the student model, which learns to mimic the teacher’s performance.
This technique enables the creation of more compact and efficient models derived from larger ones, with minimal loss of quality. For example, OpenAI likely employed distillation to develop GPT-4 Turbo, a quicker iteration of GPT-4.
Although all AI companies utilize distillation internally, some may also have applied it to rapidly advance their models to compete with cutting-edge systems. However, using distillation to replicate a competitor’s model typically breaches the terms of service for AI APIs and chat assistants.
Fine-tuning
Fine-tuning involves providing additional training to an AI model to enhance its performance for a particular task or domain, beyond its initial broad training. This usually entails incorporating new, specialized (task-specific) data.
Numerous AI startups leverage large language models as a foundation for commercial products, then augment these models’ initial training with fine-tuning based on their proprietary domain expertise to boost utility for specific industries or tasks.
(See: Large language model [LLM])
GAN
A Generative Adversarial Network (GAN) is a machine learning architecture critical to advancements in generative AI, particularly for creating realistic data such as deepfake content. GANs consist of two neural networks: a generator that produces output from training data, and a discriminator that evaluates this output. The discriminator acts as a classifier, helping the generator refine its performance over time.
The “adversarial” nature of GANs stems from a competitive setup where the two models challenge each other: the generator aims to create outputs that fool the discriminator, while the discriminator strives to identify synthetic data. This structured competition enhances AI outputs, making them more realistic without continuous human oversight. However, GANs are most effective for specialized applications, like generating realistic images or videos, rather than broad general-purpose AI.
Hallucination
“Hallucination” is the term used in the AI industry to describe instances where AI models fabricate information, producing inaccurate or entirely false outputs. This represents a significant challenge to AI quality.
AI hallucinations can lead to misleading Generative AI outputs and pose tangible dangers, such as harmful medical advice resulting from health-related queries. Consequently, most GenAI tools now include fine print advising users to verify AI-generated responses, although these disclaimers are often less conspicuous than the immediate information provided by the tools.
The fabrication of information by AI is generally attributed to deficiencies in their training data. For general-purpose Generative AI, or foundation models, this issue appears particularly challenging to overcome. There isn’t enough existing data to thoroughly train AI models to answer every conceivable question. In short, we haven’t created an omniscient AI (yet).
The prevalence of hallucinations is driving a trend toward more specialized or vertical AI models—meaning domain-specific AIs with focused expertise—to minimize knowledge gaps and mitigate disinformation risks.
Inference
Inference is the operation of an AI model, allowing it to generate predictions or derive conclusions from data it has previously encountered. Importantly, inference requires prior training; a model must first identify patterns within a dataset before it can effectively make extrapolations.
While various hardware, from smartphone processors to powerful GPUs and specialized AI accelerators, can perform inference, their performance varies significantly. Extremely large models would process predictions considerably slower on a laptop compared to a cloud server equipped with high-end AI chips.
[See: Training]
Large language model (LLM)
Large Language Models (LLMs) are the AI systems powering widely used AI assistants like ChatGPT, Claude, Google’s Gemini, Meta’s AI Llama, Microsoft Copilot, and Mistral’s Le Chat. When engaging with an AI assistant, you are interacting with an LLM that either directly processes your query or utilizes various tools such as web browsing or code interpreters.
It’s important to note that AI assistants and their underlying LLMs may have distinct names; for example, OpenAI’s large language model is GPT, while ChatGPT is its AI assistant product.
LLMs are complex deep neural networks composed of billions of numerical parameters (also known as weights; see below) that learn the intricate connections between words and phrases, constructing a multi-dimensional representation or “map” of language.
These models are developed by encoding patterns discovered across billions of books, articles, and transcripts. When an LLM receives a prompt, it generates the most probable pattern that aligns with it. Subsequently, it predicts the most likely next word, based on the preceding context, in an iterative process.
(See: Neural network)
Memory cache
Memory cache denotes a crucial process that enhances inference, which is how AI generates responses to user queries. Essentially, caching is an optimization strategy aimed at increasing inference efficiency. Since AI relies on intensive mathematical computations that consume significant power, caching reduces the number of calculations a model needs to perform by storing specific results for subsequent user queries and operations. Among various types of memory caching, KV (key-value) caching is particularly recognized. KV caching improves efficiency in transformer-based models, leading to faster results by minimizing the time and computational effort required to answer user questions.
(See: Inference)
Neural network
A neural network describes the multi-layered algorithmic architecture that forms the foundation of deep learning and, by extension, the rapid expansion of generative AI tools since the advent of large language models.
While the concept of designing data processing algorithms inspired by the human brain’s interconnected pathways originated in the 1940s, it was the more recent emergence of graphical processing hardware (GPUs) from the video game industry that truly unleashed this theory’s potential. These chips proved ideal for training algorithms with significantly more layers than previously feasible, allowing neural network-based AI systems to achieve vastly superior performance across diverse fields like voice recognition, autonomous navigation, and drug discovery.
(See: Large language model [LLM])
RAMageddon
“RAMageddon” is a playful new term highlighting a serious trend in the tech sector: a growing scarcity of Random Access Memory (RAM) chips, which are essential for nearly all modern electronic devices. With the rapid expansion of the AI industry, major tech firms and AI research labs, all competing to develop the most potent and efficient AI, are acquiring such vast quantities of RAM for their data centers that a significant shortage for other consumers and industries has emerged. This supply constraint is driving up the cost of available RAM.
This impact extends to various sectors, including gaming (where leading companies have increased console prices due to memory chip scarcity), consumer electronics (facing potentially the largest decline in smartphone shipments in over a decade due to memory shortages), and general enterprise computing (as businesses struggle to acquire sufficient RAM for their data centers). The current price hike is only expected to subside once the critical shortage ends; however, there are currently few indications that this will occur in the near future.
Training
The creation of machine learning AIs necessitates a process called training. Simply put, this involves inputting data to enable the model to identify patterns and produce valuable outputs.
At this stage of AI development, a philosophical aspect emerges: before training, the underlying mathematical structure of a learning system is merely a collection of layers and random values. The AI model truly forms and develops only through training. Fundamentally, it’s the system’s reaction to data characteristics that allows it to adjust its outputs to achieve a desired objective, whether that’s recognizing cat images or composing a haiku.
It’s worth noting that not all AI necessitates training. Rules-based AIs, like linear chatbots, which are designed to follow manually pre-defined instructions, do not require training. Nevertheless, these AI systems tend to be more limited compared to well-trained, self-learning systems.
Furthermore, training can be costly due to the vast amounts of input data required, and the volume of necessary inputs for these models has generally been increasing.
Hybrid strategies can sometimes accelerate model development and control expenses. For instance, applying data-driven fine-tuning to a rules-based AI reduces the data, compute, energy, and algorithmic complexity needed compared to building a model from the ground up.
[See: Inference]
Tokens
Human-machine communication presents distinct challenges, as people use natural language while AI programs operate via intricate data-driven algorithmic processes. In essence, tokens are the fundamental units of human-AI interaction, representing discrete data segments either consumed or generated by an LLM.
Tokens are generated through “tokenization,” a process that dissects raw data and converts it into discrete units an LLM can understand. Much like a software compiler translates human language into binary code for a computer, tokenization interprets user queries in human language for an AI program to enable it to formulate a response.
Various types of tokens exist, such as input tokens (generated from a human user’s query), output tokens (produced by the LLM as it responds), and reasoning tokens (involved in more extended, intensive tasks and processes that are part of a user’s request).
In enterprise AI, token consumption directly impacts costs. As tokens signify the volume of data processed by a model, they have become the primary method for the AI industry to monetize its services. The majority of AI companies bill for LLM usage per token. Consequently, the more tokens a business utilizes when interacting with an AI program (like ChatGPT), the greater the charges from its AI service provider (e.g., OpenAI).
Transfer learning
Transfer learning is a method where an existing, pre-trained AI model serves as the foundation for developing a new model for a distinct, yet usually related, task. This allows knowledge acquired during previous training cycles to be repurposed.
This technique can boost efficiency by streamlining model development. It’s also beneficial when data for the new target task is scarce. However, the approach has limitations: models relying on transfer learning for generalized abilities will likely need further training on specific data to achieve optimal performance within their particular domain.
(See: Fine tuning)
Weights
Weights are fundamental to AI training, dictating the significance assigned to various features (or input variables) within the training data, thus influencing the AI model’s output.
In simpler terms, weights are numerical values that highlight the most relevant aspects of a dataset for a particular training objective. They function by multiplying inputs. Model training usually starts with randomly assigned weights, which are then iteratively adjusted as the model strives to produce an output that more accurately aligns with the desired target.
For instance, an AI model designed to predict housing prices, trained on historical real estate data for a specific area, would incorporate weights for features like the number of bedrooms and bathrooms, property type (detached or semi-detached), and presence of parking or a garage.
Ultimately, the weights assigned by the model to each input indicate their proportional impact on a property’s value, as determined by the provided dataset.
This article is updated regularly with new information.