The field of artificial intelligence (AI) is vast and intricate, often characterized by specialized terminology and jargon. To navigate this complex landscape, a comprehensive glossary becomes essential for understanding the core concepts and technologies driving the AI industry. This compilation aims to define some of the most crucial words and phrases frequently encountered in AI discussions, with regular updates planned as researchers continue to innovate and identify emerging safety considerations.

At the forefront of advanced AI aspirations is Artificial General Intelligence (AGI), a term that remains somewhat ambiguous. Generally, AGI refers to AI systems that are more capable than the average human across many, if not most, tasks. Definitions vary slightly among leading organizations; OpenAI’s CEO Sam Altman describes AGI as the "equivalent of a median human that you could hire as a co-worker," while OpenAI's charter defines it as "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind conceptualizes AGI as "AI that’s at least as capable as humans at most cognitive tasks." The ongoing debate highlights the challenging nature of defining true general intelligence in AI.

A foundational component of modern AI is the Neural Network, a multi-layered algorithmic structure inspired by the interconnected pathways of neurons in the human brain. While the concept dates back to the 1940s, the recent proliferation of graphical processing units (GPUs), initially driven by the video game industry, significantly unlocked its potential. GPUs are adept at training algorithms with numerous layers, enabling neural network-based AI systems to achieve superior performance in diverse areas like voice recognition, autonomous navigation, and drug discovery.

Deep Learning is a specialized subset of self-improving machine learning that utilizes these artificial neural networks. Its multi-layered structure allows algorithms to identify complex correlations in data, surpassing the capabilities of simpler machine learning models such as linear models or decision trees. Deep learning models can autonomously identify important data characteristics, learn from errors, and refine their outputs through iterative adjustments. However, these systems require substantial datasets, often millions of data points, and typically longer training times, leading to higher development costs.

The operational backbone of AI models relies heavily on Compute, a term that broadly refers to the vital computational power necessary for AI operations. This processing power fuels the AI industry, enabling the training and deployment of sophisticated models. "Compute" often serves as a shorthand for the underlying hardware infrastructure, including Graphics Processing Units (GPUs), Central Processing Units (CPUs), Tensor Processing Units (TPUs), and other specialized components that form the bedrock of contemporary AI.

The development of machine learning AIs fundamentally involves Training. This process entails feeding data into a model to allow it to learn patterns and subsequently generate useful outputs. Before training, an AI model is essentially a mathematical structure of layers and random numbers. It is through training that the system responds to data characteristics, adapting its outputs towards a specific objective, whether it's identifying images or composing text. While not all AI requires training (e.g., rules-based chatbots), well-trained self-learning systems are generally more capable. Training can be expensive due to the large volumes of inputs required, though hybrid approaches combining data-driven fine-tuning with rules-based AI can help manage costs and development complexity.

Following training, Inference is the process of putting an AI model into action—running it to make predictions or draw conclusions from previously unseen data. Inference is impossible without prior training, as a model must first learn patterns within a dataset. Various hardware, from smartphone processors to high-end GPUs and custom AI accelerators, can perform inference, though their efficiency varies significantly, especially for very large models.

Among the most widely recognized AI models today are Large Language Models (LLMs), which power popular AI assistants like ChatGPT, Claude, and Gemini. LLMs are deep neural networks comprising billions of numerical parameters, or Weights. These weights are core to AI training, determining the importance given to different features in the training data and thus shaping the model's output. Initially assigned randomly, weights adjust during training as the model strives to match target outputs more closely. LLMs learn relationships between words and phrases by encoding patterns from billions of texts, articles, and transcripts, generating the most probable next word in response to a prompt.

An AI Agent represents a more autonomous type of AI tool that leverages AI technologies to execute a sequence of tasks on a user's behalf, extending beyond the capabilities of a basic AI chatbot. These tasks can range from filing expenses and booking reservations to writing and maintaining code. The concept implies an autonomous system that can draw upon multiple AI systems to accomplish complex, multistep objectives, though its precise definition and underlying infrastructure are still evolving.

To improve the quality of results, especially in logic or coding contexts, LLMs can employ Chain of Thought reasoning. This technique involves breaking down a complex problem into smaller, intermediate steps. While it may take longer to arrive at an answer, the outcome is typically more accurate. Reasoning models are often developed from traditional LLMs and optimized for chain-of-thought thinking through reinforcement learning.

Diffusion is a technology central to many generative AI models that produce art, music, and text. Inspired by physics, diffusion systems progressively "destroy" data structures by adding noise until the original form is obliterated. Unlike irreversible physical diffusion, AI diffusion systems learn a "reverse diffusion" process, enabling them to reconstruct the original data from noise, thereby generating new content.

Another framework in generative AI is the Generative Adversarial Network (GAN), which has been instrumental in creating realistic data, including deepfake tools. GANs consist of two competing neural networks: a generator that creates outputs based on training data, and a discriminator that evaluates these outputs. The "adversarial" setup pushes the generator to produce increasingly realistic data while the discriminator improves its ability to identify artificially generated content, optimizing outputs without human intervention. GANs are most effective for narrower applications, such as generating realistic images or videos.

Fine-tuning refers to the process of further training an existing AI model to optimize its performance for a highly specific task or domain. This typically involves feeding the model new, specialized, task-oriented data. Many AI startups utilize large language models as a base, then fine-tune them with their own domain-specific knowledge to enhance utility for a particular sector or application.

Distillation is a technique employed to transfer knowledge from a larger "teacher" AI model to a smaller, more efficient "student" model. Developers send queries to the teacher model, record its outputs, and use these to train the student model to mimic the teacher’s behavior with minimal loss of quality. This method can create faster, more compact models, as seen with OpenAI’s GPT-4 Turbo. While commonly used internally by AI companies, using distillation from a competitor's model typically violates terms of service.

Transfer Learning is a technique where a pre-trained AI model serves as a starting point for developing a new model designed for a different, yet related, task. This allows knowledge gained during previous training cycles to be reapplied, leading to efficiency savings and proving particularly useful when data for the new task is limited. However, models relying on transfer learning often still require additional domain-specific training to achieve optimal performance.

Despite their advancements, AI models can suffer from Hallucination, the industry term for generating incorrect or fabricated information. This poses a significant challenge to AI quality, as hallucinatory outputs can be misleading and potentially dangerous, particularly in sensitive areas like medical advice. The problem is often attributed to gaps in training data, especially for general-purpose foundation models, and is contributing to a trend towards more specialized, domain-specific AI models to mitigate disinformation risks.

To enhance inference efficiency, AI systems employ Memory Cache, an optimization technique designed to reduce redundant mathematical calculations. By saving particular computations for future user queries, caching, such as KV (key-value) caching in transformer-based models, cuts down on algorithmic labor, leading to faster results and lower power consumption during inference.

Central to human-AI communication are Tokens, which serve as the basic discrete segments of data processed or produced by an LLM. Through a process called "tokenization," raw data, including user queries, is broken down into these digestible units for the AI program. There are various types, including input tokens (from user queries), output tokens (generated by the LLM's response), and reasoning tokens (for longer, more intensive tasks). Token usage is also the primary mechanism for monetizing AI services, with companies charging based on the volume of tokens processed.

Finally, the immense computational demands of the burgeoning AI industry have led to a significant trend known as RAMageddon. This term describes an increasing shortage and escalating cost of Random Access Memory (RAM) chips, which are crucial for almost all tech products. Major tech companies and AI labs are acquiring vast quantities of RAM for their data centers, creating a supply bottleneck that impacts other industries like gaming, consumer electronics, and general enterprise computing, with prices expected to remain high until the shortage alleviates.