AGI
Short for Artificial General Intelligence, a hypothetical AI system capable of understanding and solving any intellectual task that a human can. Unlike today's specialized AI, AGI would match human-level reasoning across all domains. It remains a long-term research goal that has not yet been achieved.
Algorithm
A step-by-step set of instructions that a computer follows to solve a specific problem. In AI, algorithms describe how a model learns from data and makes decisions. Well-known examples include sorting, search, and optimization algorithms.
Attention
A mechanism in neural networks that allows the model to focus on the most relevant parts of an input. Attention is the core building block of the Transformer architecture and enables capturing relationships between distant words in a text. Without attention, modern language models would not be possible.
Autonomous Agents
AI systems that independently plan, execute, and make decisions to complete tasks without a human directing every step. They combine language models with tools such as web search or code execution. Autonomous agents can handle complex, multi-step tasks on their own.
Batch Size
The number of training examples a model processes at once before updating its parameters. A larger batch size can make training more stable but requires more memory. Choosing the right batch size affects both training speed and model quality.
Bias
Systematic distortions in AI models that arise from unbalanced or unrepresentative training data. Bias can cause a model to favor or disadvantage certain groups of people. Detecting and reducing bias is one of the greatest challenges in AI development.
Chain-of-Thought
A prompting technique that guides a language model to lay out its reasoning step by step before providing a final answer. This significantly improves response quality on complex tasks. Chain-of-Thought makes the model's reasoning process transparent and verifiable.
Chatbot
A program that communicates with people in natural language, often through a chat interface. Modern chatbots are powered by large language models and can deliver context-aware, detailed responses. They are used in customer service, education, and many other fields.
Computer Vision
A subfield of AI focused on enabling computers to understand and interpret images and video. Applications range from facial recognition and medical image analysis to autonomous driving. Computer vision frequently relies on deep neural networks for image recognition.
Context Window
The maximum amount of text, measured in tokens, that a language model can process in a single request. A larger context window allows the model to read longer documents and consider more conversation history. Modern models offer context windows of hundreds of thousands of tokens.
Dataset
A structured collection of data used to train, validate, or test AI models. The quality and diversity of a dataset largely determine how well a model learns and generalizes. Large, carefully curated datasets are a prerequisite for high-performing AI.
Deep Learning
A branch of machine learning that uses neural networks with many layers to recognize complex patterns in data. Deep learning has enabled breakthroughs in image and speech recognition. It forms the foundation of most modern AI applications.
Diffusion Model
A class of generative AI models that create images by gradually removing noise from a random signal. Diffusion models power popular image generators such as Stable Diffusion and DALL-E. They have dramatically improved the quality of AI-generated imagery.
Embedding
A mathematical representation of data such as text, images, or audio as a numerical vector in a high-dimensional space. Similar content ends up close together in the embedding space, enabling comparisons and search. Embeddings are the foundation for semantic search and recommendation systems.
Emergent Behavior
Capabilities that an AI model exhibits even though they were not explicitly trained, appearing spontaneously once the model reaches a certain size. Examples include logical reasoning or solving tasks not present in the training data. Emergent behavior is not yet fully understood.
Epoch
One complete pass through the entire training dataset during model training. Multiple epochs are common so the model can learn the data thoroughly. However, too many epochs can lead to overfitting, where the model memorizes the training data instead of generalizing.
Few-Shot Learning
A technique where an AI model is shown only a handful of examples so it can understand and perform a new task. Unlike traditional training, the model does not need thousands of examples. Few-shot learning is especially effective with large language models via the prompt.
Fine-Tuning
The process of further training a pre-trained AI model with additional, specialized data to optimize it for a specific task. Fine-tuning is significantly cheaper and faster than training from scratch. It is commonly used to adapt general models to particular domains.
Foundation Model
A large AI model pre-trained on massive amounts of data that serves as a base for many different applications. Examples include GPT, LLaMA, and Gemini. Foundation models can be adapted to specific tasks through fine-tuning or prompt engineering.
GANs
Generative Adversarial Networks consist of two neural networks competing against each other: a generator creates data and a discriminator judges its authenticity. Through this interplay, the generator learns to produce increasingly realistic data. GANs were pioneering for the creation of synthetic images and videos.
GPT
Stands for Generative Pre-trained Transformer, a family of language models developed by OpenAI. GPT models are pre-trained on large text corpora and can then generate, translate, summarize text and much more. They are among the most well-known and influential AI models.
Gradient Descent
A mathematical optimization method that iteratively adjusts a model's parameters to minimize errors. The model calculates which direction to adjust its weights to produce better predictions. Gradient descent is the central learning algorithm behind most neural networks.
Base Model
A synonymous term for Foundation Model, referring to a large, generally pre-trained AI model that serves as the starting point for specialized applications. Base models are valued for their versatility and broad knowledge base. They can be adapted through fine-tuning or prompt engineering.
Hallucination
When an AI model generates convincing-sounding but factually incorrect or fabricated information. Hallucinations occur because language models compute probabilities over word sequences rather than possessing genuine world knowledge. They are a well-known challenge when using AI for fact-based tasks.
Hyperparameter
Settings defined before training an AI model that control the learning process, such as learning rate, batch size, or number of layers. Unlike regular parameters, hyperparameters are not learned by the model itself. Their selection has a major impact on the performance of the finished model.
Inference
The process of applying an already-trained AI model to new input data to produce predictions or results. Inference is what happens when you ask a chatbot a question or have an AI analyze an image. It is typically much faster than training.
AI Agent
An AI system that autonomously pursues goals, makes decisions, and interacts with its environment. AI agents can use tools, search the internet, or execute code to accomplish tasks. They represent an important step toward more autonomous AI systems.
Artificial Intelligence
The broad field of computer systems that can perform tasks normally requiring human intelligence, such as language understanding, image recognition, or decision-making. AI encompasses many sub-fields, from simple rule-based systems to complex deep learning. It is one of the most transformative technologies of our time.
Artificial Neural Networks
Computing models inspired by the biological brain, consisting of interconnected artificial neurons. They learn by adjusting the strength of connections between neurons. Artificial neural networks are the fundamental building blocks of deep learning and modern AI systems.
Large Language Models (LLM)
AI models with billions of parameters trained on vast amounts of text that can understand and generate natural language. They are the backbone of chatbots, translation services, and AI assistants. Well-known examples include GPT, Claude, and LLaMA.
Latent Space
An abstract, high-dimensional space where an AI model represents the essential features of its input data in compressed form. Similar concepts end up close together in latent space, allowing the model to recognize patterns and relationships. Latent spaces are central to generative models and embeddings.
LoRA
Low-Rank Adaptation is an efficient fine-tuning method for large AI models that only modifies a small subset of parameters. This makes fine-tuning possible with significantly less computing power and memory. LoRA has made adapting large language models accessible to a much broader developer community.
Machine Learning
A subfield of AI in which computers learn from data without being explicitly programmed. Instead of rigid rules, algorithms independently recognize patterns in data and improve with experience. Machine learning includes methods such as supervised, unsupervised, and reinforcement learning.
Multimodal
Describes AI models that can process multiple types of input simultaneously, such as text, images, audio, and video. Multimodal models understand connections between different media types, for example that an image and its caption belong together. They enable more natural and versatile interactions with AI.
Natural Language Processing
A branch of AI concerned with processing and understanding human language by computers. NLP covers tasks such as text classification, translation, summarization, and sentiment analysis. Modern NLP systems are almost exclusively based on Transformer models.
Neural Network
A computing model made up of layers of artificial neurons that transforms input data into desired outputs. Each neuron receives signals, processes them, and passes the result to the next layer. Neural networks form the foundation for nearly all modern AI applications.
Open Weights
Refers to AI models whose trained weights are publicly available for anyone to download and use. Open weights enable transparency, independent research, and local usage without a cloud connection. Well-known examples include LLaMA, Mistral, and Stable Diffusion.
Overfitting
A training problem where a model memorizes the training data rather than learning general patterns. An overfitted model performs excellently on training data but fails on new, unseen data. Techniques such as regularization and dropout help prevent overfitting.
Parameter
The internal numerical values of an AI model that are learned and adjusted during training. Large language models have billions of such parameters that collectively determine how the model responds to inputs. The parameter count serves as a rough indicator of a model's capability.
Prompt
The input or instruction a user sends to an AI model to trigger a specific response or action. A prompt can be a question, a command, or the beginning of a text. The form and quality of the prompt significantly influences the quality of the answer.