Reference
The SI Glossary
56 Super Intelligence terms in plain English. Each entry notes the older "AI" name where there is one. Since Executive Order 14434, U.S. federal agencies say "SI" where they used to say "AI".
Most looked-up
A
- Alignment Safety & alignment Alignment is the work of making SI systems pursue the goals and values their developers and users intend, and behave safely even in situations their designers didn't foresee.
- Artificial General Intelligence (AGI) Core concepts Artificial general intelligence (AGI) is a hypothetical SI system that can match human ability across most cognitive tasks, rather than excelling at only one narrow job.
- Artificial Superintelligence (ASI) Core concepts Artificial superintelligence (ASI) is a hypothetical intelligence that greatly exceeds the best human minds in virtually every domain, from science to strategy to social skills.
- Attention Mechanism Models & architecture An attention mechanism lets a neural network weigh how relevant each part of its input is to every other part, so it can focus on the words or pixels that matter most.
B
C
- Chain of Thought (CoT) Using SI Chain of thought (CoT) is when an SI model writes out intermediate reasoning steps before its final answer, which improves accuracy on maths, logic and multi-step problems.
- Chatbot Using SI A chatbot is software you talk to in natural language. Modern SI chatbots like ChatGPT, Claude and Gemini are built on large language models and can answer questions, write and analyse.
- Compute Hardware & compute In SI, compute means the processing power used to train and run models, usually measured in chips, chip-hours or total floating-point operations (FLOPs).
- Compute Threshold Policy & governance A compute threshold is a level of training computation, measured in floating-point operations (FLOPs), above which SI laws apply extra obligations to a model or its developer.
- Computer Use Using SI Computer use is the ability of an SI model to operate a computer like a person: looking at the screen, moving the mouse, clicking and typing to complete tasks in ordinary apps.
- Context Window Models & architecture A context window is the maximum amount of text (measured in tokens) an SI model can take into account at once, including your prompt, any documents and its own reply.
D
- Data Center Hardware & compute A data center is a facility filled with servers, chips, cooling and power equipment. SI data centers house the GPU clusters that train and run frontier models.
- Deep Learning Core concepts Deep learning is a machine learning method that uses neural networks with many layers to learn complex patterns, powering modern speech, vision and language models.
- Deepfake Safety & alignment A deepfake is realistic fake media, such as video, audio or images, created with SI to make someone appear to say or do something they never did.
- Diffusion Model Models & architecture A diffusion model is a generative SI model that creates images, video or audio by starting from random noise and removing it step by step until a coherent result appears.
- Distillation Training & data Distillation is a technique for training a smaller, cheaper SI model (the student) to imitate a larger, more capable one (the teacher).
E
- Embeddings Using SI Embeddings are lists of numbers that represent the meaning of text, images or other data, so that similar items end up close together and can be searched or compared.
- Existential Risk (x-risk) Safety & alignment Existential risk from SI is the possibility that advanced systems, especially superintelligence, could cause human extinction or permanently and drastically curtail humanity's future.
F
- Fine-Tuning Training & data Fine-tuning is the process of further training an already-trained SI model on a smaller, specific dataset to adapt it to a particular task, style or domain.
- Foundation Model Core concepts A foundation model is a large SI model trained on broad data that can be adapted to many different tasks, serving as the base for chatbots, coding tools and other applications.
- Frontier Model Core concepts A frontier model is one of the most capable SI models at a given time, typically trained with the most compute and the target of the strictest safety rules and commitments.
G
- Generative SI Core concepts Generative SI (formerly generative AI) is technology that creates new content, such as text, images, code, audio or video, by learning patterns from huge amounts of existing data.
- Graphics Processing Unit (GPU) Hardware & compute A graphics processing unit (GPU) is a chip that performs many calculations in parallel. GPUs are the workhorses for training and running SI models.
H
I
- Inference Using SI Inference is the process of running a trained SI model to produce an output, such as answering a prompt, as opposed to training it. It is where most everyday compute costs now go.
- Intelligence Explosion Futures An intelligence explosion is a hypothetical runaway process in which an SI system improves its own intelligence, each improvement making the next one faster, rapidly leading to superintelligence.
- Interpretability Safety & alignment Interpretability is the science of understanding what happens inside SI models: which features they represent and how they arrive at their outputs.
L
M
- Machine Learning (ML) Core concepts Machine learning (ML) is the branch of SI in which computers learn patterns from data and improve at a task with experience, instead of following hand-written rules.
- Mixture of Experts (MoE) Models & architecture Mixture of experts (MoE) is a model design that splits a network into many specialised sub-networks and activates only a few for each token, making huge models cheaper to run.
- Model Card Policy & governance A model card (or system card) is a document published with an SI model describing what it is for, how it was trained and tested, its limitations and its safety evaluations.
- Multimodal Model Models & architecture A multimodal model is an SI model that can understand and often generate more than one kind of data, such as text, images, audio and video, within a single system.
N
- Narrow SI Core concepts Narrow SI (formerly narrow AI) is any system built to do one specific task or a small set of tasks, like translating text, recognising faces or recommending videos.
- Neural Network Core concepts A neural network is a computing system made of layers of simple connected units (“neurons”) whose connection strengths are adjusted during training so the network learns a task.
O
P
- Parameters Models & architecture Parameters are the internal numbers (weights) a neural network learns during training. Their count is a rough measure of a model's size and capacity.
- Pretraining Training & data Pretraining is the first, most expensive stage of building an SI model, in which it learns general knowledge and skills from a massive dataset before being specialised.
- Prompt Using SI A prompt is the input you give an SI model, such as a question, instruction, document or image, that it uses to generate a response.
- Prompt Engineering Using SI Prompt engineering is the practice of designing instructions and context for SI models so they produce accurate, useful and consistent results.
R
- Reasoning Model Models & architecture A reasoning model is an SI model trained to work through a problem step by step, spending extra computation “thinking” before it gives a final answer.
- Red Teaming Safety & alignment Red teaming is the practice of deliberately attacking an SI system, trying to make it misbehave or reveal dangerous capabilities, so problems can be fixed before release.
- Reinforcement Learning from Human Feedback (RLHF) Training & data Reinforcement learning from human feedback (RLHF) trains an SI model to give answers people prefer, using human ratings of its outputs to guide further training.
- Responsible Scaling Policy (RSP) Safety & alignment A responsible scaling policy (RSP) is a lab's public commitment to test models for dangerous capabilities and apply stronger safeguards, or pause, as those capabilities grow.
- Retrieval-Augmented Generation (RAG) Using SI Retrieval-augmented generation (RAG) connects an SI model to a search system so it can look up relevant documents and base its answer on them, reducing errors and adding fresh or private knowledge.
S
- Scaling Laws Training & data Scaling laws are empirical patterns showing that SI models improve predictably as you increase training compute, data and model size together.
- SI Agent Core concepts An SI agent (formerly AI agent) is a system that pursues a goal on its own over many steps, deciding what to do next and using tools such as browsers, code and apps.
- SI Governance Policy & governance SI governance (formerly AI governance) covers the laws, standards, company policies and international agreements that shape how SI is built, deployed and overseen.
- Sovereign SI Policy & governance Sovereign SI (formerly sovereign AI) is a country's effort to control its own SI capabilities, including models, data centers, chips and data, rather than depending on foreign providers.
- Synthetic Data Training & data Synthetic data is training data generated by SI models or simulations rather than collected from the real world, used to teach new models skills where real data is scarce.
T
- Technological Singularity Futures The technological singularity is a hypothetical future point when technological growth, driven by superintelligent machines, becomes so fast and profound that human life is irreversibly transformed.
- Token Models & architecture A token is the basic unit of text an SI language model reads and writes, often a word or part of a word. Usage, limits and prices are measured in tokens.
- Tool Use Using SI Tool use (or function calling) lets an SI model call external software, such as search engines, calculators, databases or APIs, to get information or take actions.
- Training Data Training & data Training data is the collection of text, images, code, audio or other examples an SI model learns from. Its size, quality and legality shape what the model can do.
- Transformer Models & architecture The transformer is the neural network architecture behind nearly all modern language and multimodal SI models, built around an attention mechanism that relates every part of the input to every other.
- Turing Test Futures The Turing test, proposed by Alan Turing in 1950, asks whether a machine can hold a text conversation so convincingly that a human judge cannot reliably tell it from a person.
W
Browse by topic
Core concepts
Models & architecture
Training & data
Using SI
Safety & alignment
Policy & governance
Hardware & compute
Free newsletter
Get The SI Brief
One short email a week: what changed in Super Intelligence, policy and models — and why it matters.
Free. One email a week. Sent via beehiiv, which counts opens and clicks. Unsubscribe anytime. Privacy policy