AI Glossary

Comparing two versions of something by randomly showing each to different users and measuring which performs better.

Systematically removing parts of a model to understand each component's contribution to performance.

An experiment that systematically removes or changes components of an AI system to understand each part's contribution.

A mathematical function in a neural network that determines whether a neuron should fire. Common ones include ReLU and sigmoid.

A training strategy where the model identifies which unlabelled examples would be most valuable to learn from next.

Educational technology that adjusts content difficulty and sequence based on a student's performance and needs.

Deliberately crafted inputs designed to fool or manipulate an AI system into making mistakes.

An AI system that can perceive its environment, make decisions, and take actions to achieve goals.

AI that can independently plan, make decisions, and take actions to complete multi-step tasks with minimal human guidance.

Artificial General Intelligence. A hypothetical AI system that can match human-level intelligence across all cognitive tasks, not just specific ones.

The European Union's comprehensive regulatory framework for artificial intelligence, the world's first major AI law.

Software libraries and tools that make it easier to build AI agents capable of using tools and taking actions.

A systematic evaluation of an AI system's performance, fairness, safety, and compliance with regulations.

A semiconductor specifically designed or optimised for running AI workloads efficiently.

Urban development projects integrating AI throughout city infrastructure for traffic, energy, safety, and public services.

An AI designed primarily for ongoing social interaction, emotional support, or entertainment.

An AI assistant that works alongside a human, providing suggestions and handling routine tasks while the human stays in control.

A committee within an organisation or government tasked with reviewing and guiding the ethical use of AI.

The movement and initiatives applying AI to address social, environmental, and humanitarian challenges.

The policies, standards, and oversight structures for managing how AI systems are developed and deployed.

Applying AI for crop monitoring, yield prediction, pest detection, and precision farming.

Using AI to personalise learning experiences, automate grading, and provide intelligent tutoring.

The application of AI in banking, trading, risk assessment, fraud detection, and financial planning.

Legal questions about who is responsible when an AI system causes harm or makes a mistake.

The ability to understand, use, and think critically about AI technologies and their implications.

An organisation's level of sophistication in adopting, integrating, and benefiting from AI technologies.

How prepared an organisation, sector, or country is to adopt and benefit from artificial intelligence.

Research focused on ensuring AI systems behave as intended without causing unintended harm.

Singapore's national programme to build AI capabilities, catalyse AI adoption, and grow the local AI ecosystem.

An AI system that provides personalised educational instruction, adapting to a student's learning pace and style.

Exaggerating or falsely claiming that a product uses AI when it does not, or overstating AI capabilities for marketing purposes.

Historical periods when interest, funding, and progress in AI research declined significantly.

A product or company built from the ground up with AI at its core, rather than adding AI to an existing system.

A set of step-by-step instructions that tells a computer how to solve a problem or complete a task.

Holding organisations responsible for the decisions their AI systems make, especially those affecting people's lives.

Using AI and algorithms to execute financial trades automatically based on predefined strategies and market data.

Alibaba's cloud computing division, which developed the Qwen (Tongyi Qianwen) family of AI models.

Ensuring AI systems pursue goals that genuinely match human intentions, values, and well-being.

Google DeepMind's AI system that predicts protein structures with remarkable accuracy, a breakthrough for biology.

Google DeepMind's AI that defeated world champion Go players, a milestone in AI history.

The systematic analysis of data to discover patterns, extract insights, and support decision-making.

Adding metadata or labels to data points, such as drawing boxes around objects in images or tagging sentiment in text.

Identifying unusual patterns or outliers in data that do not conform to expected behaviour.

An AI safety company founded by former OpenAI researchers. Creator of the Claude family of AI assistants.

Application Programming Interface. A standardised way for different software systems to communicate with each other.

Restrictions on how many requests can be made to an API within a given time period.

The broad field of computer science focused on creating systems that can perform tasks typically requiring human intelligence.

The ASEAN Guide on AI Governance and Ethics, a voluntary framework for responsible AI across Southeast Asian nations.

Artificial Superintelligence. A theoretical AI that surpasses human intelligence in every domain. Remains purely speculative.

Application-Specific Integrated Circuit. A chip custom-designed for a particular task, such as AI inference, offering high efficiency.

A single attention mechanism within a multi-head attention layer, each learning to focus on different aspects of the input.

The component that lets a model focus on the most relevant parts of its input when making predictions.

Area Under the Curve. A single number summarising model performance from the ROC curve. Closer to 1 is better.

A neural network that learns to compress data into a smaller representation and then reconstruct it, useful for denoising and anomaly detection.

Automated Machine Learning. Tools that automate the process of building, selecting, and tuning machine learning models.

Operating independently without human control. Applies to vehicles, drones, robots, and software agents.

AI systems that operate with minimal or no human intervention over extended periods.

Cars, trucks, drones, or other vehicles that use AI to navigate and drive without human intervention.

A model that generates output one token at a time, with each new token depending on all previously generated tokens.

A digital representation of a user or AI character in virtual environments, games, or communication platforms.

The process of calculating how much each weight contributed to an error, then adjusting them to improve accuracy.

China's leading search and AI company, developer of the Ernie large language model series.

A technique that normalises inputs to each layer during training, making networks train faster and more stably.

Processing large volumes of data in groups at scheduled intervals rather than in real-time.

The number of training examples processed together before updating the model's weights.

Statistical approaches that update predictions as new evidence becomes available, based on probability theory.

A strategy for efficiently finding the best settings or parameters by building a model of the objective function.

A text generation strategy that keeps track of multiple possible outputs simultaneously and selects the overall best one.

A standardised test or dataset used to compare the performance of different AI models on specific tasks.

Bidirectional Encoder Representations from Transformers. Google's influential model that understands text by reading in both directions.

When an AI system produces unfair or skewed results, often because its training data reflects existing human prejudices.

The balance between a model being too simple (high bias) and too complex (high variance). Finding the sweet spot is key.

Datasets so large or complex that traditional tools struggle to process them, often described by volume, velocity, and variety.

An AI model whose internal decision-making process is opaque and difficult for humans to understand.

Byte Pair Encoding. A common tokenisation method that builds a vocabulary by repeatedly merging the most frequent pairs of characters.

Chinese technology company behind TikTok and Douyin, with major investments in AI for content recommendation and creation.

The total greenhouse gas emissions caused by training and running an AI model.

When a neural network trained on new data loses its ability to perform previously learned tasks.

AI methods for determining cause-and-effect relationships rather than just correlations in data.

A prompting technique that asks the AI to show its reasoning step by step, improving accuracy on complex problems.

An AI-powered software application that conducts conversation with users through text or voice.

OpenAI's conversational AI product built on GPT models. Popularised the use of AI chatbots worldwide from late 2022.

Research from DeepMind showing that training smaller models on more data is often more efficient than simply making models larger.

A philosophical thought experiment arguing that a computer following rules does not truly understand language.

Insufficient global supply of advanced semiconductors, constraining AI development and deployment.

Breaking long documents into smaller, manageable pieces for processing by AI models with limited context windows.

Using AI to identify customers likely to stop using a product or service, enabling proactive retention efforts.

Continuous Integration and Continuous Deployment. Automated processes for testing and releasing software updates frequently.

A machine learning task where the model assigns inputs to predefined categories, like spam vs. not spam.

Anthropic's family of AI assistants, designed with a focus on safety, helpfulness, and harmlessness.

Using AI to optimise patient recruitment, data analysis, and monitoring in pharmaceutical testing.

Contrastive Language-Image Pre-training. An OpenAI model that learns to connect images and text descriptions.

An AI model whose architecture, weights, and training data are kept proprietary and accessed only through an API.

Using remote servers hosted on the internet to store, manage, and process data instead of a local computer.

Grouping similar data points together without predefined labels. Used for customer segmentation and anomaly detection.

Convolutional Neural Network. A type of neural network designed for processing grid-like data such as images.

AI writing computer code based on natural language descriptions or partial code inputs.

AI systems designed to simulate human thought processes such as reasoning, understanding, learning, and decision-making. Unlike narrow AI that focuses on a single task, cognitive AI integrates multiple capabilities to interpret complex, ambiguous information in ways that mirror human cognition.

A recommendation technique that predicts what a user might like based on the preferences of similar users.

The processing power needed to train and run AI models. A key bottleneck and cost factor in AI development.

Policies controlling access to the large-scale computing resources needed to train powerful AI models.

The field of AI that enables computers to interpret and understand visual information from images and videos.

A table showing how often a classification model correctly and incorrectly predicts each category.

In data privacy, the requirement to get explicit permission from individuals before collecting or using their personal data.

An approach where AI is trained to follow a set of written principles, using AI feedback to improve its own behaviour.

Packaging software and its dependencies into isolated units (containers) that run consistently across different environments.

Technologies and standards for proving whether content is genuine, AI-generated, or manipulated.

Systems that screen AI outputs to block harmful, offensive, or inappropriate content before it reaches users.

Using AI to review and filter user-generated content on platforms to enforce community guidelines.

The maximum amount of text an AI model can consider at once. Larger windows allow the model to work with longer documents.

Enabling AI models to learn new tasks or information over time without forgetting what they previously learned.

A training method where the model learns by comparing similar and dissimilar pairs of examples.

AI systems designed to engage in natural, human-like dialogue across text and voice channels.

Using natural language questions to search for information, with the AI understanding intent rather than just matching keywords.

Legal debates around whether AI-generated content can be copyrighted and whether AI training on copyrighted material is lawful.

The task of determining which words in a text refer to the same entity, like linking 'she' to 'Dr. Chen'.

A measure of how similar two vectors are by comparing the angle between them, regardless of their magnitude.

A loss function commonly used for classification tasks that measures the difference between predicted and actual probability distributions.

Relating information across different data types, such as matching images to text descriptions.

A technique for assessing how well a model generalises by testing it on different subsets of the data.

NVIDIA's parallel computing platform that allows software to use GPU acceleration for AI and scientific computing.

Training a model by starting with simpler examples and progressively introducing more complex ones.

OpenAI's AI system that generates images from text descriptions.

Artificially expanding a training dataset by creating modified versions of existing data, such as flipping or rotating images.

Documentation describing a dataset's composition, collection methods, intended uses, and known limitations.

A facility housing computer servers and networking equipment used to store, process, and serve data and AI workloads.

When the statistical properties of real-world data change over time, causing AI model performance to degrade.

Building and maintaining the systems and infrastructure needed to collect, store, and process data at scale.

A cycle where more users generate more data, which improves the AI product, which attracts more users.

Policies and processes for managing data quality, access, security, and compliance across an organisation.

The process of annotating raw data with meaningful tags or categories so it can be used for supervised learning.

A centralised storage system that holds vast amounts of raw data in its original format until needed.

An automated workflow that moves data from source systems through processing steps to its final destination.

An attack where an adversary intentionally corrupts training data to make the resulting AI model behave incorrectly.

Cleaning, transforming, and preparing raw data before feeding it into an AI model for training.

Laws and practices safeguarding personal information from misuse, unauthorised access, or breach.

An interdisciplinary field combining statistics, programming, and domain expertise to extract insights from data.

The principle that data is subject to the laws and regulations of the country where it is collected or stored.

A structured storage system optimised for querying and analysing large volumes of processed, organised data.

A model that makes predictions by following a tree-like series of yes/no questions about the data.

A type of machine learning using neural networks with many layers to learn complex patterns from large amounts of data.

AI-generated or AI-manipulated media, typically video or audio, that convincingly depicts something that did not happen.

A Chinese AI lab that gained attention for producing highly capable open-weight models at significantly lower training costs.

Using AI to predict future customer demand for products or services based on historical data and trends.

Making AI tools and capabilities accessible to non-experts, smaller organisations, or underserved communities.

AI inferring the distance of objects from the camera in a 2D image, creating a sense of 3D understanding.

A mathematical technique that adds carefully calibrated noise to data, allowing useful analysis while protecting individual privacy.

An AI system that generates images by starting with pure noise and gradually removing it, guided by text prompts or other inputs.

An AI-powered virtual character with realistic human appearance, used in customer service, entertainment, or education.

A virtual replica of a physical object, process, or system that is continuously updated with real-world data.

Techniques for reducing the number of input variables in a dataset while preserving important information.

Training a smaller, faster student model to replicate the behaviour of a larger, more capable teacher model.

Splitting the training of a large AI model across multiple GPUs or machines to speed up the process.

AI systems specialised in understanding, extracting, and processing information from documents.

An unmanned aerial vehicle (UAV) that can be remotely controlled or fly autonomously using AI.

A training technique where random neurons are temporarily disabled to prevent the network from relying too heavily on any single path.

Using AI to identify potential new medicines and accelerate pharmaceutical research and development.

Technology that can be applied for both beneficial civilian purposes and potentially harmful military or surveillance uses.

Running AI directly on devices like phones, cameras, or sensors instead of sending data to the cloud.

AI models and methods designed to achieve good performance while using less compute, energy, and data.

Converting text, images, or other data into numerical vectors that capture their meaning, allowing AI to compare and reason about them.

AI systems that exist within and interact with the physical world through robot bodies or other physical forms.

Unexpected capabilities that appear in AI models as they scale up, not explicitly programmed or anticipated.

AI that attempts to identify human emotions from facial expressions, voice, or text. A controversial application.

An architecture where one network compresses input into a representation (encoder) and another generates output from it (decoder).

A specific URL or address where an API can be accessed to send or receive data.

The significant electricity required to train and run AI models, raising environmental sustainability concerns.

Combining multiple AI models to produce better results than any single model alone.

One complete pass through the entire training dataset during the training process.

Baidu's family of large language models, one of the leading LLMs developed in China.

AI designed and used in ways that align with moral principles, prioritising fairness and transparency.

Extract, Transform, Load. The process of pulling data from sources, converting it into a useful format, and storing it for analysis.

The possibility that very advanced AI could pose a fundamental threat to human civilisation. A topic of active debate.

The ability to understand and describe how an AI system reached a particular decision or output.

Government restrictions on selling or sharing certain AI technologies, hardware, or data with other countries.

A balanced metric that combines precision and recall into a single number between 0 and 1.

AI that identifies or verifies a person by analysing the unique features of their face.

Ensuring AI systems treat different groups of people equitably and do not discriminate.

The process of selecting, transforming, or creating input variables to improve a model's performance.

A centralised system for storing and serving the processed input data (features) used by machine learning models.

Training AI across many devices or servers without centralising the raw data, preserving privacy.

Providing a model with a small number of examples in the prompt to help it understand the desired task or format.

The ability to learn new concepts or tasks from just a small handful of examples.

Highly detailed or specific, as opposed to broad or coarse. Often refers to classification or permissions.

The ability to precisely adjust specific aspects of an AI system's behaviour or outputs.

Adapting a pre-trained AI model with specialised data to improve its performance on a particular task or domain.

Cloud platforms that allow customers to fine-tune foundation models on their own data without managing infrastructure.

Financial technology. Companies using technology to improve and automate financial services.

Floating Point Operations Per Second. A measure of computing power, often used to describe AI hardware performance.

A self-reinforcing business cycle where each component strengthens the others, creating accelerating momentum.

A large AI model trained on broad data that can be adapted for many different specific tasks.

Field-Programmable Gate Array. A chip that can be reprogrammed after manufacturing, offering flexibility for AI workloads.

Using AI to identify suspicious transactions or activities that may indicate fraud.

The most capable AI models at the cutting edge of performance, typically requiring enormous compute to train.

A specific capability where an AI model can invoke predefined functions or APIs as part of generating a response.

Generative Adversarial Network. Two neural networks competing against each other: one creates fake data, the other tries to detect the fakes.

General Data Protection Regulation. The EU's landmark data privacy law that has influenced privacy regulations worldwide, including across Asia.

Google DeepMind's family of multimodal AI models, integrated across Google's products and services.

AI systems that create new content such as text, images, code, audio, and video by learning patterns from large datasets. Also known as gen AI. Prominent examples include large language models like GPT and Claude, and image generators like DALL-E and Midjourney.

AI that creates new content such as text, images, music, video, or code, rather than just analysing existing data.

AI applied to geographic and location-based data for mapping, urban planning, and environmental monitoring.

The world's largest platform for hosting, sharing, and collaborating on software code.

An AI-powered coding assistant built on LLMs that suggests code completions and generates code from natural language.

Indonesian super-app leveraging AI across ride-hailing, payments, food delivery, and logistics services.

Google's AI research lab, formed by merging Google Brain and DeepMind. Behind Gemini, AlphaFold, and AlphaGo.

Generative Pre-trained Transformer. OpenAI's family of large language models that generate text by predicting the next token.

OpenAI's multimodal large language model released in 2023, capable of processing both text and images.

Graphics Processing Unit. Powerful chips originally designed for graphics that are now essential for training and running AI models.

Southeast Asian super-app using AI for ride-hailing, food delivery, payments, and fraud detection.

The optimisation method used to minimise a model's errors by adjusting weights in the direction that reduces loss.

A neural network designed to work with data structured as graphs, such as social networks or molecular structures.

Research and practices aimed at reducing the environmental impact of developing and running AI systems.

The verified, correct answer or label that a model's predictions are compared against.

Connecting AI outputs to verified sources of information to reduce hallucinations and improve factual accuracy.

Gated Recurrent Unit. A simplified version of LSTM that is faster to train while achieving similar performance.

Safety constraints and rules built into AI systems to prevent harmful, inappropriate, or off-topic outputs.

When AI generates confident-sounding but factually incorrect or fabricated information.

A platform and community hub for sharing, discovering, and deploying open-source AI models and datasets.

AI systems designed to include human oversight, review, or approval at critical decision points.

AI operates autonomously but a human monitors and can intervene if something goes wrong.

Combining traditional keyword search with semantic vector search for more accurate results.

The pattern of inflated expectations followed by disillusionment and eventual productive adoption of a new technology.

Settings chosen by developers before training begins, such as learning rate or batch size, that control how training works.

A massive cloud computing provider like AWS, Microsoft Azure, or Google Cloud that operates at enormous scale.

AI generating natural language descriptions of what is happening in an image.

AI creating new images from text descriptions, sketches, or other inputs.

AI identifying what is depicted in an image, such as recognising objects, scenes, or activities.

AI dividing an image into distinct regions, assigning each pixel to a specific object or category.

A formal evaluation of the potential effects of an AI system on individuals, society, and the environment before deployment.

An LLM's ability to learn new tasks from examples provided within the prompt, without any additional training.

The portion of a model's context window used to provide examples and instructions that guide the model's behaviour.

When a trained AI model processes new input and produces output. The actual moment the AI does its work.

The ongoing expense of running a trained AI model to process real-world requests, often the largest operational cost.

AI filling in missing or selected parts of an image with contextually appropriate content.

Training a model to follow natural language instructions more reliably, making it more useful as an assistant.

The degree to which a human can understand the internal workings and reasoning of an AI model.

Internet of Things. A network of physical devices embedded with sensors and software that connect and exchange data over the internet.

A prompt designed to bypass an AI model's safety guardrails and make it produce restricted content.

JavaScript Object Notation. A lightweight data format commonly used for structured communication between systems and AI APIs.

A popular clustering algorithm that groups data into K clusters based on similarity to cluster centres.

An organised repository of information that AI systems can query to provide accurate, factual answers.

Training a smaller, faster model to mimic the behaviour of a larger, more capable model.

A structured database that stores information as a network of interconnected entities and their relationships.

An open-source platform for automating the deployment, scaling, and management of containerised applications.

A popular open-source framework for building applications powered by large language models, especially agents and RAG systems.

Predicting the next word in a sequence based on the words that came before. The fundamental task behind LLMs.

The delay between sending a request to an AI system and receiving a response. Lower latency means faster results.

The compressed, abstract representation of data that AI models work with internally. A lower-dimensional map of complex data.

A technique that normalises the inputs across features within a single training example, commonly used in transformers.

A public ranking of AI models based on their performance on standardised benchmarks.

How much a model adjusts its weights with each training step. Too high causes instability, too low causes slow learning.

Light Detection and Ranging. A sensor technology that uses laser pulses to create detailed 3D maps of the surrounding environment.

Meta's family of open-weight large language models, widely used by researchers and developers worldwide.

Large Language Model. Software trained on massive amounts of text data to understand and generate human-like language.

Low-Rank Adaptation. An efficient fine-tuning method that trains only a small number of additional parameters instead of the full model.

A mathematical formula that measures how far off a model's predictions are from the correct answers. The model tries to minimise this.

Development platforms requiring minimal hand-written code, using visual tools and pre-built components instead.

Long Short-Term Memory. A type of RNN designed to remember information over long sequences, solving the vanishing gradient problem.

A subset of AI where software improves at tasks by learning from data rather than following explicit programming.

Using AI to automatically translate text from one language to another.

A training task where parts of an image are hidden and the model learns to predict the missing regions.

A training task where random words are hidden and the model learns to predict them from surrounding context. Used by BERT.

An attack that determines whether a specific data point was used in training a model, potentially compromising privacy.

An AI system's ability to store and recall information from previous interactions or long-running tasks.

Meta's artificial intelligence research division, creator of the open-weight Llama language models.

Teaching AI systems to learn how to learn more efficiently, often called 'learning to learn'.

An architecture where software is built as a collection of small, independent services rather than one large application.

An independent AI image generation service known for producing highly artistic and stylised visuals.

A French AI company and its family of efficient open-weight language models known for strong performance relative to their size.

A system architecture where multiple specialised AI models collaborate, each contributing their expertise.

An architecture that routes each input to specialised sub-models (experts), activating only a fraction of the total parameters for efficiency.

Machine Learning Operations. Practices for deploying, monitoring, and maintaining AI models in production reliably.

Massive Multitask Language Understanding. A popular benchmark testing AI knowledge across 57 academic subjects.

A competitive advantage that protects a business from rivals and is difficult to replicate.

A trained AI system that has learned patterns from data and can make predictions or generate outputs based on new inputs.

A standardised document that describes an AI model's intended use, performance, limitations, and ethical considerations.

When AI trained on AI-generated data gradually degrades in quality and diversity over successive generations.

A standard for connecting AI models to external tools, data sources, and services in a consistent way.

When an AI model's predictions become less accurate over time as the real world changes around it.

Systematically testing an AI model's performance, capabilities, and limitations using standardised methods.

An attack that recreates a proprietary AI model by systematically querying it and analysing its outputs.

A curated collection of pre-trained AI models offered by cloud providers, allowing developers to easily test and deploy various models.

A platform where developers can discover, compare, and deploy AI models from various providers.

A centralised repository for storing, versioning, and managing trained AI models.

The infrastructure and process for making a trained AI model available to handle real-world requests.

Multiple AI agents working together, each with specialised roles, to solve complex problems collaboratively.

Running multiple attention operations in parallel, allowing the model to capture different types of relationships simultaneously.

An AI interaction spanning multiple back-and-forth exchanges, where context from earlier turns informs later responses.

AI that can process and generate multiple types of content such as text, images, audio, and video.

Converting different types of data (text, images, audio) into a shared numerical space for comparison.

Training AI to understand and connect information across different types of data such as text, images, and audio.

Minimum Viable Product. The simplest functional version of a product, built to test the core idea with real users.

National AI Research Resource. Proposed US initiative to democratise access to AI compute and data for researchers. Similar programmes exist across Asia.

Identifying and classifying key elements in text such as people, organisations, locations, and dates.

AI designed to perform a single specific task well, like playing chess or recognising faces. All current AI systems are narrow AI.

AI that excels at one specific task but cannot generalise to other domains. Describes all current AI systems.

When a product becomes more valuable as more people use it, such as social media platforms or messaging apps.

Using AI to automatically design the structure of neural networks, rather than relying on human intuition.

A computing system loosely inspired by how brain cells connect, used to find patterns in data. The building block of deep learning.

Natural Language Processing. The field of AI focused on enabling computers to understand, interpret, and generate human language.

AI platforms that allow users to build and deploy models through visual interfaces without writing any code.

Scaling data to a standard range to help AI models train more effectively and fairly.

Neural Processing Unit. A specialised chip built into devices like smartphones to run AI tasks efficiently on-device.

The leading manufacturer of GPUs used for AI training and inference. A central player in AI hardware.

AI locating and identifying multiple objects within an image, drawing bounding boxes around each one.

Optical Character Recognition. AI that reads and converts text in images or scanned documents into editable digital text.

AI processing that happens locally on a user's device, offering faster response times and better privacy.

Representing categorical data as binary vectors where only one element is 1 and the rest are 0.

Continuously updating a model as new data arrives, rather than retraining from scratch in batches.

A formal, structured representation of concepts and the relationships between them within a specific domain.

A foundation model whose weights are publicly available, enabling widespread research and commercial use.

Software whose source code is freely available for anyone to view, use, modify, and distribute.

An AI model whose code and trained weights are publicly available for anyone to use, study, and modify.

AI models whose learned numerical parameters (weights) are publicly shared, though the training code or data may not be.

The American AI company behind GPT, ChatGPT, DALL-E, and Sora. One of the most influential AI labs globally.

Coordinating multiple AI models, agents, or services to work together on a complex workflow.

AI extending an image beyond its original borders, generating new content that matches the existing scene.

When a model memorises its training data too closely and performs poorly on new, unseen data.

Running multiple computations simultaneously, essential for training large AI models quickly.

The internal settings an AI model learns during training. More parameters generally means more capable but more resource-hungry.

Principal Component Analysis. A technique for simplifying complex data by reducing the number of variables while keeping the most important patterns.

Personal Data Protection Act. Data privacy laws adopted across several Asian countries including Singapore, Thailand, and Malaysia.

The simplest type of neural network, consisting of a single layer. The building block of more complex networks.

A measurement of how well a language model predicts text. Lower perplexity means the model is more confident and accurate.

Tailoring content, recommendations, or experiences to individual users based on their data and behaviour.

An AI's ability to break down complex goals into a sequence of steps and execute them in the right order.

AI detecting the position and orientation of a person's body parts in images or video.

A method for giving transformer models information about the order of elements in a sequence, since they process everything in parallel.

The initial phase where an AI model learns general knowledge from a large, broad dataset before being specialised.

Of all the items the model predicted as positive, the percentage that actually were positive.

Using AI, sensors, and data to optimise farming practices at a granular level for better yields and sustainability.

Using AI and genomic data to tailor medical treatments to individual patients.

Using AI and statistical methods to analyse current and historical data to forecast future outcomes.

Using AI to predict when equipment will fail so maintenance can be performed proactively.

AI that not only predicts what will happen but also recommends specific actions to take.

Building privacy protections into AI systems from the beginning, not as an afterthought.

The stage where a product satisfies a strong market demand, with users actively wanting and paying for it.

The text input or instruction you give to an AI model to get a desired response.

Connecting multiple AI prompts in sequence, where the output of one becomes the input for the next.

The practice of crafting effective instructions and inputs to get better, more reliable results from AI tools.

An attack where malicious instructions are hidden in input text to manipulate an AI into unintended behaviour.

Tracking the origin and history of data or AI-generated content to verify its authenticity and source.

Removing unnecessary connections or neurons from a neural network to make it smaller and faster without significant quality loss.

Quantized Low-Rank Adaptation. Combines quantization with LoRA to fine-tune large models on consumer hardware.

Compressing an AI model to use less memory and run faster by reducing the precision of its numerical values, with some quality trade-off.

AI systems that can read text and answer questions about it, or retrieve answers from a knowledge base.

Alibaba Cloud's family of large language and multimodal models, also known as Tongyi Qianwen.

Retrieval-Augmented Generation. A technique where AI looks up relevant real-world information before generating an answer, reducing hallucinations.

An ensemble method that combines many decision trees to make more accurate and stable predictions.

A system that determines the order in which search results, recommendations, or content are presented to users.

Processing and responding to data immediately as it arrives, with minimal delay.

An AI model's ability to think through problems logically, draw conclusions, and explain its thinking process.

Of all the actually positive items, the percentage the model correctly identified.

AI that suggests relevant content, products, or actions to users based on their behaviour and preferences.

Deliberately testing an AI system by trying to make it fail, produce harmful outputs, or reveal vulnerabilities.

An AI evaluating its own outputs for errors and self-correcting before delivering a final answer.

A machine learning task where the model predicts a continuous numerical value, like a house price.

Techniques used to prevent overfitting by adding constraints or penalties during training.

A structured set of rules, guidelines, and standards governing how AI can be developed and used.

A supervised testing environment where companies can trial innovative AI products under relaxed regulations.

Training an AI by letting it interact with an environment and rewarding good outcomes while penalising bad ones.

The full term for RLHF. Training AI to produce outputs that humans rate as helpful, harmless, and honest.

Rectified Linear Unit. A simple activation function that outputs zero for negative inputs and the input itself for positive ones. Very widely used.

Training AI to automatically discover useful representations of raw data, rather than manually designing features.

A second-stage process that refines initial search results by applying a more sophisticated relevance model.

A deep neural network architecture using skip connections that allow information to bypass layers, enabling much deeper networks.

Developing and deploying AI with careful consideration for ethics, fairness, safety, and societal impact.

The practice of reporting AI vulnerabilities or safety issues privately to developers before making them public.

The principle of increasing AI capabilities in a measured way with appropriate safety measures at each stage.

Representational State Transfer. A common architectural style for web APIs that uses standard HTTP methods.

The process of finding and fetching relevant information from a database or knowledge base to assist AI generation.

When an AI finds unexpected ways to maximise its reward signal that do not align with the intended goal.

An AI trained to predict how helpful or harmful a response is, used to guide the training of other models.

The principle that individuals affected by AI decisions should be able to receive a meaningful explanation of how the decision was made.

Regulating AI according to the level of risk it poses, with stricter rules for high-risk applications.

Reinforcement Learning from AI Feedback. Using another AI model to provide training feedback instead of humans.

Reinforcement Learning from Human Feedback. Training AI using human ratings of its outputs to make it more helpful, harmless, and honest.

Recurrent Neural Network. A neural network designed for sequential data like text or time series, processing one element at a time.

The field of designing, building, and programming physical machines that can sense, decide, and act in the real world.

Receiver Operating Characteristic curve. A graph showing the trade-off between true positive and false positive rates.

Robotic Process Automation. Software bots that automate repetitive, rule-based business tasks like data entry and form filling.

A standardised test specifically designed to evaluate how safely an AI model behaves.

South Korean conglomerate investing heavily in AI chips, on-device AI, and smart device integration.

A controlled, isolated testing environment for safely experimenting with new AI technologies or regulations.

Using AI to interpret satellite photos for agriculture, environmental monitoring, urban planning, and disaster response.

Research showing predictable relationships between model size, data volume, compute budget, and AI performance.

Software Development Kit. A collection of tools, libraries, and documentation that helps developers build applications using a particular platform.

Software that indexes and retrieves relevant information from a large collection of data in response to queries.

A mechanism where each element in a sequence learns how much to attend to every other element, enabling rich contextual understanding.

Vehicle technology that uses AI sensors and algorithms to navigate roads without a human driver.

A training method where the model generates its own labels from the data, such as predicting masked words in a sentence.

Searching by meaning rather than exact keywords, using embeddings to find conceptually related content.

A training approach using a small amount of labelled data combined with a large amount of unlabelled data.

The base material and industry for manufacturing the chips that power AI and all modern computing.

Combining data from multiple sensor types (cameras, LIDAR, radar) to create a more complete understanding of the environment.

Using AI to determine whether a piece of text expresses positive, negative, or neutral emotions.

Assigning a numerical value to text representing how positive or negative the expressed sentiment is.

Sequence-to-sequence. A model architecture that transforms one sequence into another, used in translation and summarisation.

Using software to model real-world environments so AI agents can learn and be tested safely before physical deployment.

An AI system that processes each input independently within a fixed, non-evolving state, without retaining memory or context from prior interactions. Every request is treated as a fresh transaction, making the system stateless. This contrasts with multi-turn or agentic AI systems that maintain conversational history or evolving world models.

A shortcut that allows data to bypass one or more layers in a neural network, helping with training very deep networks.

Urban areas that use AI, IoT, and data analytics to improve infrastructure, services, and quality of life.

Japanese conglomerate and major technology investor with significant stakes in AI companies globally.

A mathematical function that converts a set of numbers into probabilities that sum to one, used for classification outputs.

OpenAI's text-to-video AI model capable of generating realistic video clips from text descriptions.

National initiatives to develop domestic AI capabilities independent of foreign providers, ensuring technological self-reliance.

When an AI satisfies the literal requirement of its objective but violates the spirit of what was intended.

AI that identifies and processes human speech, converting it to text or commands.

AI that converts spoken audio into written text. Also known as automatic speech recognition.

An open-source AI image generation model developed by Stability AI, widely used and modified by the community.

Processing data continuously as it flows in, rather than in static batches.

AI generating responses in specific formats like JSON, tables, or code, rather than free-form text.

AI technique that applies the visual style of one image (such as a painting) to the content of another.

AI condensing longer text into shorter versions while retaining the key information.

A mobile application that combines many services (messaging, payments, shopping, transport) into one platform, common across Asia.

Training a model on labelled data where both inputs and correct outputs are provided.

Using AI to improve efficiency, reduce costs, and manage risks across the supply chain.

A classification algorithm that finds the best boundary to separate different categories of data.

Coordinated behaviour of many simple agents producing complex collective intelligence, inspired by bees or ants.

Artificially generated data used to train AI when real data is scarce, expensive, or contains privacy concerns.

Any media content (text, images, audio, video) that is created or significantly modified by AI.

Instructions provided to an AI model at the start of a conversation that set its behaviour and constraints.

Hidden instructions given to an AI model that define its behaviour, personality, and constraints before the user interacts with it.

Breaking a complex task into smaller, manageable sub-tasks that an AI agent can handle individually.

A hierarchical classification system that organises concepts into categories and subcategories.

The accumulated cost of shortcuts or quick fixes in software that will need to be addressed later.

A setting that controls how creative or random an AI's outputs are. Higher temperature means more varied responses.

Chinese technology conglomerate with significant AI research spanning gaming, social media, and enterprise applications.

AI producing new text based on a prompt or input, ranging from completing sentences to writing entire articles.

AI models that generate images based on written descriptions or prompts.

AI that converts written text into natural-sounding spoken audio.

AI systems that generate video content from text descriptions.

Term Frequency-Inverse Document Frequency. A method for measuring how important a word is to a document relative to a collection.

The amount of data or number of requests an AI system can process in a given time period.

Data points collected at regular intervals over time, such as stock prices, weather readings, or sensor data.

The cost structure of AI APIs based on the number of tokens processed in inputs and outputs.

The maximum number of tokens an API call can process, combining both input and output.

Another term for context window. The maximum number of tokens a model can process in a single interaction.

Software that breaks text into smaller units (tokens) that an AI model can process. Different models use different tokenisation strategies.

The process of breaking input data (usually text) into smaller units called tokens that an AI model can process.

Small chunks of text, such as words or word fragments, that AI language models break text into for processing.

Enhancing AI responses by giving the model access to external tools like calculators, search engines, or databases.

An AI model's ability to call external tools, APIs, or software to accomplish tasks beyond text generation.

A text generation method that limits the AI to choosing from only the K most likely next words.

Also called nucleus sampling. A method that limits word choices to the smallest set whose combined probability exceeds a threshold P.

Tensor Processing Unit. Google's custom chip designed specifically for machine learning workloads.

The process of teaching an AI model by feeding it data and adjusting its internal settings until it performs well.

The one-time expense of training an AI model, which can range from thousands to hundreds of millions of dollars for frontier models.

Applying knowledge gained from one task to a different but related task, saving time and data.

The neural network architecture behind most modern AI language models. Uses attention mechanisms to process all parts of the input simultaneously.

AI that is reliable, transparent, fair, and respects privacy, meeting the expectations of its users and society.

Taiwan Semiconductor Manufacturing Company. The world's largest contract chip maker, producing most of the world's advanced AI chips.

A test proposed by Alan Turing where a machine passes if a human cannot distinguish it from another human in conversation.

When a model is too simple to capture the patterns in the data, performing poorly on both training and new data.

Training a model on unlabelled data to discover hidden patterns or groupings without being told the answers.

Variational Autoencoder. An autoencoder that learns a probability distribution over the compressed representation, enabling generation of new data.

The challenge of ensuring AI systems internalise and act according to human values and preferences.

A database optimised for storing and searching high-dimensional numerical representations (embeddings) of data.

Finding the most similar items in a database by comparing their numerical representations (vectors).

Systems for tracking changes to code or documents over time, allowing collaboration and rollback.

A transformer architecture adapted for processing images by treating image patches as tokens.

An AI model that can understand both images and text together, answering questions about visual content.

Vision Transformer. The original paper and model that successfully applied transformer architecture to image classification.

The complete set of tokens that an AI language model can recognise and work with.

AI software that uses speech recognition and NLP to respond to voice commands, such as Siri or Alexa.

AI that can replicate a specific person's voice from a small audio sample.

AI systems that identify hidden markers in content to determine whether it was generated by AI.

Embedding invisible or visible markers in AI-generated content to identify its origin and help detect synthetic media.

The numerical values inside a neural network that are adjusted during training to improve accuracy.

OpenAI's open-source automatic speech recognition system that can transcribe and translate audio in multiple languages.

An early and influential method for creating word embeddings where similar words have similar numerical representations.

An AI's internal representation of how the world works, allowing it to predict outcomes and simulate scenarios.

Explainable AI. The field of research focused on making AI systems more transparent and understandable to humans.

You Only Look Once. A fast object detection algorithm that processes an entire image in a single pass.

Asking a model to perform a task it was not specifically trained on, without providing any examples.