Artificial intelligence has changed dramatically over the past decade. Systems that were once mainly associated with recognising patterns, classifying images or making predictions can now generate content, interpret multiple forms of information, reason through complex problems, operate software tools and increasingly carry out multi-step tasks.
What is AI?
Artificial intelligence, or AI, is the broad field of creating computer systems capable of performing tasks associated with human intelligence, including perception, language, learning, reasoning, problem-solving, decision-making and action.
AI is not one technology. Machine learning, deep learning, generative AI, large language models, computer vision, reasoning systems and autonomous agents are all parts of the wider field of artificial intelligence.
Terms such as AI, machine learning, GPT, agentic AI, AGI and ASI are often used as though they describe the same thing. They do not.
Some describe methods for building artificial intelligence. Others describe model architectures, behaviours or levels of capability. Recursive Self-Improvement describes a potential mechanism for improving AI, while the technological singularity describes a hypothetical consequence of extremely rapid technological development.
| Term | What it describes | Simple meaning |
|---|---|---|
| AI (Artificial Intelligence) | The overall field | Computer systems performing tasks associated with intelligence |
| Machine Learning | A method for building AI | Systems learning patterns from data |
| Deep Learning | A form of machine learning | Learning using multilayer neural networks |
| ANI (Artificial Narrow Intelligence) | A capability category | AI specialised in a limited task or domain |
| LLM (Large Language Model) | A type of AI model | A large model trained to process and generate language |
| GPT (Generative Pre-trained Transformer) | A family and approach to AI models | Generative models based on Transformer technology |
| Generative AI | A capability | AI that creates new text, images, audio, video, code or other content |
| Multimodal AI | A capability | AI that works across more than one type of information |
| Reasoning AI | A capability | AI designed to perform more deliberate problem-solving |
| Agentic AI | A system behaviour | AI that can pursue goals by planning and taking multiple actions |
| Embodied AI | A system behaviour | AI that perceives and acts within a physical or virtual environment |
| RSI (Recursive Self-Improvement) | A possible improvement mechanism | AI improving the processes or tools it uses to improve itself |
| AGI (Artificial General Intelligence) | A proposed capability level | Broad and adaptable intelligence across many intellectual tasks |
| ASI (Artificial Superintelligence) | A hypothetical capability level | General intelligence substantially exceeding human intellectual capability |
| Technological Singularity | A hypothetical technological event or period | Technological change becoming extremely rapid and difficult to predict |
The change since roughly 2016 has involved far more than improvements in computing speed.
AI has moved from systems that were largely associated with recognising and predicting to systems that can generate, communicate, reason, use tools, and increasingly act.
The pace of adoption has accelerated alongside technical capability. Stanford University's 2026 AI Index reports that generative AI reached nearly 53% population-level adoption within three years. It also reports that 88% of surveyed organisations were using AI in 2025.1
Agent capability has also improved rapidly. On OSWorld, a benchmark that evaluates AI agents performing real computer tasks, reported task success increased from roughly 12% to around 66%. The same evidence also shows why caution is necessary: even leading systems still fail a substantial proportion of structured tasks.1
| Period | Major development | What AI increasingly became able to do |
|---|---|---|
| Around 2016 | Deep learning and narrow AI | Recognise, classify and predict |
| 2017 to 2020 | Transformers and large-scale pre-training | Learn representations transferable across more tasks |
| 2020 to 2022 | Large generative models | Generate increasingly sophisticated text and code |
| 2022 to 2024 | Conversational generative AI | Interact with people through natural language at mass scale |
| 2023 to 2025 | Multimodal and reasoning systems | Work across text, images, audio, and complex problems |
| 2025 to 2026 | Agentic systems | Use tools, operate software, and complete multi-step tasks |
| Emerging direction | Embodied, adaptive and increasingly autonomous AI | Perceive, reason and act continuously within environments |
Symbolic AI is an approach in which knowledge, rules, and logical relationships are explicitly represented in software.
Instead of learning every behaviour from data, a developer can define rules such as:
IF condition A is present AND condition B is present, THEN perform action C.
Expert systems in medicine, engineering and other specialist fields often used this type of approach.
Symbolic systems can be valuable where rules need to be clear, constrained, or auditable. Their limitation is that the real world contains enormous numbers of possible situations, making it impractical to manually specify every rule.
Symbolic AI has not disappeared. Modern AI systems can combine learned models with explicit software rules, databases, search systems, and other deterministic components.
Machine learning is a branch of AI in which computer systems learn patterns from data rather than having every rule explicitly programmed by a human.
Imagine trying to teach a computer to recognise an abnormality on a medical image.
One approach would be to attempt to describe every possible visual rule. Machine learning offers another approach: provide examples and allow an algorithm to learn statistical relationships that help distinguish one category from another.
| Type | How it works | Example |
|---|---|---|
| Supervised learning | Learns from labelled examples | Classifying an image as normal or abnormal |
| Unsupervised learning | Finds patterns or groups without explicit labels | Grouping similar cases within a dataset |
| Self-supervised learning | Learns from structure already present in large datasets | Pre-training modern language and vision models |
| Reinforcement learning | Learns through actions, outcomes, and reward signals | Games, robotics, optimisation and model training |
Deep learning is a form of machine learning based on neural networks containing multiple layers of computation.
Deep learning drove major advances in computer vision, speech recognition, language processing and prediction during the 2010s.
Many systems from this period became exceptionally capable at perception and classification.
A system could outperform humans at a specific task while still having very limited capability outside that task.
ANI stands for Artificial Narrow Intelligence. It describes AI that is specialised in a particular task or limited range of tasks.
Examples of narrow AI include:
Narrow intelligence can be extremely powerful. A system can be better than almost any human at one defined activity without possessing broad, adaptable human-like intelligence.
A major technical turning point occurred in 2017 with the publication of Attention Is All You Need by Vaswani and colleagues.
The researchers introduced the Transformer, a neural network architecture based on attention mechanisms rather than the recurrent architectures that had commonly been used for sequence processing.2
Transformers proved highly effective for language tasks and later became foundational to many modern large language models and generative AI systems.
Generative AI is artificial intelligence capable of creating new content based on patterns learned during training.
Earlier predictive systems were often designed to classify or forecast. Generative systems can instead produce something new.
Generative AI can create or manipulate:
A useful distinction is:
Predictive AI often asks: What is this, or what is likely to happen?
Generative AI can ask: What can be created from this information?
LLM stands for Large Language Model. An LLM is an AI model trained on large quantities of language data so that it can model relationships between tokens, words, concepts, and sequences and generate useful responses.
Modern LLMs can perform tasks including:
The term language model can now be slightly misleading because many advanced models can also work with images, audio and other forms of information.
GPT-3, described in the 2020 paper Language Models are Few-Shot Learners, demonstrated that scaling a language model could produce increasingly broad task performance without separately training a new model for every individual task.3
GPT stands for Generative Pre-trained Transformer.
OpenAI describes GPT models as generative pre-trained transformers.4
The model can generate outputs rather than only classify existing data.
The model undergoes broad training before it is used or further adapted for specific applications.
The model uses Transformer technology, which grew from the attention-based architecture introduced in 2017.
GPT is not another word for AI.
GPT describes a family and approach within the much larger field of artificial intelligence.
ChatGPT is also not a synonym for AI. It is an AI application. OpenAI publicly introduced ChatGPT on 30 November 2022 as a conversational system capable of responding to follow-up questions and interacting through dialogue.5
Humans do not experience the world entirely through text.
We simultaneously process language, vision, sound, movement, space, physical interaction, and context.
Multimodal AI is artificial intelligence capable of working across more than one type of information or modality.
For example:
text + image + audio + video
A multimodal system could examine an image, consider accompanying written information, and then discuss its observations conversationally.
In immersive learning, the concept becomes more interesting because a system could potentially interpret:
Reasoning AI refers to systems designed to devote additional computation to solving a problem, evaluating intermediate possibilities and producing a considered response.
This has contributed to large improvements in tasks involving mathematics, science, coding and complex analytical work.
It is important not to interpret impressive performance on one benchmark as evidence that an AI is uniformly intelligent.
Stanford's 2026 AI Index describes a continuing jagged frontier in which systems can achieve exceptional performance on some difficult tasks while remaining unreliable on apparently simpler ones.1
Generative AI changed what computers could produce.
Agentic AI is beginning to change what computers can do.
Agentic AI refers to artificial intelligence systems that can pursue goals through sequences of actions, often using planning, memory, external tools, and feedback from the environment.
A conventional chatbot interaction can be simplified as:
Prompt → Response
An agentic system can operate more like:
Goal → Plan → Action → Observe Result → Revise Plan → Take Another Action → Complete Task
An AI agent might:
An LLM can be an important component of an AI agent, but an LLM and an AI agent are not the same thing.
| Generative AI | Agentic AI |
|---|---|
| Primarily creates outputs | Primarily pursues objectives |
| Often responds to a prompt | Can plan a sequence of actions |
| May generate text, images, audio or code | May use software, APIs, tools, and other systems |
| Interaction may end after the response | Can observe results and adapt subsequent actions |
| Can form one component of an agent | Can combine models, memory, tools, rules and permissions |
Embodied AI refers to AI that perceives an environment and takes actions within that environment.
Robotics is an obvious example, but embodiment does not have to mean a physical robot.
An AI agent can also act within a simulated or virtual 3D environment.
Google DeepMind's SIMA research explores generalist agents that can follow natural-language instructions and perform tasks within different 3D virtual environments.6
This distinction could become particularly important for education and simulation.
Instead of merely asking:
What would you do next?
an intelligent system could observe:
What did the learner actually do next?
That creates a much richer source of evidence about competence.
RSI stands for Recursive Self-Improvement. It describes the idea that an AI system could improve the mechanisms, processes, or tools through which it improves itself.
AI systems are already used to assist with AI research and development, including:
That does not automatically constitute strong recursive self-improvement.
In a stronger RSI scenario, an AI could become better at improving its own capabilities, then use those improved capabilities to make the next round of improvements more effectively.
AI 1 improves itself → AI 2 is more capable → AI 2 improves itself more effectively → AI 3 is more capable → repeat
Recursive self-improving software has been discussed in AI research for many years. Yampolskiy's 2015 paper examined different forms of self-improving software as well as computational limits that could restrict recursive improvement.7
More recent work also distinguishes bounded self-refinement, which already occurs in practical AI systems, from open-ended autonomous recursive self-improvement.8
Strong, open-ended RSI therefore remains a proposed future mechanism rather than an established route to unlimited intelligence.
AGI stands for Artificial General Intelligence. It generally refers to artificial intelligence with broad and adaptable capability across many different intellectual tasks rather than expertise in only one narrow domain.
There is no single universally accepted test or threshold for determining when AGI has been achieved.
Researchers from Google DeepMind have proposed describing AGI according to levels of both performance and generality, rather than treating it as a simple binary threshold.9
| ANI (Artificial Narrow Intelligence) | AGI (Artificial General Intelligence) |
|---|---|
| Specialised intelligence | Broad general capability |
| Strong within particular tasks or domains | Adaptable across many different intellectual tasks |
| Clearly exists today | The definition and threshold remain debated |
| Can be superhuman at one narrow task | Would demonstrate broad capability rather than isolated excellence |
AGI does not necessarily imply consciousness, emotion or sentience. Intelligence and subjective experience are separate questions.
Autonomy is also separate from general intelligence. A narrow system can be highly autonomous, while a broadly capable model can operate under strict human control.
ASI stands for Artificial Superintelligence. It describes a hypothetical artificial intelligence whose general intellectual capabilities substantially exceed those of humans.
ASI therefore describes a proposed level beyond AGI.
ANI (Artificial Narrow Intelligence) → AGI (Artificial General Intelligence) → ASI (Artificial Superintelligence)
| Term | Full name | Meaning | Status |
|---|---|---|---|
| ANI | Artificial Narrow Intelligence | AI specialised in limited tasks or domains | Exists today |
| AGI | Artificial General Intelligence | Broad intelligence across many intellectual tasks | Definition and threshold remain debated |
| ASI | Artificial Superintelligence | General intelligence substantially exceeding human capability | Hypothetical |
Importantly, AGI does not automatically imply ASI.
A generally capable AI could encounter limits involving computing power, energy, data, algorithms, economics, hardware, regulation or physical constraints.
A 2026 Google DeepMind paper examining possible transitions from AGI to ASI identifies several potential pathways, including further scaling, new AI paradigms, recursive improvement, and large-scale multi-agent systems. The authors also discuss possible frictions and bottlenecks rather than assuming that movement from AGI to ASI would be automatic.10
The technological singularity is a hypothetical point or period at which technological progress becomes so rapid and transformative that subsequent developments become extremely difficult for humans to predict.
The singularity is not another type of AI.
It describes a proposed technological event or transition.
ASI and the singularity are therefore related concepts in some theories, but they are not the same thing.
| Stage | Meaning |
|---|---|
| AGI (Artificial General Intelligence) | A broadly capable artificial intelligence |
| RSI (Recursive Self-Improvement) | One possible mechanism through which an advanced AI could accelerate its own improvement |
| ASI (Artificial Superintelligence) | A hypothetical level of general intelligence substantially beyond human capability |
| Technological Singularity | A hypothetical period of technological change that becomes extremely rapid and difficult to predict |
This sequence is a possible scenario, not an established law of AI development.
AGI might not lead to recursive self-improvement.
Recursive self-improvement could encounter significant technical limits.
ASI could theoretically emerge through mechanisms other than RSI.
Society could also experience major technological transformation without a sudden singularity occurring.
| Concept | Full form | What it means | Current status |
|---|---|---|---|
| AI | Artificial Intelligence | The overall field of creating computer systems capable of intelligent behaviour | Established field |
| Symbolic AI | Symbolic Artificial Intelligence | AI using explicit rules, logic and representations | Exists today |
| Machine Learning | Not usually abbreviated in general writing | Systems that learn patterns from data | Exists today |
| Deep Learning | Not usually abbreviated in general writing | Machine learning using multilayer neural networks | Exists today |
| ANI | Artificial Narrow Intelligence | AI specialised in limited domains | Exists today |
| LLM | Large Language Model | A large model trained to process and generate language | Exists today |
| GPT | Generative Pre-trained Transformer | A family of generative models based on Transformer technology | Exists today |
| Generative AI | Generative Artificial Intelligence | AI that creates new content | Exists today |
| Multimodal AI | Multimodal Artificial Intelligence | AI that processes or generates more than one form of information | Exists today |
| Reasoning AI | Reasoning Artificial Intelligence | AI designed for more deliberate problem-solving | Exists today |
| Agentic AI | Agentic Artificial Intelligence | AI capable of pursuing objectives through multiple actions | Exists today and developing rapidly |
| Embodied AI | Embodied Artificial Intelligence | AI that perceives and acts within physical or virtual environments | Exists today and developing rapidly |
| RSI | Recursive Self-Improvement | AI improving the processes or mechanisms it uses to improve itself | Bounded forms exist; strong open-ended RSI remains hypothetical |
| AGI | Artificial General Intelligence | Broad intelligence across many different intellectual tasks | Definition and threshold remain disputed |
| ASI | Artificial Superintelligence | General intelligence substantially exceeding human intellectual capability | Hypothetical |
| Singularity | Technological Singularity | A hypothetical period of extremely rapid and difficult-to-predict technological change | Hypothetical |
The most significant development in artificial intelligence may not be any single model.
It is the convergence of capabilities.
Historically, computer systems separated functions that humans experience as parts of one continuous process.
One system recognised an image. Another processed speech. Another searched a database. Another calculated an answer. Another controlled a machine.
Those boundaries are increasingly disappearing.
perception + language + knowledge + reasoning + memory + tools + action
Once those capabilities are combined, AI begins to look less like a conventional piece of software waiting for one instruction and more like an intelligent participant within a larger system.
The convergence of multimodal, reasoning, agentic, and embodied AI is particularly significant for immersive learning.
A traditional digital tutor mainly receives what the learner explicitly tells it.
An intelligent system inside a simulation can potentially observe performance itself.
For example, it could detect that a learner:
The AI could then use those observations as evidence.
It could potentially change the simulation, generate a new scenario, provide targeted guidance, present remediation, increase difficulty, compare performance over time or decide that intervention is not yet necessary.
This creates an important shift.
AI moves from delivering information to observing, interpreting and responding to what a learner actually does.
In immersive learning, that could eventually allow AI to build richer models of competence based on behaviour rather than relying only on quiz answers or completion data.
AI stands for Artificial Intelligence. It is the broad field concerned with creating computer systems capable of tasks associated with intelligence, including learning, reasoning, perception, language, and decision-making.
No. Machine learning is one method used to create artificial intelligence. AI is the broader field.
ANI stands for Artificial Narrow Intelligence. It describes AI specialised in a particular task or limited domain.
LLM stands for Large Language Model. An LLM is trained on large quantities of language data and can process and generate language.
GPT stands for Generative Pre-trained Transformer. GPT models are generative models built using Transformer technology.
No. GPT is one family and approach within artificial intelligence. AI also includes machine learning, robotics, computer vision, reinforcement learning, symbolic reasoning and many other technologies.
Yes. ChatGPT is an application that uses artificial intelligence models. ChatGPT is not another name for the entire field of AI.
Generative AI is artificial intelligence capable of creating new content, including text, images, audio, video, and computer code.
Multimodal AI is artificial intelligence capable of processing or generating more than one type of information, such as text, images, audio and video.
Agentic AI refers to AI systems capable of pursuing goals through multiple actions, often using planning, memory, tools, APIs, and feedback from the environment.
Generative AI primarily creates outputs. Agentic AI can take a sequence of actions to achieve an objective. A generative AI model can be one component inside an agentic system.
RSI stands for Recursive Self-Improvement. It refers to the idea that an AI could improve the processes or tools through which it improves itself.
AI already assists with coding, optimisation, model evaluation and AI research. Bounded forms of self-refinement also exist. This is different from an autonomous system repeatedly redesigning itself and achieving unrestricted increases in general intelligence. Strong open-ended RSI remains hypothetical.
AGI stands for Artificial General Intelligence. It broadly refers to AI capable across many different intellectual tasks rather than a narrow specialist domain.
There is no universally accepted definition or test for AGI. Whether a particular system qualifies therefore depends heavily on the definition and performance threshold being used.
ASI stands for Artificial Superintelligence. It describes a hypothetical AI whose general intellectual capability substantially exceeds that of humans.
No. AGI describes broadly capable general intelligence. ASI describes a hypothetical level of general intelligence substantially beyond human capability.
The technological singularity is a hypothetical period in which technological progress becomes so rapid and transformative that subsequent developments become extremely difficult to predict.
No. ASI describes a hypothetical level of intelligence. The technological singularity describes a hypothetical technological transition or period of rapid change.
No. AGI does not automatically imply ASI. Possible limits include computing resources, energy, hardware, available data, algorithms, physical constraints and the ability of AI systems to improve themselves.
The development of artificial intelligence is no longer simply the story of computers becoming better at answering questions.
It is increasingly the story of different components of intelligent behaviour being combined.
The unresolved question is how far this convergence can continue.
Can artificial intelligence become genuinely general?
Can generally capable AI substantially improve its own capabilities?
Would AGI eventually lead to ASI?
Could recursive self-improvement accelerate technological progress enough to contribute to a singularity?
We do not yet know.
What we can observe is that AI has moved rapidly from specialised pattern-recognition systems towards increasingly capable systems that combine perception, generation, reasoning, memory, tools and action.
The next decade may therefore be defined not simply by more powerful AI, but by AI systems that participate continuously within the digital, simulated and physical environments around us.
Artificial General Intelligence, Artificial Superintelligence, Recursive Self-Improvement and the technological singularity remain areas in which terminology, thresholds and possible future pathways are actively debated. They should not be treated as interchangeable concepts.