From artificial intelligence to AGI: a working taxonomy of AI, machine learning and deep learning

Clients rarely ask for machine learning or deep learning by name. They ask for AI, and the word arrives carrying very different expectations: a chatbot, a prediction, an automation or, increasingly, something close to a digital employee. When the vocabulary is loose, so is the scope, and projects are estimated against the wrong problem.

This note has a narrow aim. It reviews how each term is defined in the literature, arranges the terms into one working taxonomy, and shows how that taxonomy changes the questions a team should ask before building anything.

The question of machine intelligence was framed by Turing as a behavioural test rather than a definition [1]. The phrase artificial intelligence itself was introduced in the 1955 proposal for the Dartmouth summer project, which conjectured that every aspect of learning or intelligence could in principle be described precisely enough for a machine to simulate it [2].

Machine learning entered the vocabulary through work on programs that improved at checkers with experience [3]. Mitchell later gave the definition still used in teaching: a program learns from experience E with respect to a task T and a performance measure P if its performance at T, measured by P, improves with E [4]. Modern textbooks treat the field through the lens of rational agents that perceive and act in an environment [5].

Deep learning describes representation learning with many layers, where each layer learns features at a higher level of abstraction from the one below [6] [7]. Foundation models is a more recent term for very large models trained on broad data that can be adapted to many downstream tasks [8].

We propose reading the terms as nested sets rather than competing labels. Artificial intelligence is the outer set: any system that performs a task we would otherwise call intelligent, including hand-written rules. Machine learning is the subset that improves from data instead of rules. Deep learning is the subset of machine learning that uses many-layered neural networks. Foundation models, including large language models, are the subset of deep learning trained at very large scale on broad data.

Artificial general intelligence does not fit inside this chain as another technique. It names a capability threshold. Legg and Hutter define intelligence as an agent's ability to achieve goals across a wide range of environments [9], and Morris and colleagues propose levels of AGI that combine depth of performance with breadth of generality [10]. Under both views, current systems are strong on some tasks and unreliable on others, so they fall short of general intelligence.

Two formal statements anchor the taxonomy. Mitchell's definition makes learning measurable: a program learns when its performance at a task, judged by a chosen measure, improves with experience [4]. The definition matters in practice because it forces a team to name the task, the measure and the experience before building anything.

Legg and Hutter go further and define a universal measure of intelligence for an agent as its expected performance across all computable environments, each weighted by its simplicity [9]. The measure is not computable in practice, but it captures why generality is the hard part: performing well in one environment says little about performance across many.

We applied the taxonomy as a scoping checklist to iBOY work that uses AI. The method is qualitative: for each system we recorded which layer of the taxonomy does the work, what data it depends on, and where a person stays in the loop. No user study or benchmark was run for this note.

GreenSignal AI, an iBOY product, uses large language models to simulate how a synthetic audience of 10,000 users built from a persona responds to campaign variants. Lumirva AI presents machine intelligence through a product interface and API. iBOY's internal studio tools use language models to draft research and messages that a person approves before anything leaves the studio.

The difference between the outer two layers of the taxonomy fits in a few lines of code. A rule is written by a person and does exactly what it says. A learned model is fitted from labelled examples and generalises, for better or worse, to cases nobody wrote down.

The listing below routes a message both ways. The rule is transparent but brittle; the logistic regression learns its weights by gradient descent on the log loss, which is the simplest instance of the training loop used, at vastly larger scale, in deep learning [7].

Naming the layer changed the questions. When a request was really a rule, such as routing an enquiry by service, a rule was cheaper, explainable and sufficient. When the task depended on patterns in historical records, classical machine learning needed clean, labelled data that clients often did not yet have. When the task was language heavy, a foundation model was the practical choice, and the open questions moved to grounding, evaluation and review.

In every system we examined, the reliable pattern was the same: the model prepares and a person decides, especially for anything that reaches a customer, spends money or makes a commitment.

The taxonomy is deliberately coarse. Reinforcement learning, probabilistic models and hybrid systems sit awkwardly inside the chain, and the boundary between a large deep learning model and a foundation model is a matter of scale and use rather than a sharp line. The findings come from a small set of studio projects and should be read as design guidance, not as generalisable evidence.

The AGI discussion is also time bound. Definitions of general intelligence remain contested, and claims that a product offers AGI today should be treated with caution.

Reading AI, machine learning, deep learning and foundation models as nested sets, and AGI as a threshold rather than a technique, gives teams a clearer first conversation. It points to the cheapest approach that can work and puts the hard questions of data, evaluation and human review on the table early.