Module 3 — Narrow AI, general AI: what research says
No topic in the field generates as much public confusion as this one. On one side, announcements of "imminent general intelligence"; on the other, systems that fail at tasks a child gets right. This module gives you the reference points to hold an informed position.
All production AI is narrow
An AI system deployed today excels in the exact scope it was built for, and nowhere else. The model that beats go champions cannot play checkers without retraining. The bank fraud detector does not detect insurance fraud. The vision model trained on chest X-rays does not read mammograms.
This narrowness is not a growing pain that will fix itself: it follows from the very principle of statistical learning. A model captures the regularities of its training data. Its competence is bounded by the distribution of that data. Outside that distribution, it does not "reason" to compensate — it extrapolates, often badly, and always with the same apparent confidence.
In a human, excellence in one domain comes with general capabilities: a radiologist can also read a form, explain their reasoning, recognize that a case exceeds their competence. A model has none of these side capabilities — and this is the most frequent intuition error among decision-makers: assuming that a system strong at one task is "generally intelligent". It is not.
What large language models change — and what they do not
Large language models (LLMs) blur the boundary in appearance: the same model translates, summarizes, writes code, answers questions about law and cooking. Is that not generality?
What actually changed. A single training task — predicting the continuation of text — on a significant share of all existing text produces a system usable across a very wide range of tasks expressible as text. That is an unprecedented generality of interface, and it has real economic value: where ten specialized models were needed, one foundation model can often serve as the common base.
What did not change. This versatility remains bounded by the distribution of text data. LLMs stay weak where training text is not enough: reliable arithmetic, long-horizon planning, strict factual consistency, physical environments, genuinely novel situations. And they inherit an awkward property: they produce a fluent answer even when they have no basis to answer — the hallucination phenomenon, covered in detail in the LLM course.
The honest phrasing: LLMs are systems that are narrow at one very broad task (modeling text), not general intelligences. The nuance sounds academic; in practice it separates solid use cases from projects headed for disappointment.
AGI: what exactly are we talking about?
"Artificial general intelligence" (AGI) refers to a hypothetical system with cognitive capabilities comparable to a human's across all domains: learning a new task from few examples, transferring knowledge between domains, reasoning about the physical and social world, knowing what it does not know.
Three reference points for reading the debates.
There is no consensual operational definition. Without a shared measurable criterion, announcements that "AGI arrives in 20XX" are neither verifiable nor refutable. Serious researchers propose test batteries; none is authoritative.
Expert predictions diverge radically — from "within ten years" to "not in this century" — and the field's history shows such predictions have always been unreliable, in both directions: the optimistic deadlines of 1965, but also the skepticism about the game of go, which fell ten years earlier than predicted.
The important questions do not wait for AGI. Bias, safety, concentration of power, job displacement, disinformation: all these issues are raised by today's narrow systems. The AGI debate, real in research, sometimes serves as a smokescreen that diverts attention from problems already here. Module 10 returns to this.
The reading grid for your projects
This distinction is not theoretical: it provides an immediate feasibility criterion.
Requests that assume generality — "an assistant that handles all of customer service", "a system that understands the company" — must be recut into narrow tasks: classify incoming requests, suggest replies for the ten most frequent motives, extract fields from forms. Each narrow task is feasible, measurable and improvable. The open-ended scope produces seductive demos and disappointing deployments.
Key takeaways
- All production AI is narrow: competence is bounded by the training distribution and does not transfer.
- LLMs bring a generality of interface (anything expressible as text), not general intelligence; their structural weaknesses remain.
- AGI has no consensual operational definition; the societal stakes are raised by today's systems, no need to wait.
- Project reflex: recut every "general" request into narrow, bounded, measurable tasks.
In the next module, we open the hood: how, concretely, a machine learns — training, loss function and gradient descent, with no math prerequisites.