Loading lesson…
AI Foundations / What AI Is
What today's systems can and cannot do.
Reviewed by Yuvaraj
Almost every AI system in use today has something in common: it does one kind of thing. Understanding why is the fastest way to see through hype.
Narrow AI (also called weak AI) is built and trained for a specific task. It can be superhuman at that task and completely helpless outside it.
This is not a temporary limitation of a few products. Every deployed AI system today is narrow AI, including the large, impressive ones. They can be broad, a modern language model handles translation, summarizing, coding, and question-answering, but their competence still comes from the data and objective they were trained on, and it degrades outside that.
Broad is not the same as general
A large language model looks general because language touches everything. But ask it to control a robot arm, or to reason reliably about a game it never saw, and the breadth stops. Breadth within one modality (text) is not cross-domain generality.
A model learns a mapping from inputs to outputs that fits its training distribution, the kinds of examples it actually saw. Two consequences follow directly:
Answer from memory before revealing, retrieval practice is what builds durable recall.
A program that plays chess at superhuman level but can do nothing else is an example of:
Artificial General Intelligence (AGI), matching human ability across essentially any task, exists in production today.
Ask about this lesson, or about anything in AI. Answers cite the lessons they draw on.
Finished this lesson?
Mark it complete to earn XP, keep your streak, and schedule a review.
That single fact, competence is tied to the training distribution, is the engine behind every narrow-AI story on this page.
Picture a model trained only on daytime, clear-weather dashcam footage to spot pedestrians. On its test set it scores 99%. Deploy it at night, in rain, and accuracy can collapse, not because the model "broke," but because rainy-night frames are out-of-distribution. The lesson: a benchmark number is only a promise about inputs that resemble the benchmark.
Artificial General Intelligence (AGI) would be a single system that learns and reasons across essentially any task a human can, transferring knowledge from one domain to another the way people do.
Status: does not exist (KNOWN)
AGI does not exist today. Whether it is achievable, and if so when, is an open research and engineering question with genuine disagreement among experts. Treat any confident claim that AGI is "here" or "arriving next year" as marketing until you see evidence.
Artificial Superintelligence (ASI), a system far beyond human ability across the board, is a further step still, and entirely speculative. It is a topic for careful discussion, not an established capability.
| Narrow AI | General AI (AGI) | Superintelligence (ASI) | |
|---|---|---|---|
| Exists today? | Yes | No | No |
| Scope | One task or domain | Any task a human can do | Beyond human across the board |
| Example | Spam filter, image classifier | None | None |
| How to treat claims | Verify on the task | Skeptical, open question | Speculative |
When you read that an AI "understands" or "reasons like a person," run it through three questions before believing the headline:
Those questions turn a vague impression into something you can evaluate, a habit this whole platform is built around. If a claim cannot survive them, it is hype.