Module 1: What Is AI Really?

4. Types of AI: Narrow vs General, Weak vs Strong

Description

In this lesson you'll classify types of artificial intelligence by their scope and their capabilities. It isn't just theory: these labels show up in articles, reports and conversations about regulation, risks and expectations. Knowing what "narrow AI" means versus "general AI" (or "weak" versus "strong") lets you read and speak precisely and not confuse what exists today with what is hypothesis or science fiction.

When someone says "ChatGPT is AGI" or "AI already surpasses humans", knowing whether they're talking about narrow or general (and weak or strong) helps you evaluate the claim. For an AI Engineer, almost everything you build is narrow AI: systems that are very capable at specific tasks, not "general minds". This lesson gives you the vocabulary and the criteria to classify and to communicate well.

Why classify types of AI

Classifying isn't an academic exercise; it has practical consequences:

Communication: When you talk to product, business or regulatory people, using "narrow AI" or "AGI" precisely avoids misunderstandings. "This model is narrow" means "it does X and Y well, not any task"; "AGI" means "it doesn't exist today, it's a hypothesis".
Expectations: If you assume an LLM is "almost AGI", you'll expect it to do any intellectual task and you'll be frustrated when it fails at facts, long reasoning or domains outside its training. If you treat it as narrow AI that's very capable at language, you'll design the system with limits in mind (RAG, verification, fallbacks).
Regulation and standards: Laws like the EU AI Act distinguish risk levels by type of application; knowing whether your system is narrow (classification, recommendation, generation in a bounded domain) helps you locate requirements (transparency, evaluation, human oversight).
Product design: Choosing the task (classification, generation, RAG, agent) and the model (API, fine-tuning, open source) means assuming you're building narrow AI. You don't design for "AGI"; you design for "doing X well in domain Y".

Narrow AI vs General AI

Narrow AI (ANI — Artificial Narrow Intelligence)

Definition: Systems that do one or a few defined tasks well, in a bounded domain. They don't generalize to any problem nor do they have broad "common sense".

Examples:

Recognizing cats in photos.
Translating between two languages.
Recommending movies.
Playing Go or poker.
Answering questions about a corpus of documents (RAG).
Generating text or images from a prompt (today's LLMs).

Key characteristic: If you change the task or the domain (e.g. from "translating" to "driving a car"), the system is useless; another one has to be designed or trained. Everything that exists today in commercial products is narrow AI.

Analogy: A chess player that beats world champions is narrow AI in that domain; if you ask it to drive a car or translate a text, it doesn't have that ability. "Narrow" doesn't mean "dumb": it means limited in scope, not in quality within that scope.

Why it matters for an AI Engineer: Your job is to choose the task (classification, generation, recommendation, etc.), the model and the way to expose it (API, RAG, agent). All of that is narrow AI by definition: a system that does one or a few things well in a bounded domain.

General AI (AGI — Artificial General Intelligence)

Definition: (Hypothetical) Intelligence that could learn and perform any intellectual task a human can do: reason, plan, adapt to new contexts, combine abilities. It isn't limited to one domain.

Examples: There are no real examples; it's a goal or a reference concept. When someone says "AGI", they usually mean a possible future, not a system that exists today.

Why it matters: In debates about risks, regulation or long-term investment, people talk about "the path to AGI". For an AI Engineer's daily work, what's relevant is that everything you build today is narrow AI: systems that are very capable at specific tasks, not "general minds". When you read "we're close to AGI" or "this model is a step toward AGI", remember that AGI isn't defined operationally and there's no consensus on when (or whether) it will be reached.

Analogy: AGI would be like a human who can learn any intellectual task: reason, plan, adapt to new contexts, combine abilities. No such system exists today; LLMs are very capable at language and related tasks, but not at "any intellectual task".

Quick comparison

Aspect	Narrow AI (ANI)	General AI (AGI)
Scope	One or a few tasks/domains	Any human intellectual task
Existence	What we have today (GPT, Claude, etc.)	Doesn't exist; a hypothesis or long-term goal
Generalization	Limited to the trained domain	Theoretically broad
Use in practice	Products, APIs, integrations	Debate, research, expectations

Weak AI vs Strong AI

These terms are used mainly in philosophy and debate; in industry you hear "narrow" vs "general" more often.

Weak AI

Definition: Systems that simulate intelligent behavior or replicate it in bounded domains, without assuming they "understand" or have consciousness. They do the task, but no mind or real comprehension is attributed to them.

Example: An LLM that generates very coherent text but doesn't "know" in the sense a human knows; it only predicts sequences of tokens.

Everything that exists today in products is considered weak AI in this sense: very powerful tools, but without consciousness or comprehension in the strong philosophical sense.

Why it matters for an AI Engineer: You don't need to settle the philosophical debate; you do need to know that when someone says "the model understands" or "doesn't understand", they're mixing two things: (1) whether the system does the task well (measurable) and (2) whether it has "real understanding" (there's no consensus and no way to measure it). In practice, you design systems that behave well; you don't assume consciousness or comprehension.

Strong AI

Definition: (Hypothetical) Systems that would have a mind, consciousness or real understanding, not just a simulation of behavior. It's a philosophical concept; there's no consensus on how to measure it or whether it's possible.

Relationship with AGI: Sometimes "strong AI" is used as a synonym for AGI; other times "strong" emphasizes consciousness and "general" emphasizes breadth of tasks. For practical work it's enough to know that strong AI doesn't exist today and that the debate is theoretical and long-term.

Other useful classifications

By type of task

Classification: Assigning a label or category (spam/not spam, type of image).
Regression: Predicting a numeric value (price, temperature).
Generation: Producing new content (text, image, audio).
Reasoning/planning: Given goals and constraints, deciding steps (still very bounded in current systems).

By type of input/output

Text → text: Translation, summarization, chat (LLMs).
Text → image: Generating images from a description.
Image → text: Describing images (vision + language).
Multimodal: Several modalities (text + image + audio) in a single system.

By how it's trained or used

Supervised: Learns from labeled examples (you'll see this in Module 2).
Unsupervised: Learns patterns without explicit labels.
Reinforcement: Learns through rewards/penalties in an environment.
Foundation models / LLMs: Large pre-trained models that are then adapted or used as-is (fine-tuning, RAG, prompt engineering).

You don't need to memorize all of it; it's enough to have the idea that there are many ways to classify and that "narrow vs general" and "weak vs strong" are the ones that come up most in discussions of scope and expectations.

By type of deployment (useful for AI Engineering):

API / model as a service: You call a model (OpenAI, Anthropic, etc.) via API; you don't train or host it yourself.
Fine-tuning: You adapt a pre-trained model with your own data; you're still using a narrow base.
RAG (Retrieval-Augmented Generation): You connect an LLM to documents or databases to reduce hallucinations and update knowledge.
Agents: Models that plan, use tools (search, code, APIs) and execute steps; still narrow but with more autonomy within the task.

Use cases by type (examples for AI Engineering)

To connect the classifications with daily work, here are examples of what kind of system you build depending on the task:

Type of task	Example system	Narrow / General	Note
Text classification	Spam filter, sentiment detector, ticket categorization	Narrow	One or a few labels; bounded domain.
Text generation	Chatbot, summarization, translation, code generation	Narrow	Very capable at language; not "any" intellectual task.
RAG (questions over documents)	Internal support, FAQ, semantic search over docs	Narrow	Connects an LLM to your own sources; reduces hallucinations in a bounded domain.
Recommendation	Movies, products, content	Narrow	Learns from clicks and preferences; bounded domain.
Vision (classification or detection)	OCR, object detection, image moderation	Narrow	One or a few vision tasks.
Agent with tools	Agent that uses an LLM + search + code to solve a task	Narrow	More autonomous than a simple chatbot, but still limited to the domain of its tools.
LLM fine-tuning	Adjusting GPT or Llama with your own domain data	Narrow	Improves within a specific domain; it doesn't turn the model into AGI.
Multimodal (text + image)	Analyzing a document with text and images; generating an image from text	Narrow	Handles multiple modalities but is still limited to those modalities.

Key note: All the examples are narrow AI: they do one or a few tasks well in a bounded domain. Not because they're "bad" (they're very capable), but because they don't generalize to any intellectual task. When you design a system, assume it's narrow; that leads you to design for one domain and to validate that the system does that task well, not "anything at all".

In every case, the system is narrow: it does one or a few things well in a domain; not "any human intellectual task". AGI doesn't appear in the table because it doesn't exist today.

Visual summary: the narrow/general and weak/strong axes

To lock in the ideas, here's a summary of the two axes that come up most in discussions of scope and expectations:

Axis 1: Scope of tasks (narrow vs general)

Narrow (ANI): One or a few tasks in a bounded domain. Examples: translating, recommending, generating text, classifying images. Everything that exists today in products is narrow.
General (AGI): Any human intellectual task; not limited to one domain. Doesn't exist today; it's a hypothesis or long-term goal.

Axis 2: Consciousness/understanding (weak vs strong)

Weak: Systems that simulate or replicate intelligent behavior without attributing consciousness or "real understanding" to them. Everything current is weak in that sense.
Strong: Hypothetical systems with a real mind/consciousness. A philosophical concept; not realized.

Combinations:

Narrow + weak: What you build today (products, APIs, RAG, agents). Example: ChatGPT is narrow (language) and weak (no consciousness is attributed to it).
General + strong: What's debated for the future (AGI with consciousness); doesn't exist.
Narrow + strong: Conceptually possible (a system that does one task and "has consciousness"); there's no consensus and no way to measure it; it isn't relevant to practical work today.
General + weak: A system that does any intellectual task but "simulates" without consciousness; that doesn't exist today either.

When you read "narrow", "general", "weak" or "strong", use this summary to place what's being talked about.

Checklist for classifying a product or system

When you have to classify a product or system (in a report, a conversation or this exercise), use this checklist:

What task does it do? (One or a few tasks in a bounded domain → narrow; "any human intellectual task" → doesn't exist today.)
Is it personalized with data? (Yes → ML; not necessarily → it could be AI without ML in the strict sense.)
Does it generate new content (text, image, audio)? (Yes → GenAI; usually implies DL/LLMs.)
Does it use deep networks (voice, image, language)? (Yes → DL; not necessarily → it could be classic ML.)
Is consciousness or "real understanding" attributed to it? (No, in current products → weak AI; strong doesn't exist today.)
Is it "almost AGI" or "general"? (No; everything current is narrow; AGI doesn't exist today.)

Applied example: ChatGPT

Let's apply this checklist to ChatGPT to see how it's used in practice:

What task does it do? It generates text, answers questions, summarizes, translates, assists with code. Those are language tasks, not "any intellectual task". → Narrow AI.
Is it personalized with data? Yes; it was trained on data (text from the internet) and it learns patterns. → ML (and DL, because it's a Transformer).
Does it generate new content? Yes; it generates text that didn't exist as such. → GenAI.
Does it use deep networks? Yes; it's a Transformer (a deep neural network). → DL.
Is consciousness attributed to it? No; it simulates conversation but no mind or real understanding is attributed to it. → Weak AI.
Is it "almost AGI"? No; it fails at long reasoning, precise facts, coherence across thousands of lines, and tasks outside language. → Not AGI.

Full classification of ChatGPT: Narrow AI, ML, DL, GenAI (text generation), weak AI. It isn't AGI or strong AI.

Use this example as a template for classifying other products in the lesson 06 exercise.

If you answer these questions consistently, you'll have a solid classification. You don't need to know the internal technology with certainty; it's enough to apply the criteria and justify briefly.

One-page summary

Narrow AI (ANI): One or a few tasks in a bounded domain. Everything that exists today in products is narrow. "Narrow" doesn't mean "dumb"; it means limited in scope.
General AI (AGI): Any human intellectual task. Doesn't exist today; it's a hypothesis or long-term goal.
Weak AI: Simulates intelligent behavior without attributing consciousness or "real understanding". Everything current is weak.
Strong AI: Hypothetical systems with a mind/consciousness. A philosophical concept; not realized.
Different axes: Narrow/general = scope of tasks; weak/strong = consciousness/understanding. A system can be narrow and weak (e.g. a spam classifier).
By type of task: Classification, regression, generation, reasoning/planning (still bounded).
By type of deployment: API, fine-tuning, RAG, agents. All narrow.
Common mistakes: "ChatGPT is AGI" (false); "Narrow AI is dumb" (false); "Strong AI = more powerful than weak" (false); "Narrow and weak are the same" (false).
For an AI Engineer: Your job is to choose the task, the model and the way to expose it (API, RAG, agent). All of that is narrow AI by definition.
Resources: Stanford Encyclopedia, MIT Technology Review (AGI), Google (Types of AI), NIST, EU AI Act, GPT-4 Technical Report.
Design implications: When you design a system (RAG, classifier, agent), remember that everything you build is narrow AI; you don't design for AGI or for strong AI. That determines expectations (limits, evaluation, fallbacks).

Summary of implications for AI Engineering

The classifications in this lesson have direct consequences for how you work as an AI Engineer:

Everything you build today is narrow AI. Your job is to choose the task (classification, generation, RAG, agent), the model and the way to expose it (API, RAG, agent). You don't design for AGI; you design for "doing X well in domain Y".
Use "narrow" to communicate expectations. When you talk to product or business people, use "narrow" to make clear that the system does X and Y well, not "any task". That avoids wrong expectations (zero failures, infinite memory, total replacement of experts).
Don't assume consciousness or "real understanding". When someone says "the model understands" or "doesn't understand", ask whether they mean "it does the task well" (measurable) or "it has real understanding" (there's no consensus and no way to measure it). In practice, you design systems that behave well.
Classify by type of task and by deployment. Use the "Use cases by type" table in this lesson to connect the classifications with daily work (classification, generation, RAG, agents). It's all still narrow.
Avoid the common mistakes. Don't say "ChatGPT is AGI"; don't say "narrow AI is dumb"; don't confuse "strong" with "more powerful". Use this lesson's checklist when you classify a product or system.
Use the "Use cases by type" table when you define the task. When you define what your system does (classification, generation, RAG, agent), the table helps you place the type of system (narrow, by type of task and by deployment).

Use this summary when you design a system, when you communicate expectations or when you classify a product in a report.

Notes for going deeper

Why two axes (narrow/general and weak/strong)? Because "scope of tasks" and "consciousness/understanding" are different things. A system can be narrow (one task) and weak (no consciousness attributed to it); LLMs are narrow and weak. AGI would be general; strong would be with consciousness; they aren't synonyms.

What to do when someone says "the model understands"? Ask whether they mean "it does the task well" (measurable) or "it has real understanding" (there's no consensus and no way to measure it). In practice, you design systems that behave well; you don't assume consciousness or comprehension.

How to use this lesson in the bootcamp or in another guide? When you design a system (RAG, classifier, agent), remember that everything you build is narrow AI: one or a few tasks in a bounded domain. You don't design for AGI; you design for "doing X well in domain Y". That determines expectations (limits, evaluation, fallbacks).

What questions to ask when you hear "ChatGPT is AGI" or "almost AGI"? What tasks does it do well? (Language and related ones.) What tasks doesn't it do, or fails at? (Facts, long reasoning, domains outside its training.) AGI implies "any human intellectual task"; today's LLMs don't meet that.

Questions to reflect on

Why is ChatGPT narrow and not general? Because it does language tasks (generating, summarizing, translating) very well but not "any human intellectual task"; it fails at long-range coherence, facts and complex reasoning. Narrow means limited in scope, not in quality within that scope.

Why doesn't "narrow" mean "dumb"? Because narrow AI can be extremely good in its domain (e.g. playing Go, translating, generating text). "Narrow" means limited in scope (one or a few tasks), not in quality within that scope.

Why aren't weak and strong about "more or less powerful"? Because "strong" refers to consciousness/understanding (a philosophical concept); "weak" is simulation without a mind. A weak system can be far more useful in practice than a hypothetical strong one.

Why is everything you build today narrow? Because your job is to choose the task (classification, generation, RAG, agent), the model and the way to expose it (API, RAG, agent). All of that is narrow AI by definition: a system that does one or a few things well in a bounded domain.

How do you use narrow/general and weak/strong in a conversation? When you talk to product or business people, use "narrow" to make clear that the system does X and Y well, not "any task". Use "weak" only if someone asks about consciousness or "real understanding"; in practice "narrow" is usually enough.

Executive summary (for quick review)

Narrow (ANI): One or a few tasks in a domain. Everything current is narrow. It doesn't mean "dumb"; it means limited in scope.
General (AGI): Any human intellectual task. Doesn't exist today.
Weak: Simulates behavior without consciousness. Everything current is weak.
Strong: Hypothetical real mind/consciousness. Not realized.
Different axes: Narrow/general = scope; weak/strong = consciousness. Narrow + weak = what you build today.
By task: Classification, regression, generation, reasoning/planning (bounded).
By deployment: API, fine-tuning, RAG, agents. All narrow.
Mistakes: "ChatGPT is AGI" (false); "Narrow is dumb" (false); "Strong = more powerful" (false); "Narrow and weak = the same" (false).
For an AI Engineer: Choose the task, the model and the way to expose it. All narrow by definition.
Resources: Stanford Encyclopedia, MIT Technology Review (AGI), Google (Types of AI), NIST, EU AI Act, GPT-4 Report.

Common mistakes

1. "ChatGPT is AGI"

Mistake: Treating today's LLMs as if they were general intelligence.
Reality: They're narrow AI that's very capable at language and related tasks, but they fail at long-range coherence, facts, complex reasoning and tasks outside their training. They aren't AGI.

2. "Narrow AI is 'dumb'"

Mistake: Believing that "narrow" means not very capable.
Reality: Narrow AI can be extremely good in its domain (e.g. playing Go, translating, generating text). "Narrow" means limited in scope, not in quality within that scope.

3. "Strong AI = more powerful than weak AI"

Mistake: Interpreting "strong" as "more technically powerful".
Reality: "Strong" refers to consciousness/understanding (a philosophical concept); "weak" is "simulation without a mind". A weak system can be far more useful in practice than a hypothetical strong one.

4. "Narrow and weak are the same"

Mistake: Confusing narrow/general (scope of tasks) with weak/strong (consciousness/understanding).
Reality: They're different axes. A system can be narrow and weak (e.g. a spam classifier: one task, no consciousness). LLMs are narrow (language tasks) and weak (no consciousness attributed to them). AGI would be general; strong would be with consciousness; they aren't synonyms.

5. "If it's very capable at language, it's already close to AGI"

Mistake: Thinking that because GPT-4 is very capable at language (it generates coherent text, solves coding problems, etc.), it's already "almost at AGI" or "not far from being general".
Reality: Narrow doesn't mean dumb, it means limited in scope. GPT-4 does one family of tasks (language) very well but fails at long reasoning, precise facts, coherence across thousands of lines, and tasks outside its training (e.g. physical control, complex multimodal perception, real-time learning without retraining). The gap between "very capable at language" and "capable of any human intellectual task" is enormous. Don't confuse quality in one domain with general scope.

Frequently asked questions

Is ChatGPT narrow or general?
Narrow. It does language tasks (generating, summarizing, translating, answering questions) very well but not "any human intellectual task"; it fails at long-range coherence, facts, complex reasoning and tasks outside its training.

Can narrow AI be "very intelligent"?
Yes. "Narrow" means limited in scope (one or a few tasks), not in quality within that scope. A model that beats champions at Go or translates between 100 languages is narrow but very capable in its domain.

Why is AGI talked about so much if it doesn't exist?
Because in debates about risks, regulation and long-term investment it's used as a reference concept: "what would happen if there were a system with human-level scope?" For an AI Engineer's daily work, what's relevant is that everything you build today is narrow.

Are weak and strong technical or philosophical terms?
Mainly philosophical: they refer to whether the system "simulates" behavior (weak) or has "a real mind/consciousness" (strong). In industry you hear narrow vs general more often; weak/strong show up in debate and regulation.

Connection with the rest of the guide

Module 5 (LLMs): When we talk about the capabilities and limits of GPT, Claude, etc., you'll be using the idea that they're narrow AI (very capable in their domain, not "general").
Module 7 (AI Engineering): Your job is to design and integrate narrow AI into products: choosing the task, the model and the way to expose it (APIs, RAG, etc.).

Exercises

Exercise 1: Narrow or general

Classify as Narrow AI or AGI (General AI):

A model that translates between 100 languages.
A hypothetical system that can learn to do any office job without retraining.
A chatbot that answers questions and generates text but sometimes makes up facts.
A robot that sweeps the house and cooks but can't program or paint.

See solution

Narrow AI. A very well-defined domain: translation between languages.
AGI (hypothetical). "Any office job" implies broad generalization.
Narrow AI. Very capable at language, but limited to that and with known failures; not any intellectual task.
Narrow AI. Several tasks but bounded ones (cleaning, cooking); not "any" human intellectual task.

Exercise 2: One sentence for each

Explain in one sentence the difference between Narrow AI and General AI (AGI).

See answer guide

Guide: "Narrow AI does one or a few tasks well in a bounded domain (and it's what exists today); General AI (AGI) would be capable of learning and performing any human intellectual task and doesn't exist yet."

Exercise 3: Weak or strong?

A voice assistant that understands commands and controls devices but has no consciousness or understanding in the philosophical sense. Weak AI or Strong AI? Why?

See solution

Weak AI. It does the task (understanding commands, controlling devices) but no consciousness or "real understanding" is attributed to it. Every current system fits into weak AI in that sense.

Exercise 4: Classify by type of task

For each of these systems, say whether the main task is classification, regression, generation or reasoning/planning (according to this lesson's "by type of task" classification):

A spam filter that marks emails as spam or not spam.
A model that predicts the price of a house from its features.
An LLM that writes a summary from a long article.
An agent that decides which tool to call (search, calculator) to answer a question.

See solution

Classification: It assigns a label (spam / not spam).
Regression: It predicts a numeric value (price).
Generation: It produces new content (a summary).
Reasoning/planning: Given goals and constraints, it decides steps (which tool to use); still very bounded in current systems.

Exercise 5: Narrow + weak in one sentence

Explain in one sentence why it makes sense to say that a chatbot like ChatGPT is narrow and weak at the same time.

See answer guide

Guide: "It's narrow because it does language tasks (generating, summarizing, translating) well but not any human intellectual task; it's weak because no consciousness or understanding in the philosophical sense is attributed to it, it only simulates intelligent behavior in its domain."

Exercise 6: General or not?

Someone says: "A model that can translate between 200 languages, summarize texts, generate code and answer questions about documents is already almost AGI." What important nuance is missing from that claim?

See answer guide

Guide: All those tasks are language tasks (or very close to it); the model doesn't do "any human intellectual task" (driving, operating, reasoning in non-linguistic domains, adapting to entirely new contexts without retraining). AGI implies generalization to any intellectual domain, not just language. Besides, those models fail at long-range coherence, facts and complex reasoning; "almost AGI" is an exaggeration.

Quick reference: narrow vs general, weak vs strong

When you read an article, an announcement or a conversation about AI, use this table to place what kind of system is being talked about:

Typical claim	How to place it
"ChatGPT is AGI"	False: it's narrow AI that's very capable at language; it doesn't do any intellectual task, and no system considered AGI exists today.
"This model understands"	Ambiguous: it can mean "it does the task well" (measurable) or "it has real understanding" (there's no consensus and no way to measure it). In practice, everything current is weak AI.
"Narrow AI is limited / dumb"	Mistake: "narrow" means limited in scope (one or a few tasks), not in quality within that scope. It can be extremely good in its domain.
"Strong AI = more powerful than weak"	Mistake: "strong" refers to consciousness/understanding (philosophical); "weak" is simulation without a mind. A weak system can be far more useful in practice.
"Everything you build today is narrow"	Correct: products, APIs, RAG, agents are narrow AI by definition; you don't design for AGI or for strong AI.

Use this table when you evaluate headlines, when you talk to product or business people, or when you design a system (choosing the task, the model, expectations).

Summary

Narrow AI (ANI): Systems that do one or a few tasks well in a domain; it's everything that exists today in products (including LLMs).
General AI (AGI): Hypothetical intelligence with human-level scope on any intellectual task; it doesn't exist; it's a long-term goal or concept.
Weak AI: Systems that simulate or replicate intelligent behavior without attributing consciousness to them; everything current.
Strong AI: Hypothetical systems with a mind/consciousness; a philosophical concept, not realized.
Knowing these classifications helps you not confuse what we have (narrow/weak) with what's debated for the future (general/strong) and to communicate precisely in reports and conversations.

Additional resources

Stanford Encyclopedia: Artificial Intelligence — Philosophical definitions, the weak/strong debate and historical context; useful for going deeper into the narrow/general and weak/strong axes (in English). Complements the "Weak vs Strong" section of this lesson.
MIT Technology Review: What is AGI? — What AGI is understood to mean today, why there's no consensus and how the term is used in industry and regulation (in English). Useful when you hear "almost AGI" or "the path to AGI".
Google: Types of AI — Classifications from a product and cloud perspective; complements this lesson's conceptual view (in English). Connects narrow/general with types of product (classification, generation, etc.).
NIST: AI Risk Management Framework — A framework for managing AI risks; it uses classifications of types of AI in the context of regulation and standards (in English).
EU AI Act – High-level summary — A summary of the European regulatory framework; it mentions narrow vs general and risk levels by type of application (in English).
OpenAI: GPT-4 Technical Report — GPT-4's technical report; the introduction discusses capabilities and limits (narrow, not AGI); useful for connecting definitions with a specific model (in English).

← Previous3. A Brief History of AI: From Turing to ChatGPT Next →5. AI in 2024-2026: Current state, what's possible and what isn't