esc
  • WorkflowsPages
  • Case studiesPages
  • Machine Learning FoundationsLearn
  • AI EngineeringLearn
  • AI Guardrails & Red TeamingLearn
  • UI/UXLearn
  • Design SystemsLearn
  • Production System DesignLearn
  • Book a callContact
  • Download ResumeContact
Let’s talk

I’ll build the agent for it.

AI engineering / Beginner / 01.02

AI vs AGI

The difference between narrow AI and AGI, why every tool you will build with is narrow, and how to design around that.

13 min readFree

Where the last lesson left us

The last lesson, What is an AI Engineer?, ended with a promise: next, AI vs AGI, and why today's tools are narrow. Here it is.

You will hear "AI" and "AGI" used as if they were the same thing. They are not. One describes the tools you can use today. The other describes something nobody has built. Knowing the difference keeps you from expecting too much from a tool, and that saves you from building a product on a false promise.

By the end you will know what narrow AI and AGI mean, why every tool you will build with is narrow, why a chat model can still feel general, and what the Turing test is. You will also run a small narrow assistant and learn one habit that matters more than any definition: check the answers.

Key words

Narrow AI
A program built to do one task, or a small set of tasks. Also called weak AI or artificial narrow intelligence (ANI).
AGI
Artificial general intelligence. A machine that could apply its intelligence to any problem, not just one. No such machine exists today.
Strong AI
A word with more than one meaning. Sometimes it means AGI. Sometimes it means a machine that is conscious. Check which one a writer means.
Turing test
A test from 1950 where a person chats with a machine and tries to tell whether it is a machine.
Brittle
Works well inside its comfort zone and breaks in surprising ways outside it, like a glass that is fine until you drop it.
Human in the loop
A design where a person checks or approves the tool's work before it counts.

Narrow AI: one job done well

Start with the plain idea. A calculator is brilliant at sums and useless at writing a poem. A spell checker finds typos and cannot book a flight. Each is built for one job.

The Wikipedia page on weak AI says it "is focused on one narrow task". It also says narrow AI is "limited to a single, narrowly defined task. Most modern AI systems would be classified in this category." It names self-driving cars among its examples, and says simple AI programs such as autocorrection and speech recognition have worked their way into society, often unnoticed.

The same page warns about a weakness. It says the behaviour of narrow AI "can become inconsistent" and that this "brittleness" can make it "fail in unpredictable ways". Keep that word in mind. It comes back at the end.

  • One task

    Each narrow tool does a defined job: sort photos, turn speech into text, suggest the next word.

  • Strong inside, weak outside

    Ask it something within its job and it can be excellent. Ask something outside and it can fail.

  • Everywhere already

    The page calls these programs common in society, often unnoticed by the public.

AGI: any problem, not just one

The other side is a machine that is not limited to one job. The Wikipedia page on artificial general intelligence describes AGI as "a hypothetical type of artificial intelligence that matches or surpasses human capabilities across most or all cognitive tasks". The word hypothetical matters: it has not been built.

It adds that, unlike narrow AI, "an AGI system can generalise knowledge, transfer skills between domains, and solve novel problems without task-specific reprogramming". Think of the difference between a person who has only ever learned to drive and a person who can learn to drive, cook, or fix a bike when asked.

The Weak AI page gives a shorter version: AGI is "a machine with the ability to apply intelligence to any problem, rather than just one specific problem".

Careful with the word strong

People mix up the labels, so slow down here. The Weak AI page says strong AI "can be interpreted in various ways", and lists three meanings: AGI, artificial superintelligence (a machine far smarter than the average human), and artificial consciousness (a machine that has "consciousness, sentience and mind").

The AGI page notes the same split: some academic sources reserve "strong AI" for programs that would experience sentience or consciousness. So when an article says "strong AI", ask which meaning it uses. In this course we say AGI when we mean a machine that can handle any problem.

Tip

When you read a news story or a product page, replace "strong AI" in your head with one of the three meanings. The sentence often turns out to say something much smaller than it first sounds.

Why every tool you will use is narrow

Here is the honest position, and it comes from the definitions, not from opinion. AGI is a hypothetical thing. The tools you can sign up for and build with today are narrow AI by that same definition.

The Stanford Encyclopedia of Philosophy entry on Artificial Intelligence says the successes it describes "have all been in limited, narrow domains". It adds that the lack of success in "the unrestricted general case" led a small set of researchers to break away into what is now called artificial general intelligence, with the goal of building artifacts that are generally intelligent and "not just capable in one narrow domain".

So narrow is the starting point, and "general" is a goal some researchers are working toward. Product pages can use big words. Your job is to ask a plain question: what exact task was this built and tested for?

Build a tiny narrow assistant

Reading about narrow AI is one thing. Let's build one so small that you can see every rule. It answers one kind of question, unit conversions like "Convert 10 km to miles". It uses plain rules: a table of conversion numbers and a pattern that spots the question. No model, no learning.

A pattern here is a regular expression, a short text recipe that describes which questions to accept. If a question does not match, or the units are not in the table, the assistant says "I can't help with that." Run it on five questions: three it was built for, two it was not.

Python
import re

FACTORS = {
    ("km", "miles"): 0.621371,
    ("kg", "pounds"): 2.20462,
    ("miles", "km"): 1.60934,
}

def narrow_assistant(question):
    pattern = r"convert (\d+(?:\.\d+)?) (\w+) to (\w+)"
    match = re.fullmatch(pattern, question.lower().strip())
    if not match:
        return "I can't help with that."
    amount, source, target = float(match.group(1)), match.group(2), match.group(3)
    factor = FACTORS.get((source, target))
    if factor is None:
        return "I can't help with that."
    return f"{amount:g} {source} = {amount * factor:.2f} {target}"

questions = [
    "Convert 10 km to miles",
    "Convert 5 kg to pounds",
    "Convert 26.2 miles to km",
    "What is the capital of France?",
    "Convert 3 cups to litres",
]
for q in questions:
    print(q, "->", narrow_assistant(q))

Output

Convert 10 km to miles -> 10 km = 6.21 miles
Convert 5 kg to pounds -> 5 kg = 11.02 pounds
Convert 26.2 miles to km -> 26.2 miles = 42.16 km
What is the capital of France? -> I can't help with that.
Convert 3 cups to litres -> I can't help with that.

Read the output. The first three questions are inside its job, and it answers all three correctly. It never gets tired and never guesses. Then comes a question about France. That is outside its job, so it says so. The last one is the interesting case: "Convert 3 cups to litres" looks like a conversion, but cups are not in its table, so it says it cannot help.

That is narrow skill in its clearest form: perfect inside the lines, honest about the edge. The next section is about what happens when a tool does not have such clear lines.

Why chat models feel general

Now the puzzle. A chat model can draft an email, explain a joke, write code and summarise an article, all in one conversation. That feels like general intelligence. So is it AGI?

The sources do not give one answer. The AGI page says Google DeepMind researchers proposed a framework in 2023 with five performance levels of AGI: emerging, competent, expert, virtuoso and superhuman. They considered large language models such as ChatGPT or LLaMA 2 to be "instances of emerging AGI", comparable to unskilled humans. The page also notes that some researchers and industry figures claim early forms of AGI may already exist, while the topic remains "a subject of intense debate".

So calling a chat model "AGI" or "narrow" depends on whose definition you pick. For your work, a different question is more useful: what does it do reliably, and what does it get wrong? The lesson after this one explains how these models work.

Broad, but still wrong at times

Breadth is not the same as reliability. The Wikipedia page on hallucination in AI describes a hallucination as "a response generated by AI that contains false or misleading information presented as fact". It says chatbots powered by large language models "may embed plausible-sounding random falsehoods within their generated content, such as fabricated citations".

The page adds that detecting and reducing such errors poses "significant challenges" in high-stakes scenarios, naming chip design, supply chain logistics and medical diagnostics.

Think of a very well-read friend who answers instantly and sounds sure every time, even when they are guessing. That is useful for drafts and ideas. It is risky for anything where a wrong answer costs money, time or safety.

Quick check

A chat model writes a clear, confident answer. What does that tell you about whether it is correct?

A stand-in model, plus a check

Here is how that habit looks in code. The program below has a stand-in model. It is a tiny fake function that returns the same fixed text every time, so the code runs free with no account, no key and no network. It is not a real model. A real model is slower, costs money for every token it reads and writes, can give a different answer each time you ask, and can fail or refuse.

Our stand-in gives a wrong answer on purpose: 12 km is about 7.46 miles, not 8.00. A small check, using the same rule our narrow assistant used, catches it before a user sees it.

Python
def fake_model(messages):
    # Stand-in for a real model: always returns the same fixed text.
    return "12 km = 8.00 miles"

def check(answer, km):
    expected = round(km * 0.621371, 2)
    return f"{expected:.2f} miles" in answer

answer = fake_model([{"role": "user", "content": "Convert 12 km to miles"}])
print("Model said:", answer)
if check(answer, 12):
    print("Check passed: show it")
else:
    print("Check failed: do not show it, ask a human")

Output

Model said: 12 km = 8.00 miles
Check failed: do not show it, ask a human

This is a real pattern in AI products. A flexible but unreliable tool does the broad work, and a narrow, dependable rule checks the result. When the check fails, the answer is held back and a person looks at it. A check like this cannot cover every question, but it can cover the ones where a mistake hurts most.

Tip

Not everything can be checked by a rule. Where it cannot, a human in the loop is the check. Decide up front which answers a person must approve.

When will AGI arrive? Nobody agrees

You will see headlines with dates. Here is what the sources say, without adding a guess of our own.

The AGI page says there is "no single agreed-upon definition of intelligence as applied to computers", and quotes computer scientist John McCarthy writing in 2007: "We cannot yet characterize in general what kinds of computational procedures we want to call intelligent." It adds that questions about consciousness, goal-setting, emotional capabilities and reasoning "remain contested".

It also records how past predictions went. In 1965 Herbert A. Simon speculated that "machines will be capable, within twenty years, of doing any work a man can do". The page says this prediction failed to come true. It lists polls from 2012 and 2013 where the median expert estimate for being 50% confident of AGI was 2040 to 2050, and 16.5% answered "never" when asked for a 90% confidence date. It also notes that some thinkers deny it is possible, while most researchers believe it can be achieved in the future.

  • No agreed definition

    Without one, "AGI has arrived" can mean different things to different people.

  • No agreed test

    The page says several tests meant to confirm human-level AGI have been considered, such as the Turing test and a coffee test. Each is a different idea of what to measure.

  • No agreed date

    Estimates differ a lot, and the page notes a long record of early predictions that did not come true.

The practical meaning: build for what exists today and can be tested today. If a claim depends on a date or a label you cannot check, set it aside.

The Turing test, in one paragraph

The Turing test was introduced as the imitation game by Alan Turing in his 1950 paper "Computing Machinery and Intelligence". It opens with the question "Can machines think?", and because "thinking" is hard to define, Turing replaced it with a closer, clearer one: could a machine do well in a game where a judge chats with a person and a machine and tries to tell which is which. In the modern version, per the Wikipedia page, a human evaluator reads a text conversation between a human and a machine, and "the machine passes if the evaluator cannot reliably tell them apart". The results "would not depend on the machine's ability to answer questions correctly, only on how closely its answers resembled those of a human".

That last point is why the test is debated. The page says every element of the assumption that a judge can tell whether a machine is thinking by comparing its behaviour with a human's has been questioned, and that some AI researchers have questioned the test's relevance to their field. It also notes that philosopher John Searle argued against the test's ability to detect consciousness. Passing a chat test shows a machine can sound like a person. It does not show the machine is right, safe or reliable.

What this means for how you build

Put the lesson to work. You are not designing for a general mind. You are designing around a tool that is good at some things and wrong at others. Three habits cover most of it.

  1. 1

    Pick a narrow job

    "Summarise this support email in two lines" is a job you can test. "Run my customer service" is not.

  2. 2

    Check the output

    Add a rule, a test or a second look. The stand-in example showed a check catching a wrong number.

  3. 3

    Keep a person where it matters

    For money, health, safety, legal text or anything hard to undo, a human approves before it counts.

  4. 4

    Show the edge to users

    Tell people what the tool is for. Our narrow assistant said "I can't help with that". Good products do the same.

Quick check

Which plan best fits a tool that is good at some things and wrong at others?

Quick check

By the Weak AI page's definition, which statement is true of today's AI systems?

Recap

Next: how large language models work, one word at a time.

Sources

How is this lesson?

Last updated on 16 November 2026

Learn with me

Every lesson is free. I add new ones as I learn. Have a question, or want help with your own work? Book a short call.