Artificial General Intelligence (AGI)

AGI stands for artificial general intelligence. The term appears in news, in company announcements, and in arguments about where AI is heading. People use it with different meanings, which makes those arguments hard to follow.

This lesson gives a working meaning for the term and separates what is already clear about today's models from what is still an open question. After this lesson you can explain what someone is claiming when they say "AGI is here" or "AGI is far away", and what evidence would be needed either way.

1. The key word is "general"

Take the three words one at a time.

The middle word carries the meaning. To see why, compare AGI with the AI that came before.

Each system in the left column was built for its own lane. Outside that lane it can do nothing. A spam filter cannot translate. This is the narrow AI from the first lesson.

AGI describes AI with no lane at all. It would handle almost any thinking task a human can, including tasks it was never specifically trained for.

The term was written down as a category before anyone agreed that anything belonged in it. There is no single agreed definition of AGI. Experts disagree on how close current models are, and on how anyone would know when it has arrived.

2. AGI as a claim about everyday work

Without the technical words, AGI is a claim about work. Take the tasks you do on a computer this week:

• Summarise a 40-page lease.
• Reconcile the invoice folder, which means checking that the invoices match the payments.
• Fix the test that keeps failing.

Now give the same list, on the same day, to two workers: you and an AI model.

flowchart LR
    L["This week's computer tasks<br/>summarise a 40-page lease<br/>reconcile the invoice folder<br/>fix the failing test"]
    L --> Y["You"]
    L --> M["An AI model<br/>given the same list"]
    Y --> R["Compare the results"]
    M --> R

The AGI claim is that the model can do any of this computer-based work about as well as a person, and often better.

This is a description of what AGI would look like in practice. It is not a settled fact about today's models.

3. Why the term is hard to pin down

Here is a short exchange that shows the difficulty.

You:    Summarise this lease and flag anything unusual.

Model:  Here is a summary of the lease. It runs for five years and
        covers the ground floor...

You:    You skipped the break clause. And I wanted flags, not prose.

Nothing in this exchange is broken. The model produced a reasonable summary. It still was not what the person wanted. It missed one clause, and it wrote paragraphs where the person expected a list of flagged items.

The work is not finished until the person says it is. "As good as a person" therefore depends on who is judging and what they wanted. That is one reason people cannot agree on whether a model has reached AGI.

4. How much computer work a top model can already do

Give a top model a full day of ordinary computer work. The result has two parts.

For a good share of the work that is within reach, the model's result is also better than a person's. The main reason is speed. It finishes in seconds what takes a person an hour.

So current models are not fully at AGI, but a large part of screen-based work is already possible for them.

5. Asked for a total, the model writes a program

Language models are not reliable at doing large sums internally. The reason comes from how they read text: a model sees a number as a few tokens, not as a value it can calculate with.

Suppose you give a model a file called orders.csv with thousands of orders and say "add these up for me".

id,customer,amount
1041,Brandt GmbH,1240.00
1042,Okonkwo Ltd,86.50
1043,Vela SA,2310.75
1044,Brandt GmbH,905.20
1045,Haldane & Co,47.99
 ... 8,395 more rows

A capable model does not try to add 8,400 numbers itself. It writes a short program and runs it. The program below is in Python, a common programming language. You do not need to read it in detail. It opens the file and adds up the amount column.

>>> import csv
>>> rows = csv.DictReader(
...         open("orders.csv"))
>>> sum(float(r["amount"]) for r in rows)

41902.65

The steps are:

1. The model writes a short program that reads the file and adds up the amounts.
2. The software around the model runs the program on a computer.
3. The model gives you the result that the program calculated.

The answer is exact because the computer did the arithmetic. The model did no mental arithmetic at all. Its contribution was knowing that a program was the right tool and writing a correct one.

Knowing when to use a tool like this is a large part of what makes current AI useful for real work.

6. What is clear and what is still open

The question "is today's AI already AGI" splits into two parts.

The first column can be tested. Pick a task, give it to a model, and check the result.

The second column is harder. Being better at many things is not the same as being better at everything. Nobody has shown that a model can do every task a person can, work on its own without someone checking the result, and get things the way the user wants without being guided.

So one way to state the open question is: are current models truly better than a person, or better only at the parts of the work that suit them. The answer to the second column is what the term AGI stands or falls on.

Interview points

• AGI means artificial general intelligence: AI that can do almost any thinking task a person can, at human level or better, including tasks it was not trained for.
• The word "general" separates AGI from narrow AI, where each system does one job and nothing else.
• There is no agreed definition of AGI and no agreed test for it.
• A practical reading of AGI is a claim about computer work: a model handles any screen-based task about as well as a person.
• The term is hard to pin down because "done well" depends on what the person asking wanted.
• Current top models can do a large share of everyday computer work, often faster than people, and still fail at some tasks.
• Models are unreliable at large calculations done internally, so capable models write and run a program and report its result.
• Speed, breadth, and exact arithmetic through code are observable strengths. Doing everything, working unsupervised all the time, and matching the user's intent without guidance are still open.

Summary

• AGI is about generality: no fixed lane, any task.
• Earlier AI systems each did one job.
• Seen practically, AGI is the claim that a model can do your computer work as well as you.
• Top models already do much of that work, and some of it better because they are faster.
• For exact totals, a model writes a program instead of calculating itself.
• Whether models are better than people at everything, reliably and unsupervised, is not settled.