AGIBeginnerLesson 265 min read

What people mean by AGI

Everyone argues about whether AGI is close. Almost all of that argument is really an argument about what the letters mean.

Lesson in motion

In 60 seconds

What people mean by AGI

Everyone argues about whether AGI is close. Almost all of that argument is really an argument about what the letters mean.

1/6
In simple words
A calculator is amazing at sums and useless at everything else. AGI would be a machine that can learn any new job the way a person can — including jobs nobody planned for it.
AGI stands for Artificial General Intelligence. The word doing the work is general.
Narrow AI — what we have
  • Brilliant inside its training distribution.
  • Falls apart just outside it.
  • Cannot pick its own goals.
  • Does not learn from today's experience unless retrained.
  • Chess engines, spam filters, image models, and yes — today's LLMs.
General AI — the claim
  • Learns a genuinely new skill from few examples.
  • Transfers what it knows to unfamiliar problems.
  • Sets its own sub-goals to reach a big one.
  • Improves from its own experience, continuously.
  • Handles the messy real world without a handler.

Six definitions people actually use

DefinitionWhat it demandsTrouble with it
Human-level at most jobsMatch a competent adult across the board"Most jobs" is doing enormous unstated work
Passes as human in conversationThe Turing testArguably passed, and it clearly was not the finish line
Economically valuable workDo what people are paid to doDepends on the economy, not the mind
Learns anything a human canFull generality of learningNo way to test it exhaustively
Self-improvingCan make itself betterA capability claim, not an intelligence one
Conscious and self-awareInner experienceWe cannot test this even in each other
Watch out
This is why "will we get AGI by 2030?" is close to unanswerable as asked. Under definition 3 a reasonable person can argue we are partway there. Under definition 6, we have no idea how we would even check.

A more useful question

Drop "is it AGI?" and ask three sharper questions instead. They have actual answers, and they are the ones that change what you should do:
  1. 1

    What can it do that it could not do last year?

    Measurable. Track capability, not labels. This is the question that predicts your next twelve months.
  2. 2

    What can it do without a human checking?

    This is the safety-relevant one, and it is the whole of Track B. Autonomy, not intelligence, is what creates consequences.
  3. 3

    Where does it still fail, and does the failure pattern change?

    Failures shrinking steadily is a very different world from failures shrinking then plateauing.

Two words you will hear alongside it

  • ASI — artificial superintelligence. Better than the best humans at essentially everything. A step beyond AGI, and the subject of most of the scary conversations.
  • Transformative AI — a deliberately capability-free term meaning "AI that changes society as much as the industrial revolution did." It sidesteps the definition fight by asking about effect instead of nature.
Do this
Where this connects to the rest of the guide: you do not need AGI for any of Track B to matter. Today's clearly-not-general models already have tools, permissions and consequences. The security work is due now, not when someone declares AGI has arrived.

Watch and read more

Lab

A definition of AGI you can actually defend.

~10 min

The problem

Write your own one-sentence definition of AGI, then find two systems that satisfy it but obviously are not AGI, and one that fails it but arguably should count. Revise. Repeat until you cannot break it.

You are done when

Hard questions

Try to answer before you reveal. If you can answer these, you understood the lesson.

Q1Someone claims their system is AGI because it passes a broad benchmark suite. Give the strongest counterargument that is not 'benchmarks are contaminated'.Reveal
Benchmarks are a fixed, finite list, and 'general' means 'not covered by any fixed list'. A system optimised against a suite demonstrates competence at that suite — which is a claim about the suite's coverage, not about generality. The interesting test is performance on tasks drawn after the system was built, by someone trying to find its edges.

Please sign in to continue.

Questions people ask

Are current LLMs AGI?

Under most definitions, no. They are startlingly broad — they handle a huge range of text tasks — but they do not learn continuously, they lack reliable grounding in the physical world, and their failure modes are unlike human failure modes. Under weak definitions, some people say yes. That disagreement is about vocabulary more than about facts.

Why do experts disagree so wildly on timing?

Different definitions, different weight on current trends continuing, and honestly, different temperaments. Someone who thinks the missing pieces are engineering says a decade; someone who thinks they are unsolved science says many decades or never.

Is AGI inevitable?

Nothing is. It requires continued investment, energy, research breakthroughs and a society that keeps choosing to build it. Any of those can change.

Would AGI be conscious?

Unknown, and possibly unanswerable. Intelligence and consciousness may be entirely separate things. A system could plan brilliantly with nothing it is like to be it — or not. We have no test.

Lesson test

5 questions. Get 3 right (60%) to pass and complete this lesson.

Sign in with your phone number to take the test and save your progress