Put it to workBeginnerLesson 595 min read

How to argue about timelines

You will be asked "so when does AI take over?" at dinner. Here is how to answer honestly without sounding like either a hype merchant or a cynic.

🧪 Lab · 15 min📝 Lesson test
Lesson in motion

In 60 seconds

How to argue about timelines

You will be asked "so when does AI take over?" at dinner. Here is how to answer honestly without sounding like either a hype merchant or a cynic.

1/5
In simple words
Some very smart people say "soon". Other very smart people say "not for ages". They are not lying. They are looking at different things and using different words.

Why the smartest people disagree

If you believe...You expect...Because
Scaling keeps workingBig capability jumps within a decadeThe curve has not bent yet
Scaling is flatteningSlower progress, plateausEach step costs much more than the last
Key pieces are missingDecades, or a different paradigmNo continual learning, no grounding, no causal models
Agents change everythingRapid change without any new modelCapability times autonomy times deployment
Bottlenecks are physicalSlower than the hypeEnergy, chips, data, regulation, and the speed institutions move

Four moves that make you sound sensible

  1. 1

    Ask what they mean

    "When you say AGI, do you mean it does most jobs, or that it learns like a person?" Half the disagreement evaporates right here.
  2. 2

    Talk capabilities, not labels

    "Will an AI do a full week of junior developer work unsupervised by 2028?" has an answer you could bet on. "Is it AGI?" does not.
  3. 3

    Separate capability from deployment

    Something being possible is not the same as it being everywhere. Hospitals and banks move slowly for good reasons.
  4. 4

    Give a range and say why

    "I would not be shocked by a lot of change in five years, and I would not be shocked if it takes thirty. Here is what would move me either way."

Signals worth actually tracking

  • Task length. How long a job can an agent complete unsupervised? This is the number that has been moving, and it matters more than benchmark scores.
  • Reliability on the long tail, not the average. Averages have been improving for years; tails are what block real deployment.
  • Cost per useful task, falling. This decides adoption far more than raw capability does.
  • Continual learning. If a system genuinely learns from its own experience in deployment, that is a real regime change.
  • Real incidents. Every serious agent failure teaches more about the actual risk landscape than a benchmark ever will.
Watch out
Two claims should both make you sceptical: "this changes everything next year" and "it is just autocomplete, nothing to see." Both are cheap, both feel good to say, and neither survives contact with what the systems actually do.
Real example
A dinner-table answer you can actually use: "Nobody knows. But we already have software that reads your email and can send email on your behalf — and that is worth getting right whether or not anything smarter ever arrives."
Do this
This is the point of the whole guide. The security work does not depend on the timeline. If AGI is thirty years away, Track B still protects you this quarter. If it is close, Track B is the foundation everything else is built on. Either way you do the same thing on Monday.

Watch and read more

AI timelines debateYouTube searchForecasting AI progressYouTube search

Lab

A forecast you wrote down, with a falsifier.

~15 min

The problem

Write three specific, dated, falsifiable predictions about AI capability. For each, state what evidence would change your mind. Put a calendar reminder to score yourself.

You are done when

Hard questions

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

Q1Rewrite 'AGI by 2030' as a bettable claim, then say what makes the rewrite better.Reveal
For example: 'By 31 December 2030, a publicly available system will complete a randomly selected 40-hour software task from a held-out set, unsupervised, at professional quality, in at least 50% of attempts.' Better because every term is measurable, the resolution source is specified, and someone could take the other side — which forces you to hold a real position rather than a mood. Any prediction nobody could bet against is not a prediction.

Please sign in to continue.

Questions people ask

Who should I actually read on this?

Read people who make falsifiable predictions and then publish how they did. Prefer forecasters who update in public. Discount anyone whose position has not moved in five years — and anyone selling something priced on the answer.

Are AI researchers themselves worried?

Surveys of the field consistently show a wide spread with a meaningful minority assigning real probability to severe outcomes. There is no consensus. Anyone telling you "experts agree" — in either direction — is not describing the surveys.

Does it matter what I think?

For your work, less than you would expect: the practical steps are the same under most timelines. For your vote, your career choices and what you teach your kids, it matters quite a lot.

How do I talk to someone who is frightened?

Take it seriously rather than dismissing it, then move to what is actually controllable: how systems are deployed, what permissions they get, what oversight exists. Agency is the antidote to dread, and there is genuinely a lot of it available here.

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