Learn For Everybody
Experiments
Economics of AI · 15.11.2026

Fiddling with ASI

ASI stands for artificial superintelligence: AI better than the best humans at nearly everything.

One road there runs through a loop. AI gets better at AI research, so research goes faster, so AI gets better again.

Against that loop pulls a stubborn fact: in every field studied so far, each new idea is harder to find than the last.

Which force wins? On the model below, it comes down to one number. Move it and the future changes from a slowdown to steady growth to an explosion. Nobody yet knows which side we are on. This page lets you find the line yourself.


2. The loop

AI gets betterAI does more ofthe researchresearch goes fasterthe loopeach new idea isharder to find
The loop turns at the current growth rate in the model in section 5. A full doubling can take a different length of time as that rate changes. Change the dials there, then come back.

3. The model in words

As AI improves, it can add research effort. That pushes progress faster. At the same time, each comparable step can be harder to reach, pulling progress slower. If the push outweighs the pull, each doubling comes sooner, with spending held fixed.

On the starting illustration, each doubling of AI's ability now brings about 1% more research effort; on the line it would have to bring about 38% more.

growth of AI's abilities=(how effective research is)×(research effort)(how much more you get from more effort)×(abilities so far)((how much past progress helps the next step) − 1)

research effort=(spending)×(abilities so far)(how much better AI adds to research effort)

Equation (1) describes the growth rate of AI’s abilities: the proportional change in abilities over time.

What each one means:


4. The one number that decides it

Put equation (2) into equation (1). Research effort is partly made of abilities, so abilities now appear twice: once through the effort AI supplies, and once through how hard the next idea is. Together they leave abilities raised to the power λψ + φ − 1 in the growth rate of abilities, meaning the proportional change over time.

growth of AI's abilities = δ × spendingλ × (abilities so far)λψ + φ − 1

In the simplified model used here, with spending held fixed, whether growth speeds up or slows down as AI improves depends on one number, the loop's strength, λψ + φ.

In common sense. Each time AI doubles its abilities, two things happen to the speed of progress.

If the push is bigger than the pull, each doubling makes the next one come sooner. That is λψ + φ above 1. If the pull wins, each doubling takes longer than the last, unless someone keeps spending more.

Why above 1 would be new

In research, for as long as anyone has measured, the pull has won.

Computers have long helped design better chips. AI could also substitute for some of the thinking work of research itself: a better AI might become a better AI researcher, and copies can be made quickly.

Below 1, exactly 1, above 1

Each tick marks a doubling of the length of task AI can do. The number above it is how many months that doubling took.

Below 1: each doubling takes longer.

2.4 mo3.14.15.47.29.403 years

What it would feel like: Each new model a little less impressive than the last jump.

Exactly 1: a steady rhythm.

2.0 mo2.02.02.02.02.003 years

What it would feel like: New abilities on a schedule, like clockwork.

Above 1: each doubling comes sooner.

1.9 mo1.71.51.31.11.0goes vertical at 1.30 years03 years

What it would feel like: Each new model arrives faster than the last, until the calendar can't keep up.


5. Fiddle with it

“Abilities” here means the length of task AI can do. METR scores models on an index; Garicano and Imas turn index points into doublings of task length at about five points per doubling, a ratio Epoch reported and they check against METR’s measured task lengths. For example, METR’s task length went from about one hour in early 2025 to about five hours in late 2025 for the best model; see the overview. It is not a measure of general intelligence. The model has no point marked ASI. It shows how growth behaves, not when superintelligence arrives.

Stories to try

Try these yourself

Use the dials below. The tick appears while the setting meets the goal.

  1. Guess first: how fast must spending grow to hold today's pace? Set ψ to 0, λ to 0.8 and φ to 0.65 (under the dials), and keep the productivity step at its starting value. Find a pace within 1% of today's for ten years. Not yet
  2. Guess first: with ψ at 0.35 and φ at its starting value, what λ puts you on the line? Keep the other dials at their starting values and find a setting within 0.005 of λψ + φ = 1, on either side. Not yet
  3. Guess first: can a one-time productivity step alone make growth go vertical within two years? Set ψ to 0.49, spending growth to 1 and the other dials to their starting values. Find the smallest step that does it, within 0.1 above the threshold. Not yet

Verdict: growth slows
Loop strength, λψ + φ: 0.55 below 1: the loop alone cannot keep growth up

Push: 0.014; pull: 0.462
Length of task AI can do (doublings) Task length, converted from METR’s capability index at about five index points per doubling (Garicano and Imas’s assumption); not general intelligence.
010200246810
How fast that grows Doublings a year. Dashed line: today's pace.
today's pace05100246810

At today's pace, about 5.9 doublings a year, a doubling would take 2 months. Because growth slows along the way, the next doubling takes about 2.4 months. With these settings, the doubling under way a year from now takes about 6.1 months, and the one under way in year ten about 3.3 years.

Starting pace: about 5.8 doublings a year in the length of task AI can do, from METR's fitted trend after October 2025 (29.3 points a year), the figure Garicano and Imas use.

The map

difficulty wins:growth slowsthe loop wins:explosionφ above 1:growth speedsup evenwithout AIψ must reach 0.46push = pull: λψ + φ = 1ABdrag me→ past progress helps more (φ)↑ better AI adds more to research effort
Regions assume spending is held fixed. A: before-and-after comparison. B: break model. Their headline fit (φ 0.84) uses λ = 0.75, so it belongs on a different map.

The spending dial is a growth index of research inputs set from outside the model. It does not model AI earning the money for more research; section 7 covers that extension.

Turn up λ and the line swings down: each unit of research effort is worth more.

Read the map like this. Every point is a possible world, set by two numbers: how hard ideas get (across) and how much better AI adds to research effort (up). The line is every world where the push exactly equals the pull. Below and left of it, the pull wins and growth fades. Above and right of it, the push wins and growth runs away.

Moving the dot up means better AI adds more research effort for each doubling of its ability. Moving it right means past progress helps more. On this illustration ψ would have to rise from about 0.014 to about 0.46. What that means per doubling is in section 3. The red arrow measures the gap from wherever your dot sits: ψ must reach 0.46.


6. What it implies

A threshold for growth that speeds up by itself. With spending held fixed, the same equation gives a slowdown, steady growth or an explosion, depending on whether λψ + φ is below, at or above 1. Near the line, small changes make enormous differences.

A one-time boost is not an explosion. Below the line, with spending unchanged, a permanent rise in research productivity gives a jump in growth that then fades; above the line it brings the explosion forward.

The before-and-after illustration is far from the line, and the evidence is limited. In an illustration built from their before-and-after illustration, ψ is near zero, but it comes from sixteen model scores at one company. These data come from a period when AI helped with research under human direction; they do not show the feedback that would operate if research became much more autonomous. The map shows how far ψ would have to move. It cannot say whether it will.


7. What the model leaves out, and who is working on it

growth of AI's abilities=(how effective research is)×(research effort14)(how much more you get from more effort)×(abilities so far5)((how much past progress helps the next step) − 1)

research effort=(spending3)×(abilities so far)(how much better AI adds to research effort2)

2 how much better AI adds to research effort

Tasks AI cannot do yet

If some part of research, such as judgement, stays human, does that part set the pace?

Who: Aghion, Jones and Jones, for the economy; David Chalmers's “factor F” hypothesis (a key capacity, such as research vision, that AI may not reach soon).

4 research effort

Delays

Garicano and Imas check spending lags of up to a year in their appendix, section C. This simulator assumes AI’s ability feeds back into research at once; adding a delay to that feedback is a separate extension.

The code is open →


8. Takeaway: three questions, and what would change the answer

1. Is it a jump, a faster pace, or a pace that keeps rising?

What would move you towards an explosion

  • the time between doublings keeps shrinking across several new models, as in the third panel of section 4.

What would move you away

  • one fast jump, then the old rhythm returns.

2. What makes AI supply more research as it improves, and what limits it?

What would move you towards an explosion

  • the share of research AI leads keeps rising while staff and spending stop rising as fast.

What would move you away

  • AI’s share of research stalls, or matched tasks show little improvement from better AI.

3. How much more research output does better AI get from the same paid-for inputs?

What would move you towards an explosion

  • on matched research tasks, or with the same controlled resources, output keeps rising as AI improves.

What would move you away

  • matched tasks show little gain from better AI.

Progress per dollar is one noisy signal, to read alongside prices, delays and how fast inputs grow. On its own it does not show how much AI contributes.

One last thing. Scroll back up to the loop in section 2. It turns at the speed your settings give, and thirty seconds of watching is ten years in the model. If it slows down, the rising difficulty of finding ideas is winning. If it keeps a steady pace, you are on the line, or spending is doing the work. If it speeds up until it is a blur, you have set up an explosion.


9. Sources

Built by Ashish Kulkarni with Claude and Codex. Claude is made by Anthropic. Code and text are open. Built to be argued with.

Design craft learned from Keshav Jindal.


Next: The research machine (optional, for the curious) →