Learn For Everybody
Experiments

The research machine

Optional, for the curious: explore what holding a faster research pace might require.

A calculator of what research would require under stated assumptions. It does not predict whether AI will supply that research, or whether an explosion happens.

Tejas Subramaniam, an AI strategy researcher, set out a short model of AI research.

If AI has doubled Anthropic's pace, Tejas's model says that's like having about 6 times the research workforce.
On the same model, doubling the pace again would take about 30 times the original workforce.

Based on a short model by Tejas Subramaniam, using METR's estimate that Anthropic's models are improving about twice as fast as before on well-specified tasks.

The dials need JavaScript. Both views below show the default model.

Reference cases, using Tejas's starting assumptions. They do not change when you move the dials.

A step of progress takesPaceResearcher-equivalents
6 months1×1×
3 months2×about 6×
1.5 months4×about 30×

Compute beyond the old trend lowers these numbers. Harder ideas raise them. Test that with the dials below.

METR would not be surprised by 1.3× or 3×. Through this model that is about 2× to 16× the research labour. This is our calculation, not METR's.

The months are an illustration. METR measured a benchmark index, not release dates. These figures use Tejas's assumptions, and 4× is our what-if, not a forecast. These are starting figures. Holding the pace takes more each year if ideas get harder.

What these inputs produce

With these inputs, AI progress runs at 2× its old pace.

What the target requires

To reach METR's 2× with everything else as set, research labour must be 6.25× its old level.

Changing research labour changes the first result, not the second.

Old pace means METR's baseline, March 2024 to October 2025. We assume that trend would have continued. It is not an observed world without AI.

Change the inputsStarting inputs

The experiment dial holds training compute fixed. It changes the total compute pool and its split.

Target: METR's 2×. Change it in the workings.

A simplified timing picture: experiment computing is used now, while training's share shows up over W years.

Researchers 6.25× → Research lab 2.5× → Stock of methods (software) → Training run → AI capability 2×.

Compute pool 1× splits into Experiments 1× → Research lab, and Training 1× → Training run.

METR would not be surprised by 1.3× or 3×

Methods are non-rival: each training run uses the whole stock. This assumes every training run can use every method at once, with no cost of adapting it. Older models stay available even if new runs get weaker. Pipe motion shows relative rates; the numbers give their values.

How the model works

Two things make AI models better: more training compute, and better methods, which we call software. Tejas assumes software supplies two-thirds of progress. If training compute keeps its pace, doubling overall progress needs software progress to run 2.5 times as fast.

Better methods come from a research lab. The lab needs researchers, human or AI, and computing power for experiments, and he weights them equally. Add researchers without adding experiment compute and each extra one adds less. That is why 2.5 times the pace of new methods takes 2.5 × 2.5 = 6.25 times the labour.

Research labour here means human researchers plus AI doing research work, measured in researcher-equivalents. An L of 6.25 means 6.25 times the old research labour, not 6.25 extra. It does not mean 6.25 times the staff. AI raises effective research labour when it adds research capacity or lets the existing team do more. Replacing a human researcher with equal AI capacity leaves this measure unchanged. Adding research labour adds no experiment compute.

Try these

Click a row to load it into the dials. The right-hand number is the research labour needed.

To double the pace (2×)

  • Tejas's modelHis assumptions, unchanged.6.25×
  • Research productivity falls 20%The same effort finds 20% less.9.77×
  • Twice the experiment computeTraining compute stays on its old path.3.13×
  • Twice the compute everywhereExperiments and training both double. Depends on our illustrative T₀ and W. METR is confident the acceleration was not due to faster training compute scaling.2×
  • AI caused 80% of the speedupMETR's guess. This is the labour equivalent of AI's part only, and it depends on where the other causes act.4.84×
  • Research labour and experiment compute are strong complementsResearchers, human or AI, share experiment compute, so extra labour helps less (σ = 0.75). This is about research labour against experiment computing. It does not capture bottlenecks between tasks inside research, such as research judgement against coding.9.41×
  • Research labour and experiment compute are very strong complementsPast a point, more researchers, human or AI, add almost nothing, and experiment compute sets a limit (σ = 0.5). This is about research labour against experiment computing. It does not capture bottlenecks between tasks inside research, such as research judgement against coding.Never reaches 2×: approaches a ceiling of 1.67×

To double the pace again (4×)

  • Tejas's modelHis assumptions, with the target at 4×.30.25×
  • Twice the experiment computeThe labour needed halves.15.13×
  • Four times the experiment computeThe labour needed falls to a quarter.7.56×
  • Research productivity falls 20%The same effort finds 20% less.47.27×

Where this comes from

Tejas called his post a scrappy summary, his loose impression of tweets by Ryan Greenblatt, Ajeya Cotra and others. He says falling research productivity could push the figure much higher. This page builds his model so you can change its assumptions. The starting software share, Cobb-Douglas form, labour weight, A and E are his, and the 2× is METR's. The compute split, the time path and the other settings are ours.

What the number depends on

These cautions are ours.

  • The model treats AI researchers as free of compute. In practice they run on compute too, which would compete with experiments and training.
  • Tejas supposes the two-thirds split and the equal weights. Change them and the answer moves.
  • METR's benchmark index gave 1.9×, with a 95% interval of 1.5× to 2.8×. Its 2× judgement also draws on other evidence, and it says Anthropic redacted most of what it found most informative.
  • Much of METR's evidence measures well-specified tasks, and METR calls its estimate uncertain.
  • The index and effective compute are different things. Reading one as the other assumes they move together.
  • METR guesses that perhaps 80% of the acceleration in October 2026 came from AI. Tejas's model gives AI all of it.
  • Tejas says falling research productivity could raise the labour figure much further. The 'Research productivity falls 20%' case shows one example.

What did the speedup cost? Anthropic's research spending rose about 5 times over the same year. We work through what that does and does not show on a separate page: The spending puzzle.

Inside the model

This is the same model as The machine, with its parts named. Two equations do the work: one makes new methods, the other makes better models. You can change anything here and the machine changes too. Hover over or tap any symbol to see what it means, or open In plain English under a card.

About this model

Two sectors, in growth rates. Time is in years; growth is in doublings per year. Every ratio compares with the assumed continuation of METR's baseline trend at the same date.

The research sector extends the Romer and Jones setup by adding experiment compute. The ideas-getting-harder extension draws on Bloom, Jones, Van Reenen and Webb, Are Ideas Getting Harder to Find? (American Economic Review, 2020). The compute split, the bridge and the time path are this page's own extensions. The paper does not calibrate θ here.

T0, W, θ, q₀, q₁ and L0, K0 defaults are illustrative. We did not estimate them.

A simplified timing picture: experiment computing is used now, while training's share shows up over W years.

L = 6.25, E = 1, A = 1 → g/g0 = 2.5 → S → C × S → AI capability 2×. K = 1, q0 = q1 = 0.5; c = 1 doubling a year.

METR would not be surprised by 1.3× or 3×

Follow the named parts into the equations below. The two views share every input.

Numbers in substitutions are rounded; ≈ marks approximate arithmetic.

Final sector · Training run

Effective compute = C × S g0 = s × T0 c0 = (1 − s) × T0

g0 ≈ 0.67 × 3 ≈ 2 c0 = (1 − 0.67) × 3 ≈ 1 doubling/year.

In plain English

A model gets better in two ways: more training compute, or better methods. The two multiply, so doubling either one doubles effective compute. In growth rates the multiplying becomes adding: growth in effective compute equals growth in C plus growth in S. That lets us split baseline growth: s to software, 1 − s to training. Tejas supposes two-thirds and one-third. No research labour appears in this equation. Research labour comes in one step earlier, in the research lab. Methods are non-rival: each run uses all of S. This assumes every training run can use every method at once, with no cost of adapting it. Older models stay available even if new runs get weaker.

Symbols

  • C — Computing power used to train models
  • S — The stock of research methods
  • g0 — Baseline software growth in doublings/year
  • c0 — Baseline training growth in doublings/year
  • s — Software share of baseline progress
  • T0 — Baseline effective-compute growth in doublings/year

Research sector · Research lab

g = g0 × A × F(L, E)λ G = g/g0 = A × F(L, E)λ F(L, E) = La × E1 − a

With a = 0.5 and λ = 1, this is Tejas's equation: G = A × √L × √E.

F = 6.250.5 × 11 − 0.5 ≈ 2.5 G ≈ 1 × 2.51 ≈ 2.5 Cobb-Douglas a ≈ 0.5 F(1,1) = 1.

In plain English

This equation makes new methods. Its output is a pace: how fast S grows. Its inputs are research labour, human or AI, and experiment compute. With equal weights and λ = 1, neither input can do the work alone: four times the labour only doubles the pace. A scales the whole result. If ideas get harder to find, A falls and the same lab finds less. Cobb-Douglas builds in one fixed answer to how far more research labour can stand in for more experiments, so no symbol for it appears here. Choose CES below to set it yourself.

Symbols

  • g — Software growth in doublings a year
  • g0 — Software growth before the speedup
  • G — New software pace divided by old
  • A — Research productivity compared with baseline
  • L — Human and AI research labour
  • E — Experiment compute compared with baseline. Compute per unit of time
  • λ — Returns to scale in research
  • F — Combines labour and experiment compute
  • a — Labour weight against experiment compute

Set the assumptions

Target
Accounting
Research sector
Compute split
Attribution scenario

Only b × h enters the result. It is an assumption about how much of the acceleration to credit to AI, not a change of units.

Levels

Labour: 100 → 625. Compute: 100 → 100.

Levels cancel; only ratios matter. That is what the normalisation F(1,1) = 1 does.

From the target to labour

  1. Target X ≈ 2 → Xe = 1 + 1 × 1 × (2 − 1) ≈ 2.
    Why this step

    Start from the target speedup. Adjust it for how much AI caused and for how the index relates to effective compute.

  2. G = (Xe × T0 − c)/g0 = (2 × 3 − 1)/2 ≈ 2.5.
    Why this step

    Take away what training compute supplies. Software must supply the rest.

  3. Tg = (G/A)1/λ = (2.5/1)1/1 ≈ 2.5. Tg is the required value of F.
    Why this step

    Divide by research productivity, then undo the power λ. This gives what the lab's research labour and experiment compute must produce.

  4. L = (Tg/E1 − a)1/a = (2.5/11 − 0.5)1/0.5. L ≈ 6.25×.
    Why this step

    Undo the production function to find the labour that produces it.

Progress as labour rises

At 6.25× labour, progress runs at 2×. Labour uses a logarithmic axis.

Labour ratio on a log scale. METR would not be surprised by 1.3× or 3×. The dotted horizontal line marks 2×. The dot shows your forward input.

Labour to hold the pace

These figures depend on our illustrative T₀. At current settings, with T₀ = 1: L(5) = 35.4×. With T₀ = 6: L(5) = 205,000×.

Labour needed to hold the target paceTarget 2×. At year 0: 6.25×. L(5) = 1,130×. Labour uses a logarithmic axis. Muted dashes hold θ = 0.0123451×10×100×1,000×10,000×Years

Target 2×. At year 0: 6.25×. L(5) = 1,130×. Labour uses a logarithmic axis. Muted dashes hold θ = 0.

Years from now; labour ratio on a log scale. Zero labour is stated separately. The dashed reference holds θ at zero. A wall ends the line. Very large labour ratios use number words. This is a sensitivity exercise. We did not estimate these parameters for Anthropic, and the chart does not predict labour supply or model AI feeding back into research labour.

Worked example: holding the pace

Suppose AI doubles the pace. On the starting model, that is like having about 6.25 times the research workforce, counting human researchers and AI research work together. To hold that pace for five years, the workforce would need to be about 1,130 times the workforce on the old path at that date.

The yearly growth factor is how many times more research effort each year keeps the old pace. On the starting model, the settings for research difficulty and returns to effort set it. The calculation multiplies 6.25 by the yearly growth factor raised to the 15th power: three powers for each of five years. The three comes from the starting settings for how much faster software must improve. The starting factor is 1.414, the square root of two; its 15th power is about 181. Using Anthropic’s slower spending growth in 2025, about 2.8 times a year on the lowest count from the spending puzzle, as a stand-in gives a factor of 2.786, whose 15th power is about 4.7 million, giving about 30 million times the workforce on the old path. A small change in this uncertain factor makes a large difference.

These are extreme illustrations, not forecasts. They assume the doubled pace is held, training follows its set path, labour and computing combine as in the starting model, experiment computing stays a fixed share, and research gets harder at a steady rate. The second case also treats spending growth as research-effort growth, which the spending puzzle says is uncertain. More experiment computing, a change in research, or a slower pace could change the result.

Sources

Next: Overview →