The report's quantitative forecasting hinges on one estimate from Epoch AI: effective compute grows by ~10× per year.
That number is a product of three independently-tracked factors, each of which has held for roughly a decade:
| Factor | Annual rate | What it means |
|---|---|---|
| Hardware (Moore's law $) | 1.5×/yr | More FLOPs per dollar; six decades of trend. |
| Investment growth | 2.5×/yr | Tech-lab spending on compute, last decade. |
| Algorithmic efficiency | 3×/yr | Compute needed for fixed performance shrinks. |
| Product (effective compute) | ~10×/yr | One order of magnitude per year. |
$1.5 \times 2.5 \times 3 = 11.25$, rounded down to 10× as a conservative estimate.
If this rate holds for even a few more years, every other discussion in the paper inherits exponential dynamics.
Even if the rate slows substantially, the gap between today and "even modest AGI" closes faster than human institutional adaptation. This is the quantitative backbone of the report's claim that "preparing for AGI day" is the wrong planning horizon (01 - The Big Picture).
"Suppose by the time human-level AGI is available, base model progress plateaus but effective compute continues to grow at 10×/yr for a bit longer. Even if AGI were initially expensive to run, and only 1000 instances could be run, after a year it would be 10,000... after five years it would be 100 million instances, or 1 million instances a hundred times faster."
This is the key bridge to 09 - Pathway 4 — Multi-Agent Collectives: quantitative scaling alone may suffice for ASI even if no single model exceeds human level, because you can run millions of them in parallel, each faster.
The paper is careful: 10×/yr is not a law. It depends on: - Continued investment growth (no AI-spending crash) - Continued hardware $/FLOP gains (no Moore's death) - Continued algorithmic gains (no theoretical ceiling) - No deliberate slowdown (14 - Bottleneck — Deliberate Slowdown) - Sufficient data (10 - Bottleneck — Data Wall) - Sustainable energy + materials economics (11 - Bottleneck — Economics & Resources)
Each of these is a bottleneck section in the paper. Every one is a way the 10× story breaks.
The paper recommends, as one of its most practical outputs: track these three subcomponents quantitatively and continuously. Not because forecasting will get accurate — it won't — but because deviations in the components are early signals about which scenario you're in.