Section 5.1 · Growth dynamics
How fast could capability grow?
Effective compute has grown roughly an order of magnitude per year — the product of better hardware, more investment, and more efficient algorithms. Whether that translates into genuinely new capabilities is the open question.
Effective compute / yr10 ×
By end of decade10,000 ×
Largest training runs / yr4 ×
Algorithmic efficiency / yr3–6 ×
| Horizon | AGI instances | Note |
|---|---|---|
| Start | 1,000 | Even if AGI is expensive to run. |
| +1 year | 10,000 | At ~10× effective compute / yr. |
| +5 years | 100,000,000 | …or 1M instances running ~100× faster. |