"Three Types of Intelligence Explosion" is a report by Tom Davidson, Rose Hadshar, and William MacAskill, published by the Forethought Foundation on 2025-03-17. It decomposes the idea of AI accelerating AI into three nested feedback loops — software, chip technology, and chip production — and argues that each has a distinct characteristic speed and a distinct implication for how power would be distributed if an intelligence explosion occurred.
Summary of argument
The report distinguishes three types of intelligence explosion (IE), each adding a further feedback loop to the previous type and each assigned an author estimate of the likelihood that it accelerates:
| Type | Feedback loops included | Likely to accelerate? |
|---|---|---|
| Software IE | AI-driven software improvements only | ~50% |
| AI-Technology IE | + chip technology improvements | ~75% |
| Full-Stack IE | + chip production scaling | ~90% |
The three loops are argued to differ in their characteristic time lags. Software improvements operate on the scale of months, set by retraining cycles; chip technology improvements operate on the scale of years; and chip production scaling operates on a years-plus scale, set by fab construction timelines. Because of these differing lags, the authors argue the loops would activate in order, with the fast loops engaging first and the slow loops later.
On the software loop, the report argues that software could improve by roughly 13 orders of magnitude before physical limits bind, equivalent to about four GPT-sized jumps.
The authors argue that the distribution of power varies by type. Software IE is described as favoring single countries or companies, on the reasoning that whoever holds the best AI iterates fastest and does so locally. Full-stack IE is described as distributing power more broadly, because physical fab locations matter.
The report presents this decomposition as an alternative to a monolithic fast-versus-slow takeoff framing, offering instead a typology the authors intend to be usable in policy analysis. The authors connect the typology to export-control and compute-governance debates: on their account, if software IE dominates, chip export controls bite only temporarily, whereas if full-stack IE dominates, they bite much harder. These connections bear on Export Controls (AI) and Compute Governance.
Relationships
- depends-on: Scaling Laws, AI Software Progress, Semiconductor Supply Chain.
- supports: Recursive Self-Improvement (RSI), Compute Governance.
- related: The Industrial Explosion (Davidson, Hadshar) (companion paper), AI 2027 (which assumes fast software IE), Broad Timelines (which argues for wider uncertainty).