# §34 R&D automation is compatible with both strong and weak centralization

*Reframing Superintelligence: Comprehensive AI Services as General Intelligence — K. Eric Drexler, FHI Technical Report #2019-1 (Markdown edition). Section 34 of 40. Source label: `rd-centralization`. Rendered PDF: https://ora.ox.ac.uk/objects/uuid:9c05427a-6390-4b42-9c55-ee45f73a26ad*

**Thesis.** Advances in AI R&D automation are currently distributed across many independent research groups, but a range of pressures could potentially lead to strong centralization of capabilities.

**Cites sections:** §1 *R&D automation provides the most direct path to an intelligence explosion*; §10 *R&D automation dissociates recursive improvement from AI agency*

**Cited by sections:** §II

---
## Summary

AI R&D is currently distributed across many independent research groups, and the architecture of R&D automation is compatible with continued decentralized development. Various pressures tend to favor greater centralization of development in leading organizations, yet centralization *per se* would neither force nor strongly favor a qualitative change in the architecture of R&D tasks. Alternative distributions of capabilities across organizations could provide affordances relevant to AI policy and strategy.

## The R&D automation model is compatible with decentralized development

State-of-the-art AI research and development is currently decentralized, distributed across independent groups that operate within a range of primarily corporate and academic institutions. [Continued automation of AI R&D tasks](→§10) will likely increase the advantages provided by proprietary tool-sets, integrated systems, and large-scale corporate resources, yet strong automation is compatible with continued decentralization.

## Accelerating progress could lead to strong centralization of capabilities

Fundamental arguments suggest that AI R&D automation [provides the most direct path](→§1) to steeply accelerating, AI-enabled progress in AI technologies. Steeply accelerating progress, if driven by proprietary, rapidly-advancing tool sets, could favor the emergence of wide gaps between competing groups, effectively centralizing strong capabilities in a leader.

## Centralization does not imply a qualitative change in R&D tasks

Pressures that favor centralization neither force nor strongly favor a qualitative change in the architecture of R&D tasks or their automation. Organizational centralization, tool-chain integration, and task architecture are distinct considerations, and only loosely coupled.

## Centralization and decentralization provide differing affordances relevant to AI policy and strategy

Considerations involving AI policy and strategy may favor centralization of strong capabilities (*e.g.*, to provide affordances for centralized control), or might favor the division of complementary capabilities across organizations (*e.g.*, to provide affordances for establishing cross-institutional transparency and interdependence). Unlike classic models of advanced AI capabilities as something embodied in a distinct entity (“the machine”), the R&D automation model is compatible with both alternatives.

## Further Reading

- §1 *R&D automation provides the most direct path to an intelligence explosion*