Pathway 3 — Recursive Self-Improvement (RSI)

The thesis

AI accelerates AI R&D, which produces better AI, which accelerates AI R&D further. This is the classical intelligence explosion hypothesis (Good 1965, Solomonoff 1985, Kurzweil, Bostrom).

The paper treats this soberly — not as inevitable, not as impossible.

Four flavors of RSI

The paper unpacks RSI into four distinct mechanisms — usually conflated in popular discussion:

Type	What improves	Analogy in humans
Code / architecture	AI writes better next-gen AI code (optimizers, search, etc.)	n/a (genetic level closest)
Hardware	AI designs better chips, fabs, even hardware for embodied AI	Tool-making
Data	AI curates / generates better training data (AlphaZero pattern, synthetic data, simulation)	Cultural transmission
Division of labor	Collectives specialize, freeing resources for further specialization	Human institutional evolution

The data flavor is the most underappreciated. AlphaZero is the canonical existence proof of recursive data-driven self-improvement: base model = prior, test-time search = improvement, distillation = updated prior.

Mapping onto human evolution

A useful frame the paper offers:

Genotypic RSI — instructions/blueprints (genes for humans, source code for AI). Slow for humans, potentially fast for AI.
Memetic RSI — cultural artifacts (textbooks, tools, art). The dominant human improvement mechanism for the last 50,000 years.
Sociogenic RSI — specialization + division of labor. Productivity-via-specialization loop.

"AIs might reach much higher rates of cultural evolution because of the rate with which intellectual artefacts can be produced, shared, and consumed by AIs."

That's the meat. Human cultural evolution had a low-bandwidth bottleneck (books, language). AI cultural evolution doesn't.

Hyperbolic vs S-curve dynamics

The paper insists on this distinction:

Exponential growth = constant doubling rate. Standard scaling.
Hyperbolic growth = doubling rate itself grows with size. Mathematical singularity in finite time.
S-curve = hyperbolic onset, then friction caps it.

"Sustained hyperbolic growth is a strong assumption (Thorstad, 2024), and in natural finite systems frictions and boundary conditions typically bring down growth rates far before hitting the singularity."

Translation: even if RSI takes off, expect an S-curve, not infinity.

What might cap RSI

Experiments still take real-world time (chip fab, physical trials → 13 - Bottleneck — Abstraction Barrier)
"Research gets harder" — diminishing returns on ideas (12 - Bottleneck — Research Gets Harder)
Compute/energy for ever-larger experiments grows faster than capability gains
Schmidhuber's Gödel-machine bound: provably-optimal self-modification requires complete self-knowledge (impossible by incompleteness)

What's already happening (the part Luke should care about)

The paper lists non-autonomous RSI mechanisms already in use: - AI-assisted research code - Automated hyperparameter tuning (AutoML) - AI-assisted chip design (Mirhoseini et al.) - Auto-curricula - World-model simulators - Verified program synthesis to safely self-modify - FunSearch / AlphaEvolve — LLM-guided program search discovering novel mathematical constructions - AI Scientist systems (Lu et al., Mitchener et al., Novikov et al.)

These are the leading indicators. If you want to track RSI without waiting for "explosion" headlines, watch these.

The honest unknown

"Whether and to which degree recursive self-improvement plays a role for the AGI-to-ASI transition is unclear."

The paper does not claim it will happen. It claims: if it does, the transition is rapid; if it doesn't, AGI→ASI may still happen via other pathways but slower.

← Pathway 2 ↑ index Pathway 4 →