Where’s the “Intelligence Explosion”? (Ramez Naam, Noahpinion)

Author: Ramez Naam, guest post on Noah Smith’s Noahpinion
URL: https://open.substack.com/pub/noahpinion/p/wheres-the-intelligence-explosion
Published: September 27, 2026
One-line: Ramez Naam, one of the world’s leading techno-optimists, makes a data-driven skeptic’s case that the AI self-improvement loop is 5–10× too weak to cause a runaway intelligence explosion anytime soon.

Summary

The core claim. Given current data, the AI self-improvement feedback loop would need to be roughly 5–10× stronger to sustain itself, let alone run away. Naam expects incredibly rapid AI progress by the standards of nearly any other technology — but not a sudden explosion to incomprehensible superintelligence.

Taxonomy of RSI. Naam distinguishes five types of recursive self-improvement: (1) AI boosting human researcher productivity; (2) increasing autonomy with diminishing returns; (3–4) fuller autonomous loops still facing diminishing returns; (5) a runaway loop requiring accelerating returns. He argues we have solid evidence for types 1–2, none for types 3–4, and is skeptical type 5 is reachable without a major conceptual breakthrough.

Narrow superintelligence is already here. In highly verifiable domains — formal math, chess, parts of coding — AI is already superhuman. OpenAI recently published an AI-generated proof resolving the Navier–Stokes existence problem. But superhuman narrow performance does not automatically generalize.

The benchmark gap. Real AI research at OpenAI runs roughly an order of magnitude harder than METR benchmarks or the AI 2027 forecast suggest. METR’s 80%-success task horizon is ~3–4 hours; OpenAI’s internal research-task horizon in July 2026 was ~15 minutes. The AI 2027 forecast of 11 hours looks “substantially over-optimistic” against internal data.

Diminishing returns at every step. OpenAI used 124× more tokens per researcher and engineers shipped ~7× more lines of code — but ran only 1.6× more experiments. Anthropic’s internal data shows a similar gap: a geometric mean of 4× productivity uplift per researcher, but reaching 2× overall capabilities progress would require an order-of-magnitude larger uplift. More tokens ≠ more experiments ≠ better ideas.

What would change this. Naam identifies several potential accelerants: better AI judgment in identifying promising research directions, reduced dependence on physical experiments, stronger AI-to-AI knowledge transfer, or a conceptual breakthrough in learning efficiency. For now the evidence doesn’t point to any of these being imminent.

Noah Smith’s framing. Smith introduces Naam’s piece as “agnostic” on his own part — even without a Singularity, AI of 2040 will look “godlike,” and the practical policy questions may not hinge on whether a fast takeoff occurs in 2027.

Connections

Ramez Naam is one of the world’s leading techno-optimists — he predicted the solar and battery revolutions — making this a notable piece of skepticism about fast AI takeoff from within the techno-optimist camp. The piece sits alongside Singularity and Technological Singularity discussions in Kyle’s corpus, and cites Anthropic’s Claude Mythos Preview System Card directly on the productivity-to-capabilities gap.

The benchmark gap finding — real research 10–44× harder than forecasts — connects to AGI debates: if even the most optimistic forecasts (AI 2027) are this far off on internal OpenAI research, the “tight feedback loop” premise that underlies fast-takeoff scenarios may not hold.