Can a Million AI Agents Create Superintelligence?

Author: Ramez Naam (guest post, Noah Smith / Noahpinion) URL: https://www.noahpinion.blog/p/can-a-million-ai-agents-create-superintelligence One-line: Parallelizing AI agents yields diminishing returns on intelligence — not just speed — and real-world benchmark data show gains far too weak to produce the self-sustaining recursive self-improvement that RSI models assume.

Key claims

• The core distinction Naam draws is between speed (how fast agents finish a fixed task) and intelligence (whether they can solve a harder problem). More agents buy meaningful speed gains; intelligence gains from parallelization are far smaller and often plateau quickly.
• OpenAI used 124× more tokens per researcher but ran only 1.6× more experiments per researcher — the same gap between inputs and useful outputs that Naam documented in his prior post on RSI.
• On formal mathematical proofs (Lean benchmark, 24-hour runs with Anthropic’s models), the first tenfold expansion from 1 to 10 Opus 5.5 agents added 27 score points; expanding from 10 to 100 agents added just 2 — almost pure saturation after the initial gain.
• Ten Opus 5.5 agents outscored 100 Opus 5 agents and 100 Fable 5.1 agents on the same math benchmark — a smarter model with one-tenth the agents beat an older model at full scale, which is Naam’s central point: underlying model intelligence matters more than agent count.
• On OpenAI’s Navier-Stokes Millennium Problem result (run with ~10,000 concurrent agents), OpenAI researcher Noam Brown said he “wouldn’t even attribute 10% of the credit to multi-agent” — the smarter underlying model deserved the credit, not the swarm.
• GPT-5.6 Sol on the cybersecurity benchmark SEC-Bench Pro scored ~71.2 with 1 agent, 74.3 with 4, and 76.4 with 16 — each fourfold expansion bought fewer points, a sub-logarithmic scaling curve where gains are already shrinking with only 16 agents.
• The three leading software-RSI models (Chan et al. 2026, including Geoffrey Hinton and Yoshua Bengio; Eth & Davidson 2025; Davidson & Houlden 2025) assume doubling researchers multiplies research output by 2× to 2.6×. Real-world multi-agent benchmark intelligence curves show roughly logarithmic or sub-logarithmic gains — dramatically weaker than those assumptions. At a billion agents, the gap between hypothetical RSI models and observed data is orders of magnitude.
• Anthropic’s October employment report shows research agents logging 19 agent-hours per human work-hour by late August 2026. Even under generous assumptions (each agent-hour equals a human research-hour), log scaling predicts only ~2× Anthropic’s April research pace by August — not the 25× the Chan et al. exponent would imply. Naam uses this as a forward-looking test: watch the Epoch Capabilities Index slope in early 2027.
• Power-law learning curves in ML have been observed for over 30 years (since 1993). Naam treats them as a fundamental feature of machine learning rather than an obstacle that will be trivially removed; escaping them probably requires the kind of conceptual breakthrough that current agents struggle to supply.

Notable quotes

“More agents are useful for speed: they can solve many problems faster. They’re less useful for boosting AI intelligence. Multi-agent gains in solving harder problems often diminish incredibly quickly and may even saturate.”

“Removing humans from the loop doesn’t remove shared blind spots or sequential dependencies. Adding more copies of today’s models hasn’t shown the sustained intelligence gains that may be needed for repeated conceptual breakthroughs.”

“I wouldn’t even attribute 10% of the credit to multi-agent.” — Noam Brown (OpenAI), on the Navier-Stokes result

“Ten Opus 5.5 agents are able to solve problems a hundred older-model agents can’t. The stronger model does more with one-tenth as many agents.”

“A thousand copies don’t automatically supply the insight one copy is missing.”

“These are among the strongest diminishing returns we’ve looked at. Of course, these diminishing returns may change. And we have less experimental data to go off of today than we’d like. We’ll need to recalibrate as new evidence comes in.”

“In summary, for now, multi-agent scaling shows incredibly steep diminishing returns to model intelligence, or the ability to reach conceptual breakthroughs. I expect those sorts of breakthroughs to be necessary for major AI advances and for self-sustaining RSI, let alone a runaway intelligence explosion.”

How it connects

• Ramez Naam — author; this is the second in his RSI skepticism series on Noahpinion
• Noah Smith — host publication; introduces Naam as “one of the best futurists alive”
• Where’s the “Intelligence Explosion”? (Ramez Naam, Noahpinion) — part one of this series; established the RSI feedback loop baseline; this post extends it to agent parallelization
• AI Agents — the core subject; what happens when you scale agent count rather than model capability
• Artificial Intelligence — broader context
• Anthropic — provides the key empirical data: Opus 5.5 math benchmark results, agent runtime employment report, and the acknowledgment that current models lag on “ambitious research ideas”
• OpenAI — provides the Navier-Stokes 10,000-agent result and GPT-5.6 Sol cybersecurity benchmark data
• Dario Amodei — author of “Machines of Loving Grace,” which posited “a country of geniuses in a datacenter” — the framing Naam is testing empirically
• Geoffrey Hinton — co-author on Chan et al. 2026, the paper whose 2.6× researcher-doubling multiplier Naam finds implausible against real-world data
• Dwarkesh Patel — cited for emphasizing AI’s “massive population size and perfect knowledge transfer” as drivers of agent-swarm optimism; and for his interview with Noam Brown on the Navier-Stokes result