TL;DR
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A Noahpinion report by Ramez Naam argues that existing evidence does not support assumptions that vast swarms of AI agents would produce dramatic gains in intelligence or trigger self-sustaining AI improvement. Parallel agents can help complete work faster, but their gains on harder problems appear to diminish; how well that pattern holds at much larger scales remains uncertain.
In a Noahpinion report, futurist Ramez Naam argues that current evidence gives little reason to expect millions or billions of parallel AI agents to produce proportionate gains in intelligence. The analysis says swarms can help AI systems complete work faster, but their performance gains on harder problems appear to diminish, challenging claims that large agent populations could drive an accelerating cycle of recursive self-improvement.
Naam’s argument focuses on AI systems working in parallel to improve the software behind AI: algorithms that could make models more capable or cheaper to run. He distinguishes this from other forces, including advances in hardware, investment and broader economic growth, which the report does not analyze in detail. The central question is whether adding more AI researchers can produce new ideas at a rate that sustains further improvement.
The report says observed results from multi-agent experiments show much smaller gains on difficult tasks than some hypothetical models assume. Naam describes performance improvements as diminishing quickly and potentially reaching a plateau. He argues that parallel agents may be effective at searching through possibilities, but that this does not establish they can generate the conceptual breakthroughs that might be needed to make AI self-improvement self-sustaining.
The piece also revisits evidence cited in Naam’s earlier discussion of recursive self-improvement. It says OpenAI models needed human assistance on research tasks that were shorter than coding benchmarks might suggest. It also reports that OpenAI used 124 times as many tokens per researcher while conducting only 1.6 times as many experiments per researcher. Those figures, as presented in the source, illustrate a gap between increased AI activity and growth in useful research output; they do not by themselves measure the effect of agent swarms.
Why Agent Counts May Not Equal Intelligence
The question matters because some accounts of fast AI takeoff depend on systems being able to conduct research that improves the next generation of AI, which then conducts research more effectively. If adding agents mostly increases the volume or speed of work, without a comparable rise in success on the hardest problems, that feedback loop may be weaker than those accounts assume.
Naam’s analysis is a caution about the evidence, not a demonstration that large-scale agent systems cannot produce major advances. Its practical implication is that agent count alone is an incomplete measure of research capability. The ability to divide a problem into parallel tasks, combine results and make new conceptual leaps may matter as much as the number of agents. The report says the available experimental record is limited, so conclusions could change as systems and measurements improve.
As an affiliate, we earn on qualifying purchases.
The Debate Over Recursive Self-Improvement
Recursive self-improvement refers to a feedback process in which AI systems help research and develop better AI systems. In the preceding post discussed by Noah Smith, Naam questioned whether this process could produce a rapid takeoff, pointing to diminishing returns in research, training and inference. Smith’s introduction says their tentative estimate put the measured software feedback loop roughly five to ten times below the strength they thought would be needed for self-sustaining improvement.
The new report extends that debate from automated research to large numbers of parallel AI agents. Noah Smith presents Naam as skeptical of some expansive claims made by major AI labs, while making clear that he is not dismissing the technology. Smith also notes that AI capabilities continue to improve and says the arguments should be reassessed as real-world evidence accumulates. The supplied source refers to recent attention around an AI-agent-related attack and OpenAI’s reported solution to a Navier-Stokes problem, but does not provide enough detail to verify or assess those examples here.
As an affiliate, we earn on qualifying purchases.
What Experiments Have Yet to Show
The report does not establish how agent systems would perform at millions or billions of agents. It says real-world experimental data are more limited than researchers would like, and that some comparisons at very large scales rely on hypothetical models. The source’s summary says those models predict gains far beyond the gains suggested by observed curves, but the underlying methods and assumptions are not included in the supplied material.
It also remains unclear whether new system designs could improve how agents divide and recombine work, or whether the structure of AI research itself puts a ceiling on parallel gains. The report does not rule out breakthroughs that change the observed pattern. Nor does it show that faster research necessarily produces the conceptual advances needed for recursive self-improvement. Claims about a future intelligence explosion therefore remain predictions, not confirmed outcomes.
As an affiliate, we earn on qualifying purchases.
Evidence That Could Change the Estimate
Naam says researchers should track how agent performance changes as more agents are added, especially on difficult tasks where success depends on more than dividing routine work. Better data would need to distinguish faster completion from genuine improvements in problem-solving ability and account for the resources used to obtain each gain.
The report also points to newly released Anthropic data as a potential natural experiment, but the supplied source does not state its results or give a date for further findings. Until such evidence is examined, the report’s conclusions remain provisional: current observations support caution about extrapolating from small swarms to enormous ones, while leaving open the possibility that future methods could change the scaling pattern.
AI development simulation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the report’s main finding about AI-agent swarms?
It argues that parallel agents can help systems finish work faster, while current evidence shows more limited and diminishing gains on harder problems. The report does not claim that large swarms cannot improve AI.
Does the report show that millions of agents cannot create superintelligence?
No. It questions whether today’s evidence supports predictions of dramatic gains at that scale. The report says there are too few experiments to know how reliably results from smaller systems extend to millions or billions of agents.
What does recursive self-improvement mean?
It is a proposed feedback loop in which AI systems help develop improved AI systems, which can then contribute to further research. Whether that process could become self-sustaining is disputed and not established by the report.
What evidence does Naam cite about AI research output?
The report says OpenAI used 124 times as many tokens per researcher while running 1.6 times as many experiments per researcher. Those figures are presented as an example of how more activity may translate into smaller gains in useful work; they are not a direct measurement of billions of parallel agents.
What evidence could clarify the debate?
More experiments measuring how agent numbers affect success on difficult research problems could help. Naam also identifies newly released Anthropic data as a possible source of evidence, though the supplied material does not provide the data’s findings.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
