TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Futurist Ramez Naam argues that current evidence suggests parallel AI agents can make some work faster but produce diminishing gains on harder problems. His analysis challenges assumptions that agent swarms could independently drive rapid recursive self-improvement, while acknowledging that evidence is limited and the results could change.
Futurist Ramez Naam argues that adding large numbers of AI agents may speed up work without producing comparable gains in problem-solving ability, challenging a proposed route to rapid AI self-improvement. The argument, presented in a commentary cross-posted by another writer, draws on existing observations and says measured gains on difficult problems appear much lower than some theoretical models assume; the available evidence remains limited.
Naam’s analysis focuses on software research: AI agents working in parallel to improve the algorithms that make AI systems more capable or less expensive to run. It does not attempt to assess every possible source of AI progress, such as better hardware, investment or broader economic growth. The question is narrower: how much can more AI researchers, operating at once, discover?
According to the commentary, experiments indicate that parallel agents are useful for completing multiple tasks faster, but their gains in solving increasingly difficult problems can diminish quickly and may level off. The post contrasts those observations with assumptions in some studies of recursive self-improvement, where much larger agent groups are projected to produce far greater gains. Naam says the gap becomes especially striking when theoretical projections are extended from roughly 100 agents to a billion, though the supplied material does not provide the underlying chart data or enough detail to independently assess the comparison.
The argument builds on a previous Naam post about recursive self-improvement, or RSI: AI systems improving the research and coding that lead to more capable AI. That earlier analysis cited examples of AI research requiring human assistance on tasks shorter than coding benchmarks might suggest, and of a reported OpenAI effort using 124 times as many tokens per researcher while running 1.6 times as many experiments per researcher. These figures are presented in the commentary as evidence that extra AI activity does not translate directly into equivalent gains in useful research output.
How Agent Swarms Could Affect AI Progress
The distinction between speed and capability matters because some forecasts of rapid AI progress rely on systems being able to do more than run tasks concurrently. For a self-reinforcing research loop, AI systems would need to produce improvements that make the next round of research more productive. If adding agents mainly increases the volume of work, while results on difficult questions level off, parallelism alone may not supply the conceptual advances such a loop would require.
That does not show that agent swarms cannot contribute to stronger AI. Faster testing, coding or search could still support progress, and Naam’s argument concerns one mechanism rather than the overall pace of AI development. The commentary’s practical point is that predictions of an intelligence explosion should not treat the number of agents as a direct substitute for demonstrated gains in difficult research. Readers, researchers and policymakers evaluating such forecasts need to distinguish observed performance from assumptions about scaling.
As an affiliate, we earn on qualifying purchases.
The Debate Over Recursive Self-Improvement
Recursive self-improvement refers to a proposed feedback cycle in which AI helps develop better AI, potentially increasing the pace of later research. The supplied commentary says Naam previously judged the measured software feedback loop to be roughly five to ten times below what would be needed for self-sustaining improvement. That is described as a tentative estimate, not a settled measurement or consensus finding.
The writer introducing Naam’s posts says he is skeptical of some expansive claims about AI, including claims associated with major AI labs, but not dismissive of the technology. The writer also notes that examples involving large agent groups have drawn attention, including reported use of a swarm in work on the Navier–Stokes problem. The supplied material does not establish the details of that example or show that it proves a general capability gain.
Naam’s analysis likewise cautions that diminishing returns may change as systems and methods improve. It says the current data set is smaller than researchers would want and that conclusions should be revised as new measurements become available. The claim is therefore a reading of current evidence, not a prediction that future agent systems will necessarily hit the same limits.
As an affiliate, we earn on qualifying purchases.
What Current Agent Tests Cannot Show
The supplied material does not identify the full set of experiments behind Naam’s comparison, their methods, or the exact measured gains at different group sizes. It also does not include the cited images or the underlying data, so the claim that real-world results are far below model projections cannot be independently checked from this account alone. The comparison should be treated as Naam’s interpretation of evidence, not as a comprehensive finding about all multi-agent systems.
It remains unclear whether the observed diminishing returns reflect lasting limits, current model weaknesses, the way tasks were divided, or other features of the experiments. The commentary also leaves open whether future systems might coordinate more effectively or make the conceptual breakthroughs that Naam says could be needed for self-sustaining improvement. No evidence in the supplied material establishes that agent swarms will cause, or cannot cause, superintelligence.
As an affiliate, we earn on qualifying purchases.
Evidence to Watch in Future Tests
Naam says future evidence should be used to recalibrate the estimates, with attention to how performance changes as agent groups grow and tasks become harder. The commentary also points to newly released Anthropic data as a potential natural experiment, but the supplied material does not state what the data contain, when they were released, or what result Naam expects to find.
The next useful developments would be transparent, comparable tests showing how agent count affects both completion time and quality on difficult research tasks, alongside clear descriptions of the workloads and human assistance involved. Until such results are available, the central question remains unresolved: whether parallel agents can do more than accelerate existing work and reliably generate the novel ideas needed to strengthen AI’s own research loop.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the main claim about AI agent swarms?
Naam’s argument is that parallel agents may improve speed more than intelligence. He says gains on harder problems appear to diminish quickly in the evidence discussed, though the data are limited.
Does the analysis show that AI will not become superintelligent?
No. The commentary examines one proposed path: recursive self-improvement driven by software research. It does not establish that superintelligence is impossible or rule out other sources of progress.
What does recursive self-improvement mean?
Recursive self-improvement is a proposed cycle in which AI systems help develop more capable AI systems, which then contribute to further improvements. Whether that cycle can become self-sustaining is disputed.
Why is the evidence still uncertain?
The commentary says there are fewer experiments than researchers would like, and the supplied material does not include the underlying data or methods. Results may also change as models and coordination techniques improve.
What evidence could clarify the debate?
Comparable tests that report how agent numbers affect both speed and success on difficult research tasks would help. Naam also points to newly released Anthropic data as a possible source of evidence, but its contents and implications are not specified here.
Source: rss
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
