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
A Ramez Naam essay examines whether running large numbers of AI agents in parallel could power recursive self-improvement. Naam argues that available experiments show diminishing gains on harder problems, but the evidence is limited and does not settle what larger swarms can achieve.
Futurist Ramez Naam argues that large swarms of AI agents may speed up research without delivering comparable gains in problem-solving ability, challenging one proposed route to recursive self-improvement. The essay, cross-posted by its publisher, says available experiments show sharply diminishing returns as agents tackle harder tasks, while stressing that evidence remains limited and predictions about much larger systems are uncertain.
The essay distinguishes between using more agents to complete more work quickly and using them to make an AI system substantially more capable. Parallel agents can search many possible solutions or handle separate tasks, Naam writes, but observed performance gains on harder problems appear to diminish and may level off. The question is whether such systems can produce the conceptual breakthroughs needed to improve the underlying models, rather than simply increase the volume of activity.
Naam contrasts experimental observations at relatively small scales with hypothetical models that project much larger gains. The source describes a comparison at 100 agents and an extrapolation to one billion, saying real-world gain curves are far below the hypothetical projections at the larger scale. It provides no numerical values, experimental details, or common measurement method for those curves, so the comparison cannot be independently assessed from the supplied material alone.
The essay focuses on software research: AI systems improving algorithms that make AI more capable or less costly to run. It does not treat hardware, investment or wider economic growth as part of the main analysis. Naam says the returns from parallel agents could change as systems and methods improve, and calls for updating the assessment when better experimental evidence becomes available.
Why Agent Counts May Not Scale Capability
The distinction matters because some arguments for rapid AI self-improvement depend on AI systems carrying out research at far greater scale than human teams can. If adding agents mainly increases speed or the number of experiments, but yields smaller gains on the hardest research problems, a large swarm alone may not create the accelerating cycle those arguments envision.
That does not establish that recursive self-improvement is impossible or that current AI progress will slow. The essay’s narrower point is that agent count should not be treated as a direct proxy for intelligence or research breakthroughs. For readers tracking claims about superintelligence, the relevant evidence includes how well agents coordinate, whether they solve harder problems, and whether their work improves the underlying models.
As an affiliate, we earn on qualifying purchases.
The Self-Improvement Argument So Far
The essay follows a previous post by Naam on recursive self-improvement, defined as AI advances feeding into subsequent rounds of AI research. In that earlier discussion, he argued that several steps in the feedback loop face diminishing returns. The source says OpenAI models needed human assistance on research tasks shorter than coding benchmarks might suggest, and reports that OpenAI used 124 times as many tokens per researcher while running 1.6 times as many experiments per researcher. The supplied text does not specify the measurement period or full methodology for those figures.
That earlier post also described a tentative estimate that the measured software feedback loop was roughly five to ten times below the strength needed to sustain itself. Naam presented this as a calibration to revisit as data change, not a settled measurement. He further cited Anthropic’s reported view that more capable AI has helped maintain the pace of progress. The current essay extends that debate by asking whether multiplying the number of AI researchers could overcome the constraints.
The publisher introduces Naam as a skeptic of some expansive claims made by major AI labs, especially Anthropic, while saying he is not dismissive of AI. The source mentions public discussion of agent swarms, including a Hugging Face attack and OpenAI’s reported solution of a Navier-Stokes problem, but it gives too little detail to establish what role agents played in either case.
“I’m not sure how much this matters, when the capabilities of the underlying models — swarm or no swarm — are still improving at such a steady and impressive rate.”
— The essay’s publisher
As an affiliate, we earn on qualifying purchases.
What the Experiments Do Not Show
The supplied source does not identify the experiments behind the agent-scaling curves, provide their data, or explain how “intelligence” is measured across tasks. It also does not give enough information to test the comparison between observed results and projections at one billion agents. That scale is an extrapolation, not a reported experiment in the material provided.
It remains unclear whether improvements in agent coordination, task division or model capability would change the pattern. The essay itself says there is less experimental evidence than researchers would like and that diminishing returns may shift. It also does not establish whether any particular agent swarm caused the examples mentioned in the introduction, or whether parallel agents can produce rare breakthroughs even when average gains diminish.
As an affiliate, we earn on qualifying purchases.
Evidence That Could Change the Estimate
Naam says the assessment should be recalibrated as new results arrive. Relevant evidence would include experiments that report agent numbers, task difficulty, computing and token use, performance measures, and whether gains persist as tasks become more demanding. Results should also distinguish faster completion from improvements to the underlying AI system.
The essay points to newly released Anthropic data as a natural experiment it intends to examine, but the supplied source does not include the dataset, its findings or a publication date. No specific follow-up date or next milestone is given. Until those details are available, the claims about very large swarms remain a combination of small-scale observations and uncertain extrapolation.
AI research collaboration software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is recursive self-improvement?
Recursive self-improvement is a proposed feedback loop in which AI advances help systems conduct research that produces further AI advances. The essay asks whether parallel agents could make that loop strong enough to sustain itself.
Does the essay say agent swarms are useless?
No. Naam says agents can help complete many problems faster. His argument is that this speed or increased activity does not automatically translate into large gains on harder problems or new conceptual breakthroughs.
Has a billion-agent system been tested?
The supplied source describes a billion-agent comparison as an extrapolation from gain curves, not as a billion-agent experiment. It gives no underlying numerical results or methodology.
What could alter Naam’s assessment?
More detailed experiments could show whether larger swarms continue to improve performance on difficult research tasks, and whether those gains lead to better AI models. Naam says the view should be updated as new evidence emerges.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
