The author of the Interconnects AI newsletter has published an essay explaining why he has not bought into true recursive self-improvement (RSI), even as frontier labs such as OpenAI and Anthropic run thousands of concurrent agents to improve their own processes. His central claim is that this activity is better described as lossy self-improvement: it accelerates efficiency, but does not produce a rapid jump in peak intelligence. For businesses, the distinction matters because vendor promises of imminent superintelligence rest on the same evidence.
What the essay argues about RSI
The author frames the current moment as an era in which a few organizations use thousands of concurrent agents, and those organizations are the frontier AI labs themselves, in particular OpenAI and Anthropic. He notes that employees at these labs have rapidly updated their expectations for the pace of AI progress and associated risks, and that the competitive culture of the San Francisco AI scene amplifies any AI concern. Fear sells, he writes, which raises general awareness, but exaggerating risk timelines or severity has negative second-order effects. He recalls the loud AI safety debates of 2023 and 2024, when the primary risks did not arrive in the forecasted timelines. The step from anxiety about agents to extinction risks, in his view, feels very religious.
His alternative scenario rests on three points. Automatable research is too narrow to achieve a massive net acceleration in progress, given the exponential costs implied by scaling laws. Diminishing returns from adding more AI agents in parallel are real. Resource bottlenecks and politics are a major factor in building strong LLMs, and AI can do much less to accelerate that. He balances this against the possibility that labs have seen genuinely scary, specific breakthroughs that are not public yet, and holds high uncertainty on that point. Foundational, imagination-based breakthroughs are what would make him update his RSI timelines toward something more unpredictable and unstable.
The essay draws on two recent Dwarkesh podcasts, with Noam Brown and with John Schulman, Beren Millidge and Charlie O'Neill. The Noam Brown conversation, he writes, made him internalize how large a short-term acceleration mass inference capacity is, and he doubts labs can spend a constant portion of growing compute on internal R&D, especially with plans to IPO and increased scrutiny on basic economics. From the trio podcast he took agreement on the technical state of the art: current techniques solve problems we know how to state, but do not generalize magically to unknown, harder problems in most partially verifiable domains. Progress in math, he says, is an exception rather than a rule.
What this means for business
The practical consequence is that buyers should separate inference-time scaling, which is fairly predictable, from RSI, which is highly uncertain. The author expects agent swarms to be effective at clear, open problems with verifiable answers, and to help most with efficiency rather than peak intelligence, because LLM serving has clear, measurable metrics to improve. That supports better inference-time scaling and more efficient multi-agent systems, and it means cheaper intelligence at a given level is likely to accelerate. For a small company, this favors narrow, measurable automation tasks; for a large one, it favors investing in evaluation and in the human side of research rather than waiting for a step change.
Limitations remain. The author states that all scaling laws show exponential compute and resources are needed for linear improvements in intelligence, and that RSI will have a harder time improving pieces such as managing complex post-training recipes. He also notes that a recurring problem in RSI discussions is the lack of specification in intelligence: LLMs do not cross human-shaped thresholds like remote worker or AI researcher discretely, and a long tail of tasks will always exist. In science, he buys a 10x faster cycle of experiment design and testing in the near future, but not hypothesis generation and intuition building, because accelerating understanding is the key bottleneck and humans improve only marginally there. A vendor claim of full autonomy should therefore be tested against a specific, measurable task list.
The marker to watch is whether labs keep spending a constant share of their growing compute on internal R&D as they approach public markets, and whether agent swarms deliver measurable efficiency gains in serving and multi-agent systems rather than new peak capability. If efficiency gains keep arriving while peak intelligence improves only linearly, the lossy self-improvement view is confirmed, and business planning should assume cheaper, narrower automation rather than a sudden general worker.
