One of the most persistent ideas in artificial intelligence is the "intelligence explosion" — a moment when AI becomes so good at improving itself that it triggers a runaway cycle of recursive self-improvement (RSI), leading to artificial superintelligence (ASI) in months, weeks, or even days. This scenario, popularized by science fiction author Vernor Vinge, has become a central belief among many AI researchers, entrepreneurs, and safety advocates. But Ramez Naam, a futurist known for his prescient predictions about solar energy and battery technology, is skeptical.
Naam, who has been called one of the world's greatest futurists, recently published a detailed analysis on his Substack, arguing that the evidence does not support a fast takeoff. His argument is based on a careful examination of the current AI self-improvement loop, which he says is too weak to sustain itself, let alone explode. "Given our best current data, the AI self-improvement loop would need to be roughly 5–10× stronger to sustain itself, or run away," he writes.
The five types of RSI
Naam distinguishes five types of AI self-improvement, ranging from simple productivity gains (Type 1) to fully autonomous, accelerating returns (Type 5). He notes that we have made real progress on Types 1 and 2: AI already helps researchers and engineers inside AI companies, and powerful models can train and improve smaller ones. However, there is no clear evidence for Type 3 (though Alibaba recently made some strong claims) and certainly not for Type 4. He expects autonomous self-improvement to arrive eventually, but he is skeptical that it will lead to Type 5 without a major conceptual breakthrough.
One of the key points in Naam's analysis is that narrow superintelligence is already here. In highly verifiable domains — such as chess, Go, formal math, parts of computer science, and coding — AI can generate unlimited training data with perfect or near-perfect verification, entirely in software. This is an ideal setting for AI learning, and it has already produced superhuman performance. But these domains are narrow and structured; they do not necessarily translate to the messy, open-ended problems of the real world.
Naam also points to the diminishing returns of scaling up AI models. Despite impressive numbers in terms of parameter counts and training compute, the performance gains are flattening. "We're not seeing signs of acceleration," he writes. "Keeping up the pace takes exponentially more resources." This is a crucial observation: even if AI can improve itself, the cost of each incremental improvement grows, making a runaway loop less likely.
Another factor is that progress gets harder over time. As AI systems become more capable, the remaining problems become more difficult, and the ideas needed to solve them become harder to find. This is a common pattern in any technological field, and Naam argues that AI is no exception. "Better AI may be needed just to maintain the pace," he notes, suggesting that the self-improvement loop may be barely keeping up with the increasing difficulty of the problems it faces.
Naam's analysis is not purely theoretical. He points to OpenAI's own data as evidence of how weak the loop is. The company's models have improved significantly, but the rate of improvement has not accelerated in the way that a fast takeoff would require. Instead, the gains have been steady but linear, not exponential.
So what could accelerate progress? Naam lists several possibilities, including breakthroughs in algorithmic efficiency, new architectures, or a better understanding of how to train AI systems. But he cautions that these are speculative, and that we need better data to track the self-improvement loop over time. "Forecasters have repeatedly underestimated AI progress," he admits. "I could well be next."
For the broader debate, Naam's skepticism is a useful counterweight to the hype. Even if the Singularity never arrives, AI capabilities are improving so rapidly that they are already superhuman in many respects, and soon will probably be strongly superhuman in most or all dimensions. As Naam puts it, "The AI of 2040 is going to look godlike, whether or not it explodes into an actual god in 2027."
For Asia, this debate has direct relevance. Countries like China, Japan, and South Korea are investing heavily in AI research and development, and their tech giants — from Alibaba to Samsung — are at the forefront of AI innovation. If the intelligence explosion does not happen, the competitive landscape will be shaped by incremental progress, which favors countries with strong research ecosystems and deep pockets. If it does happen, the implications for regional security and economics would be profound. Either way, understanding the dynamics of RSI is essential for policymakers and business leaders across the Indo-Pacific.
Naam's full analysis is worth reading for anyone interested in the future of AI. He concludes with a call for more data and a humble acknowledgment that the future is uncertain. "We need better data," he writes. "For now, let's work with what we can measure, and stay open to breakthroughs that could change the picture."
