↵ select ↓ ↑ navigate esc close

marginal revolution read out: Are we witnessing an intelligence explosion?

marginal revolution read out ·

Listen: https://serve.buildtall.systems/0c1386c8bd435d81a4f1c30d117bd366c3c7074e5e024de8e37a85c4b86593ac.mp3

Source: https://marginalrevolution.com/marginalrevolution/2026/11/are-we-witnessing-an-intelligence-explosion.html?utm_source=rss&utm_medium=rss&utm_campaign=are-we-witnessing-an-intelligence-explosion


Welcome to marginal revolution read out. In this episode, the post Are we witnessing an intelligence explosion? asks whether the recent acceleration in AI capability growth reported by METR should be read as the start of an intelligence explosion. The post links to two pieces, one a brief commentary and the other the underlying analysis by economists Luis Garicano and Alex Imas.

The first link is to the post itself, Are we witnessing an intelligence explosion? on Marginal Revolution. It draws on a new analysis from Silicon Continent by Luis Garicano and Alex Imas, and it quotes their key accounting result. According to the quoted passage, AI capabilities improved roughly one point nine one times faster after October 2025, but research and development spending grew six point one nine times and the technical workforce three point one zero times. That means measured capability gains per dollar fell by approximately sixty-nine percent, and measured gains per technical employee-year decreased by thirty-eight percent. The authors of the underlying analysis caution that these figures are accounting and nothing causal can be inferred, but they point out that the same numbers are consistent with either congestion, where added effort becomes less productive, or with each advance getting harder. They also note that with enough congestion, the numbers could even be consistent with each advance making the next one easier, which is what an explosion would require. The Marginal Revolution author takes the broader point that the idea of an intelligence explosion still remains to be demonstrated.

The second link is the full analysis titled From Luis Garicano and Alex Imas, published on Silicon Continent. The piece asks directly whether we are approaching an intelligence explosion, and it begins by noting that the METR report used internal data from a frontier lab, Anthropic, and claimed that as of October 2026 the pace of capability growth on well-specified tasks was likely about twice what it was from March 2024 to October 2025. The authors argue that calendar-time acceleration alone does not identify an intelligence explosion. An explosion requires increasing returns to research effort, meaning that each unit of input generates more and more proportional progress as capability grows. So they set out to estimate the relationship between research investment and measured capability.

The authors divide the gains in capability by the research inputs that produced them. They use public regulatory filings to measure research and development spending, which rose from about three point five five billion dollars at an annual rate to twenty-two billion dollars, a sixfold increase. Technical employment rose from about six hundred thirty to one thousand nine hundred sixty people. When they compare this to the METR capability index, they find that progress is faster in calendar time, but each unit of progress is taking more and more resources. In other words, maintaining the same rate of improvement requires ever larger increases in inputs. This echoes the famous Bloom, Jones, Van Reenen, and Webb finding that ideas are getting harder to find across semiconductors, agriculture, and medicine.

To test whether the October 2025 acceleration signals an explosion, the authors fit a research production function of the kind used by Jones and by Aghion, Jones, and Jones. The central parameter is phi, which captures whether existing knowledge makes further advances easier or harder. If phi is below one, each additional increase in intelligence makes the next advance harder, and sustaining growth requires ever-growing inputs. If phi equals one, intelligence reproduces itself and growth continues at a constant rate even without more inputs. An explosion happens only if phi exceeds one, so that each advance makes the next one easier and growth accelerates on its own. The estimates in the paper put phi at around zero point eight four, with a range from zero point six six to zero point nine four in most specifications. That suggests ideas are becoming harder to find in AI, but inputs are growing fast enough to overcome that difficulty and produce faster progress for now.

The authors also test whether the data require a discrete break in October 2025. When they estimate the model using only the observations before October 2025 and then feed in the actual inputs afterward, the model under-predicts the acceleration, accounting for at best half of it. A model with a break estimates a one-time step in research effectiveness, rather than a continuing acceleration. The authors suggest this kind of jump is what the introduction of coding agents might deliver. It raises the level of productivity once, but it does not by itself create an intelligence explosion.

Finally, the authors extend the model to allow AI to do research by itself. An intelligence explosion is precisely the case where AI systems do the research that improves them. In this framework, an explosion needs AI to supply enough of its own research input to more than offset the increasing difficulty of further advances. Looking at the available data, the authors do not yet see convincing evidence that AI is supplying its own research input at the scale required. Both of their models therefore point to the same conclusion: we are not seeing an intelligence explosion yet.

Across the two links, the episode raises a useful distinction between scaling laws and explosions. Scaling laws can produce exponential capability growth from exponential resource growth, but an explosion requires increasing returns to resources themselves. The post and the underlying paper both suggest that current evidence, while showing faster calendar-time progress, is more consistent with harder-to-find ideas than with the self-reinforcing dynamic of an intelligence explosion.

  1. Are we witnessing an intelligence explosion?
  2. From Luis Garicano and Alex Imas