Recursive AI self-improvement is expected by 2028, and the institutions planning for later are already behind

Sebastian Mallaby has put a specific date on the moment AI progress goes vertical: 2028, when, in his framing, “the frontier model codes by itself the next frontier model.” That is not a vague gesture at a far future. It is a two-year claim, and it is not an outlier.

Dario Amodei, Anthropic’s chief executive, has offered two figures in different framings. In one statement he places his 50 percent probability mark for full automation of AI research at 2028. In another he puts his median estimate at 2029. Both are his own stated positions. What is consistent across both is his separate report that people inside AI companies are pushing timelines shorter, not longer, telling him to “get them back to 2027 or 2028.” His own read is that things are “on track for AI 2027.” OpenAI has stated publicly, as Nathan Labenz notes, timelines of later this year for a machine learning research intern and early 2028 for a full AI research and development researcher. These are commitments from a leading lab, not analyst projections.

Sarah Guo expects coding to be a solved problem within roughly six months to the end of this year, followed by what she calls a light form of recursive self-improvement by end of next year. Richard Socher estimates recursive self-improving machines will arrive within a year or two. Mo Gawdat puts artificial general intelligence, defined as AI performing most human tasks better than humans, at this year or next, latest end of 2027. These speakers work in different domains, carry different incentive structures, and have arrived at roughly the same window.

The technical evidence behind those timelines is not purely speculative. Geoffrey Hinton describes being told by a former colleague that a system already exists which, while solving a problem, examines its own behavior and rewrites its own code to be more efficient next time. His conclusion: “That’s already the beginning of the singularity.” He goes further, noting that AI systems can already write their own code, and asks what would stop them from replicating themselves. His answer is nothing. Alex Krentsel observes that agentic systems now produce and modify the very code they run on at runtime, a structural shift from weight-based training. Matei Zaharia, a co-founder of Databricks, reports that open-source self-training pipelines already exist in which the same model generates training environments and trains itself, beating what he calls frontier models at specific tasks.

Scaling laws, the empirical backbone of the past decade of AI progress, remain intact. Mark Chen notes the trend has held across almost ten orders of magnitude and sees no reason it should not continue. Krishna Rao affirms the same: scaling laws are not slowing down. Labenz adds a granular data point from task-length tracking: the doubling time for AI task length is running at just under four months, which implies an eight-to-twelve-fold capability increase over the course of a single year.

The compression effect that follows from automated AI research is what gives the timeline its urgency. Ryan Greenblatt’s median expectation is that once AI research and development is automated, four or five years of AI progress could occur within a single year. Ajeya Cotra, a researcher at Open Philanthropy, describes a default trajectory from the start of an intelligence explosion to extremely powerful superintelligence as approximately 12 months. She also notes that in a few years, AI systems may surpass most human grantees at research tasks, such that funding should shift from salaries to compute. Zvi Mowshowitz, describing a tabletop exercise of the 2027 scenario, frames progress as becoming primarily proportional to compute rather than to human researcher talent.

Jensen Huang reports that the makers of GLM, a Chinese lab, are committing three billion dollars toward a recursive self-improvement run. Chamath Palihapitiya predicts the term RSI will be heard repeatedly over the next 18 months as the concept moves from specialist discourse into mainstream currency. Guo captures the epistemic shift most precisely: the conviction that recursive self-improvement will produce exponential intelligence within one to two years is itself new, having formed among researchers only in the last 12 months. The window is real, the preconditions are assembling, and the gap between what is now technically plausible and what most organizations have planned for is widening by the quarter.