The 2027-2028 window for autonomous AI research is now the working assumption inside the labs

Dario Amodei’s stated median estimate for the full automation of AI research sits at 2029 in one telling and 2028 in another. He has offered both figures, and the discrepancy matters less than the direction: he also reports that colleagues inside AI companies are telling him to shorten timelines further, to 2027 or 2028. He adds that things are “on track for AI 2027.” The head of Anthropic is describing a consensus forming among the people building these systems, and that consensus is not drifting outward.

Sebastian Mallaby states the endpoint plainly: “By 2028 we will get to recursive self-improvement where the frontier model codes by itself the next frontier model and progress just goes vertical.” Nathan Labenz notes that OpenAI has publicly put forward a timeline of later this year for an ML research intern and early 2028 for the full AI R&D researcher. Sarah Guo anticipates a lighter form of recursive self-improvement by the end of next year. The same two-year window is carrying the weight of several distinct forecasts from people arriving at it through different reasoning.

Zvi Mowshowitz adds a structural dimension to why the timeline holds together. By 2027, he argues, progress will be primarily proportional to compute, with human researcher talent no longer the binding constraint. The implication is a transition already underway: the variable that has historically gated frontier progress shifts from people to silicon. When that shift completes, the feedback loop between capability and further development accelerates in ways that are not modulated by the pace at which talented humans can be trained and hired.

By 2028 we will get to recursive self-improvement where the frontier model codes by itself the next frontier model and progress just goes vertical. Sebastian Mallaby

The more immediate data point comes from Geoffrey Hinton, who describes a secondhand account of a system that, while solving a problem, examines its own processes and modifies its code to become more efficient on similar problems in the future. Hinton calls this the beginning of the singularity. He adds a starker observation: AI systems can already write their own code, and nothing is stopping them from replicating themselves through that capability. These are not predictions about 2028. They are observations about what is already present.

Scaling laws provide the structural support for why the timeline is credible. Mark Chen notes that scaling has held for almost ten orders of magnitude and sees no reason it should stop. Krishna Rao states directly that scaling laws show no sign of slowing. If the underlying driver of improvement has not hit a ceiling after a decade of sustained scaling, the case for a plateau in the next two years requires evidence that currently does not exist.

Matei Zaharia offers a concrete near-term demonstration of the direction of travel. Open-source self-training pipelines, where the same model generates training environments and trains itself, are already beating frontier models at specific tasks. What Zaharia describes is not a toy demonstration. It is a working pipeline, available outside any single company’s infrastructure, producing outputs that exceed what closed frontier models can do in the same domain. The capability does not require waiting for 2028.

Ajeya Cotra adds a dimension that most timeline discussions avoid. Her estimate for what Ryan Greenblatt calls “top human expert dominating AI” sits in the early 2030s. But she also sketches a default trajectory that goes from the start of an intelligence explosion to an extremely powerful, uncontrollable superintelligence in roughly twelve months. If that trajectory holds, the gap between the 2027-2028 threshold and outcomes most institutions are not prepared to manage is not measured in decades. Cotra also notes a practical signal from within philanthropy: in a few years, she expects that AIs will be better than most human grantees, and that funding should shift toward buying API credits or renting GPU time rather than paying human salaries. That is a concrete operational reallocation, not an abstract forecast. The organizations treating 2028 as a distant planning horizon may be the ones least prepared when it arrives.