The phrase “after AGI” can make something requiring explanation sound inevitable: how one capability becomes another. A system competent at many tasks does not automatically gain access to energy, experiments, organisations or physical resources. An improvement on a test does not establish how quickly it will spread, where it will encounter limits or who will be able to use it. Separating those steps allows discussion of possibilities without turning them into a calendar.
The abstract of From AGI to ASI distinguishes four routes: scaling, paradigm shifts, recursive improvement and agent collectives. It also raises bottlenecks and uncertainty about how much they would slow progress. Genewein and colleagues' report appeared on 10 June 2026, with a second version on 30 August. It sets out research questions about possible transitions without guaranteeing an arrival date. Source: arXiv · Google DeepMind
Improvement needs a mechanism
Separating routes invites different tests. If the hypothesis is that more resources produce greater capability, the relevant questions concern additional performance and its cost. If improvement depends on a new idea, ask which problem the idea solves that the previous method did not. One curve cannot establish both things without information about what is changing. A spectacular result may open an investigation before explaining its own cause.
Recursive improvement deserves particular conceptual care. Suppose a system helps propose modifications to another. That would produce an accumulating chain only if the proposals were useful, could be evaluated and were incorporated with sufficient resources. Designing a variant, verifying improvement and deploying it are different operations. Difficulty at any stage could alter the process's speed. This sequence is a way to examine the hypothesis, rather than a prediction that improvement must stop.
Collectives introduce another question: how much extra performance comes from combining capabilities and how much is lost to coordination? In a hypothetical research organisation, many agents could explore alternatives, but somebody must decide which results are reliable, which conflict and what test resolves the disagreement. Increasing participation multiplies available work; it can also multiply claims needing validation. A competent organisation has to deal with both, rather than treating every new output as an established contribution.
Institutional transition is not included
The ability to produce good answers does not settle who should have authority to decide for others, even if a technical route develops quickly. A system improving research can expand what an institution knows; deciding how to distribute benefits, accept risks or impose measures still requires public reasons. No performance leap by itself identifies who should answer to affected people or which objections they must be able to raise.
Preparing institutions should therefore not depend exclusively on predicting a date. Securing evidence access, review capacity and ways to stop particular uses may be reasonable while capabilities are observed. Those measures have costs and need justification through the problems they solve. Their advantage is allowing learning and adjustment without requiring an authority to predict the entire technological trajectory correctly in advance.
The limit of this recommendation also makes the report interesting: listing plausible routes does not determine their probability, duration or combination. Readers can begin by demanding a bridge between stages: what improves, through which mechanism, supported by which test and under what conditions it could fail. That bridge explains more than the promise that one grand acronym will be followed by an even larger one.