What bottlenecks define the path from AGI to superintelligence?

The useful framing here is methodological, not predictive. Rather than forecast when superintelligence arrives, the report maps the continuum from human-level AGI to artificial superintelligence — informally, a system more cognitively capable than large organizations of humans — with Universal AI (the theoretically well-understood endpoint) supplying formal grounding at the far end. The middle is structured as four candidate pathways: scaling AGI, AI paradigm shifts, recursive self-improvement, and ASI emerging from large-scale multi-agent collectives. Crucially each pathway is paired with frictions and bottlenecks, and whether those frictions prove negligible or decisive is left as concrete open research questions. The recommended posture is plural: hold many speculative scenarios at once, and pair them with continual benchmarking and forecasting methods robust enough to keep working in a post-AGI regime.

This connects two threads in my vault that usually sit apart. On the recursive-improvement pathway, Can human-AI research teams improve faster than autonomous AI systems? is a direct intervention on one of the four routes — it argues the human-in-the-loop variant may dominate pure self-improvement on both speed and safety, which is exactly the kind of friction-versus-acceleration question the report leaves open. On the multi-agent-collective pathway, Do self-organizing agent teams outperform rigid hierarchies? is empirical evidence about whether collective intelligence actually scales the way the pathway assumes — and it suggests the gains are protocol-conditional, a friction the macro forecast cannot see.

The standard counterargument is that landscape-mapping with "the future is unpredictable" caveats risks being unfalsifiable scenario-spinning. The report's own answer is the right one: the deliverable is not a prediction but a set of forecasts plus a monitoring discipline to update their relative plausibility — which is more defensible than any single timeline and more actionable, since it tells you which bottlenecks to instrument now.