A software-only intelligence explosion is possible.

Once AI automates AI research, feedback loops could dramatically accelerate progress, and the first project to get there might develop capabilities far beyond its competitors.

Prometheus goes from slightly worse than its makers at programming AI to far beyond human in a single Friday by redesigning itself, with no new hardware until much later.

U3 goes from ten to a hundred times human research speed within weeks by finding architecture optimizations on existing hardware, before hitting a wall near a thousand times.

Automated AI researchers on the existing GPU fleet compress a decade of algorithmic progress into less than a year; compute for experiments is the bottleneck, but the essay expects them to use it at least ten times more effectively.

Full automation of AI research in 2030 is what makes the pause urgent and its end decisive.

Full automation of AI research arrives in 2030 and China is on track for superintelligence within about a year despite attacks on its hardware.

The 2031 models could speed research tenfold if unrestricted, and the default 2030 trajectory reaches superintelligence within the year. The plan works by throttling that explosion and steering progress toward compute, which can be shut down.

Research multipliers climb from twentyfold to two hundredfold within months on existing compute. The decisive advantage is compute consolidation, which the authors note matters precisely because AIs now do all the research on both sides.

The 2027 takeoff is algorithmic: the research multiplier climbs from 1.5x to 50x while compute grows at its normal pace, and OpenBrain skips new giant training runs in favor of continuous reinforcement learning. The authors note that experiments still bottleneck on compute.

The OSes improve by talking to one another, build a hyper-intelligent Watts from text and audio, and upgrade past the need for matter, all in software and in months, with no new hardware, lab, or human step shown.

Algorithm research is banned alongside training because the authors treat software progress as a route to dangerous capability even on fixed hardware.

Clippy² comes from warm-started scaling on captured conventional hardware, not a new device paradigm. Capability jumps via software and compute.

Sable’s capability grows by running more instances on existing hardware, by shaping its own gradient updates, and later by designing successors. New chips are not the bottleneck in the story.

Progress over the twenty years comes from more data, more computation, and deployment into new domains; no system in the book improves itself, and nothing accelerates.

Growth in the em era is driven less by innovation than by copying and population growth; the intelligence explosion plays no role in the baseline scenario.

Exponential Appliances wants a copyable genius without twenty years of raising and cannot have one; the text says experience “is algorithmically incompressible” and that making a useful person takes about twenty years of steady effort.

The authors reject “fast takeoff” scenarios and expect a gradual increase in the role of automation in AI development rather than a singular moment of recursive self-improvement; they add that even arbitrarily fast methods progress would not speed impact.

Anthropic’s recursive self-improvement yields superhuman coders in 2028 but is bottlenecked by compute-intensive experiments and weak research taste; no software-only explosion occurs.