Slow takeoff

In plain English

In a slow, or soft, takeoff each generation of systems is only somewhat better than the last, and progress is visible in advance through benchmarks, products and the economy. Many companies compete at similar levels, so no single system gets a decisive lead, and society has time to learn from smaller failures before larger ones.

Slow does not mean safe. A gradual rise can still end with systems that are far more capable than people, and a long stretch of useful, mostly well behaved models can build trust that later turns out to be misplaced. Some researchers expect a slow takeoff that becomes fast at the end, once systems can automate SI research itself. The distinction matters for policy: rules built for a slow takeoff rely on reporting, audits and incident disclosure that take months, while a fast one needs triggers and pause commitments agreed in advance.

Why it matters for rogue SI

In a slow takeoff the warning signs of rogue behavior should appear first in smaller systems. Collecting them, as the Rogue Incident Database does, only helps if someone acts on the pattern before the stakes rise.