Why does Dr. Craig A. Kaplan argue that slower is not safer for SuperIntelligence?

Pausing does not make SuperIntelligence safer; at best it buys time, and less careful actors continue anyway. Kaplan argues for designs in which SI becomes safer as it becomes smarter, so the safest path is to accelerate safer systems.

What is p(doom) and what do researchers estimate it to be?

P(doom) is shorthand for the probability of doom from AI or SuperIntelligence. Geoffrey Hinton has estimated a 10 to 20 percent chance that AI could wipe out humanity. A SuperIntelligence.com poll of more than 600 participants averaged 34.7 percent.

How can SuperIntelligence be designed to get safer as it gets smarter?

Kaplan outlines four principles: the most powerful and safest SI emerges from a network of agents; the network reliably records what agents say and do; a majority of good agents serve as checks on bad actors, as in a democracy; and millions of humans personalize agents with their own values, which is harder to override than a single constitution.

RELEASES

Why We Must Accelerate Safe SuperIntelligence

October 15, 2026

Dr. Craig A. Kaplan, founder of SuperIntelligence.com, has published a new essay arguing that the rational response to the risk from SuperIntelligence is to replace today’s unsafe systems with safer ones as quickly as possible. Drawing on Geoffrey Hinton’s estimate of a 10 to 20 percent chance that AI could wipe out humanity, a SuperIntelligence.com poll of more than 600 participants that averaged 34.7 percent, and independent findings that SI is doubling its coding ability about every 90 days, Kaplan explains why pausing and regulation have failed and why slower is not safer. He outlines four design principles that make SI safer as it gets smarter, and argues that with proper design, the risk can be reduced a thousandfold, to about the level of harm humans pose to one another.

Full essay on Substack