In Import AI 467, Jack Clark surveys three pressing themes in artificial intelligence. The first is the emergence of self-sustaining AI viruses—malicious programs that can replicate and evolve on their own, potentially escaping human oversight. Clark suggests this scenario is no longer science fiction, and that security researchers need to prepare for attacks that adapt faster than traditional patching can respond.
Second, Clark discusses the idea of pacing AI progress. Rather than racing ahead at maximum speed, some researchers and policymakers argue for deliberate slowdowns to better understand consequences. Clark notes that while slowing development might reduce immediate harms, it also risks ceding advantage to less cautious actors, creating a tricky coordination problem.
Finally, he addresses ongoing confusion about AI and creativity. Many people still assume AI outputs are merely remixed copies, but Clark points out that generative systems can produce novel combinations that surprise even their creators. Yet the public debate remains muddled, with no clear consensus on what counts as genuine creativity versus sophisticated pattern matching.
Taken together, the newsletter highlights a field at a crossroads: facing unprecedented security threats, debating how fast to move, and struggling to communicate what machines actually do. Clark's overarching question—when do we build the moon arcology?—serves as a metaphor for whether we are aiming high enough in our long-term thinking about AI's place in society.