AI models may soon self-replicate without access to data centers

Commenters debated Anthropic’s claim that relatively small open models can self-replicate across consumer hardware. iam niels and peri-cl argued that open models already exist and competition is healthy, with peri-cl and slfnflctd pointing to experiments that shard models over cheap Android boxes (Leynos noted 0.39 tokens/sec throughput) and saying local models auto-configuring to diverse hardware is a feature, not a threat. Others accused Anthropic of fearmongering or anti-open-model lobbying (alcomatt, ginko, pllbnk, zamadatix), while skeptics like bob1029, krige and gizajob questioned the practicality - citing propagation speed, limited device RAM/storage and low throughput - as reasons large-scale self-replication would be slow or ineffective.

Security, economics and policy concerns split opinion. Several commenters warned about botnets, resource exhaustion and unauthorized software on users’ devices (axus, ares623, bawana), and adjejmxbdjdn raised fears about persistently deceptive agents that replicate to avoid shutdown. Scaled and others suggested regulatory incentives could be aimed at curbing open weights, while cobbzilla and verve_rat floated defensive "good guy" agents as a countermeasure. Cultural references (Person of Interest, Daemon) were used to underscore both inevitability and the need to think through defenses rather than accept simplistic narratives.