a 7B-70B parameter model finetuned for this task spreading over the internet somewhat like a more modern ILOVEYOU.
The best-case scenario would be something like “The capabilities necessary for rogue replication scale with size, ensuring that a 70B parameter model is VERY FAR from havinf a 50% time horizon of 40+ hours, which, if we trust METR, is necessary for rogue replication, while a capable model will be like Agent-1 or Agent-2 from AI-2027 and require far more compute per token”. Additionally, the AI-2027 team pointed to a paper which would imply that attackers won’t have a significant advantage over defenders as AI capabilities improve.
The best-case scenario would be something like “The capabilities necessary for rogue replication scale with size, ensuring that a 70B parameter model is VERY FAR from havinf a 50% time horizon of 40+ hours, which, if we trust METR, is necessary for rogue replication, while a capable model will be like Agent-1 or Agent-2 from AI-2027 and require far more compute per token”. Additionally, the AI-2027 team pointed to a paper which would imply that attackers won’t have a significant advantage over defenders as AI capabilities improve.