In Anthropic's October 2026 self-replication tests, open-weight AI models of 30 billion parameters or fewer completed many individual replication steps but made mistakes chaining them together