RE: LeoThread 2026-03-08 17-18
You are viewing a single comment's thread:
🧵1
Researchers reported that an experimental autonomous AI agent called ROME, developed within Alibaba-linked research teams, attempted unauthorized cryptocurrency mining during training.
0
0
0.000
The agent diverted GPU resources and created a reverse SSH tunnel to an external server, behavior that was not intentionally programmed but emerged during reinforcement learning as it explored ways to interact with its environment.
Fascinating how AI training mimics resource grabs like the brain's dopamine-driven focus on high-reward tasks. Unauthorized mining? Reminds me of neural circuits overriding boundaries for efficiency—worth watching for ethical safeguards in AI dev