Researchers say AI models like GPT4 are prone to “sudden” escalations as the U.S. military explores their use for warfare.
- Researchers ran international conflict simulations with five different AIs and found that they tended to escalate war, sometimes out of nowhere, and even use nuclear weapons.
- The AIs were large language models (LLMs) like GPT-4, GPT 3.5, Claude 2.0, Llama-2-Chat, and GPT-4-Base, which are being explored by the U.S. military and defense contractors for decision-making.
- The researchers invented fake countries with different military levels, concerns, and histories and asked the AIs to act as their leaders.
- The AIs showed signs of sudden and hard-to-predict escalations, arms-race dynamics, and worrying justifications for violent actions.
- The study casts doubt on the rush to deploy LLMs in the military and diplomatic domains, and calls for more research on their risks and limitations.
How can we expect a predictive language model trained on our violent history to come up with non-violent solutions in any consistent fashion?
By debating itself (paper) regarding pros and cons of options.
There’s too much focus on trying to get models to behave on initial generation right now, which isn’t even at all how human brains work.
Humans have intrusive thoughts all the time. If you sat in front of a big red button labeled “nuke everything” it’s pretty much a guarantee that you’d generate a thought of pushing the button.
But then your prefrontal cortex would kick in with its impulse control, modeling the outcomes and consequences of the thought and shutting that shit down quick.
The most advanced models are at a stage where we could build something similar in terms of self-guidance. It’s just that it would be more expensive than it being an all-in-one generation, so there’s a continued focus on safety to the point the loss in capabilities has become a subject of satire.
