We imagine {that a} sensible strategy to fixing AI security issues is to dedicate extra time and sources to researching efficient mitigations and alignment strategies and testing them towards real-world abuse.
Importantly, we additionally imagine that bettering AI security and capabilities ought to go hand in hand. Our greatest security work up to now has come from working with our most succesful fashions as a result of they’re higher at following customers’ directions and simpler to steer or “information.”
We might be more and more cautious with the creation and deployment of extra succesful fashions, and can proceed to reinforce security precautions as our AI techniques evolve.
Whereas we waited over 6 months to deploy GPT-4 to be able to higher perceive its capabilities, advantages, and dangers, it could generally be essential to take longer than that to enhance AI techniques’ security. Due to this fact, policymakers and AI suppliers might want to be sure that AI improvement and deployment is ruled successfully at a worldwide scale, so nobody cuts corners to get forward. This can be a daunting problem requiring each technical and institutional innovation, however it’s one which we’re wanting to contribute to.
Addressing issues of safety additionally requires intensive debate, experimentation, and engagement, together with on the bounds of AI system habits. Now we have and can proceed to foster collaboration and open dialogue amongst stakeholders to create a protected AI ecosystem.
