Conceptual Research on Safe AI
A research center at Carnegie Mellon University
Welcome to the center for Conceptual Research on Safe AI (CoRe Safe AI) at CMU.
Early work on AI safety was either concerned with the safe deployment of AI of the time — say, keeping early self-driving cars from crashing — or theoretical in nature, trying to anticipate systems that did not yet exist, based on abstract concepts such as agents, the pursuit of power to achieve objectives, and intelligence explosions.
Now that we have highly capable broad AI systems, concepts such as “AI alignment” have taken very concrete instantiations, such as the fine-tuning of LLMs-as-chatbots. This has allowed for concrete experimental work on such topics to take place, which is an important advance for the field and one that should also inform theoretical work. Still, focusing entirely on today’s systems, advanced as they may already be, risks taking our attention away from challenges to arrive in the future.
The center pursues modern ways to think conceptually about AI safety, informed by the latest advances but with an eye towards the future, aiming to identify important aspects that are not yet visible today but that we should expect in the (possibly near) future.
More detail to follow. If you are interested in AI safety you may also be interested in the Carnegie AI Safety Initiative (CASI).