What it means that new AIs can “reason”

OpenAI's latest language model, o1 (nicknamed Strawberry), uses a "think, then answer" approach that significantly improves its performance. On challenging tasks in physics, chemistry, and biology, o1 performs similarly to PhD students. It also excels in math and coding, scoring 83% on an International Mathematics Olympiad qualifying exam compared to GPT-4's 13%.

This improvement intensifies potential risks, with o1 scoring "medium" risk for capabilities related to weapons of mass destruction. While it can't walk beginners through developing deadly pathogens, it could assist experts in planning known biological threats.

The rapid progress in AI capabilities highlights the need for better evaluation methods and safety measures to harness benefits while mitigating risks.