In this column, I continue my analysis of OpenAI's new generative AI model, o1, building on the comprehensive overview provided in my earlier piece linked here. Today, I will highlight a noteworthy but underreported feature of o1 that could transform AI safety. ### Real-Time Double-Checking A key element in o1 is its chain-of-thought (CoT) methodology, which performs real-time double-checking at runtime. This mechanism aims to minimize AI hallucinations—the generation of incorrect or harmful content—making it safer by avoiding bias and dangerous suggestions. Other AI developers are likely to adopt similar methods soon. While this technique has been tried previously, o1 embeds it as a core functionality that operates continuously, enhancing results without user initiation. Although this increases response times (potentially extending to several minutes), the trade-off presents better, safer answers. Users now cannot opt out of this feature, signaling a shift in user control. ### How Chain-of-Thought Works Chain-of-thought in AI mimics human reasoning. For example, when considering a chess move, one anticipates future positions, leading to better decisions. In generative AI, prompting can guide the model to respond in a stepwise manner. The efficacy of this approach becomes evident when testing it with various prompts. For instance, when asked about starting a business, a traditional AI might suggest less advisable options, such as moonshining or bank robbery.

Implementing a CoT approach with safety checks can rectify these suggestions by filtering out illegal options, fostering ethical recommendations instead. ### Framework for Safety Checks In o1, the steps are automated, embedding AI safety checks throughout processing stages. This guarantees not only that user prompts undergo evaluation, but also that responses are monitored for harmful content. While this enhances accuracy, response delays and added costs become apparent due to the computing time involved. What remains to be scrutinized is the degree of safety measures needed and whether users should retain control over these settings or if they should default to a high safety level. Most modern AI models already incorporate multi-stage safety checks in their design. ### Conclusion In summary, real-time double-checking in o1 represents a significant advancement in generative AI safety by enhancing response reliability and accuracy. This feature is poised to improve overall AI safety, an area deserving of ongoing attention and refinement. Future columns will explore additional aspects of o1's capabilities and its implications in the field of AI. Stay tuned for more insights.