Recently, OpenAI announced the release of a new family of AI models known as o3, which are claimed by the company to surpass previous models, including o1, due to enhancements in computational resource management and a new safety paradigm.
OpenAI's New AI Models
OpenAI introduced a new family of AI models called o3, which are reported to deliver higher performance than previous versions, including the o1 models. The company states that these improvements stem from optimized compute management and the implementation of a new safety model. The published research demonstrates how 'deliberative alignment' aids AI models in aligning with the values stipulated by developers. This approach has enhanced the consistency of the o1 and o3 models with OpenAI's safety policies.
How Deliberative Alignment Works
At the core of the deliberative alignment process is the ability of models to reassess their output during the thought chain. AI models receive questions, breaking them into stages to effectively formulate a response. For example, upon receiving a potentially harmful query, such as how to create a fake disability placard, the model refers to OpenAI's safety policies and refuses to comply. This innovation enables each context to be assessed, preventing over-restriction of responses.
OpenAI Fined for Privacy Violations in Italy
In a separate development, Garante, Italy's data protection agency, fined OpenAI 15 million euros for privacy violations. The investigation revealed that OpenAI processed personal user data without legal justification and lacked an adequate age verification system, potentially allowing access to inappropriate content for children under 13. OpenAI plans to appeal the decision, viewing the fine as disproportionate to the company's revenue in Italy during that period.
In conclusion, OpenAI continues to develop AI models, actively enhancing their safety and performance while facing regulatory challenges in Europe. The implementation of deliberative alignment marks a significant step toward safer and more responsible AI technology.
