On December 20, 2024, OpenAI unveiled a new family of AI reasoning models known as o3, which the company claims surpasses the capabilities of its predecessor, o1.

The o3 model is anticipated to be publicly available in 2025, with OpenAI asserting that its innovative approach could set a standard for future AI reasoning models to align with human values.

OpenAI's strategy for training the o-series models involves utilizing synthetic data generated by AI, which minimizes the need for extensive human input and addresses latency issues.

These models employ a 'chain-of-thought' technique, allowing them to break down prompts into smaller components by generating follow-up questions.

While the o-series models are designed to emulate human reasoning, they fundamentally operate by predicting the next token in a sequence rather than engaging in actual cognitive processes.

The o1-preview model has demonstrated superior performance in safety benchmarks, outpacing competitors like GPT-4o and Claude 3.5 Sonnet in resisting jailbreak attempts.

OpenAI's novel approach, termed deliberative alignment, enhances safety during inference, a phase where users interact with the models, differing from traditional safety measures applied during pre-training and post-training.

Deliberative alignment is crucial for ensuring that AI models remain consistent with the values of their human developers when generating responses.

OpenAI aims to balance the moderation of responses to potentially unsafe prompts while avoiding excessive refusals that could hinder the model's utility in legitimate inquiries.

An example from OpenAI's research illustrates the effectiveness of this approach, as the model successfully declined a request to forge a disabled parking placard by referencing its safety policy.

Research has shown that the implementation of deliberative alignment has significantly reduced the rate of 'unsafe' responses from model o1 while enhancing its performance on benign queries.

As AI models become increasingly prevalent, the discourse surrounding AI safety remains contentious, with critics arguing that some safety measures may resemble censorship.