OpenAI previews o-3 models in 12th day of 'shipmas' wrap-up

The “12 days of OpenAI” wrapped up on Friday with a preview of its newest o3 and o3 mini reasoning AI models to be launched in the New Year.

Open AI’s o-series 3 models have been optimized for coding, with the o3 mini said to be a faster and distilled model of the full version. The company said it plans to launch the o3 mini by the end of January 2025, and the full-version o3 model soon after.

For those looking to get their hands on the o3 models sooner, CEO Sam Altman said the company is opening up early access applications for safety and security researchers to test both frontier models starting Friday.

Today, we shared evals for an early version of the next model in our o-model reasoning series: OpenAI o3 pic.twitter.com/e4dQWdLbAD

undefined OpenAI (@OpenAI) December 20, 2024

What sets these large language models apart from previous models is the introduction of a new training technique OpenAI calls “deliberate alignment.”

According to the company, the term refers to “a training paradigm” that enables the models to “use chain-of-thought (CoT) reasoning to reflect on user prompts, identify relevant text from OpenAI’s internal policies, and draft safer responses.”

This allows the models to precisely adhere to OpenAI’s safety policies without requiring human intervention or interaction.

These improvements in capabilities can also be leveraged to improve safety. Today we’re releasing a paper on deliberative alignment that shares how we harnessed these advances to make our o1 and o3 models even safer to use. https://t.co/tX8GkVOVdA

undefined OpenAI (@OpenAI) December 20, 2024

The AI start-up says deliberate alignment will help prevent models from complying with “malicious prompts, over-refusing benign inquiries, and falling victim to jailbreak attacks.”

This is because these models are trained with the reasoning skills to 'think' through more complex safety scenarios instead of responding instantly, which previously has led to poor data efficiency and decision boundaries, OpenAI said.

“We find that o1 dramatically outperforms GPT-4o and other state-of-the art LLMs across a range of internal and external safety benchmarks,” it said.

Microsoft-backed OpenAI released o1 AI models in September which are designed to have more processing time for user queries to solve hard problems.

As part of OpenAI’s “12 days” of product announcements, OpenAI has introduced safety improvements to its upgraded o1 System Card, a new $200 per month ChatGPT Pro plan, released its text-to-video generator Sora Turbo to the public, rolled out its ChatGPT search engine to all users, and for the first time launched a 1-800-ChatGPT text and call line to answer user questions.