Did you know OpenAI Shares First Glimpse of its New Frontier Reasoning Models o3 and o3-Mini

 

The makers of ChatGPT are back and this time around, they’re previewing another set of frontier reasoning models called o3 and o3-mini. Many did expect the unveiling of the latest reasoning design and they weren’t disappointed.

The company shared how no release for the models is on the cards yet as they continue training the system. However, it wants the research community to get active and experiment with it after submitting applications before the public release.

No date for the global release was shared but the fact that the tech giant jumped from o1 to o3 and skipped o2 as a whole is interesting. As per reports, they didn’t wish to conflict with another telecom firm in the UK called O2. After all, you can only imagine the trademark issues coming forth.

The word reasoning continues to gain widespread attention in the world of AI and it’s a simple term really. It involves breaking down instructions into small tasks to produce the most strong solutions. Such models often depict the effort needed to get the reply instead of just coming up with the final verdict without any justification.

As per the firm, o3 is comprehensive and therefore internal testing phases prove its performance to surpass that witnessed in the past. It’s also designed to beat out its predecessor in all coding tasks. What was even more interesting was how it beat out the company’s own chief scientist in aspects like competitive programming.

The model just aced one of the biggest math competitions, attaining a noteworthy result of 87.7%. This was one of the biggest and best standouts worth a mention.

The company also shared more on the matter including how it didn’t just provide yes/no rules that we see in most AI models for deliberative alignment. Instead, it provided active reasoning related to if the request was in line with the firm’s safety standards. So as a whole, the safety aspect is so much greater than what was previously witnessed on GPT-4.