In light of competition from various entities, including the Chinese AI firm DeepSeek, OpenAI is revising the communication approach of its latest AI model, o3-mini, to better share its step-by-step “thought” process.
On Thursday, OpenAI announced that both free and premium users of ChatGPT, the company’s AI chatbot platform, will experience an enhanced “chain of thought” feature that offers deeper insights into the model’s reasoning and the basis for its answers. Those subscribing to premium plans using o3-mini in the “high reasoning” setting will also benefit from this improved output, as stated by OpenAI.
“We’ve introduced an upgraded [chain of thought] for o3-mini to make it simpler for users to understand the model’s reasoning,” an OpenAI representative informed TechCrunch via email. “With this update, you can track the model’s reasoning process, which will enhance your confidence in its responses.”
Reasoning-centric models like o3-mini conduct thorough self-verification before providing outputs, thus overcoming certain common pitfalls often encountered by other models. However, this meticulous process usually results in slightly longer response times, ranging from a few seconds to several minutes.
DeepSeek’s R1 model, which functions similarly to o3-mini, openly displays its entire reasoning process, a method that numerous AI researchers consider advantageous. This transparency not only facilitates enhanced understanding of the model but also significantly improves user experience in various situations, signaling when the model might be following the correct or erroneous path.
Previously, OpenAI chose not to fully disclose the reasoning steps for o3-mini or its earlier versions, o1 and o1-mini, largely for competitive reasons. Users were only presented with summaries of the reasoning, which occasionally contained inaccuracies.
When we shared information about 🍓 prior to the release of o1-preview, witnessing the CoT in real-time was often the “aha” moment that clarified its potential impact. These are not the raw CoTs, but this marks a significant stride toward transparency, and we’re thrilled to share this with the public.
— Noam Brown (@polynoamial) February 6, 2025
Although OpenAI has not yet showcased the complete reasoning steps for o3-mini, the company claims to have struck a “balance”: o3-mini has the capability to “think freely” and subsequently arrange its “thoughts” into more comprehensive summaries.
“To enhance clarity and safety, we’ve implemented an additional post-processing phase where the model reviews its raw thought chain, eliminating any unsafe content and simplifying complex concepts,” the OpenAI representative elaborated. “Moreover, this post-processing feature allows non-English speakers to receive the reasoning in their preferred language, fostering a more accessible and user-friendly environment.”
During a recent Reddit AMA, Kevin Weil, OpenAI’s Chief Product Officer, hinted at forthcoming changes.
“We are planning to reveal significantly more than what is currently shown — [displaying the model’s reasoning process] is on the horizon,” he remarked. “While the specifics are still to be determined, sharing the full chain of thought could lead to competitive undermining, yet we recognize that many users (particularly power users) desire it, so we’re committed to finding the right balance.”
TechCrunch offers a newsletter dedicated to AI! Sign up here to receive it every Wednesday in your inbox.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
