Under pressure from competitors, including Chinese AI firm DeepSeek, OpenAI is changing the way its latest AI model, o3-mini, communicates its step-by-step “thought” process.
On Thursday, OpenAI announced that free and paid users of ChatGPT, its AI chatbot platform, will be able to see an updated “thought chain” that shows in more detail the model’s “reasoning” steps and how it arrives at an answer to a question. The company announced that it will now be able to According to OpenAI, users of the Advanced ChatGPT program who use o3-mini in an “Advanced Reasoning” configuration will also be able to see this updated view.
‘We’ve introduced an updated [thought chain] in o3-mini designed to help people understand how their models think,’ an OpenAI spokesperson told TechCrunch in an email. ‘This update will allow us to track the model’s reasoning and improve the clarity and credibility of its responses.
Reasoning models like o3-mini help avoid some of the pitfalls that typically cause models to fail because they thoroughly fact-check before producing results. As a tradeoff, inference models take longer to arrive at a solution. It usually takes a few seconds to a few minutes.
DeepSeek’s R1 model is a “reasoning” model, similar to o3-mini, and shows its entire thought process. Many AI researchers consider this to be the preferred method. Not only does the inference step make the model easier to study, it also provides a better user experience in some cases and helps show whether the model is on the right track or the wrong track.
OpenAI has chosen not to show the full inference steps for o3-mini and its predecessors o1 and o1-mini, in part for competitive reasons. Instead, users are only shown a summary of the reasoning steps, which is sometimes incorrect.
While OpenAI still doesn’t show o3-mini’s full reasoning steps, the company says it has “found a balance”: o3-mini can “think freely” and then organize its “thoughts “ into a more detailed summary.
The OpenAI spokesperson went on to say: ‘To improve clarity and security, we added an additional post-processing step where the model reviews the original thought chain, removes unsafe content, and simplifies complex ideas. ‘Additionally, this post-processing step allows non-English speaking users to receive thought chains in their native language, creating a more accessible and user-friendly experience.
In last week’s Reddit AMA (‘Ask Me Anything’), OpenAI’s chief product officer Kevin Wale said that these changes are coming.
We’re trying to show more than we are right now. He said: ‘[Demonstrating the thought process of the model] is coming soon. We’re undecided about everything, but we know that showing the whole chain of thought will lead to competitive refinement, but we also know that people (at least advanced users) want it, so we’ll find the right balance.