OpenAI has updated the “chain of thought” feature of its o3-mini AI model to make it easier for users to understand how it thinks. This comes in the wake of the release of DeepSeek-R1, a rival reasoning model that also shows the thought process behind its responses.
Reasoning models are designed to break down their decision-making processes step-by-step, and therefore take longer to generate responses. Such explanations make it easier to understand why a particular response was given, allowing users to see why their prompt may or may not have resulted in the desired answer. They also allow AI researchers to identify potential biases or errors, and improve its reasoning capabilities.
The chain of thought summary behind an o3 response. Image: OpenAI
In an X post, OpenAI said it has released an “updated chain of thought in OpenAI o3-mini for free and paid users, and in o3-mini-high for paid users.” o3-mini-high is a paid variant of o3-mini with deeper reasoning capabilities and more detailed thought processes at the cost of slower response times.
Why is OpenAI only unveiling its reasoning processes now?
Prior to this update, OpenAI models o3-mini, o1, and o1-mini only gave users access to chain of thought summaries rather than the full reasoning; the company said this was “to provide a balanced trade-off between speed and accuracy,” though it has also referred to “competitive advantage” in the past as a factor in the decision. Users have been threatened with bans when they tried to jailbreak o1 into giving up its full thought process.
However, as DeepSeek’s open-source R1 shows the whole story behind its responses without concern, OpenAI has decided to, not quite lift the lid, but move it ajar. An OpenAI rep told eWeek that, in the update, “the model’s raw (chain of thought) remains hidden as it’s hard to understand,” but is instead presented in a way that is “easy to read.”
The new, more in-depth reasoning summary will go through a post-processing step that simplifies any overcomplicated explanations, removes any “unsafe” reasoning explanations, and translates it into the user’s native language, the rep told eWeek. “With this update, you will be able to follow the model’s reasoning, giving you more clarity and confidence in its responses,” they added in an email.
In a recent Reddit thread, Kevin Weil, OpenAI’s chief product officer, wrote that “showing all chain of thought leads to competitive distillation, but we also know people (at least power users) want it, so we’ll find the right way to balance it.” This is the result of that balance.
Related: Sam Altman Says OpenAI Has Been on “The Wrong Side of History” Regarding Open Source