OpenAI’s o1 model doesn’t show its thinking, giving open source an advantage

OpenAI introduces new inference paradigm in large language models (LLM) o1 modelwhich recently underwent a major upgrade. However, while OpenAI has a strong lead in inference models, it may lose some ground open source competitors is emerging rapidly.

Models like o1, sometimes called Large Reasoning Models (LRM), use extra inference time calculation cycles to “think” more, check their responses and correct their answers. This enables them to solve complex reasoning problems that would be difficult for a traditional LL.M. and makes them particularly useful for tasks such as coding, mathematics and data analysis.

However, developers have received mixed reactions to o1 in recent days, especially after the updated version was released. Some people post examples of o1 doing incredible things, while others post examples of o1 doing incredible things. expressed frustration The model’s confusing response. Developers have experienced various issues, including making illogical changes to the code or ignoring instructions.

o1 Details kept confidential

Part of the confusion is due to OpenAI’s secrecy and refusal to show details of how o1 works. The secret behind the success of LRM is the additional tokens generated by the model when it reaches its final response, called the model’s “ideas” or “inference chains.” For example, if you prompt Classic LLM to generate code for a task, it will generate the code immediately. In contrast, LRM will generate reasoning tokens to examine the problem, plan the code structure, and generate multiple solutions before issuing a final answer.

o1 hides the thought process and only shows the final response and a message that shows how long the model thought and possibly a high-level overview of the reasoning process. This is partly to avoid response clutter and provide a smoother user experience. But more importantly, OpenAI treats the inference chain as a trade secret and hopes to make it difficult for competitors to copy o1’s capabilities.

The cost of training new models continues to grow while profit margins fail to keep pace, forcing some AI labs to become more secretive in an effort to extend their lead. Even Apollo studies did this model red teamis not granted access to its reasoning chain.

This lack of transparency has led to various speculations by users, including accusations that OpenAI degrades model quality to reduce inference costs.

Open source model is fully transparent

On the other hand, open source alternatives such as Alibaba’s Qwen with questions and Marco-o1 Show the complete reasoning chain of their model. Another option is Shenxun R1which is not open source but still reveals inference markers. Viewing the inference chain enables developers to address prompts and find ways to improve the model’s response by adding additional instructions or contextual examples.

Visibility into the inference process is especially important when you want to integrate the model’s responses into applications and tools that expect consistent results. Additionally, control over the underlying model is important in enterprise applications. Private models and the scaffolding that supports them (such as safeguards and filters to test their inputs and outputs) are constantly changing. While this may result in better overall performance, it may break many prompts and apps built on top of them. In comparison, the open source model provides developers with complete control over the model, which is a more powerful option for enterprise applications, where performance on very specific tasks is more important than general skills.

QwQ and R1 are still in preview, with o1 leading the way in accuracy and ease of use. For many uses, such as issuing general ad hoc prompts and one-time requests, o1 is still a better choice than open source alternatives.

But the open source community will soon catch up to private models, and we expect more to emerge in the coming months. They can be a suitable alternative in situations where visibility and control are critical.