OpenAI is preparing to launch a new model, the o3-mini, after the success of DeepSeek, a new AI company which has upset OpenAI employees as it has been seemingly trained on a fraction of the computing power of OpenAI’s models, and uses data from OpenAI’s models.
The o3-mini model is being launched early, and is cheaper, and faster, and more efficient, aimed at crushing DeepSeek.
Former and current OpenAI employees speaking to Wired say there is an internal power struggle at OpenAI between the research and product groups, as well as an ongoing feud between the teams working on advanced reasoning and those working on chat.
Employees would like OpenAI to build a unified chat product, rather than relying on a drop-down menu in ChatGPT to decide whether to use GPT-4o or o1.
OpenAI spent years experimenting with reinforcement learning to fine-tune the model which became o1, and DeepSeek built its model off this work.
Experts in the field believe that whilst DeepSeek may have accelerated the model development process, in the end, the winner will be the company with the most chips.