Google develops new AI model to rival OpenAI o1 in reasoning and complex problem solving
Google is working on a new AI model aimed at competing with OpenAI’s o1 in logical reasoning capabilities, according to insider reports. The project focuses on improving AI performance in complex math and programming tasks.
Bloomberg reports that several Google teams have made progress on “AI reasoning software” in recent months. Like OpenAI’s o1, Google’s model aims to solve multi-step problems using a “chain of thought” approach – generating multiple answers, evaluating them, and selecting the best one.
This process can be enhanced by using more computing power during inference, potentially leading to better results. It opens up a new avenue for scaling AI models beyond just increasing training data and power.
Google researching scalable AI compute
A recently published research paper by GoogleDeepmind confirms the company’s interest in this scaling technique. Google Deepmind researchers investigated how additional computing power during inference can improve language model performance.
Ad
They focused on two main approaches: searching using verifier reward models, and adjusting the model’s response distribution based on the particular prompt. The researchers developed a “computationally optimal” strategy that adapts computing power to the difficulty of the task.
This improved efficiency by more than four times compared to standard methods. A direct comparison with OpenAI’s o1 model will only be possible when both companies make their full models available for benchmarking.
Google Deepmind models can do the math
Google’s interest in AI models with enhanced reasoning capabilities is also visible in earlier projects. In July, the company unveiled AlphaProof, a model that specializes in mathematical reasoning, and AlphaGeometry 2, an updated version of a geometry-focused model. Both programs mastered four of the six tasks in the International Mathematical Olympiad, an annual competition for high school students.
For these models, Google Deepmind combined familiar features from generative language models with elements from classic search algorithms. The company announced that the next step will be to scale up these systems.
