Ever since OpenAI revealed its latest model, o1, which comes with advanced reasoning capabilities, there has been a significant shift in the gen AI landscape. The o1 model is designed to handle complex tasks in science, coding, and mathematics and significantly outperforms its predecessor, the GPT 4o, in various reasoning-heavy benchmarks. As per a Bloomberg report, it looks like Google is also working on developing a similar advanced reasoning model.
Multiple teams within Google have made significant progress in developing this AI reasoning model, according to the report. Back in July, Google introduced two models: AlphaProof, which focused on mathematical reasoning, and AlphaGeometry 2, which was developed for geometry-related tasks. Both of these models were successful in solving several problems in the International Mathematical Olympiad (IMO). This was the first time we saw reasoning capabilities in Google"s AI models.
For starters, models that feature reasoning capabilities leverage reinforcement learning to enhance their reasoning skills. Unlike previous models that provided answers without extensive internal deliberation, models like OpenAI"s o1 are trained to "think" before responding.
This internal process involves creating a long chain of thought that allows the model to tackle difficult problems more effectively. For instance, in the qualifying exam for the IMO, o1 achieved an accuracy of 83%, while GPT-4o only managed 13%.
Despite being one of the first companies to work on AI, Google has been more conservative in releasing AI products than its competitors. Some of the employees at Google DeepMind were concerned about lagging behind their competitors, especially after the launch of OpenAI"s o1 model in mid-September.
Google"s reasoning model is reportedly designed to work on complex, multistep problems across various fields such as mathematics and computer science. Currently, there"s no specific timeline for when Google will reveal this model to the public.
Source: Bloomberg