Google has recently introduced a new AI model called Gemini 2.0 Flash Thinking Experimental, which focuses on reasoning capabilities but is still in the experimental phase. Initial tests indicate that there is room for improvement.
The Gemini 2.0 Flash Thinking Experimental is available on Google's AI prototype development platform, AI Studio. According to its model card, it is best suited for multi-modal understanding, reasoning, and programming, and can tackle complex problems in fields such as coding, mathematics, and physics.
Logan Kilpatrick, the product lead for Google AI Studio, mentioned in a post on the X platform that Gemini 2.0 Flash Thinking Experimental is Google's initial attempt in the field of reasoning. Jeff Dean, Chief Scientist at Google DeepMind, also noted in his post that the model enhances its reasoning abilities through "thinking." Dean pointed out that increasing the computation time for reasoning processes can yield better results.
Built on the recently released Gemini 2.0 Flash model, Gemini 2.0 Flash Thinking Experimental is designed similarly to other reasoning models like OpenAI's o1. Unlike most AI models, reasoning models can effectively perform self-fact-checking, thereby avoiding common errors. However, they typically require more time (from a few seconds to several minutes) to generate answers.
Upon receiving a question, Gemini 2.0 Flash Thinking Experimental pauses for a while, considers multiple related questions, and explains its reasoning process during this period. After some time, the model provides what it deems to be the most accurate answer.
It is worth noting that the performance of Gemini 2.0 Flash Thinking Experimental is not always perfect. For example, when asked how many "R"s are in the word "strawberry," the model incorrectly answered "two."
Following the release of OpenAI's o1 model, other AI labs have also launched various reasoning models, including Google. In early November, DeepSeek, an AI research company funded by quantitative traders, released a preview of its first reasoning model, DeepSeek-R1. The same month, Alibaba's Qwen team introduced a model that they claim is the first "open" model to challenge o1.
Reports indicate that multiple teams at Google are working on developing reasoning models. In November, it was reported that at least 200 researchers at Google are focused on this technology.
The rise of reasoning models is partly due to the search for new methods to improve generative AI. As the approach of scaling up models through brute force becomes less effective, the AI field needs new breakthroughs.
However, not everyone believes that reasoning models are the best way forward. These models are often costly and require significant computational power to run. While they perform well in benchmark tests, it remains unclear whether reasoning models can maintain this rate of progress.
Google's release of the Gemini 2.0 Flash Thinking model is part of its efforts in "agent-based" AI, and OpenAI has also made the full version of its o1 reasoning model available to ChatGPT subscribers.
