Google has unveiled a new AI model focused on "reasoning, " though it's still in the experimental phase and appears to need refinement based on initial testing. Named Gemini 2. 0 Flash Thinking Experimental, the model is accessible through AI Studio, Google's platform for AI prototyping. It's described as particularly suited for multimodal understanding, reasoning, and coding, and is capable of tackling complex problems in areas like programming, math, and physics. Logan Kilpatrick, head of product for AI Studio, referred to Gemini 2. 0 Flash Thinking Experimental as Google's "first step" in their reasoning technology journey. Jeff Dean, the chief scientist at Google DeepMind, mentioned that the model is designed to strengthen its reasoning abilities through a process of using "thoughts. " Dean noted that increasing the amount of computation time during inference shows promising results, which refers to how much processing power the model uses to consider a question. The model is built on the recently announced Gemini 2. 0 Flash model and resembles OpenAI's o1 and other reasoning models that can effectively fact-check themselves and avoid common AI errors. However, they typically require more time—from seconds to minutes—to reach conclusions. Gemini 2. 0 Flash Thinking Experimental evaluates a prompt by pausing to consider related cues and explains its reasoning before delivering what it believes is the most accurate answer.

But the model sometimes falters, as demonstrated when it incorrectly stated that the word "strawberry" contains two R's. The rise of reasoning models is not limited to Google. Following the release of o1, numerous AI labs have developed their own models. In November, DeepSeek, an AI research company financed by quant traders, introduced its reasoning model, DeepSeek-R1. Alibaba's Qwen team also launched an “open” alternative to o1 around the same time. Bloomberg reported in October that Google has multiple teams working on reasoning models, and a November report from The Information disclosed that over 200 researchers at Google are dedicated to this technology. The surge in reasoning models is partly driven by the quest for innovative ways to enhance generative AI, as traditional "brute force" methods for scaling models have become less effective. Despite their potential, reasoning models have their critics due to their expensive computational requirements and the uncertainty around their long-term capability to consistently improve.