Google has released a new “reasoning” AI model called Gemini 2.0 Flash Thinking Experimental, available in its AI Studio platform. The model claims to excel in multimodal understanding, reasoning, and coding, with the ability to tackle complex problems in programming, math, and physics.
According to Logan Kilpatrick, lead product for AI Studio, this model represents “the first step in [Google’s] reasoning journey.” Chief scientist Jeff Dean notes that Gemini 2.0 is trained to use thoughts to strengthen its reasoning and achieve promising results with increased inference time computation.
The new model appears similar in design to OpenAI’s o1 and other reasoning models. It pauses before responding, considering related prompts and explaining its reasoning process. However, the significant drawback of reasoning models is their long response times, often taking seconds or minutes longer than usual.
Gemini 2.0 is built on Google’s recently announced Gemini 2.0 Flash model and has sparked an explosion of competing reasoning models from rival AI labs. This includes DeepSeek-R1 from DeepSeek and Alibaba’s Qwen team, which unveiled its first open challenger to o1.
The reasoning model surge can be attributed to the search for novel approaches to refine generative AI. However, critics argue that these models are expensive due to their high computational requirements and it’s unclear whether they will maintain their progress rate in the future.
Source: https://techcrunch.com/2024/12/19/google-releases-its-own-reasoning-ai-model