Google releases 'reasoning' AI model to take on OpenAI's o1
Google has unveiled a new artificial intelligence (AI) model, called Gemini 2.0 Flash Thinking Experimental, which is available through AI Studio - Google's AI prototyping platform. According to the tech giant, it is "best for multimodal understanding, reasoning, and coding." The AI model can solve complex problems in a range of disciplines including programming, math, and physics.
Google's reasoning journey and model training
In a post on X, Logan Kilpatrick, the head of product for AI Studio, described Gemini 2.0 Flash Thinking Experimental as "the first step in [Google's] reasoning journey." Separately, Jeff Dean, Chief Scientist at Google DeepMind, said the model is "trained to use thoughts to strengthen its reasoning," in his post. He added promising results when increasing inference time computation—the computing power used to operate the model during question consideration.
How the new reasoning model works
Gemini 2.0 Flash Thinking Experimental's unique features
The Gemini 2.0 Flash Thinking Experimental model builds on Google's recently launched Gemini 2.0 Flash model and is similar to OpenAI's o1 and other reasoning models. Unlike these types of AI, reasoning models can self-verify their results, helping them sidestep common pitfalls that often hinder AI models. However, this also means they tend to take longer—usually seconds to minutes more—to arrive at conclusions.
How the new model processes information
When given a prompt, Gemini 2.0 Flash Thinking Experimental takes a moment to respond, looking at a few related prompts and explaining its thought process along the way. Ultimately, it summarizes what it thinks is the most accurate response. However, being in the experimental stage, there have been cases where the model's responses weren't completely accurate—like incorrectly counting the number of 'R's in "strawberry."
Reasoning models: A growing trend in AI research
The launch of OpenAI's o1 has triggered a wave of reasoning models from rival AI labs, including DeepSeek and Alibaba's Qwen team. Back in October, Bloomberg reported that Google had several teams working on reasoning models. The Information later revealed in November that at least 200 researchers at the company were working on this tech.
