Google has introduced a new type of artificial intelligence (AI) model that they are calling a “reasoning” AI model. However, this model is still in the experimental phase, and based on some initial tests, it seems there are areas where it could be improved.
The new model is named Gemini 2.0 Flash Thinking Experimental, which is quite a lengthy name. It can be accessed through Google’s AI Studio, a platform designed for creating and testing AI applications. According to a description provided for this model, it is particularly good at understanding complex information, solving problems, and programming tasks. This means it can tackle challenging questions in areas like math, physics, and coding.
Logan Kilpatrick, who is in charge of AI Studio, mentioned on a social media platform called X that this new model is just the beginning of Google’s efforts to develop reasoning capabilities in AI. Jeff Dean, who is the chief scientist at Google DeepMind (the research division focused on AI), added that this model is designed to enhance its reasoning abilities by simulating thought processes.
Dean also noted that they see promising results when they allow the model more time to think through its answers, which means it uses more computing power to come up with its responses.
Gemini 2.0 Flash Thinking Experimental is built on an earlier version called Gemini 2.0 Flash and shares similarities with other reasoning models, like OpenAI’s o1. What sets reasoning models apart from typical AI is their ability to double-check their own answers, which helps them avoid some common mistakes that regular AI might make.
However, there is a downside: reasoning models usually take longer to provide answers, often taking several seconds to a few minutes to come up with a solution.
Read Also: In 4 Months, AI Coding Assistant Cursor Raised $100M At A $2.5B Valuation Led By Thrive, Sources Say
When the model is given a question, it pauses for a moment to think about related questions and explain its reasoning. After some time, it summarizes what it believes is the best answer. For example, when I asked it how many “R’s” are in the word “strawberry,” it incorrectly responded with “two.”
This shows that the model still has some issues, especially with simple tasks like counting letters in words.
Since the release of OpenAI’s o1 model, many other AI companies have started developing their own reasoning models. For instance, in early November, a company called DeepSeek introduced its own reasoning model called DeepSeek-R1. Around the same time, Alibaba’s team also announced a new model that they claimed was the first open competitor to o1.
Reports from October indicated that Google has multiple teams working on reasoning models, with at least 200 researchers dedicated to this technology.
The surge in developing reasoning models is partly due to the need for new ways to improve generative AI. As one of my colleagues pointed out, traditional methods of simply making models bigger are not producing the improvements they used to.
However, not everyone believes that reasoning models are the best solution moving forward. They can be quite costly to operate because they require a lot of computing power. Although they have shown good performance on tests so far, it remains uncertain whether they will continue to improve at the same rate in the future.
