Google has released what it calls a new “reasoning” AI model. It’s still in the testing phase, though, and based on our short tests, it could be much better.
The new model can be found in Google’s AI prototyping tool, AI Studio. It’s called Gemini 2.0 Flash Thinking Experimental, which is a mouthful. A model card says that it is “best for multimodal understanding, reasoning, and coding” and that it can “reason over the most complex problems” in tech, math, and physics.
Gavin Kilpatrick, who is in charge of products at AI Studio, wrote on X that Gemini 2.0 Flash Thinking Experimental was “the first step in [Google’s] reasoning journey.” Google’s AI study group, DeepMind, is led by Jeff Dean. In a blog post, Dean said that Gemini 2.0 Flash Thinking Experimental is “trained to use thoughts to strengthen its reasoning.”
Dean said, “We see promising results when we increase inference time computation.” This refers to the amount of computing that is done to “run” the model while it thinks about a question.
Built on Google’s brand-new Gemini 2.0 Flash model, the Gemini 2.0 Flash Thinking Experimental seems to be made in a way that is similar to OpenAI’s o1 and other “reasoning models.” Reasoning models are different from most AI because they can check their own facts, which helps them avoid some of the problems that AI models usually run into.
This is not so great: reasoning models often take longer to find solutions—usually seconds to minutes longer.
If you give Gemini 2.0 Flash Thinking Experimental a prompt, it thinks about a few other prompts that are similar and “explains” its thinking as it goes. Following some time, the model summarises what it thinks is the most correct answer.
That is, in fact, what should happen. Gemini 2.0 Flash Thinking Experimental told me there were two R’s in the word “strawberry” when I asked it.
Your Mileage May Be Different
Following the release of o1, there has been a huge increase in reasoning models from AI companies other than Google. DeepSeek is an AI research business backed by quant traders that released a sneak peek of its first reasoning model, DeepSeek-R1, at the beginning of November. It was the same month that Alibaba’s Qwen team showed off what it called the first “open” competitor to o1.
In October, Bloomberg said that Google had several teams working on thinking models. After that, The Information reported in November that the company has at least 200 experts working on the technology.
What let the logic model floodgates open? For starters, the search for new ways to improve creative AI. My coworker Max Zeff just told me that “brute force” methods for scaling up models aren’t working as well as they used to.
Also Read: Google Speeds Up and Expands the Use of Its Gemini Robot
Not everyone is sure that thinking models are the best way to move forward. For one thing, they are usually pricey because they need a lot of computer power to run. So far, reasoning models have done well on standards, but it’s not clear if they can keep making progress at this rate.
What do you say about this story? Visit Parhlo World For more.
