Google debuts reasoning-optimized Gemini 2.0 Flash Thinking Experimental model – SiliconANGLE

Google LLC today released a new artificial intelligence model, Gemini 2.0 Flash Thinking Experimental, that’s optimized for reasoning tasks.

The company says the algorithm can tackle problems across fields such as programming, physics and math. It’s based on another Google model, Gemini 2.0 Flash, that debuted earlier this month. The latter algorithm is positioned as a midrange option that balances response speeds with output quality.

In a demo video, Google showed Gemini 2.0 Flash Thinking Experimental tackling a logic puzzle that required it to analyze a photo of four billiard balls. It developed a correct answer after deducing that the photo needs to be flipped. The model developed the solution through a process in which it tried several different approaches one after another.

The algorithm’s release is not unexpected. Earlier this year, Bloomberg reported that Google had assigned several AI research teams to building reasoning-optimized AI models. The Information later put the number of staffers working on the project at more than 200.

Google’s reasoning models reportedly use an approach known as chain-of-thought reasoning to carry out processing. The technique breaks down tasks into simpler substeps, which can improve AI output quality. The method was introduced by Google researchers in a 2022 paper.

Chain-of-thought reasoning also powers o1, OpenAI’s rival series of reasoning models. One of the LLMs in the lineup, o1-preview, successfully completed a qualifying exam for the U.S. Math Olympiad. In internal tests, it also answered a set of science questions better than a group of experts with doctorate degrees.

Earlier this month, OpenAI released a paid ChatGPT plan with an upgraded version of o1-preview. The new model can answer relatively simple programming questions with 75% fewer errors than its predecessor. It’s also better at solving math problems. 

The launch of Gemini 2.0 Flash Thinking Experimental should create more competition for o1. Google plans to offer its new model through AI Studio, a service that developers can use to access the company’s Gemini series of LLMs. 

“Built on 2.0 Flash’s speed and performance, this model is trained to use thoughts to strengthen its reasoning,” said Google Chief Scientist Jeff Dean. “And we see promising results when we increase inference time computation!”

The company’s efforts to develop neural networks that can reason previously produced a pair of AI systems called AlphaGeometry and AlphaProof. They’re designed to solve geometry problems and generate mathematical proofs, respectively.

AlphaGeometry is based on an LLM from the Gemini series that has been trained on a large collection of mathematical data. AlphaProof, in turn, combines a language model with AlphaGo, a neural network that Google previously trained to play board games. Together, the two AI systems achieved the same score in this year’s International Mathematical Olympiad test as a silver medalist.

Image: Google