Google releases Gemini 2.0 Flash Thinking, an AI model that introduces a thought process to enhance inference, and exceeds OpenAI's o1-preview and GPT-4o in various tests



Google has released ' Gemini 2.0 Flash Thinking ', a model trained to generate 'thought processes' in Gemini 2.0 Flash. It is capable of more powerful inference than the original Gemini 2.0 Flash model.

Gemini 2.0 Flash Thinking Modes | Gemini API | Google AI for Developers

https://ai.google.dev/gemini-api/docs/thinking-mode?hl=ja




The Gemini 2.0 Flash model on which Gemini 2.0 Flash Thinking is based was just released on December 11, 2024. Among the Gemini 2.0 model family, Gemini 2.0 Flash is a model that emphasizes response speed, but in terms of performance it surpasses the Gemini 1.5 Pro, the top model of the previous generation.

Google announces 'Gemini 2.0', a high-speed, lightweight model that overwhelmingly surpasses the previous generation of high-end models - GIGAZINE



On December 19, 2024, Google released a trial version of Gemini 2.0 Flash Thinking, which is equipped with a 'thought process' generation function in Gemini 2.0 Flash.

A demo by Noam Shazier, a key figure in the Gemini project, looks like this.




When I asked a probability question, 'What are the odds when flipping a coin until you get either heads, heads, or tails, heads?', the 'Thinking' column appeared and the model's thinking began to be displayed.



The correct answer, '2:3,' is returned in about 30 seconds.



Logan Kilpatrick, head of Google AI Studio, is trying out a slightly trickier puzzle.




They were shown images of four billiard balls numbered '7,' '9,' '11,' and '13,' and asked, 'How can you make the total 30 using only three of them?'



The Gemini 2.0 Flash Thinking model tried adding up all the possible combinations, but after finding that none of them worked, it said, 'Is there a way to interpret the image representation? The balls have numbers printed on them. Wait a minute... Can you turn the numbers upside down? If you turn a 9 upside down, it looks like a 6.' It then correctly solved the puzzle by finding the three numbers '6,' '11,' and '13' that make up 30.



In

the Chatbot Arena, where AI performance is blind tested by humans, it took first place, beating out OpenAI's o1-preview and GPT-4o, as well as many other models.




Gemini 2.0 Flash Thinking is available in Google AI Studio , allowing you to test responses.

As a test, we typed in 'Tell me how to eat a donut hole.' After going through a thought process that included 'giving a straightforward yet humorous answer' and 'considering whether the word 'how to eat' is meant literally or figuratively,' the system responded with 'Well, unfortunately, there aren't any edible holes in donuts!'


Related Posts:

in Software,   Web Service, Posted by log1d_ts