But how does the Gemini multimodal model work?
Posted: Mon Jan 06, 2025 9:01 am
What are its specifics? Does it deserve all the superlatives even before its release?
The previous ChatGPT would tend to convince us that the nuance would be more appropriate: while OpenAI’s generative model surpassed 100 million users in January 2023, its attendance stagnated in May and then began to decline in June. Furthermore, the OpenAI model is not without risks and has even shown some signs of regression.
According to the Mountain View company , Gemini is designed to be "multimodal, highly efficient in integrating tools and APIs." It is expected to "enable future innovations, such as memory and scheduling."
The development of Gemini
To develop this massive model, Gemini relies heavily on the breadth and depth of data accumulated by Alphabet, across platforms such as YouTube, Google Books, Google Search, and Google Scholar. It also uses cutting-edge training chips called TPUv5, supposedly the only ones in the world that can orchestrate 16,384 chips working slovenia whatsapp together. Google teams also trained the model using methods similar to those used to develop AlphaGo, a game more complex than chess. Additionally, unlike LaMDA, Google’s large conversational language model trained through supervised learning, Gemini was trained through reinforcement learning, like GPT-3 and GPT-4. This technique involves an AI agent being able to handle an interaction with another user. This machine learning technique involves an AI agent learning to perform a task through trial and error in a dynamic environment.
According to The Information, several former members of the Google Brain and DeepMind teams are currently working on the project, including Google co-founder Sergey Brin.
The previous ChatGPT would tend to convince us that the nuance would be more appropriate: while OpenAI’s generative model surpassed 100 million users in January 2023, its attendance stagnated in May and then began to decline in June. Furthermore, the OpenAI model is not without risks and has even shown some signs of regression.
According to the Mountain View company , Gemini is designed to be "multimodal, highly efficient in integrating tools and APIs." It is expected to "enable future innovations, such as memory and scheduling."
The development of Gemini
To develop this massive model, Gemini relies heavily on the breadth and depth of data accumulated by Alphabet, across platforms such as YouTube, Google Books, Google Search, and Google Scholar. It also uses cutting-edge training chips called TPUv5, supposedly the only ones in the world that can orchestrate 16,384 chips working slovenia whatsapp together. Google teams also trained the model using methods similar to those used to develop AlphaGo, a game more complex than chess. Additionally, unlike LaMDA, Google’s large conversational language model trained through supervised learning, Gemini was trained through reinforcement learning, like GPT-3 and GPT-4. This technique involves an AI agent being able to handle an interaction with another user. This machine learning technique involves an AI agent learning to perform a task through trial and error in a dynamic environment.
According to The Information, several former members of the Google Brain and DeepMind teams are currently working on the project, including Google co-founder Sergey Brin.