Google Unveils Gemini: A Multimodal LLM Outperforming GPT-4 with Advanced Features
On December 6th, Sundar Pichai announced Gemini, an LLM (Large Language Model) equipped with multimodal reasoning capabilities, touted as "the most powerful and versatile model we have ever built." Google response to GPT-4 comes in three sizes: Gemini Nano, Gemini Pro, and Gemini Ultra. According to a document shared by Dylan Roussel on X, Google is expected to soon unveil the new name for its chatbot, along with Gemini Advanced, powered by Gemini Ultra 1.0.
As a reminder, the Nano version runs on Android phones, and a refined version of Gemini Pro enhances Bard's capabilities, particularly in reasoning, understanding, planning, summarization, and code. Initially accessible in English in over 170 countries and territories, Google recently announced its availability in all countries where Google Bard is deployed, with support for all associated languages, including French.
Gemini Ultra was initially announced for early 2024, and Google had shared its performance metrics.
Scoring 90.0% compared to GPT-4's 86.4%, Gemini Ultra is the first model to surpass human experts in Massive Multitask Language Understanding (MMLU), utilizing a combination of 57 subjects such as mathematics, physics, history, law, medicine, and ethics to test both world knowledge and problem-solving abilities.
Google also compared Gemini Ultra to GPT-4 on various benchmarks, including code and reasoning, with Gemini Ultra outperforming, except for common-sense inference in HellaSwag.
Planned Google Modifications: According to the shared document, initially dated February 7th but revised to February 8th, the modifications to the Bard/Gemini chatbot will make it more user-friendly, reducing visual distractions and simplifying navigation.
Gemini Advanced: Gemini Advanced, a paid service, will allow subscribers to access additional features through Gemini Ultra 1.0. Google states that Gemini Advanced will be "much more capable of handling highly complex tasks such as coding, logical reasoning, following nuanced instructions, and creative collaboration."
The document does not specify the subscription cost but mentions it will be available in over 150 countries and territories. It will support multiple languages but will be fully optimized only for English queries.
Gemini App: The Gemini app will be available on the Google Play Store for Android and integrated into the Google app on iOS. Initially, it will be available only in the United States but will be deployed in other countries soon, excluding the UK, Switzerland, and EU countries, which may have to wait due to regulations, particularly concerning privacy protection.