Google has introduced its newest AI model, Gemini, featuring three tiers: Nano, Pro, and Ultra. AI model, Gemini outshines OpenAI’s GPT, reportedly outshines OpenAI’s GPT in the majority of industry-standard benchmarks, according to sources within the company.
Insiders at Google have disclosed that Gemini Ultra surpasses ChatGPT in 30 out of 32 widely recognized academic benchmarks designed to test Large Language Models (LLM). This development could potentially position Google ahead in the AI race, where OpenAI’s GPT model has been widely acclaimed as the most robust and popular in the field.
According to reports from Insider, Gemini Ultra achieved a remarkable 90% score in the MMLU (massive multitask language understanding) benchmark, marking the first instance of a model outperforming human experts. The MMLU benchmark includes 57 subjects, spanning areas like math, physics, history, law, medicine, and ethics, evaluating both world knowledge and problem-solving capabilities.
While Gemini Nano and Pro AI models have already been integrated into Google’s products, Gemini Ultra is slated for release in early 2024. Gemini Nano powers AI features in Google’s Pixel 8 Pro smartphone, and Gemini Pro is accessible through the company’s Bard chatbot. Additionally, Gemini Pro will be made available to enterprise customers and AI studio developers through Google’s Vertex AI program.
Eli Collins, Vice President of Product at DeepMind, the Google division responsible for developing the AI platform, has emphasized that Gemini Ultra possesses the ability to comprehend “nuanced” information across various formats, including text, images, audio, and code.
Although Collins mentioned that some of the data used to develop the application was derived from publicly available web content, specific details about the sources employed to train the AI were not explicitly disclosed.
This breakthrough in AI capabilities could potentially reshape the landscape, positioning Google’s Gemini Ultra as a formidable competitor against existing models like ChatGPT.