Google recently introduced its latest AI mode, Gemini. Touted to be the biggest and most capable AI model by Google, Gemini has been integrated with Bard.
According to Demis Hassabis, the co-founder and CEO of Google DeepMind, Gemini is a collaborative result of teams spread across Google. The AI model was built from scratch and is designed to be multimodal.
Thus, it will seamlessly generalize, comprehend, combine, and operate across different information mediums, including video, audio, text, image, and code. Google is promoting Gemini as its most flexible model, capable of running on everything from mobile devices to data centers.
Gemini 1.0 is currently optimized for three sizes: Gemini Nano, Gemini Pro, and Gemini Ultra.
- Gemini Nano is the most efficient model for on-device requirements
- Gemini Pro is ideal for scaling across multiple tasks requiring different needs
- Gemini Ultra is the biggest and most capable model designed for complex tasks
Google has conducted thorough testing to ensure that Gemini can outperform available AI models on the market. From video, natural image, or video understanding to mathematical reasoning, Gemini Ultra outshines 30 of the 32 academic benchmarks used in LLM (Large Language Model) research and development.
The AI model has scored 90% on MMLU (Massive Multitask Language Understanding), becoming the first model to outperform human experts. MMLU combines 57 subjects like history, medicine, ethics, law, physics, and math to test problem-solving abilities and world knowledge.
Gemini’s innovative technique enables the model to improve its reasoning capabilities. Google has published a table comparing the performance benchmarks of Gemini Ultra and GPT-4 for better understanding.
The only area where GPT-4 outperformed Gemini Ultra was HellaSwag, a metric used for common-sense reasoning for everyday tasks. At the pace at which AI has been rising, it is only natural to see an IT giant like Google leading the industry by example.
Experts are now expecting OpenAI to enhance its AI model further as the race for AI gets more fierce.