NEWS  AI  December 7, 2023  Reading time: 2 Minute(s)

Max (RS editor)

Google has unveiled its latest AI endeavor, Gemini, a product of the fusion of two renowned research teams, Brain Team and DeepMind. This unified force, known as Google DeepMind, is now making waves with the introduction of three versions of Gemini, marking a significant stride in the realm of artificial intelligence.

Gemini, described as a state-of-the-art AI architecture, distinguishes itself by being natively multimodal, a departure from the conventional approach of stitching together separate components for different modalities. According to Google, Gemini was meticulously crafted, pre-trained on diverse modalities, and fine-tuned with additional multimodal data, resulting in an unprecedented understanding and reasoning capacity across various inputs.

The three versions of Gemini cater to different needs. Gemini Ultra, the largest and most capable, targets highly complex tasks, while Gemini Pro is designed for versatility across a range of devices. The third iteration, Gemini Nano, emphasizes efficiency for on-device tasks. Google hints at potential future sizes, suggesting ongoing innovation.

 GOOGLE GEMINI ULTRA / GPT4 COMPARISON 

Performance-wise, Gemini Ultra reportedly outshines in 30 out of 32 academic benchmarks for current state-of-the-art language models (LLMs). It even surpasses OpenAI's GPT-4 across various categories, except for commonsense reasoning in everyday text tasks.

Acknowledging the heightened capabilities comes a responsibility for enhanced safety measures. Google emphasizes additions to its AI Principles policy, incorporating novel research, adversarial testing, collaboration with external experts, and the integration of safety classifiers to filter out negative content.

Gemini 1.0 is now rolling out to various products, with Bard utilizing a fine-tuned version of Gemini Pro. The Pixel 8 Pro, powered by Gemini Nano, enhances features like Summarize in the Recorder app and Smart Reply in Gboard for WhatsApp. Additionally, Gemini is set to improve Google's Search Generative Experience, reducing latency by 40% in English in the US.

While Gemini Ultra undergoes extensive trust and safety checks, Google plans to make it available for early experimentation to select customers, developers, and partners in the coming year. The dawn of Gemini heralds a new era in AI, promising unparalleled multimodal capabilities and setting the stage for transformative advancements.

 IMAGES AND VIDEO BY GOOGLE/DEEPMIND