01.07.24

Unveiling Gemini: Google DeepMind's Most Advanced AI Yet

DeepMind’ team is excited to introduce Gemini, their most capable and general AI model, to date. Gemini is a groundbreaking AI model developed through extensive collaboration across Google, including contributions from Google Research. Unlike previous models, Gemini is multimodal from the start. This means it can seamlessly understand and integrate various types of information — text, code, audio, image, and video.

Versatility

One of Gemini's current standout features is its flexibility. It runs efficiently on everything from massive data centers to mobile devices. The first version, Gemini 1.0, is optimized into three sizes to cater to different needs:

• Gemini Ultra: The largest model, perfect for highly complex tasks.
• Gemini Pro: Ideal for a wide range of tasks, offering excellent scalability.
• Gemini Nano: The most efficient model for on-device tasks.

Unmatched Performance

Gemini has been rigorously tested against various benchmarks, and the results are impressive. Gemini Ultra, for instance, exceeds the technology level in 30 out of 32 tests in large language model research. It even outperforms human experts on the Massive Multitask Language Understanding (MMLU) test, scoring 90.0%.

Advanced Reasoning

Gemini's reasoning capabilities are particularly noteworthy. It's designed to think more carefully before answering hard questions, significantly improving its performance. For example, it achieves a 59.4% score on the new MMMU benchmark, which involves complex multimodal tasks. MMMU benchmark includes 11.5K thoroughly collected multimodal questions demanding college-level subject knowledge to answer.

A True Multimodal Marvel

Unlike other models that stitch together different components for various modalities, Gemini was pre-trained to be multimodal. This means it can naturally understand and reason for text, images, audio, and more. Whether explaining complex subjects like math and physics or understanding nuanced information, Gemini works with that.

Programming Prowess

Gemini's coding skills are top-notch. It understands, explains, and generates high-quality code in languages like Python, Java, C++, and Go. In competitive programming, a specialized version called AlphaCode 2 significantly outperforms its predecessor, solving nearly twice as many problems and ranking better than 85% of competition participants.

Reliability and Efficiency

Gemini 1.0 was trained on Google's AI-optimized infrastructure using custom-designed Tensor Processing Units (TPUs). These TPUs make Gemini faster and more efficient than earlier models. Google is also introducing Cloud TPU v5p, the most powerful and scalable TPU system yet, accelerating Gemini's development even further.

Building with Responsibility

At Google, responsible AI development is a priority. Gemini has undergone the most comprehensive safety evaluations of any Google AI model to date. Safety features include dedicated classifiers to filter out harmful content and robust measures to address challenges like factuality and bias.

Bringing Innovation to You

Gemini is being rolled out across various Google products. Bard, Google's conversational AI, now uses Gemini Pro for enhanced reasoning and understanding. Gemini Nano powers new features on the Pixel 8 Pro, and in the coming months, Gemini will be integrated into more Google services like Search, Ads, Chrome, and Duet AI.

Developers can access Gemini Pro via the Gemini API in Google AI Studio or Google Cloud Vertex AI starting December 13. Gemini Ultra will be available to selected customers and partners for early experimentation before a broader rollout in 2025.

AI Innovation Ahead

Gemini represents a significant milestone in AI development, marking the start of a new era at Google. It advances science to transform everyday life. As innovation continues and models are responsibly enhanced, the goal remains to make the world a better place with AI.

Read More: The Next Step in AI Development: GPT-5