AI has been the focus of my life’s work, as for many of my research colleagues. Ever since programming AI for computer games as a teenager, and throughout my years as a neuroscience researcher trying to understand the workings of the brain, I’ve always believed that if we could build smarter machines, we could harness them to benefit humanity in incredible ways.
This promise of a world responsibly empowered by AI continues to drive our work at Google DeepMind. For a long time, we’ve wanted to build a new generation of AI models, inspired by the way people understand and interact with the world. AI that feels less like a smart piece of software and more like something useful and intuitive — an expert helper or assistant.
Today, we’re a step closer to this vision as we introduce Gemini, the most capable and general model we’ve ever built.
Gemini is the result of large-scale collaborative efforts by teams across Google, including our colleagues at Google Research. It was built from the ground up to be multimodal, which means it can generalize and seamlessly understand, operate across and combine different types of information including text, code, audio, image and video.
Introducing Gemini: our largest and most capable AI model
Gemini is also our most flexible model yet — able to efficiently run on everything from data centers to mobile devices. Its state-of-the-art capabilities will significantly enhance the way developers and enterprise customers build and scale with AI.
We’ve optimized Gemini 1.0, our first version, for three different sizes:
- Gemini Ultra — our largest and most capable model for highly complex tasks.
- Gemini Pro — our best model for scaling across a wide range of tasks.
- Gemini Nano — our most efficient model for on-device tasks.
State-of-the-art performance
We’ve been rigorously testing our Gemini models and evaluating their performance on a wide variety of tasks. From natural image, audio and video understanding to mathematical reasoning, Gemini Ultra’s performance exceeds current state-of-the-art results on 30 of the 32 widely-used academic benchmarks used in large language model (LLM) research and development.
With a score of 90.0%, Gemini Ultra is the first model to outperform human experts on MMLU (massive multitask language understanding), which uses a combination of 57 subjects such as math, physics, history, law, medicine and ethics for testing both world knowledge and problem-solving abilities.
Our new benchmark approach to MMLU enables Gemini to use its reasoning capabilities to think more carefully before answering difficult questions, leading to significant improvements over just using its first impression.
Advanced coding
Our first version of Gemini can understand, explain and generate high-quality code in the world’s most popular programming languages, like Python, Java, C++, and Go. Its ability to work across languages and reason about complex information makes it one of the leading foundation models for coding in the world.
Gemini Ultra excels in several coding benchmarks, including HumanEval, an important industry-standard for evaluating performance on coding tasks, and Natural2Code, our internal held-out dataset, which uses author-generated sources instead of web-based information.
Gemini can also be used as the engine for more advanced coding systems. Two years ago we presented AlphaCode, the first AI code generation system to reach a competitive level of performance in programming competitions.
Using a specialized version of Gemini, we created a more advanced code generation system, AlphaCode 2, which excels at solving competitive programming problems that go beyond coding to involve complex math and theoretical computer science.
When evaluated on the same platform as the original AlphaCode, AlphaCode 2 shows massive improvements, solving nearly twice as many problems, and we estimate that it performs better than 85% of competition participants — up from nearly 50% for AlphaCode. When programmers collaborate with AlphaCode 2 by defining certain properties for the code samples to follow, it performs even better.
We’re excited for programmers to increasingly use highly capable AI models as collaborative tools that can help them reason about the problems, propose code designs and assist with implementation — so they can release apps and design better services, faster.
See more details in our AlphaCode 2 technical report.
More reliable, scalable and efficient
We trained Gemini 1.0 at scale on our AI-optimized infrastructure using Google’s in-house designed Tensor Processing Units (TPUs) v4 and v5e. And we designed it to be our most reliable and scalable model to train, and our most efficient to serve.
On TPUs, Gemini runs significantly faster than earlier, smaller and less-capable models. These custom-designed AI accelerators have been at the heart of Google’s AI-powered products that serve billions of users like Search, YouTube, Gmail, Google Maps, Google Play and Android. They’ve also enabled companies around the world to train large-scale AI models cost-efficiently.
Today, we’re announcing the most powerful, efficient and scalable TPU system to date, Cloud TPU v5p, designed for training cutting-edge AI models. This next generation TPU will accelerate Gemini’s development and help developers and enterprise customers train large-scale generative AI models faster, allowing new products and capabilities to reach customers sooner.
