According to Google, Gemini Pro is the same as the GPT-3 released by OpenAI a year ago5 The AI model competes with a product that claims to be the winner in six out of eight benchmarks. And the more compact nano version (with 1.8 billion and 32500 million parameters) are optimized specifically for Android app development, and they are distilled from the larger Gemini model.
Gemini claims to be the most powerful and versatile model available, with the first version of Gemini 10 can understand and manipulate different types of information, including text, images, audio, and **.
Gemini is natively designed as a multimodal model, pre-trained on different modalities at the beginning, and then fine-tuned with additional multimodal data to improve performance. Currently, Gemini 10 is trained to recognize and understand different types of information such as text, images, audio, and ** at the same time, and can answer questions involving complex topics, which also makes Gemini excel in explaining and reasoning about complex problems such as mathematics and physics.
And that's not all, Gemini 10 is able to understand and interpret the world's commonly used programming languages such as Python, J**A, C++ and Go, and produce high-quality programs. It is worth mentioning that Google DeepMind launched the AI** generation system AlphaCode 2 years ago, and now through a special version of Gemini, it has established a more advanced AlphaCode 2, which is not only good at programming, but also can deal with competitive programming problems related to mathematics and theoretical computer science.
The Google DeepMind team used the TPU V4 and V5E (Tensor Processing Unit) developed by Google to accelerate machine learning Xi to train Gemini 1 on a large scale0。There are three types of models: Ultra, Pro, and Nano, depending on the size of the model.
The mid-range Gemini Pro is able to beat the GPT-35. Expandable for a variety of tasks;Gemini Nano is used for specific tasks and mobile devices.
The Gemini Ultra is the largest and most powerful model designed for highly complex tasks, while the Gemini Nano is the most efficient model for processing the tasks on the device.
Although it has not been officially announced, according to internal sources, Gemini has trillions of parameters, and the computing power used for training is even five times that of GPT-4.
Since it is a model that has been used to harden the GPT-4, Gemini has of course undergone the most rigorous tests.
Google evaluated the performance of both models on a variety of tasks and was pleasantly surprised to find that Gemini Ultra surpassed GPT-4 on 30 of the 32 commonly used academic benchmarks, from natural imagery, audio, comprehension to mathematical reasoning
And in the MMLU (Massive Multitasking Language Understanding) test, the Gemini Ultra scored at 90A high score of 0%, surpassing human experts for the first time.
The MMLU test includes 57 subjects, such as mathematics, physics, history, law, medicine, and ethics, and is designed to examine the world's knowledge and problem-solving skills.
In each of these more than 50 different subject areas, Gemini is as good as the best experts in these fields.
Google's new benchmark for MMLU allows Gemini to use reasoning more carefully before answering complex questions, a significant improvement over relying solely on intuitive responses.
In a blog post published on the same day, Google said that the Gemini models were rigorously tested and evaluated how well they performed on a variety of tasks.
From natural images, audio, and comprehension, to tasks like mathematical reasoning, Gemini Ultra outperformed current SOTA results in 30 of the 32 academic benchmarks widely used in large language model development.
In addition, Gemini Ultra scored a high of 90 in MMLU (Massive Multi-Task Language Understanding Dataset).0%, surpassing human experts for the first time. The MMLU dataset contains 57 subjects such as mathematics, physics, history, law, medicine, and ethics, which are used to test the knowledge and problem-solving ability of large models.
The new approach to the MMLU test set allows Gemini to use its reasoning power to think more carefully before answering a difficult question, which is a significant improvement over Gemini's performance compared to answering questions based on their first impressions alone.
Google also specifically announced a comparison with GPT-4, the strongest large language model currently in Onpeai, in terms of capabilities in all aspects, and the results showed that in terms of text processing, in addition to the MMLU score of 90%, it exceeded GPT-4's 86In addition to 4%, Gemini Ultra scored higher than GPT-4 in reasoning, math, **, and other abilities.
In terms of multimodality, Gemini also surpasses the capabilities of GPT-4 in all aspects, including image and audio.
According to Jeff Dean, Google's chief scientist and head of artificial intelligence, the Gemini model has reached an astonishing level in terms of multimodal model inference capabilities.
Disclaimer: 1. This number does not make any statements or warranties on the availability, accuracy, timeliness, validity or completeness of any information released, and hereby declares that it does not assume any responsibility or any consequences that may arise from the information.
2. The content of this number is non-commercial and non-profit, and the content of ** does not mean that it agrees with its views and is responsible for its authenticity, nor is it intended to constitute any other guidance. We are not directly or indirectly responsible for any inaccuracies or errors in the information published or any information published.
3. Some of the information, materials, texts, etc. in this number are on the Internet, and all of them have been indicated. If you find that there is a work that infringes your intellectual property rights and personal legitimate rights and interests, please contact us, and we will modify or delete it in time.