US search giant Google LLC has officially launched Gemini, its latest artificial intelligence (AI) model that is expected to surpass the capabilities of OpenAI Inc.’s GPT-4.
Developed by the company’s DeepMind Technologies Ltd., the new large language model (LLM) has been improved to work with various sizes and use cases.
Gemini’s suite of multimodal tools is a key feature that makes it unique from earlier models.
Instead of text, the technology trained on images, video, and audio. Its multimodality allows for comprehending and processing a range of data types as inputs and outputs.
It is also built to perform specialized coding and understand complicated physics problems, which it can turn into mathematical formulas. Such advanced problem-solving skills have not been seen in traditional LLMs.
In addition to Gemini 1.0, the technology is offered in three variants, the Ultra, Pro, and Nano, with the former demonstrating an understanding more advanced than a professional human, according to Google.
In a multimodal language understanding (MMLU) test, the tech firm stated that Gemini Ultra logged a 90% accuracy, beating human expert performance.
Google chief scientist Jeff Dean said the AI model’s Ultra version was the first to achieve such a performance on MMLU across 57 subjects.
The search titan currently offers users access to the fined-tuned type of Gemini Pro via its chatbot Bard. The Ultra is set to be released in 2024 with Google’s upgraded AI assistant, Bard Advanced.
The company eventually looks to have Gemini available in more than 170 languages and use the LLM to aid its Pixel phone lineup and Search Generative Experience (SGE).
Gemini’s Potential to Rival OpenAI’s GPT-4
Gemini represents proof of Google’s continued commitment to being an AI-first company.
The firm has accelerated the launch of its AI technology and allocated resources to a few new AI efforts to try to reduce the noise around ChatGPT and reinvent itself as the world’s leading AI company.
Companies involved in AI are sort of in a “tit-for-tat arms race,” according to former chief executive Oren Etzioni. While there is no reason to question Gemini’s ability to outperform GPT-4 on certain benchmarks, Etzioni believed GPT-5 may perform better.
Google’s AI capabilities have shown improvement, and their flexibility may open the doors to new applications across various domains.
Still, Professor Alexei Efros said Gemini, like GPT-4, may still show limitations in understanding practical complexities, expressing his uncertainty over gaining the information they need about Gemini as they do not know what is inside such proprietary models.
For now, more real-world assessments need to be conducted to determine the actual performance levels of Google’s latest creation.