Google's latest innovation in the field of artificial intelligence made a surprise debut today, despite reports that it would postponed to 2024. This language model, born in the shadow of ChatGPT, presents itself as a breakthrough in the technological landscape. With cutting-edge features and promises of revolutionary integrations, Gemini will certainly be talked about. Here's what Google had to say about the new language model.
Features of Gemini, Google's ambitious response to OpenAI
Gemini comes in three variants: Nano, Pro e Incredibly , each designed for specific purposes. This diversification demonstrates Google's versatility and ambition to make AI accessible to all.
- Gemini Nano is a lighter version, ideal for use offline on Android devices like the Google Pixel 8 Pro
- Gemini Pro, more robust, is intended for enhance Google's AI services, including Bard
- Gemini Ultra, the most powerful of the three, is designed for data centers and business applications
One of the most interesting features of Gemini is its capacity multimodal. Unlike other models that focus only on text, Gemini can understand and interact with video and audio. A bit like started doing ChatGPT a few months ago. This integrated approach promises to revolutionize the way we interact with technology, opening new frontiers in machine learning.
According to the information provided by Google, in addition to its capabilities, the linguistic model shines for efficiency. Trained on Tensor Processing Units from Google, offers superior performance at a reduced cost. With the introduction of TPU v5p, Google aims to further improve efficiency in training and running large-scale models.
Gemini Nano marks Google's debut in the mobile device sector, promising to bring advanced AI features to top Android smartphones, like the Pixel 8 Pro. There won't be a dedicated app to access Nano, but Google intends integrate it into existing functions like summary in Recorder app and smart reply in Gboard for WhatsApp. Thanks to the AICore platform, developers will be able to take advantage of Gemini Nano, paving the way for exciting new applications.
Going up the level, we find Gemini Pro, available for developers and enterprise customers starting December 13 via the new API in Google AI Studio or Google Cloud Vertex AI. Google Bard, a platform for text-based prompts, already leverages Gemini Pro, with plans to extend support to other modes. Currently available in 170 countries, Google plans to expand Bard with the Pro model into more regions and languages.
Moving on to Gemini Ultra, we find ourselves faced with the most exciting model of all. However, we will have to wait a little longer before we can use it, as Google is completing rigorous safety and reliability checks. Ultra will initially be available to a select group of customers, developers and partners for preliminary feedback, before being released to developers and enterprise customers early next year. This model is also expected to be integrated into a new advanced Bard experience.
In terms of performance, Google Gemini has achieved impressive results, surpassing 90% in the MMLU test and beating existing AI models in numerous academic tests. Although Gemini Pro outperforms GPT-3.5, it is in turn outperformed by GPT-4, while Ultra has a marginal advantage over GPT-4. This puts Google's model in a competitive position in the field of AI.
As for security, Google claims that Gemini was developed with a strong emphasis on responsibility and on safety, although open questions remain about its impact on privacy, ethics and employment.
