Google continues to train Gemini, its next AI model. While continuing to train the model, intended to overshadow ChatGPT 4, the giant allowed a handful of developers to test a very first version of artificial intelligence.
Last May, Google lifted the veil on PaLM 2, a new version of the PaLM linguistic model, intended to animate Bard. In the process, the Mountain View giant mentioned the arrival of an even more sophisticated version of its artificial intelligence, Gemini.
Developed by the Deepmind subsidiary, this language model is presented as the most advanced in the world. Not surprisingly, Google aims to compete with GPT-4, the latest version of GPT, the linguistic model behind the paid version of ChatGPT. For Demis Hassabis, co-founder and CEO of DeepMind, Gemini is largely capable of overwriting GPT-4.
Also read: Google will connect Bard to all its applications and services
Where is Gemini, Google’s super-powerful AI?
According to information gleaned by The Information, Google is progressing by leaps and bounds in the development of Gemini. The search giant has just given access to a first version of AI to a small group of companies. This access, confined to a handful of external developers, indicates that Google ” is on the verge of “ to include Gemini “in its services to consumers and to sell it to businesses”, says the media, citing three anonymous sources close to the case. If the leak is correct, we wouldn’t have to wait long before Gemini is incorporated into Bard, Gmail, Google Docs, Google Maps and even YouTube.
This first version is described as relatively large from a parameters point of view. However, this is not yet the most powerful version of Gemini developed. Google has probably developed several iterations of the model, which will be used according to needs and constraints. This is also the case with PaLM 2. Parameters determine the processing capacity of an AI model. The more parameters supported, the more precise answers a model will be able to provide to its interlocutors. At the moment, it is still unknown how many parameters are supported by Gemini.
Apparently, the company plans to offer Gemini to professionals through Google Cloud Vertex AI, the former Google AI Platform. This machine learning platform, which brings together all of Google’s cloud services, makes it easier to deploy and maintain AI models.
Training in progress
While providing early access to developers, Google continues Gemini training, the article says. Clearly, machine learning engineers feed the model with a corpus of data composed of tokens, which represent a word, a sentence, or a character. The leak does not specify when Google plans to conclude the training period of its artificial intelligence. For the moment, we do not know when Gemini could be made accessible to Internet users.
Obviously, Google is not the only giant to orchestrate a counterattack against ChatGPT. Determined to gain a place of choice in the AI race, Meta is currently working on a new AI model capable of matching, or even surpassing, GPT-4. This ambitious model would be ready during next year. At the same time, the open source world continues to produce a host of alternatives, such as the powerful Falcon 180B.