Just one week after the announcement of Gemini Ultra 1.0Google surprised everyone with the announcement of Gemini Pro 1.5.
If the first AI language model, active since last Thursday, has demonstrated enormous potential, its evolution is certainly no exception. For Google, in fact, Gemini 1.5 will be able to work like its predecessor, but using much fewer resources for data processing.
This was possible thanks to the use of an architecture Mixture-of-Experts (MoE) which sees multiple AI models working simultaneously. The MoE philosophy, beyond the purely performance side, also allows AI to be trained more easily.
The rapid launch, however, from many points of view makes little sense. Ultra, in fact, has been incorporated into the service (for a fee) Google One, capturing customers’ attention. The arrival of a more advanced model, however, has certainly made subscribing to this platform less attractive, even from a simple user perception.
Despite a risky commercial move, Gemini 1.5 represents a new generation of LLM which can process up to 1 million tokens, with Google detailing how this allows “Achieving the longest context window of any large-scale foundation model so far“.
Gemini 1.5 is based on a MoE architecture to optimize its performance
When we talk about tokens, we are referring to fragments of a word. This, therefore, means that the new AI model is able to deal with even very long texts, offering relevant and comprehensive answers. To understand why the Mountain View giant focuses so much on token management, just understand that GPT-4 Turbo Of OpenAIat the moment, can handle it 128.000.
One million tokens, according to a rough estimate, could correspond to approximately 700.000 parole. During the presentation, Google stated how its model can perform “Complex reasoning about large amounts of information” giving an example of analysis relating to a 402 page transcript of the Apollo 11 mission to the Moon.
As you might guess, Gemini 1.5 is not currently available to the public, being offered free of charge only to developers and corporate customers through AI Studio e Vertex AI by Google.
The company also warned testers of possible problems such as prolonged latency times due to the embryonic nature of the AI model. There is currently no known public launch window.