Google Gemini
Google DeepMind’s original intent for their Gemini suite of large language models (LLMs) was to make them multimodal. The integrated suite has a unified user interface (UI) that can handle text, pictures, code, and audio.
The Google Bard LLM, PaLM 2, was superseded by Gemini in December 2023. Google changed the name of Bard to Gemini in February 2024.
Different Subscription Models
The free version of Gemini can be accessed through a web browser for desktop users. The free version is also available to mobile users who download the Gemini app for Android or the Google app for iOS.There is a premium subscription plan that allows users to access a more advanced version of Gemini. The service that provides this model is known as Google One AI Premium. It seems to be a split from the company’s current cloud storage service, Google One, and represents a new membership tier.
Monthly fees for Gemini Advanced’s Google One AI Premium plan are $19.99 as of this writing. It provides a discounted yearly plan in addition to a free two-month trial.
Entire Google Gemini Network
“Gemini will support a complete ecosystem — from the products that billions of people use every day, to the APIs and platforms helping developers and businesses create.” This statement was made by Sundar Pichai, CEO of Google and Alphabet.
Google is rebranding and combining numerous of its other products and services that focus on artificial intelligence to reflect this approach. Gmail, Docs, Sheets, Slides, and Meet will all be able to use Gemini Advanced, while Duet AI will be rebranded as Gemini for Workspace moving forward.
Artificial Intelligence Models from Gemini
Gemini was initially expected to be accessible in four sizes—Gecko, Otter, Bison, and Unicorn—as predicted by Zoubin Ghahramani, VP of Google DeepMind.
Many had hoped that Gecko would be a breeze to carry around in a mobile device’s pocket. It was believed that Otter would do well on many different types of unimodal tasks. Predictions indicated that Bison will perform well on a small subset of multimodal activities. It was anticipated that Unicorn would be well-suited for many multimodal tasks. At this point in time, it appears as though the only sizes of Gemini will be the mobile Nano, the desktop/browser Gemini Pro, and the premium subscriber Gemini Advanced (often dubbed Gemini Ultra).
The Gemini Method
Rumor has it that the Google Pathways architecture is used by Gemini’s AI models. This AI design starts with teaching a set of independent machine learning (ML) models how to do a given task. The modules are linked to create a network after training the network modules can provide a single output type or multiple output types when coordinated. Decoders provide outputs in various modalities depending on the encoded inputs and the work at hand, whereas encoders transform various data types into a common language on the back end.
Methods Used to Train Gemini AI
Supervised learning: Patterns learned from labelled training data were used to train Gemini AI modules to anticipate new data’s outputs.Without labelled examples, the Gemini AI modules were taught to find structures, patterns, or correlations in the data on their own.
With the use of reinforcement learning, the Gemini AI modules were able to refine their decision-making processes via a series of iterative trials, learning to maximize rewards and reduce penalties.According to certain specialists in the field, Google trained the Gemini modules on Cloud TPU v5e chips using reinforcement learning with human feedback (RLHF) extensively. The processing power of the chips utilized to train Chat GPT is five times lower than that of TPUs, asserts Google.
Google has been mum regarding the specific datasets used to train the AI models used by Gemini. But it’s safe to assume that Google’s programmers recycled data from training PaLM 2 and used the LangChain architecture.Web documents, books, code, pictures, audio, and video would have originally taught the Gemini foundation models if this is so. How well this method performs in comparison to training a base model for one mode and progressively adding other modes is an open question. (According to Google, Gemini models are naturally multimodal, and both methods would back this assumption.)
The Origin
Speculation in the media has it that Gemini stands for “Generalized Multimodal Intelligence Network Interface,” but no one has been able to verify this.
Google Bard suggests that the integrated LLM suite was likely named after the constellation Gemini, which derives from the ancient Greek fable of Castor and Pollux, which is the zodiac sign that inspired it. In response to a question, Google Gemini confirmed the idea and said that this is consistent with Google’s practice of naming products after celestial bodies.
