Soon, There Will be a New Sheriff in Town: Introducing Gemini, Google's Multimodal AI

According to brand new information from Google DeepMind CEO in 2016, an artificial intelligence program called AlphaGo from Google DeepMind's lab did the impossible by winning against a champion player at the board game Go. This surprising move showcased the potential of AI, causing many to question its capabilities. Now, Demis Hasabis, the DeepMind co-founder and CEO, is revealing their latest project: Gemini. The engineers at DeepMind are using techniques from AlphaGo to create an AI that aims to dominate the competition, including GPT, on this episode of AI Focus.

What is Gemini?

DeepMind's Gemini was announced this year at Google's developer conference. It is a large language model that is still in development and is anticipated to rival GPT-4, OpenAI's latest large language model. Gemini will focus on tool and API integrations, allowing for wider collaboration, and it will boast improved memory and planning capabilities. It is set to replace Google's current language model, Palm 2, which already powers their chatbot and the do-it AI workspace. However, Gemini is not your average AI model. It is designed to be multimodal, capable of processing images, text, and other kinds of data like graphs and maps. These impressive capabilities will be transferred to every service already powered by Palm 2, providing a more well-rounded and in-depth user experience.

The Power of Multimodal Models

The potential of multimodal models is exemplified in a recent example shared by the Google Research blog. It demonstrates how a model can extract features from a video to create a summary and answer follow-up text questions. With Gemini, not only will it be capable of generating text, but it can also create accompanying images, all in the exact style desired by the user. This amalgamation of capabilities, melding GPT-4 with the convenience of generating personalized visuals, promises to elevate the user experience to new heights. From writing stories to providing images, Gemini has the potential to transform various creative processes.

Four Sizes: Gekko, Otter, Bison, and Unicorn

Gemini will be available in four different sizes, mirroring the release of Palm 2. While the exact specifications of each size have not been specified, Google's CEO did mention that they will be similar to Palm 2's options. Palm 2 is available in four sizes: Gekko, Otter, Bison, and Unicorn. Gekko is the lightweight option, suitable for mobile devices. With Gemini, users will have a range of choices to fit their specific needs, whether it's a compact AI solution or an all-encompassing linguistic powerhouse.

The Secret Weapon: Pairing Generative AI with AlphaGo Techniques

Gemini's secret weapon lies in the integration of generative AI with AlphaGo's sophisticated techniques. DeepMind pioneered the reinforcement learning technique that AlphaGo utilized to master complex games like Go. Reinforcement learning involves software making repeated attempts and receiving feedback to improve its performance, similar to training a dog with rewards and corrections. Additionally, AlphaGo employed a method called tree search to remember and explore moves on the board, further enhancing its strategic abilities.

This combination of generative AI, focused on language capabilities, and AlphaGo's reinforcement learning techniques sets the stage for Gemini to be an AI that can plan and solve problems. DeepMind's CEO, Demis Hasabis, highlighted the significance of this pairing, stating, "At a high level, you can think of Gemini as combining some of the strengths of AlphaGo-type systems with the amazing language capabilities of the large models." This integration has the potential to revolutionize large language models, such as GPT, by infusing them with highly principled algorithms and decision-making abilities based on human feedback.

The Journey to Mastery: Training, Fine-Tuning, and Safety

DeepMind is currently training Gemini, a process that is expected to cost hundreds of millions of dollars. But the cost is not the only hurdle they face. DeepMind believes that training is just the beginning. After the training phase, the AI model will move into the fine-tuning and safety phase, where it will undergo further optimization and ensure ethical and responsible use.

While the exact release date of Gemini has not been specified, DeepMind's focus and dedication suggest that it will be available soon. As the competition in the AI industry intensifies, with companies like OpenAI and others pushing the boundaries of AI capabilities, Gemini's arrival will likely be met with eager anticipation.

The Powerhouses: Google DeepMind and AI Innovation

Google, often regarded as the starting point for some of the brightest AI minds in the world, has joined forces with DeepMind to challenge OpenAI's dominance in the field. DeepMind, known for its powerful algorithms and sophisticated AI creations, caught Google's attention in 2014 after developing software that used reinforcement learning to master video games. Over the years, DeepMind has harnessed this technique to acquire progressively more human-like skills, culminating in the development of AlphaGo. This revolutionary machine defeated human champions in the game of Go, which was considered one of the most complex games to master.

DeepMind's success with AlphaGo underscored the importance of algorithms in machine learning. Contrary to popular belief that machine learning heavily relies on big data and computation power, DeepMind demonstrated that principled algorithms play a crucial role. In fact, AlphaGo Zero, a later iteration of AlphaGo, performed at a much higher level using an order of magnitude less computation by employing more principled algorithms.

Shaping the Future of AI

With Gemini on the horizon, Google DeepMind aims to reshape the AI landscape once again. By combining the strengths of AlphaGo-type systems with the language capabilities of large models, Gemini is set to revolutionize the capabilities of large language models. The integration of reinforcement learning techniques and the ability to learn from human feedback propels Gemini to new levels of problem-solving and creativity.

As the AI industry celebrates Gemini's imminent arrival, it is an exciting time for AI enthusiasts, researchers, and developers alike. Will Gemini change the world as we know it? Only time will tell. Stay tuned for the latest updates on Gemini, and be sure to subscribe to AI Focus to stay informed about all the advancements in the world of AI.