Google’s DeepMind subsidiary unveiled a new suite of AI technologies called Gemini, adding a multimodal spin to the existing LaMDA and PaLM 2 models. December 6, 2023, marked the debut of this new family of AI engines. Google is not banking on a single model; instead, it’s releasing a spectrum. Gemini’s offerings include Gemini Pro, Gemini Deep Think, Gemini Flash, and Gemini Flash Lite, each catering to distinct use cases.
Gemini exists in three size tiers: Ultra, Pro, and Nano. The lightweight Nano can function on a single device, while the Ultra is designed for data center operations. All are built on the same foundational research, which originated from the merger of Google’s AI division with the London-based DeepMind lab. This common link allows the same underlying research to be applied across the spectrum, from the Nano to the Ultra.
Different sizes and models are built for seamless integration, allowing Google to exert control over hardware, model, and distribution channels such as Google Cloud, Android, and Search. Gemini is multimodal, processing language, images, audio, video, and code natively rather than as an add-on feature.
At the forefront of Google’s strategic deployment is Gemini as a platform for enterprise solutions, powering customer service chatbots and virtual assistants for the next generation of human-machine interaction. Competitors in the field, including OpenAI’s GPT-4 and Anthropic’s Claude, follow different strategies, like selling access via APIs/subscriptions or open-sourcing. With Gemini, Google is wagering on vertical integration and a tiered family of models that can slot into everything from a phone to a cloud server.
The Gemini line stems from competitive and technical pressures. Competitors have set the curve with their AI models, and as researchers at DeepMind have pushed the boundaries of large language models for years, Gemini represents the culmination of that effort. The Nano sizes are suitable for applications like customer service chatbots, which require speed and reliability. Contrarily, the Deep Think is aimed at roles like scientific research assistants, demanding deep reasoning and long context. By segmenting the market before it fully segments itself, Google caters to a variety of specialized deployments.
However, the complexity of managing four model variants across three sizes poses a risk, leading to multiple training pipelines, inference stacks, and safety evaluations. The reward is a precise fit for developers and enterprises, providing precisely the model they require without excess. The success of the Gemini family hinges on execution, with Google looking to make breakthroughs routine across its ecosystem.
Background and why it matters
Google DeepMind began in London in 2010, founded by Demis Hassabis, Shane Legg, and Mustafa Suleyman with a stated mission to build general-purpose artificial intelligence. The founders met at University College London and started by teaching AI to play classic Atari games with no programmed knowledge of how they worked. After repeated trial and error, the systems learned on their own. Google acquired the lab in 2014, and in April 2023 it was merged with Google Brain to form the current Google DeepMind.
In plain terms, Google DeepMind is an artificial intelligence research laboratory and an Alphabet subsidiary headquartered in London. It builds neural network models trained through reinforcement learning, a method in which software improves by acting and receiving feedback. The company introduced neural Turing machines in 2014, then drew worldwide attention in 2016 when its AlphaGo program defeated a Go world champion. A later system, AlphaZero, outperformed leading programs in Go, chess, and shogi after only days of self-play.
DeepMind has also pushed AI into the sciences. In 2020, AlphaFold made a major advance in predicting how proteins fold, a long-standing challenge in biology. Over time, the lab has produced models for game-playing, mathematics, and algorithm discovery. It now develops large language models and other generative AI systems, including tools for turning text into images, video, and music.
This history matters because the Gemini models are a direct product of that long-running research program. DeepMind’s founding goal was a flexible AI that could handle many kinds of tasks, and Gemini is built from the merged work of the London lab and Google Brain. For readers, the competition between DeepMind and its rivals shapes which AI tools reach everyday products such as search engines, cloud services, and mobile phones. The stakes extend beyond convenience: these systems are becoming part of how people get information, solve problems, and interact with software.
