In the whirlwind of Google’s recent Gemini announcements, the tech giant continues to push the boundaries of artificial intelligence. Following the rebranding of its AI bot Bard to Gemini and the introduction of Gemini Advanced, Google has now unveiled Gemini 1.5, marking the next phase in the evolution of its AI model.
Gemini 1.5, according to Google, represents a significant advancement in AI architecture. Built upon a foundation of Transformer and Mixture-of-Experts (MoE) architecture, this new iteration promises enhanced efficiency and performance compared to its predecessors.
Unlike traditional Transformer architectures, which operate as single large neural networks, MoE models like Gemini 1.5 are designed with multiple “expert” neural networks. These expert networks allow for more targeted activation of pathways within the model, resulting in improved efficiency tailored to specific tasks.
The initial release of Gemini 1.5, dubbed Gemini 1.5 Pro, is positioned by Google as the pinnacle of versatility. Comparable to GPT-3.5 in its broad applicability across various tasks and devices, Gemini 1.5 Pro boasts a standard 128,000 token context window, equivalent to that of GPT-4.
However, Google is not resting on its laurels. Testing is currently underway on expanded context windows, with trials reaching up to one million tokens. This expanded capacity enables Gemini 1.5 Pro to handle more extensive prompts, accommodating up to one hour of video, 11 hours of audio, or codebases exceeding 30,000 lines – a substantial leap from previous capabilities.
In a demonstration of its prowess, Google showcased Gemini 1.5 Pro’s ability to comprehend vast amounts of data and provide nuanced insights. Analysing a 402-page transcript from the Apollo 11 mission, the model accurately identified a sketch depicting Neil Armstrong’s historic moonwalk and his iconic quote, showcasing its capacity for contextual understanding and inference.
According to Google, Gemini 1.5 Pro surpasses its predecessor, 1.0 Pro, on 87% of testing benchmarks while rivaling the performance of 1.0 Ultra, the engine driving Gemini Advanced. Moreover, its “in-context learning” capability allows the model to adapt and apply new information without additional prompts, further enhancing its utility and versatility.
For developers eager to explore Gemini 1.5, Google offers access to trials through AI Studio and Vertex AI, with a waitlist available for interested parties. Once widely available, the base version of Gemini 1.5 will be accessible for free, albeit with a 125,000 token window. However, access to the expanded one million token window will likely come at a cost, though pricing details remain undisclosed at present.
As Google continues to refine its AI capabilities with Gemini 1.5, the potential for innovation and advancement in various fields is substantial. Whether it’s deciphering complex data sets or extracting insights from multimedia content, Gemini 1.5 heralds a new era of AI-driven possibilities, propelling us further into the realms of technological discovery and ingenuity.
