How Google's Gemini 1.5 AI model fares against its predecessor
US tech giant Google has unveiled its upgraded AI model, Gemini 1.5, a mere two months after the debut of the original Gemini. This new version is now accessible to developers as well as enterprise users, with a full consumer release anticipated in the near future. Google sees great potential in Gemini as a multifaceted tool for businesses and individuals alike.
Improvements and techniques in Gemini 1.5
The general-purpose Gemini 1.5 Pro is said to rival the high-end Gemini Ultra, and outperforms the Gemini 1.0 Pro in 87% of benchmark tests. This model employs a "Mixture of Experts" (MoE) technique, which only activates part of the model when processing queries. This results in faster and more efficient performance for both users and Google.
Enormous context window for larger queries
Google CEO Sundar Pichai is enthusiastic about Gemini 1.5's enormous context window, capable of handling larger queries and processing more information at once. This window can manage up to 1 million tokens, in comparison to GPT-4's 128,000 tokens and the current Gemini Pro's 32,000 tokens. Pichai simplifies this by stating it's equivalent to "10 or 11 hours of video" or "tens of thousands of lines of code." He also claims the company is testing a 10 million token context window.
Potential business applications and availability
Pichai believes that the bigger context window will be a game-changer for businesses, as it allows for more personal context during queries. He envisions filmmakers using Gemini to forecast reviews and companies analyzing financial records with the AI model. Initially, Gemini 1.5 will be up for grabs by developers and business users, via Google's Vertex AI and AI Studio. Eventually, it will replace Gemini 1.0 as the standard version with a 128,000-token context window.