Just recently, Google made significant headlines by rebranding Google Bard to Gemini, introducing a Gemini app named Gemini Advanced, and unveiling a new premium AI plan. Adding to its recent flurry of announcements, Google has now revealed another advancement in AI technology.
Google introduced its latest cutting-edge model, Gemini 1.5, on Thursday. Despite the recent launch of Gemini 1.0 in December, this new model showcases substantial enhancements over its predecessor, such as an extended context window, improved comprehension, and overall superior performance.
The sophistication of this model prompted Google CEO Sundar Pichai to highlight that the initial release of Gemini 1.5, known as 1.5 Pro, achieves a comparable quality to the previously introduced 1.0 Ultra, the company’s most advanced large language model (LLM), while utilizing less computational resources.
Pichai expressed his optimism about the potential unlocked by longer context windows, stating, “They will enable entirely new capabilities and help developers build much more useful models and applications.”
To achieve this heightened performance, Gemini 1.5 leverages a revamped version of the Mixture-of-Experts (MoE) architecture. This architecture enables the model to learn and selectively activate the most pertinent pathways within its neural network, thereby enhancing efficiency, as stated in the press release.
Google asserts that Gemini 1.5 Pro can handle up to one million tokens in production, a substantial increase from the original 32,000 tokens accommodated by Gemini 1.0. This advancement is significant because the model’s context window, representing the volume of information it can assimilate, is defined by tokens. Consequently, a model’s ability to process more tokens enhances the quality and depth of its responses.
According to Google, 1.5 Pro exhibits the capability to process extensive volumes of information in a single session, encompassing tasks like analyzing up to 1 hour of video, 11 hours of audio, and codebases exceeding 30,000 lines of code or 700,000 words.
During a demonstration, Google showcased how 1.5 Pro efficiently processed a 44-minute silent Buster Keaton film and promptly responded to various questions, including multimodal queries, as demonstrated in the accompanying video.
Furthermore, the model showcased impressive performance in benchmark tests, outperforming 1.0 Pro in 87% of Google’s benchmark evaluations for developing LLMs. Gemini 1.5 Pro excelled in assessments like Needle In A Haystack (NIAH) and Machine Translation from One Book (MTOB), which evaluate the model’s precision and learning capabilities.
To address user concerns regarding the safety of Gemini 1.5 Pro, Google emphasized that extensive evaluations have been conducted to ensure the secure and responsible deployment of this advanced model.
Google is introducing 1.5 Pro with a one million token context window in a restricted preview to developers and enterprise clients through AI Studio and Vertex AI at no charge. Upon the model’s broader release, Google intends to offer 1.5 Pro with pricing options ranging from the standard 128,000 token context window to one million tokens.
