Google has made a significant announcement by introducing Gemini, its latest AI model that competes with the GPT-4 from OpenAI. This release coincides with the December Feature Drop for Google Pixel smartphones. Gemini stands out as a multimodal AI model, showcasing proficiency across text, images, audio, video, and code. The model comes in three sizes – Ultra, Pro, and Nano, enabling it to function effectively on both mobile devices and data centers. Google asserts that Gemini represents a leap forward in their AI technology.
A comparative analysis between Gemini and GPT-4 shows that Gemini Ultra excels in seven out of eight text-centric evaluations over OpenAI’s offering. These evaluations encompass reasoning, mathematics, and coding capabilities. Moreover, Gemini has demonstrated superiority over human experts in MMLU tasks that evaluate problem-solving skills and general knowledge proficiency. Furthermore, in multimodal assessments, Gemini outperformed in all 10 image, video, and audio benchmarks. It is worth noting that independent validation of these results remains pending.
Google attributes Gemini’s remarkable performance to its multimodal architecture. By efficiently handling images and text without the need for external Optical Character Recognition (OCR) systems, Gemini streamlines processes. Unlike conventional methods that stitch together different modes after training, Google trained Gemini across modalities from the outset.
Addressing concerns regarding AI safety and accountability, Google underscores Gemini’s design principles, emphasizing safety classifiers that mitigate violence, stereotypes, and ensure factual accuracy. Nonetheless, the efficacy of these measures will require real-world validation.
Regarding multilingual capabilities, there is uncertainty about Gemini’s performance in languages beyond English. Some reports indicate limitations in multilingual proficiency. Presently, Gemini is confined to English, as affirmed by reports. However, Google may broaden language support for Gemini in the future.
Unexpectedly, Google has initiated the rollout of Gemini, with the Pro variant being integrated into their ChatGPT counterpart, Bard. This specialized version is tailored for enhanced reasoning, planning, and comprehension. Google envisions introducing Bard Advanced, providing users access to premium models and features. Details on the availability and pricing of the Advanced version are pending. Gemini is also being embedded in the Google Pixel 8 Pro as part of the December Feature Drop. The Nano iteration of Gemini will power functionalities such as Summarize in the exclusive Recorder app and a developer preview of Smart Replies in Gboard. Initially, Smart Replies will be accessible in WhatsApp before expanding to other messaging platforms.
Google has outlined plans to extend Gemini’s integration to other products like Search, Ads, Chrome, and Duet AI. The company has commenced Gemini testing within its Search Generative Experience (SGE), resulting in a 40% latency reduction for US English searches. Android developers can enroll for a preview of Gemini Nano to incorporate the AI into their applications. This preview, part of Google’s AICore app, will extend to a broader range of devices in the following months. The AI functionalities will leverage machine learning capabilities present in processors from Qualcomm, Samsung S.LSI, MediaTek, and Google.
While Gemini Pro and Nano are already accessible, Google continues to fine-tune the Ultra version of the AI. Comprehensive safety evaluations with input from industry collaborators are required for this version. Google aims to offer Gemini Ultra to select partners initially, with plans to expand availability to developers and enterprise clients at the start of the next year. With the release of Gemini, Google asserts its commitment to directly competing with OpenAI’s GPT-4, presenting a sophisticated multimodal AI model that has demonstrated promising performance across various assessments.
