Google Unveils Its Biggest and Most Advanced AI Model Gemini

Google Unveils Its Biggest and Most Advanced AI Model Gemini:

Google is set to introduce its most expansive artificial intelligence model, addressing growing queries about its AI monetization strategy. The new model, named Gemini, encompasses three distinct tiers: Gemini Ultra, the largest and most advanced; Gemini Pro, designed for versatile task applications; and Gemini Nano, tailored for specific tasks and mobile use.

Initially, Google intends to offer Gemini licenses to customers through its Google Cloud platform for integration into their respective applications. From December 13 onwards, developers and enterprise users can leverage Gemini Pro via the Gemini API accessible through Google AI Studio or Google Cloud Vertex AI. Moreover, Android developers will have access to Gemini Nano for mobile application development. Google plans to harness Gemini’s capabilities within its own products, such as the Bard chatbot and the Search Generative Experience (SGE), which employs conversational text to respond to search inquiries (SGE’s widespread availability is pending).

Enterprises can leverage Gemini for elevated customer service interactions through chatbots and personalized product suggestions. Additionally, it facilitates trend identification for advertising purposes. Content creation, including marketing campaigns and blog content, is another application, along with productivity tools for meeting summarization or code generation.

Google showcased Gemini’s capabilities, illustrating its ability to capture a chart screenshot, analyze extensive research material, and update the chart accordingly. Another instance highlighted Gemini’s prowess in scrutinizing math homework photos, identifying correct answers, and flagging errors.

In a recent blog post, the company emphasized Gemini Ultra’s groundbreaking performance, surpassing human experts in Massive Multitask Language Understanding (MMLU). This achievement spans 57 subjects encompassing fields like mathematics, physics, history, law, medicine, and ethics, assessing both global knowledge and problem-solving. Google claims Gemini Ultra excels in understanding intricate nuances and reasoning within complex subjects.

In a recent blog post, CEO Sundar Pichai emphasized Gemini’s collaborative development across various Google teams, including Google Research. Pichai highlighted Gemini’s multimodal nature, designed from scratch to seamlessly comprehend and integrate diverse information types like text, code, audio, images, and video.

Google’s chatbot Bard is set to integrate Gemini Pro for enhanced reasoning, planning, and understanding, starting today. An upcoming version, “Bard Advanced,” slated for early next year, will harness the power of Gemini Ultra, marking a significant upgrade akin to ChatGPT but under Google’s umbrella.

This update arrived eight months following the initial Bard launch by the search giant and a year after OpenAI introduced ChatGPT based on GPT-3.5. While OpenAI rolled out GPT-4 in March of this year, Google’s executives declined direct comparisons between Gemini Pro and GPT-3.5, remaining tight-lipped about Gemini’s performance against GPT-4.

However, according to a white paper released by Google on Wednesday, Gemini’s Ultra model showcased superior performance over GPT-4 in certain benchmarks, signaling a notable achievement for the technology.

When questioned about potential fees for accessing “Bard Advanced,” Sissie Hsiao, Google’s general manager for Bard, emphasized their focus on creating a quality user experience and mentioned the absence of finalized details regarding monetization plans.

During a press briefing, Eli Collins, vice president of products at Google DeepMind, acknowledged the likelihood of Gemini possessing unique capabilities compared to current-generation Large Language Models (LLMs). However, he indicated that understanding Gemini Ultra’s distinct capabilities is an ongoing process.

Reports suggested Google delayed Gemini’s launch due to readiness concerns, echoing past experiences of the company’s turbulent AI tool rollouts earlier in the year.

Addressing the delay queries, Collins explained that the extended testing duration is inherent when assessing more advanced models. He highlighted Gemini as the most rigorously tested AI model developed by the company, boasting “the most comprehensive safety evaluations” among Google’s models.

Despite its extensive scale, Collins highlighted Gemini Ultra’s cost-effectiveness, stating that it not only delivers enhanced capabilities but also operates more efficiently. He emphasized advancements in computational efficiency during the model training phase, enabling Google to achieve higher efficiency in training such models.

Collins mentioned that the company plans to release a technical white paper detailing the model but indicated they wouldn’t disclose the model’s perimeter count. Earlier this year, CNBC’s investigation revealed that Google’s PaLM 2 large language model, the company’s most recent model at the time, utilized nearly five times the text data for training compared to its predecessor LLM.

Additionally, on the same day, Google unveiled its next-generation tensor processing unit (TPU) for training AI models. The TPU v5p chip, already adopted by Salesforce and startup Lightricks, boasts improved performance for its cost compared to the previously announced TPU v4 in 2021. However, Google refrained from providing performance details in comparison to market leader Nvidia.

This chip launch follows recent moves by cloud competitors Amazon and Microsoft, who showcased their own customized silicon geared toward AI applications.

During Google’s third-quarter earnings conference call in October, investors showed increased interest in understanding how the company plans to translate AI advancements into tangible profits. This highlights growing investor curiosity regarding Google’s AI monetization strategies.

In August, Google initiated an “early experiment” named Search Generative Experience (SGE), offering users a glimpse into a generative AI-driven search experience. Given the significance of search as a major revenue source for Google, SGE aims to introduce a more conversational approach, aligning with the era of chatbots. Despite this, SGE remains in its experimental phase and hasn’t yet been released to the wider public.

Investors have sought a timeline for SGE since its announcement at Google I/O, the annual developer conference in May. However, during the recent Gemini announcement, details regarding SGE were notably sparse. Executives remained vague about the public launch plans, mentioning an integration of Gemini into SGE “in the next year.”

In Wednesday’s blog post, CEO Sundar Pichai highlighted the monumental nature of this new model era, describing it as one of the most substantial science and engineering endeavors undertaken by the company. Pichai expressed genuine enthusiasm for the future and the opportunities that Gemini is poised to unlock for individuals worldwide.