Key Points
- Google announced the launch of its latest, most powerful AI model, Gemini, which has three sizes.
- Executives said Tuesday that Gemini Pro outperformed OpenAI’s GPT-3.5 but dodged questions about how it stacked up against GPT-4.
- The company is planning to license Gemini to customers through Google Cloud for them to use in their own applications. It will also power consumer-facing Google AI apps like the Bard chatbot and Search Generative Experience.
On Wednesday, in response to growing demands that it reveal its plans to monetize artificial intelligence, Google unveiled what it claims to be its most advanced and comprehensive AI model to date.
The three-tiered Gemini language model will be available in three sizes: the ultra-capable Gemini Ultra, the versatile Gemini Pro, and the smaller Gemini Nano, designed for specialized applications and mobile devices.
For the time being, the business intends to license Gemini to clients via Google Cloud so that they can include it into their own apps. With the release of the Gemini API in Google AI Studio and Google Cloud Vertex AI on December 13, developers and enterprise customers will have access to Gemini Pro. Developers working on Android apps will also have the option to use Gemini Nano. Another Google product that will run on Gemini is the Bard chatbot. Another one is Search Generative Experience, which is still in its early stages but aims to answer search questions with conversational-style prose.
It might help businesses find trends to advertise their products on, improve customer care interactions with chatbots and product suggestions, and more. Gemini has a variety of potential uses, including content production for marketing campaigns or blog posts, productivity tools for generating code or meeting summaries, and more.
As an example, the business provided a screenshot of a graphic that Gemini could then use to analyze hundreds of pages of data and update the chart. Finding the right answers and highlighting the wrong ones in a photo of someone’s arithmetic assignment is another example.
A blog post from the company on Wednesday announced that Gemini Ultra had surpassed human experts on MMLU, an assessment of global knowledge and problem-solving abilities that uses 57 different subjects including math, physics, history, law, medicine, and ethics. Theoretically, it is capable of grasping subtleties and reasoning in intricate domains.
In a blog post published on Wednesday, CEO Sundar Pichai stated that Gemini is the product of extensive teamwork throughout Google, including those at Google Research. “It was designed from the beginning to be multimodal, so it can comprehend, process, and merge various forms of data such as text, code, audio, images, and videos. It can also generalize and effortlessly work across these formats.”
Gemini Pro will be utilized by Google’s chatbot Bard to augment its skills, including sophisticated reasoning, planning, comprehension, and more, as of today. The company’s management announced Tuesday on a conference call with reporters that “Bard Advanced,” utilizing Gemini Ultra, will be launched early next year. Bard, its ChatGPT-like chatbot, has never been updated more than this.
This update follows eight months of Bard’s first introduction by the search giant and one year of ChatGPT on GPT-3.5 from OpenAI. The Sam Altman-led startup released GPT-4 in March of this year. On Tuesday, executives said that Gemini Pro performed better than GPT-3.5, but they avoided discussing how it fared against GPT-4.
Google published a white paper on Wednesday stating that GPT-4 was beaten in several benchmarks by Gemini’s Ultra variant.
When asked about potential subscription fees for “Bard Advanced,” Sissie Hsiao, Google’s general manager for Bard, stated that the company is now focused on providing a positive user experience and has not yet shared any information about how it wants to monetize the service.
Google DeepMind VP Eli Collins said, “I suspect it does” when asked during a press briefing whether Gemini possesses any additional skills when compared to current generation LLMs; nevertheless, the company is still trying to determine what exactly Gemini Ultra’s novel capabilities are.
Reports surfaced that Google delayed the release of Gemini due to a lack of readiness, echoing the company’s tumultuous distribution of its AI tools earlier this year.
When questioned why there had been a delay, Collins explained that it takes more time to evaluate the more complicated models. Among Google’s AI models, Collins claims Gemini has “the most comprehensive safety evaluations” and has been through more testing than any other.
According to Collins, Gemini Ultra is far more cost-effective to maintain than its largest variant. “It’s more efficient, it’s more capable,” he remarked. “During the training of Gemini, we still need a lot of computing power, but we’re becoming considerably better at training these models.”
Collins stated that on Wednesday, the business will publish a technical white paper that delves more into the methodology, but that the perimeter count will remain under wraps. The most recent artificial intelligence model from Google, PaLM 2 large language model, utilized nearly five times as much text input for training as its predecessor, LLM, according to a report by CNBC earlier this year.
Google also unveiled its next-gen tensor processing unit (GPU) for AI model training on Wednesday. Salesforce uses the TPU v5p
Google claims that, compared to the TPU v4 announced in 2021, which Lightricks and other startups have started adopting, provides superior value for money. But it didn’t say how they fared in comparison to Nvidia, the industry leader.
Only weeks after Amazon and its cloud computing rivals
Microsoft and its partners featured bespoke silicon designed to combat AI.
Google officials were peppered with investor inquiries over the company’s plans to convert artificial intelligence into tangible profits during the October third-quarter results conference call.
Since search is a significant source of revenue for Google, the search engine introduced a “early experiment” in August named Search Generative see (SGE). This feature allows users to see how a search engine powered by generative AI would look and feel. The outcome reflects the era of chatbots with its increased conversational tone. Nevertheless, it has not yet been released to the public as it is still under the experimental phase.
Ever since the experiment was unveiled at Google’s annual developer conference, Google I/O, in May, investors have been requesting a schedule for SGE. On Wednesday, Gemini said that it would be integrated into SGE “in the next year,” although officials were ambiguous about when the public debut would take place and barely brought up SGE in the statement.
In a blog post published on Wednesday, Pichai stated, “This new era of models represents one of the biggest science and engineering efforts we’ve undertaken as a company.” “I can’t wait to see what the future holds, and the doors that Gemini opens for people all over the world.”
Amanda Byers is a graduate of Columbia, where she played volleyball and annoyed a lot of professors. Now as Zobuz’s entertainment and Lifestyle Editor, she enjoys writing about delicious BBQ, outrageous style trends and all things Buzz worthy.