Just this week, reports claimed that Google postponed the launch of itsbig new AI model called Gemini, which is supposed to take on OpenAI’s advanced GPT-4 model. It looks like that was vastly incorrect, as Google has just launched Gemini out of the blue today, right alongside theDecember Feature Drop for its Pixel phones. The AI is available in multiple sizes, with one of them even capable of running locally on theGoogle Pixel 8 Pro.
Inits announcement, Google says that Gemini is its most capable and advanced model yet. In contrast to many other solutions out there, it’s multimodal, meaning that it can work across text, images, audio, video, and code. It can combine, understand, and operate across these information types seamlessly. Gemini can also scale for different environments. It’s available in three sizes (Ultra, Pro, and Nano), which allows it to run on both phones and data centers.
Google is going right for its biggest competitor
Google isn’t shying away from comparing Gemini directly with GPT-4. The company compared Gemini Ultra to Open AI’s latest model in a number of benchmarks, and Google’s product came away as better in seven out of eight text-focused tests. The tests include reasoning, math, and coding abilities. The company also ran multimodal benchmarks, with its AI coming out at the top in all 10 image, video, and audio benchmarks the company used. Google also says that Gemini outperforms human experts on MMLU tasks (massive multitask language understanding), which combines 57 subjects to test world knowledge and problem-solving abilities. These numbers will still have to be put to the test by independent researchers, though.
Google says it achieved these promising results because Gemini is multimodal by design. For example, thanks to Gemini natively working with both images and text, it doesn’t need assistant from OCR systems (object character recognition), which are usually used to read text from images and documents to make them machine-readable. Google also says it trained Gemini on different modalities from the start. The standard approach is to stitch separate modes together after training.
Following ChatGPT’s recent faux pas with itspilling raw training data via a rather simple attack, Google is quick to claim that its AI is built “with responsibility and safety at the core.” The company used an array of techniques to avoid harm, including safety classifiers to avoid violence and stereotypes as well as ensuring factual correctness. Only real-world testing will show how well these measures work, though.
Regarding text, it’s unclear how good Gemini is in languages other than English. In the Independent report that claimed that Gemini was postponed, one of the reasons cited were concerns about poor multilingual performance. Gemini is only rolling out in English for now, which would fit this report.
The company is already rolling out Gemini to Bard and the Pixel 8 Pro
In a surprise move, Google is already rolling out Gemini today. The Pro version of the model is coming toGoogle’s ChatGPT competitor Bard. Google says it’s a “specifically tuned version of Gemini Pro in English for more advanced reasoning, planning, understanding and more.” Next year, Google will introduce Bard Advanced, which “gives you first access to our most advanced models and capabilities.” It’s unclear if the Advanced version will be paid, which would mirror OpenAI’s strategy with ChatGPT. As mentioned, Gemini is only available for the English version of Bard.
Gemini is also coming to the Google Pixel 8 Pro as part of the December Feature Drop for Pixel phones. It uses the Nano variant of the AI, which will power features like Summarize in the Pixel-exclusive Recorder app and a developer preview of Smart Replies in Gboard. The latter feature is coming first to WhatsApp, though Google says it will expand to more communication apps next year.
In the next few months, Google will also make Gemini available for more products, including Search, Ads, Chrome, and Duet AI. The company has also revealed that it’s already started testing Gemini in Search for its Search Generative Experience (SGE). In its tests, it’s been able to reduce latency by 40% in English in the US.
December’s Pixel Feature Drop is here with Gemini Nano, Watch Unlock, and more
The Pixel 8 Pro gets exclusive access to Google’s GPT-4 rival, but there’s plenty of AI to go around
While Gemini Pro and Nano are already rolling out, Google is still optimizing the Ultra version of its AI. The most advanced model still needs more safety-testing, including from industry partners. In the process, the company will open up Gemini Ultra to select partners, with the big rollout coming “early next year” for developers and enterprise customers.
