Google has quite unexpectedly announced Gemini, a comprehensive system of AI models that Mountain View hopes will be able to compete with the current leader GPT-4 and help chatbot developer Bard catch up with OpenAI in the current phase of the generative AI race.
Briefly about Google Gemini
Previous reports about the postponement of the Gemini launch turned out to be false – since yesterday, the technology has been available to users of the Bard chatbot and Google Pixel 8 smartphones. Gemini version 1.0 is the next generation of the artificial intelligence model and, in fact, the further evolutionary development of LaMDA technology in tandem with all the related developments of Google’s star DeepMind division. First of all: Google Gemini was originally designed as a multimodal model capable of analyzing text, images, video, audio, and code. It is available in three versions for different devices and computing power: Gemini Nano, Gemini Pro, and Gemini Ultra.
Gemini Nano is already running on Google Pixel 8 smartphones – it runs locally on the device itself and works without an Internet connection. The Gemini Pro version can already be tested in the Bard chatbot, which supports queries in English in more than 170 countries, including Ukraine. In 2024, it is planned to be implemented directly into Google search. The most powerful Gemini Ultra will be the basis of the Bard Advance chatbot, which is planned to be launched in early 2024. Google assures that Gemini Ultra already outperforms human experts in tests on 57 subjects, including math, physics, history, medicine, law, and others.
Google Gemini vs OpenAI GPT-4. Demo and comparison of features
The first comparisons and tests have already appeared, which allow us to get a first impression of the new model’s capabilities and generally show the superiority of Google Gemini over OpenAI GPT-4 in almost all key metrics. Below are the results of comparative tests of Google Gemini and OpenAI GPT-4 in various benchmarks and the most common scenarios, including math and programming.
As you can see, Google Gemini outperforms OpenAI GPT-4 and demonstrates higher accuracy in all but one of the tests, HellaSwag, which assesses the ability of artificial intelligence to understand the context and logic of real-world events. In other words, a common sense test.
We all remember the embarrassment at the Google Bard presentation, when a chatbot’s mistake in advertising cost the parent holding Alphabet a lot of money – it quickly lost $100 billion in market value. But since then, a lot of water has flowed and Google has significantly improved Bard. Perhaps the most visual and stunning demonstration of Google Gemini’s capabilities is this video, which has already gone viral and is actively spreading on social media.
There are already first impressions on social media – users note the faster query processing speed and better accuracy of Google Gemini compared to OpenAI GPT-4. However, it is important to keep in mind that all these are individual cases and subjective impressions, so these results should be taken accordingly and not be abstracted and pushed to the globe.
And while these demos certainly look convincing, it’s worth remembering that these are just the first tests. And, as always, more independent tests of the hallucination tendency that currently plagues all models without exception, from unbiased parties free of emotional bias, are needed to better understand. And only with time, when more examples emerge, will it be possible to answer whether Google has finally entered the race and Gemini looks as good in practice as it does on paper.
It is anyone’s guess whether Google, OpenAI, or perhaps someone else will eventually create a full-fledged AGI. But we have something to write about, and you have something to discuss.
