The announcement of Gemini — the "killer" GPT-4 was released on the Google DeepMind website. According to the CEO of the company, the new neural network is "the beginning of a new era of AI at Google".

Gemini - Google DeepMind
Gemini is built from the ground up for multimodality — reasoning seamlessly across image, video, audio, and code.

Brief introduction

In fact, Gemini consists of several neural networks at once. They differ in size. For example, the smallest Nano version is able to work offline directly on your smartphone. Larger versions will be integrated into all the company's services, including the Bard chatbot.

If we look at the hard numbers, then Gemini is comparable or superior to GPT-4 on most test benchmarks. In the image and video mode, Google's neural network is significantly better than its competitor from OpenAI. However, the latter is not very surprising given the company's expertise in these areas.

As mentioned in one of our previous articles, the powerful AI developments of large IT companies are closed for the purpose of earning money. That's what Google did, which promises to add its neural network to Google Cloud for paid use by developers.

Why is Gemini not a revolution?

Firstly, in the official press release, Google has compared Ultra, the biggest version of its family of neural networks, with GPT-4, a model that OpenAI completed about a year ago. Of course, in terms of metrics, Gemini Ultra outperforms GPT-4 in most tasks. But if we take Pro, the second-largest model, into account, we see a big gap between OpenAI's one-year-old neural network and Google's recent Gemini "revolution". It’s probably worth keeping silent about the quality of the Nano version.

Gemini Ultra and Pro performance on text benchmarks from the official report

To be fair, Google's models have impressive performance. But I would venture to guess that GPT-5, which will likely be released next year, will significantly outperform the new Gemini.

Google’s best Gemini demo was faked | TechCrunch
Google’s new Gemini AI model is getting a mixed reception after its big debut yesterday, but users may have less confidence in the company’s tech or

Also, more and more articles began to appear on the Internet in which the GPT "killer" was destroyed. For example, it was confirmed that demonstration videos, where the neural network works beautifully and quickly with pictures and videos, were edited. However, this is not at all surprising: cherry picking has long dominated this industry.

💡
Cherry picking is a logical fallacy that consists of pointing to isolated cases that support a certain point, while ignoring a significant portion of related cases that may contradict that point.

Trained by the experience of reading works with beautiful titles and incredible demos (specially selected), I was somewhat skeptical about Gemini and did not sing praises. And, apparently, for good reason. Did Google understand that it would be discovered? Yes I guess. But was the game worth the candle? Let's see.

Share this post