Fresh Google Gemini, GPT-4 and mathematics

Introduction

So, Google finally decided to release the Gemini language model without waiting for the New Year, and, of course, promising a revolution. It is superior to all publicly available models, and in some places superior to people. Its special feature is multimodality (in particular, the ability to work with images and videos) in almost real-time mode, which is quite impressive demonstrations.

Fortunately, the matter was not limited to just demonstrations; Google connected a new language model to its Bard chatbot, and it You can also try. But there are several nuances. The Gemini Pro version is available there, and all the “miracles on turns” have been announced for Gemini Ultra, we will see it… someday. In addition, all this is available in “170 countries in English (?!)”. Accordingly, the European Union did not enter, and this can only be understood by the fact that the PaLM2 icon appears next to the answers, and not “two iridescent stars” like Gemini (this is the name of the constellation Gemini, by the way).

It must be said that some of the ability of large language models to also solve mathematical problems was a surprise, almost the same as most of their other skills, however. Here is a very illustrative Google gif about how, as the number of parameters increases, the model “grows” a sense of humor, for example:

And if “mastery” of texts and meanings in text models is still somehow expected, then mathematics is right there! But, once it appeared in some form, users immediately began to make demands on this ability so that it could pass the Unified State Exam in mathematics for them, and the like. And the next step, which the developers promised, is precisely the ability of AI to “turn to a specialist.” Microsoft first taught how to access the Internet through Bing GPT-4, and this was a big step (if only the search there was normal…) Further, they promised such features as queries to specialized mathematical engines and the ability to execute real code in a virtual environment.

There will still be many comparisons of Gemini and GPT-4 on the network, there are tables, there are already custom videos from popular YouTubers (Mark Robert, a former NASA engineer who made the world’s largest toy gun, filmed Gemini helping him come up with a script for the video). Well, I just love mathematics.

Comparison

So let’s get started. Testing will take place in Russian. It is known that during the “first steps” of GPT, results in Russian were weaker, obviously because the number of educational texts using Russian was much smaller than the English corpus. But then something happened, very similar to the addition of an additional model for translation, and the “reasoning” seemed to remain in one universal language – English. At least, with the naked eye, the quality of the results has become indistinguishable for GPT. We hope Google follows the same path.

The task came up on one blog platform, for which thanks to the user elicaster. And so that everything would be completely adult, she turned out to be just very multimodal – that is, a rather low-resolution picture.

Microsoft Bing immediately raised its paws and asked that “I would like a task in text” (I don’t know, to be honest, why it has an image recognition function). But Google still showed “multimodality” and recognized the text quite well. True, with the formulas he came up with something like this:

В задаче указано, что интенсивность поступления деталей в первые 30 минут работы цеха растет по закону:

a