Gemini

Gemini is the name of a series of computer programs from Google that can process language. You type or speak a question, and the program answers in full sentences. Such programs are called language models: they have learned from enormous amounts of text which words typically follow one another. Gemini can not only read text, but also process images, audio recordings, and videos. It is developed by Google DeepMind, the company’s research division for artificial intelligence. The first Gemini models were introduced in late 2023, and new versions have appeared regularly ever since.

Google’s Answer to ChatGPT

Until 2022, Google was practically without competition in the field of search. Then the company OpenAI released the chatbot ChatGPT, and suddenly millions of people were asking their questions to a chat program instead of a search engine. For Google, this was a serious threat, since the company makes most of its money from advertising alongside search results. Gemini is the direct response to that.

That is why Gemini is closely watched on the stock market. When Google unveils a new version, it’s not just about technology, but about market share. Analysts regularly compare Gemini’s results with those of GPT from OpenAI and Claude from Anthropic. These comparisons run through standardized test tasks, so-called benchmarks, such as math problems or programming tasks.

Reach is also important. Google can build Gemini directly into products that billions of people already use: Search, Gmail, Android, YouTube. No competitor has this advantage on this scale. Whether this actually translates into profit remains uncertain, since operating such models costs a great deal of electricity and computing power.

Built for Images and Audio from the Start

Many older AI systems could only handle text. If they needed to understand an image, a second program was placed in front of them to translate the image into a text description. Gemini was trained from the outset on text, images, audio, and video simultaneously. Such systems are called multimodal. In practice, this means you can upload a photo of your homework and ask a question about it right away.

Technically, Gemini is based on an architecture called the Transformer, which was developed in 2017 partly by Google researchers. The model breaks down every input into small building blocks and calculates which blocks are important to one another. It then predicts, piece by piece, what should come next. During training, billions of such predictions are made and each one is slightly corrected each time.

Gemini comes in several sizes, because different tasks require different amounts of computing power. Small variants with names like Flash or Nano are fast and cheap, and some even run directly on a phone. Large variants like Pro or Ultra are more powerful, but slower and more expensive. One feature Google particularly promotes is the large context window: Gemini can process very long texts at once, up to several hundred pages in top-tier versions.

Gemini in Everyday Life and in the Headlines

Most often, people encounter Gemini without realizing it. In Google Search, summarized answers now often appear at the top instead of plain links; these are written by Gemini. In Gmail, it suggests replies; in Google Docs, it rephrases text. On many Android phones, it has replaced the old Google Assistant. There is also a dedicated app and the website gemini.google.com.

In business news, the name usually comes up in three contexts: new versions, comparisons with OpenAI, and mishaps. In 2024, Gemini drew criticism because its image generation produced historically inaccurate depictions, prompting Google to temporarily disable the feature. Such errors point to a fundamental problem: language models sometimes invent convincing-sounding false information. Experts call this hallucination.

A common misconception, by the way, is that Gemini is a search engine. It is not. It does not search the internet, but generates text from what it has learned, even though it can draw on search results when needed. There is also a risk of confusion with the American cryptocurrency exchange Gemini founded by the Winklevoss twins, which has nothing to do with Google’s AI.