Google Announces Gemini, the AI It Says Beats GPT-4

Nearly a year after ChatGPT sent Google scrambling, the company unveiled a program built from the ground up to handle text, images, audio and video together. The most powerful version isn't available yet.

Google announced Gemini today, the AI program it has been building since it merged its two research labs in April, and made a claim it has been unable to make all year: that on most standard tests, its program beats OpenAI's GPT-4. The best version, which Google calls Gemini Ultra, will not be available until early next year. A middle version goes into Google's Bard chatbot today, and a small one will run on phones.

What Google is emphasizing is that Gemini was built as a single program that handles words, pictures, sound, and video from the start, rather than a text program with other abilities bolted on. A demonstration video shows it watching a person draw a duck and commenting as the drawing takes shape, spotting a magic trick with cups, and inventing a game from a map. Google has acknowledged the video was edited for length and that the responses were not as fast as shown, which has drawn some criticism.

On one benchmark, a test of knowledge across 57 subjects, Google says Gemini Ultra is the first program to score better than human experts. Whether the benchmarks translate into a better chatbot is something nobody outside Google can check until Ultra ships.

The announcement matters less for the numbers than for the fact of it. Google invented the technology, was caught flat-footed when a smaller company shipped it first, and has spent a year being asked whether it had lost its edge. Today's answer is that it is at least at the front of the pack. The race it announced in February now has two runners it's hard to separate.