Google shows how it made that too-good-to-be-true Gemini AI demo [Gallery]

Ben Schoon | Dec 8 2023 - 7:45 am PT

Google this week launched Gemini, it’s new AI model that powers Bard and will do a whole lot more in the future. And, to show of the capabilities of Gemini, Google produced a truly impressive demo that, really was too good to be true. And, well, it was, but Google is showing off exactly how it was made.

Google’s video demo of Gemini shows real-time video and audio from a user interacting with Gemini, and Gemini responding in real-time. Google is rather up-front with a disclaimer on the video saying that sequences have been shortened, but some have taken issue with there being a little more editing trickery than that. The voice demo in the video is, in fact, not at all used by Gemini, and is rather based on the text prompts that were used with Gemini to make the demo, and even the video was only used in still frames. Google confirmed this in a statement to Bloomberg.

While Google’s demo certainly feels like it was designed to imply Gemini is more impressive than it actually is, a behind-the-scenes blog post that shows how Google created the demo shows that, really, this is still a pretty impressive product.

In the post, Google shows the stills and text prompts that were used to get Gemini’s replies. The first few include showing Gemini recognizing the three separate hand gestures of Rock, Paper, Scissors, only to put that all together in the fourth prompt, and going on to identify a pattern in the user’s gameplay in later stills.

It’s all worded differently from the video, but still shows Gemini’s capability in reasoning.

A further Twitter/X post from Oriol Vinyals, a lead for Google DeepMind, adds that the video was made “to inspire developers” of potential user experiences powered by Gemini.

Really happy to see the interest around our “Hands-on with Gemini” video. In our developer blog yesterday, we broke down how Gemini was used to create it. https://t.co/50gjMkaVc0

We gave Gemini sequences of different modalities — image and text in this case — and had it respond… pic.twitter.com/Beba5M5dHP
— Oriol Vinyals (@OriolVinyalsML) December 7, 2023

Google is already using Gemini Pro to power Bard, as well as local experiences on the Pixel 8 Pro with Gemini Nano. The more powerful Gemini Ultra, meanwhile, arrives next year.