Google's highly anticipated Gemini AI model is facing skepticism and criticism after it was revealed that the most impressive demo of Gemini was essentially fabricated. The video titled "Hands-on with Gemini: Interacting with multimodal AI" garnered a million views, showcasing the AI's capabilities in understanding and combining language and visual elements.
The demo featured interactions with Gemini, demonstrating its flexibility and responsiveness to various inputs, such as narrating the evolution of a duck sketch, responding to voice queries about a toy duck, tracking a ball in a cup-switching game, and recognizing shadow puppet gestures. The video created an impression of a highly capable and intuitive AI model.
However, it was later revealed that the video was not authentic. Google admitted to creating the demo by capturing footage to test Gemini's capabilities, then prompting Gemini using still image frames and text prompts. While Gemini may be capable of generating responses shown in the video, it did not perform the showcased interactions live and as implied.
The discrepancies include carefully tuned text prompts with still images, selected and shortened to misrepresent the actual mode of interaction. For instance, the video shows Gemini quickly recognizing hand gestures in a game of Rock, Paper, Scissors, but the documentation reveals that the model needs to be shown all three gestures simultaneously and prompted to respond.
Other examples, like the reordering of sketches of planets and determining the correct order of sticky notes, also deviate from the video's portrayal. The video lacks transparency about the model used, whether it is the current Gemini Pro version or the upcoming Ultra version. Source: Axel, Axel, Axel, Axel, Axel, Axel, Axel, Axel
While some may consider these discrepancies trivial, they raise questions about the trustworthiness of Google's demonstrations. The video, titled "Hands-on with Gemini," implies that the interactions showcased are genuine, leading viewers to believe in the AI's capabilities. Google's attempt to inspire developers may have backfired, eroding trust in the company's claims about its AI models.