Published on GoogleLarge-scale natural language model "Gemini"Since it was made available to Android platform developers on December 12 and will also be used in its own services, related reports have alleged that the video used by Google to show the intelligence of "Gemini" was actually fake, and the actual reaction performance of "Gemini" was made to appear faster than normal through editing.
TechCrunch website saidComparing the training methods on Google's developer blog with the video content Google uses to showcase Gemini's intelligence, it can be found that Google actually uses some editing techniques to make Gemini's overall response speed appear faster.
Oriol Vinyals, vice president of research at Google DeepMind, emphasized that the Gemini demonstration video is primarily intended to let developers understand the technical effects of this model, and hopes to promote the emergence of more software services that use Gemini technology.
However, even though Google states that Gemini will be able to improve its accuracy and speed through continuous learning as more people use it, using editing techniques to reinforce the impression of Gemini's rapid response could lead many people to believe the model has had impressive performance from the outset. While this may be necessary from a marketing perspective, it could pose challenges for developers in their design (and training may still require time).
Google has already begun deploying the "Gemini" model in its services, and has begun using the smaller Gemini Nano on the Pixel 8 Pro, and has opened it up to Android developers for testing. It is also preparing to provide the larger Gemini Pro to developers and businesses, and plans to launch the largest and more complete Gemini Ultra early next year. Perhaps by then we will see the actual appearance of the "Gemini" with better execution efficiency.
