Google’s newly unveiled AI video synthesis model, Veo 3, has introduced a groundbreaking feature: synchronized audio tracks. This marks a significant advancement in AI video generation, as previous models from 2022 to 2024 produced silent, short clips. Veo 3 now generates eight-second, high-definition videos complete with voices, dialog, and sound effects.

Following the launch, the immediate question on many minds was: how well does Veo 3 simulate actor Will smith eating spaghetti?

The “spaghetti benchmark” gained notoriety in March 2023, when an early, somewhat unsettling AI-generated video using the open-source ModelScope was widely shared. The video became a popular example, even prompting Smith himself to create a parody in February 2024.

While ModelScope’s output was memorable, other models like Gen-2 from Runway, which was not yet publicly available, had already demonstrated superior video synthesis.Though, the ModelScope video’s peculiar nature made it a lasting point of reference for comparing AI video generation progress.

Veo 3’s Spaghetti Simulation: A Crunchy Affair

AI app developer Javi Lopez recently put Veo 3 to the test, generating a “Will Smith eating spaghetti” video and sharing the result on X.The video’s audio, however, has an unexpected twist: the faux Smith seems to be crunching rather than slurping the spaghetti.

The faux Smith appears to be crunching on the spaghetti.

This “crunching” phenomenon likely stems from a quirk in Veo 3’s sound effect application. The AI model,trained on vast datasets,may have encountered numerous instances of chewing sounds associated with crunching noises. Generative AI models function by identifying and replicating patterns,so skewed training data can lead to unexpected outcomes.