Google’s AI Will Smith Struggles With Crunchy Spaghetti
▼ Summary
– Google launched Veo 3, an AI video model that generates synchronized audio tracks, a first for major AI video generators.
– Veo 3 can produce 8-second HD clips with voices, dialogue, and sound effects, unlike earlier silent and short AI videos.
– The “Will Smith eating spaghetti” benchmark originated in 2023 as a viral example of poor AI video quality, later parodied by Smith himself.
– Veo 3’s attempt at the spaghetti test resulted in a glitch where the AI added crunching sounds, likely due to overrepresented chewing examples in training data.
– AI models like Veo 3 rely on pattern-matching from training data, and imbalances in data can lead to unusual outputs like incorrect sound effects.
Google’s latest AI video generator, Veo 3, marks a significant leap forward by adding synchronized audio to synthetic clips—but its attempt at recreating Will Smith’s infamous spaghetti-eating moment reveals some amusing quirks in the technology.
The newly launched model can produce eight-second high-definition videos complete with voices, dialogue, and sound effects, a first for mainstream AI video tools. Yet when put to the test with one of the internet’s favorite benchmarks—Smith awkwardly chewing spaghetti—the results were unexpectedly crunchy.
This peculiar test stems from a viral 2023 clip generated by an open-source model called ModelScope, which showcased hilariously distorted AI-generated footage of the actor. Though more advanced systems like Runway’s Gen-2 existed at the time, the spaghetti video became a cultural touchstone for tracking AI’s progress in video synthesis. Earlier this year, Smith even poked fun at the meme himself.
Developer Javi Lopez recently fed the prompt “Will Smith eating spaghetti” into Veo 3, sharing the output on social media. While the visuals showed improvement, the audio took an odd turn—the synthetic Smith appeared to be munching on crispy noodles rather than twirling soft pasta.
The glitch likely stems from training data imbalances. AI models rely on patterns in the examples they’re fed, and if certain associations—like chewing sounds paired with crunchy foods—dominate the dataset, the output can skew strangely. In this case, the system defaulted to a crackling sound effect, turning spaghetti into an unintended snack.
Though Veo 3 represents a major step forward, moments like these highlight the ongoing challenges in refining generative AI. Even as the technology advances, unexpected hiccups—like a blockbuster actor crunching on tender pasta—remind us that synthetic media still has room to grow.
(Source: Ars Technica)
