Testing Veo 3: Convincing Video, Generated Audio and Physics Failures

Matthew Berman opens with a playful generated imitation and asks collaborator Alex not to replace him. He introduces Google’s Veo 3, released in May 2025 with native sound and dialogue generation. The clips in this video are examples, not a controlled measure of all model outputs.

Berman begins with generated street interviews featuring people leaving a bar. He finds the scenes persuasive, but says the dialogue does not match every part of his prompt. Two samples respond to the idea that their characters are not real with overlapping dialogue.

He tries to generate two trains racing toward each other on the same track and colliding. The output does not follow the requested setup: in one example the trains are on separate tracks and explode without the specified collision, despite convincing-looking sound and effects.

A Rubik’s Cube prompt produces attractive surfaces and sounds, but the movements are not physically convincing. In another generation a piece appears to change or detach, and a more detailed request still fails to show a consistent solution.

Berman observes odd internal gears and a frame change in later cube attempts. He then shows a generated version of a familiar meme. Visual appeal alone, he notes through these examples, does not mean the system carried out the requested action.

The video includes a paid Hostinger Horizons promotion. Its claims about turning prompts into deployed web apps and the first-month discount are sponsor claims and a historical offer, not an independently verified assessment of production readiness.

Berman shows a short 1980s-style scene by creator Javi Lopez, who says it took less than two hours using Veo 3, Magnific for upscaling, music tools and CapCut for editing. Berman likes the atmosphere but notices that a book disappears; the time estimate is the creator’s report and the scene uses multiple tools.

Alex introduces his own shrimp fried rice clip, where the character looks more like a lobster and another version appears to toss a wok without touching it. He prefers the first generation, illustrating variation across samples.

For a horror prompt about aliens in a dark alley, Alex likes one clip’s lighting and finds the sound effects stronger in the second. He says he accidentally generated matching prompts with Veo 2 first and prefers the selected Veo 3 results; this is a small anecdotal comparison.

He asks for a cowboy riding a motorcycle-horse hybrid through a ring of fire. Some outputs convey the idea, but others produce an odd ring, a motorcycle with only a horse tail, or a jump without a clear physical cause.

Berman discusses Google Flow as a tool for arranging and extending generated clips into scenes. Google described Flow as combining Veo, Imagen and Gemini; assembling longer stories remains a filmmaking process rather than evidence that a whole movie emerged in one generation.

A Jurassic Park-style travel video shows visitors and dinosaurs. Berman likes the people but finds some dinosaurs more like animatronics than realistic animals; another feathered creature looks better to him. Dinosaur appearance varies by species, so this scene is not evidence for a single look across all dinosaurs.

Alex tries a T. rex chasing tourists. In one sample the people run toward the animal and it appears to pass through them; other samples show more coherent movement, dramatic camera work or gates opening. One unusually tall person is another visible inconsistency.

Berman shows Ethan Mollick’s fictional product-review examples and jokes about his own job. The clips imitate familiar YouTube presentation styles and review imaginary products, but he hears speech pauses that feel slightly unnatural.

Berman says he showed the AI-generated opening imitation to his wife without explaining it and she initially thought someone had copied the channel. This is one anecdote about apparent realism, not evidence that generated media is generally undetectable.

The fictional reviews continue with implausible products and a flying experience. Berman also plays an example of generated singing and captions that appear to match the words, while noting the overall performance is not perfect.

A clip depicts Darth Vader, prompting Berman to comment on copyright. The appearance of a recognizable character does not establish that a use is authorized or resolve copyright questions; he then plays additional fictional street interviews with varied accents.

A boy-band prompt produces catchy music and dancing, but a performer suddenly appears in the background. Berman says it is impressive while remaining far from perfect, and shows a second variation.

He tests physical sequencing with a Rube Goldberg machine. Several elaborate scenes have the look of marbles, dominoes and gears but lack a continuous cause-and-effect chain; in one, motion appears to go in an unexpected direction.

One later Rube Goldberg attempt is his favorite: a marble travels down a wooden slide and a box opens. He still notices color changes. This selected example illustrates a better-looking sequence, not verified physical simulation.

Alex asks for a time-lapse of New York City being built from LEGO. He says bricks appear or fly into place rather than being assembled naturally; the second result is less convincing to him.

A prompt for a visual representation of a fourth dimension yields abstract portals and effects rather than a clear explanation. Alex likes the visual effects but acknowledges the output largely skips the requested concept.

A hamburger assembly prompt produces appetizing ingredients, but pieces appear held in place and the sequence does not finish. A second version looks better yet still omits the top bun and shifts scenes.

Alex tests a cel-shaded realistic person and likes the style, then shows a generated stand-up joke that made him laugh. These short samples demonstrate aesthetic possibilities without establishing consistent performance across prompts.

A solar-system prompt generates striking views, but Alex notices a planet appearing abruptly in front of a moving camera. He considers the result visually appealing despite the continuity problem.

Berman asks for a photorealistic Portal-inspired scene using an expanded prompt. The setting and futuristic portal gun appear, but the character does not travel through as requested and the expected relationship between portals is not convincingly shown.

A second portal clip has impressive visual detail but only briefly suggests a reflection. Berman distinguishes the attractive look from adherence to the actual instruction.

He samples Meta Puppet’s satirical film This Is Plastic, assembled using Flow and Veo 3. The excerpt depicts a fictional plastic baby and a later commercial; its narrative is satire, not evidence for medical claims mentioned by fictional characters.

Berman then shows excerpts of an edited alien-invasion film with fictional broadcasts and personal reactions. He is struck by the integration of multiple situations and social-media conventions; the film is presented as a curated longer creation.

A fictional video call and social-media livestream include visual markers such as viewer counts and floating reactions. Berman praises the familiar interface details, but the scenes and interface are generated fiction rather than real footage of an emergency.

He highlights the interaction of objects and especially the addition of audio as a major strength, then closes with another highly detailed generated clip. His overall impression is enthusiastic, while the earlier failed tests show the limits of these particular samples.