Generative AI Has a Three-Minute Problem

Something curious is happening in the generative AI space: companies are measuring progress in minutes and seconds rather than quality or usefulness. Google just announced that Lyria 3 Pro can now generate three-minute songs, up from 30 seconds. Meanwhile, OpenAI quietly shut down its Sora video generation app this week, barely months after launch. The timing isn't coincidental—it's symptomatic of a fundamental miscalculation about what users actually want from generative media tools.

The push for longer outputs seems logical on paper. After all, most songs are longer than 30 seconds, and most videos need more than a few clips to tell a story. But this logic ignores a critical reality: duration alone doesn't make generated content useful. A three-minute AI-generated song isn't inherently more valuable than a 30-second one if neither is particularly good, memorable, or suited to a specific purpose. In fact, longer outputs often just mean more time spent sorting through mediocre material.

This duration obsession reveals how generative AI companies are still thinking like engineers solving technical challenges rather than product designers solving user problems. Extending generation length is a measurable engineering achievement—you can benchmark it, publicize it, and use it to demonstrate progress to investors. But it doesn't necessarily address the actual friction points that prevent these tools from becoming indispensable.

Consider why Sora failed to gain traction. The problem likely wasn't that videos were too short, but that the workflow, use cases, and output quality didn't align with what creators actually needed. Similarly, Lyria 3 Pro's three-minute songs might be technically impressive, but musicians and content creators are probably more interested in controllability, style consistency, and integration with existing creative workflows than raw duration.

The market is starting to send signals about this disconnect. Reddit CEO Steve Huffman's announcement about "verifying humanness" and Baltimore's lawsuit against xAI over Grok deepfakes both point to a growing concern: we're drowning in AI-generated content of questionable value. Adding duration to the mix doesn't solve the quality, trust, or utility problems—it potentially exacerbates them.

Interestingly, the companies finding traction in adjacent spaces are focusing on different metrics entirely. Amazon's acquisition of Fauna Robotics and their kid-size humanoid robots isn't about making bigger robots—it's about finding the right form factor for specific tasks. Lucid Bots raised $20 million for window-washing drones not because they made drones that could clean for longer periods, but because they solved a real problem efficiently.

The generative AI industry needs to learn from these examples. Instead of racing to generate longer outputs, companies should be asking harder questions: What duration actually serves the use case? What level of control do users need? How does this integrate into existing workflows? When is AI generation the wrong tool entirely?

Until the industry shifts from celebrating technical milestones to delivering genuine utility, we'll continue seeing impressive demos that fail to become essential tools. Three-minute AI songs might be an engineering achievement, but they're not necessarily a product achievement—and the market is increasingly able to tell the difference.