Intelligence Secures $7.9M to Solve AI's Taste Problem

The Game That Wasn't Fun

Grace Li and her college friends faced a peculiar engineering challenge in early 2025: their AI game engine could generate playable games, but every output felt hollow. The mechanics worked. The graphics rendered. Players could move and interact. Yet something fundamental was missing, a quality that resisted algorithmic measurement.

The problem wasn't technical capability. It was taste. And taste, they realized, couldn't be automated away.

That insight led to Design Arena, a platform now used by 5.3 million people to rank AI-generated visual content. On Monday, Intelligence, the company behind the tool, announced a $7.9 million seed round led by Index Ventures, with participation from Conviction, A*, Valkyrie, and others. The round arrives as the AI industry grapples with a bottleneck that pure compute can't solve: understanding what humans actually want from generative outputs.

Ranking Engines Instead of Building Them

For consumer users, Design Arena functions as a sophisticated model router wrapped in a ranking interface. Users enter prompts into a ChatGPT-style window, select their desired format (websites, images, or roughly a dozen other visual categories), and specify a style preference. The platform then presents a series of head-to-head comparisons, forcing users to choose between outputs until a clear hierarchy emerges.

The interface is simple. The data it generates is not.

Behind each ranking session sits a stream of preference signals that frontier labs are willing to pay for. According to Li, Intelligence is currently generating $60 million in annual recurring revenue by selling access to this human-led evaluation data. The business model turns every user choice into training signal, every aesthetic judgment into a data point that helps models understand the gap between functional and desirable.

Users don't particularly care which models they're ranking, Li notes. They want the best output, regardless of provenance. That indifference is precisely what makes their feedback valuable. Unlike traditional user testing, where participants know they're evaluating a specific product, Design Arena users are optimizing for their own satisfaction. The result is preference data that's harder to game and more reflective of real-world aesthetic judgment.

Geography of Taste

Intelligence tracks more than binary preferences. Because users must log in to receive outputs, the platform can map how aesthetic tastes shift across regions and evolve over time. Li points to one pattern that emerged from the data: web dashboards designed for Asian markets tend toward maximalist layouts, dense with information and visual hierarchy. Western preferences skew cleaner, more minimal.

These geographic signatures matter for models aiming to serve global markets. A design that tests well in San Francisco may feel sparse and incomplete in Seoul. A layout optimized for Mumbai might overwhelm users in Stockholm. At DailyTechWire, we've tracked similar localization challenges across consumer apps, but Intelligence is productizing that insight at model training scale.

The temporal dimension is equally revealing. As users interact with more AI-generated content, their preferences shift. Early adopters tolerate rougher edges. Mainstream users expect polish. Tracking these changes in real time gives model developers a moving target, one that automated benchmarks struggle to capture.

The Benchmark Problem

Automated evaluation has dominated AI development for good reason: it scales. You can run a model against thousands of test cases in hours, generating clean metrics that compress model performance into a single number. But those benchmarks are increasingly vulnerable to gaming, as demonstrated by a recent breach at Hugging Face that exposed how easily leaderboard positions can be manipulated.

Human evaluation doesn't scale as easily, but it's harder to fake. A person comparing two logo designs or two website mockups brings contextual judgment that resists simple optimization. They notice when composition feels off, when color choices clash, when a layout serves the wrong audience. That holistic assessment is what Design Arena monetizes.

The challenge is building a sustainable business around it. Human feedback platforms face a fundamental tension: they need enough volume to be useful to model developers, but each evaluation costs time and attention from real users. That balance has proven difficult to strike.

The Cautionary Tale

Yupp attempted a similar model and raised $33 million from a16z crypto before shutting down earlier this year. The platform also secured frontier model customers and claimed over 1.3 million users, but couldn't translate that traction into a viable long-term business. The failure suggests that user volume alone isn't sufficient. The feedback has to be valuable enough that labs will pay recurring revenue for it, and the user experience has to be compelling enough to maintain engagement without burning through acquisition budgets.

Intelligence appears to have threaded that needle, at least for now. The $60 million ARR figure, if accurate, implies the company has locked in substantial enterprise contracts. That's a different trajectory than Yupp's, which struggled to convert attention into revenue before running out of runway.

Other players in the space are also finding traction. LM Arena, which applies a similar ranking methodology to text-based model outputs, raised $150 million in a Series A just four months after launching its paid product. The round, announced in January, suggests investors see human-in-the-loop evaluation as a durable category, not a temporary arbitrage.

The Week After Graduation

Li traces Intelligence's origin to a specific moment: roughly a week after closing their first major deal with a frontier lab. That compressed timeline, from college project to enterprise contract, reflects how acute the taste problem has become for AI companies. Models can generate endless variations. They can optimize for technical metrics. But they can't yet tell you which design will resonate with a human audience.

That gap represents both a market opportunity and a broader question about the trajectory of AI development. If taste remains stubbornly resistant to automation, then human-in-the-loop systems like Design Arena become permanent infrastructure rather than temporary scaffolding. Models get better at rendering, at composition, at technical execution. But the final judgment, the decision that separates good from great, stays human.

For now, Intelligence is betting that judgment is worth $7.9 million in seed capital and the attention of Index Ventures. The company's ability to convert 5.3 million users into $60 million in ARR suggests the bet has merit. Whether that model sustains as generative AI matures, and as automated evaluation improves, will determine if Design Arena remains essential infrastructure or becomes a footnote in the industry's search for scalable taste.