There's a business built entirely on asking one question: "which AI output is better?"
One startup burned $33 million chasing it and shut down not long after. Another asked the same question, grew revenue 12x in six months, and just raised $7.9 million more.
Everyone picks AI models by benchmark score, right?
On benchmarks like MMLU and GSM8K, frontier models are already near-perfect or beating human experts. But how much does that score actually track how a model feels to use?
Grace Li and her co-founders started this business right before graduating from Harvard in 2025. The team was originally building an AI game engine, and ran into one problem. Technically the games worked fine — people just didn't find them fun. Li called it "the missing bottleneck for a lot of these models to make improvements in the design space." Passing a benchmark and actually pleasing a human turned out to be two completely different problems.
So the team pivoted entirely. Instead of building more games, they built a platform to collect human taste at scale. That became DesignArena. A user types one prompt, several AI models each generate a website, image, presentation, or logo, and people vote on which one's better. Selling that vote data to frontier AI labs is the business model. Li says they closed their first deal a week after launching, and "the rest is kind of history."
So why did Yupp die while DesignArena took off?
"Show people two AI outputs and make them vote" isn't a new idea. A startup called Yupp built the exact same thing. It raised $33 million in a round led by a16z crypto's Chris Dixon, built a platform comparing over 800 models, and pulled in 1.3 million users. And then it shut down.
The reason it failed is the interesting part. Big labs like OpenAI and Anthropic already had their own RLHF pipelines in place, and companies like Scale AI had already locked up the data-labeling market. The enterprise customers Yupp needed already had their own feedback loops. Going broad meant it never became irreplaceable to anyone.
Meanwhile Arena — formerly LMArena, and before that Chatbot Arena, which compares text responses — went the opposite direction. It started in April 2023 as a side project by three UC Berkeley researchers, and by January 2026 it raised $150 million at a $1.7 billion valuation. Its paid "AI Evaluations" service had an annualized run rate north of $30 million at the time, and it's since crossed $100 million. CEO Anastasios Angelopoulos put it this way: "we don't feel like we can responsibly deploy AI if we don't know what value it's providing to humans."
| Yupp | DesignArena · Arena | |
|---|---|---|
| Scope | 800+ models, everything, everywhere | Design-specific / text-specific — a real niche |
| Enterprise fit | Customers already had feedback loops | Direct deals feeding frontier labs data they lacked |
| Funding · outcome | Raised $33M, then shut down | $7.9M seed → 12x ARR in 6mo / $150M Series A |
Same "just ask people" model, wildly different outcomes — and the reason comes down to one thing. Did you sell data that frontier labs actually couldn't get anywhere else, in a lane nobody else could replace you in? DesignArena leans on this hard, tracking how taste shifts by region and time zone — Asian users' web dashboard votes, for instance, skew far more maximalist than European or North American ones.
How to actually use this
Here's the part that doesn't stop at the funding news: DesignArena is free to try right now. A few ways to put it to work instead of just reading about it.
- Get a real read on your AI drafts
Type one prompt into designarena.ai and several models generate a website, image, or presentation at once. If you can't tell whether your output is actually good, put it up for an A/B vote there. - Pick tools by leaderboard, not benchmark score
There are separate real-user vote rankings by category — websites, logos, presentations. Instead of trusting a spec sheet to tell you if a model has "taste," check the actual votes. - Build a tiny A/B habit on your own team
Whenever you have two draft options for a landing page or a piece of copy, make "which one's better" a quick Slack poll or form. It's the same mechanism DesignArena runs at millions of votes — just at team scale. - Check the fine print before you upload anything
Drafts you upload to platforms like this can become taste data sold to frontier labs. If it's client-confidential work, check the terms first.
Worth digging into
The original TechCrunch piece The story that broke this, with the full Grace Li interview. techcrunch.com
DesignArena itself Try a prompt and browse the category leaderboards directly. designarena.ai
Y Combinator profile Founder backgrounds, team size, and batch info. ycombinator.com
Why Yupp shut down A breakdown of what killed the near-identical competitor. techbuzz.ai
Arena's $150M raise How the same model became a unicorn in the text-evaluation market. siliconangle.com
