This is really good! It would be really cool to somehow get human designs in the mix to see how the models compare. I bet there are curated design datasets with descriptions that you could pass to each of the models and then run voting as a "bonus" question (comparing the human and AI generated versions) after the normal genAI voting round.
This would be extra interesting for unique designs - something more experimental, new. As as for now even when you ask AI to break all rules it still outputs standard BS.
How about adding "mobile"? A lot of the time models tend to default to designs that don't make sense on mobile, even when instructed to design it as such.
I tried the vote and both results always suck, there's no option to say neither are winners. Also it seems from the network tab you're sending 4 (or 5?) requests but only displaying the first two that respond, which biases it to the small models that respond more quickly which usually results in showing two bad results
Yes — great point. We originally waited for all model responses and randomized the vote order, but that made it a very bad user experience -- some models, especially open-source ones, took over 4 minutes to respond, leading to a high voter drop-off rate.
To preserve the voter experience without introducing bias, our current approach waits for the slowest model within each binary comparison — so even if one model is faster, we don’t display until both are ready. You're right that this does introduce some bias for the two smallest models, and we'd love to hear suggestions for how to make this better!
As for the 5th request: we actually kick off one reserve model alongside the four randomly selected for the tournament. This backup isn’t shown unless one of the four fails — it’s not the fastest or lowest-latency model, just a randomly selected fallback to keep the system robust without skewing results.
nice! Training models using reward signals for code correctness is obviously very common; I'm very curious to see how good things can get using a reward signal obtained from visual feedback
But this could be a legitimate way to design apps in general if you could tell the models what you liked and didn't like.
To preserve the voter experience without introducing bias, our current approach waits for the slowest model within each binary comparison — so even if one model is faster, we don’t display until both are ready. You're right that this does introduce some bias for the two smallest models, and we'd love to hear suggestions for how to make this better!
As for the 5th request: we actually kick off one reserve model alongside the four randomly selected for the tournament. This backup isn’t shown unless one of the four fails — it’s not the fastest or lowest-latency model, just a randomly selected fallback to keep the system robust without skewing results.