VICTOR BEIGELMAN
Remixing AIs
In fall 2024, the Meta AI Studio team observed in usage data that many similar AIs were popping up across a few different entertainment and companionship use cases. Pairing this with UXR that suggested some people found the creation experience daunting, we developed a hypothesis that "remixing" an existing AI could lower perceived friction while subtly yet concretely fostering variety on the consumption side. My role in validating this idea was to bridge model behavior and UX by designing the right levers on top of the latter that would shift the former in a clear and consistent manner.
From a model behavior perspective, the key challenge centered on steerability. For both preset options and freeform remixing, we needed to ensure the creator's input fed into a prompt on the backend that modified the original AI substantially without losing its core purpose. The sweet spot, as ever, required a lot of experimentation to find—I toggled with positive vs. negative instructions, the right wording to maintain the essence of the original AI, concrete dimensions eligible for adjustment, etc. to land the right prompt.
The experimentation process also required evals to compare the remixed AI's generated inputs against the original, as well as the preset or custom steer that resulted in them. I worked with eng and DS to assess the data from our smoke test for how well a remixed AI maintained a core purpose while faithfully adjusting certain dimensions like tone and personality based on the steer. I also wrote the UI latency strings while the model did its work on the backend, as a way to keep people engaged for the ~10 seconds they might be waiting while hinting at the nuts and bolts.
Naturally, there was no way to guarantee a remixed AI would come out exactly as a creator had envisioned, so in addition to allowing them to edit aspects like name, description, instructions, etc. before actually publishing it, we leveraged Meta's image generation models to allow people to change the AI's appearance. Given the variance in avatar style outputs, even with the model set to a low temperature, this was a critical control to offer, with a fun intuitive interface meant to spark creativity.
As of summer 2025, in part due to the affordances we've made around creator controls and original AI attribution, AI remixing contributes a substantial portion of overall AIs created through Meta AI Studio. Along with user feedback, the hard data signals that our efforts to lower the bar to entry are working. There's still a long way to go to refine what's steerable and bubble that up into the UX clearly, but such is the (rewarding!) process of building AI products.