The Lab · In-house research
We build our own models
AI Studio began as the best way to use every frontier model. The Lab is how we go further: a small research group of machine learning engineers and data scientists, building the models and methods behind the app, trained on signal no one else has.
“Every day, thousands of creators show us what a good result looks like, by what they keep, what they retry and what they share. The Lab turns that signal into models.”
The Lab at AI Studio
Three programs
One runs today. The other two open, in this order, as the founding research team is hired.
-
01
Active
Preference Learning at Scale
We train our own reward models on the behavioural signal of more than 1,000,000 creators: systems that learn to judge images and video the way our audience does. They work ahead of the user, scoring candidates before results appear, catching weak generations for automatic retries, and learning which of the models in production will serve a given prompt best, per credit spent.
The same program runs our evaluation harness. Every new model release is benchmarked within days of shipping, with paired comparisons, confidence intervals and category-level breakdowns, so what we route to is a measurement, never a guess.
Reward modelingPaired-comparison statisticsLearned model routingFrontier model evaluation
-
02
Planned
Custom Model Post-Training
Generic models produce generic results. For the tools our creators use most, like photo restoration, hairstyles, consistent characters and signature styles, we intend to post-train our own: fine-tunes and adapters of open-weight image and video models, trained on datasets we curate in-house.
The aim is a family of models that exist nowhere else, each tuned for exactly one job, measured against the frontier in our own harness and shipped straight into the app. This program has not started yet; it is the first one we open once the team is in place.
LoRA and adapter trainingCurated in-house datasetsCharacter and identity consistencyTask-specialized checkpoints
-
03
Planned
Efficient Inference and On-Device Generation
Speed is a creative feature. The plan is step distillation, quantization and self-hosted serving, so that our highest-volume tools eventually run on our own models: faster, at a fraction of the cost, with near-instant draft previews before a full-quality render.
Further out sits the Lab’s long-horizon goal, and we call it an ambition rather than a project: models compact enough to run on Apple silicon itself. Generation on the iPhone in your hand, private, instant and offline. That is a multi-year target, measured from the day Program 02 ships.
Step distillationQuantization and servingSelf-hosted inferenceOn-device generation
Now hiring
The Lab is hiring
Senior roles in machine learning, statistics and ML systems are open now. A small team, real data, and results in the app within weeks. All roles are remote.