Sabi trains image models on personal taste from brain signals
The startup says a wearable EEG cap can turn passive neural responses into rewards, though the results have not been independently replicated.
Sabi has introduced “reinforcement learning with neural feedback,” or RLNF, a method that uses signals from a wearable EEG cap to personalize image generation without asking people to rate outputs.
In the company’s demonstration, a face changes while a participant looks at it. Sabi says arousal signals from the brain become an implicit reward, allowing the model to move toward that person’s aesthetic preferences. The team used a relatively simple embedding and k-nearest-neighbor system instead of training a separate reward model, and says randomized rewards did not produce the same improvement.
The approach is meant to address a limitation of population-level preference tuning: individual taste can differ sharply from the average preferences learned from large datasets. If the method holds up, passive feedback could give consumer AI products another route to personalization beyond prompts, ratings and preference histories.
The evidence is still preliminary and comes from Sabi itself; no independent replication or peer-reviewed paper was available with the announcement. The company also highlights the risks. Neural data could reveal private preferences, face-optimization systems may worsen body dysmorphia or narrow beauty standards, and services would need unusually strong consent, encryption and on-device processing. Sabi currently offers a waitlist rather than a broadly available product.
