Sabi Unveils RLNF: Brain Signals Replace Human Ratings in AI Training

For the past few years the standard way to make models understand people has been RLHF: ask humans to rate outputs, then train on those ratings. A brain-computer interface startup now wants to drop the rating step entirely and read your brain instead.

The company is Sabi. Co-founder Rahul Chhabra has shown what it calls RLNF, Reinforcement Learning with Neural Feedback, which uses brain signals captured by a wearable EEG device as the reward in place of explicit human scores.

A face that drifts toward you

Sabi's demo is easy to grasp. An AI-generated face sits on screen. The viewer wears Sabi's device and says nothing, clicks nothing, just watches. Over time the face morphs toward whatever the viewer seems to prefer.

The logic is simple enough. The brain responds differently to things we like and things we don't, and the model treats those responses as feedback and keeps adjusting. Where RLHF relies on a person consciously deciding on a score, RLNF taps an earlier, more direct layer of reaction, one people may not be able to put into words themselves.

The idea is not new in research. Academic work has used error-related potentials in EEG as implicit RL rewards, and other teams have used fNIRS signals to judge agent performance. But Sabi is the first to package it into a consumer wearable and tell the story with a face that changes in real time.

A beanie that reads your mind

Sabi's hardware is what makes the pitch hang together. The Bay Area company builds a noninvasive BCI that fits high-density EEG sensors into a beanie or a baseball cap, paired with an in-house "brain foundation model" meant to turn imagined speech into text and AI commands. This month it closed a $50 million seed round led by Khosla Ventures, with Accel and DST Global among the participants.

Seen that way, RLNF looks like Sabi reaching for a bigger story than thought-to-text. If brain signals can serve as a training reward, everyone wearing a Sabi device becomes a labeler who never has to say a word.

Compelling, and a little unsettling

The appeal is obvious. Human ratings are slow, expensive, subjective and prone to fatigue, while neural signals are continuous, instant and nearly free. For aesthetic or emotional preferences that are hard to describe anyway, reading the reaction directly could beat a survey.

The problems are just as obvious. EEG is noisy and varies a lot between people, and "I like this" may be hard to separate from "this grabbed my attention." Optimizing models against subconscious reactions could also push AI toward whatever is most stimulating and addictive, a road recommendation algorithms have already traveled. And once training data comes straight from your brain, nobody has yet said where the privacy line sits.

A face that drifts toward you is a lovely demo. It is still a long way from training frontier models, and the questions Sabi has to answer are not only technical ones.