Emotion intelligence · Private beta
Amphora gives generative AI a biological basis for feeling. Our fMRI prediction engine maps any visual input to predicted brain activation — turning emotional outcome from a guess into a specification.
No spam. Unsubscribe anytime. Privacy policy
Experiment 1 · May 26, 2026
Using TRIBE v2 — Meta AI's fMRI encoder — as the reward signal, we steered a 3B parameter language model toward higher predicted cortical activation in Broca's area over 200 RL training steps. The entire language network responded. Here's the data.
Cortical surface — base · LoRA · Δ
Reward · loss · KL over 200 steps
20 cortical ROI activations — base vs LoRA
BOLD timeseries — all 20 regions
The model learned to write differently. Without being told how, it shifted from scene-setting prose to dialogue-driven narrative — exactly the style that predicts higher BOLD in Broca's, Wernicke's, and STS regions.
Read the full write-up →Experiment 2 · May 29, 2026
An independent LLM judge, MMLU regression tests, style robustness checks, syntactic complexity metrics, and MAUVE — five experiments run on the same step-200 checkpoint. The brain reward signal and human writing quality aligned on every measured axis.
The brain's language network is sensitive to the same properties that make text compelling. The reward signal was neuroscientific — it had no knowledge of human preferences. Yet the two objectives aligned on every axis.
Read the eval suite →Progress Notes · Members only
We ran an independent LLM judge, MMLU, style robustness, syntactic complexity, and MAUVE against the step-200 checkpoint. The brain reward signal and human writing quality aligned on every measured axis.
We used TRIBE v2 as the reward signal in an RL training loop. After 200 steps, the model had learned to produce text that drives 150% higher predicted cortical activation in Broca's area.
The generative AI industry has spent enormous effort making images more realistic. It has not given AI any mechanism to understand what an image does to the person looking at it.
The core of what we're building — a lightweight model that takes any visual input and outputs predicted brain activation across emotional and perceptual regions — is working.
What we're building
Submit any visual asset — a photo, AI render, ad creative — and receive a map of predicted brain activation across the regions that drive emotional and perceptual response. Not vague labels. Precise neuroscience.
Integrate our API into any generative AI pipeline. Set a target emotional profile — aspiration, intimacy, awe, tension — and let the system iterate until the output scores for the intended feeling. Emotional coherence becomes a specification, not an afterthought.
Know whether your image activates the precuneus for aspiration, the amygdala for threat and awe, or the fusiform cortex for intimacy and recognition — before it goes live. A precise, scientifically grounded signal for designers, AI systems, and brand strategists.